EUFIRELAB

EVR1-CT-2002-40028

D-08-06

http://eufirelab.org

EUFIRELAB:
Euro-Mediterranean Wildland Fire Laboratory,
a “wall-less” Laboratory
for Wildland Fire Sciences and Technologies
in the Euro-Mediterranean Region

Deliverable D-08-06

Towards a Euro-Mediterranean
Wildland Fire Danger Rating System
AUTHORS by PARTNER

P014: Giovanni BOVIO, Raffaella MARZANO

P023: Inmaculada AGUADO, Emilio CHUVIECO, Jesús MARTÍNEZ
Héctor NIETO, Javier SALAS
P027: Wanda BEROLO, Pierre CARREGA, Dennis FOX, Nuno GERONIMO,
Jean-Guillaume ROBIN
P030: Israel GÓMEZ, Pilar MARTÍN, Javier MARTÍNEZ-VEGA, Lara VILAR
P036: Ioannis GITAS, Michael KARTERIS, Spyros TSAKALIDIS

December 2006

The views expressed are purely those of the writers and may not, in any circumstances, be regarded as stating an
official position of the European Commission
EUFIRELAB

CONTENT LIST
1 Scope and objectives...............................................................................................................................................1
1.1 First phase.....................................................................................................................................................1
1.2 Second phase................................................................................................................................................2
1.3 Third phase....................................................................................................................................................2
2 Proposed methods for wildland fire danger estimation at European scale .............................................................3
3 Review of fire danger rating System developed in MEGAFIRES project................................................................4
3.1 Introduction....................................................................................................................................................4
3.2 Use of meteorological indices and satellite data ...........................................................................................4
3.3 Figures...........................................................................................................................................................6
3.4 Tables............................................................................................................................................................8
3.5 Analysis of long-term fire risk on a European level .......................................................................................9
3.5.1 Selection of risk variables....................................................................................................................9
3.5.2 Techniques to estimate large fire occurrence ...................................................................................10
3.5.3 Figures...............................................................................................................................................12
3.5.4 Tables................................................................................................................................................14
4 Review of fire risk estimation method developed for the SPREAD project ...........................................................15
4.1 Input data for fire risk mapping....................................................................................................................16
4.2 Human danger ignition ................................................................................................................................16
4.3 Probability of Ignition and Fuel Moisture Content .......................................................................................17
4.3.1 Fuel Moisture Content of live fuels ....................................................................................................17
4.3.2 Fuel Moisture Content of dead fuels .................................................................................................17
4.3.3 Probability of Ignition related to the FMC ..........................................................................................17
4.4 Propagation Danger ....................................................................................................................................18
4.4.1 Average Rate of Spread and Flame Length......................................................................................18
4.4.2 Propagation danger versus RoS and FL...........................................................................................18
4.5 Wildland Fire Danger Assessment ..............................................................................................................19
4.6 Figures.........................................................................................................................................................19
5 Review of the European Forest Fire Information System-Risk Forecast ..............................................................22
5.1 Introduction..................................................................................................................................................22
5.2 Meteorological danger indices.....................................................................................................................23
5.3 FPI Model. ...................................................................................................................................................23
5.3.1 General presentation .........................................................................................................................23
5.3.2 Input data...........................................................................................................................................24
5.3.3 Computation of FPI............................................................................................................................26
5.4 Long-term indices........................................................................................................................................26
5.4.1 Probability of Fire Occurrence...........................................................................................................26
5.4.2 Likely Damage...................................................................................................................................27
5.5 Figures.........................................................................................................................................................28
5.6 Tables..........................................................................................................................................................29
6 Basic structure and characteristics of the proposed Euro-Mediterranean Wildland Fire Risk Index ....................31
6.1 General presentation...................................................................................................................................31
6.2 Figures.........................................................................................................................................................33
7 Components of the fire risk system: Ignition Danger Index...................................................................................34
7.1 Fuel Moisture...............................................................................................................................................34
7.1.1 Introduction........................................................................................................................................34
7.1.2 Live fuels (satellite information).........................................................................................................34
7.1.3 Study case in Central Spain ..............................................................................................................35
7.1.4 Probability of ignition related to live fuel moisture content ................................................................39
7.1.5 Probability of Ignition in Dead Fuels (Meteorological Index).............................................................40
7.1.6 Figures...............................................................................................................................................44
7.2 Probability of Ignition (Human Factors).......................................................................................................49
7.2.1 Introduction........................................................................................................................................49
7.2.2 Estimation of human ignition danger at regional scale: the case of Alpes-Maritimes (France) ........50
7.2.3 The third model: statistical approach B. ............................................................................................53
7.2.4 Model comparison and conclusion. ...................................................................................................53
7.2.5 Study Area and Wildland Fire Database ...........................................................................................54
7.2.6 Methodology: Geographically Weighted Regression (GWR), the linear and logistic case ...............55
7.2.7 Results and discussion......................................................................................................................55

D-08-06.doc
EUFIRELAB

7.2.8 Conclusion.........................................................................................................................................56
7.2.9 Figures...............................................................................................................................................57
7.2.10 Tables................................................................................................................................................61
7.3 Human ignition danger in Southern Europe based on fire occurrence maps .............................................63
7.3.1 Fire Occurrence Hot Spot Areas .......................................................................................................63
7.3.2 Figures...............................................................................................................................................65
7.3.3 Tables................................................................................................................................................67
8 Propagation danger index .....................................................................................................................................68
8.1 Average rate of spread and Flame Length..................................................................................................68
8.2 Propagation danger (PD) ............................................................................................................................69
8.3 Figures.........................................................................................................................................................69
9 Vulnerability index..................................................................................................................................................70
9.1 Population vulnerability ...............................................................................................................................70
9.1.1 Introduction........................................................................................................................................70
9.1.2 Example of population vulnerability mapping in Piemonte Region (Italy) .........................................72
9.1.3 Estimation of population vulnerability at Euro-Mediterranean scale .................................................73
9.1.4 Figures...............................................................................................................................................76
9.1.5 Tables................................................................................................................................................80
9.2 Vulnerability related to environmental value ...............................................................................................82
9.2.1 Presentation ......................................................................................................................................82
9.2.2 Figures...............................................................................................................................................83
9.3 Potential soil erosion ...................................................................................................................................85
9.3.1 Introduction........................................................................................................................................85
9.3.2 Objectives..........................................................................................................................................85
9.3.3 A European scale model to predict potential erosion risk .................................................................86
9.3.4 An operational model for the Mediterranean context ........................................................................89
9.3.5 Figures...............................................................................................................................................92
9.3.6 Tables................................................................................................................................................99
10 Euro-Mediterranean Wildland Fire Risk Index.....................................................................................................100
10.1 General presentation.................................................................................................................................100
10.2 Figures.......................................................................................................................................................101
11 References ..........................................................................................................................................................102
12 Annex: Mapping post-fire soil erosion risk...........................................................................................................114
12.1 Introduction................................................................................................................................................114
12.2 The impact of forest fires on soil erodibility ...............................................................................................114
12.3 Site description ..........................................................................................................................................115
12.4 Methods.....................................................................................................................................................116
12.4.1 Mapping soil erosion risk.................................................................................................................116
12.4.2 Soil erosion factors ..........................................................................................................................116
12.4.3 Partial validation of the soil erosion risk map ..................................................................................117
12.5 Results.......................................................................................................................................................117
12.5.1 Distribution of the soil erosion factors .............................................................................................117
12.5.2 The soil erosion risk map ................................................................................................................117
12.5.3 Partial model validation ...................................................................................................................118
12.6 Discussion .................................................................................................................................................118
12.7 Conclusions ...............................................................................................................................................119
12.8 Figures.......................................................................................................................................................120
12.9 Tables........................................................................................................................................................122

D-08-06.doc
EUFIRELAB

SUMMARY

The basis and structure have been defined taking

This deliverable is strongly related and in a sense
into account the proposals and suggestion of the
the succession of the deliverables D-08-02 (Wildland
projects analysed in the previous chapter.
fire danger and hazards: a state of the art), D-08-05
In this chapter, it has been determinant the scale of
(Common methods for mapping the wildland fire
work (global) and the availability of homogenous
danger), and D-08-03 (Towards a Euro-Mediterranean
information for entire area of study.
Wildland Fire Danger Rating System: basis, structure).
The structure of the system follows a hierarchic
Deliverable D-08-03 addresses the design of the
scheme with three sub-indices concerning the fire
basis and structure of a prototype of a Euro-
danger associated to ignition, propagation and
Mediterranean Wildland Fire Danger Rating System.
vulnerability.
Deliverable D-08-06 addresses the final design of
the system. In chapters 7 to 9, the variables that would have to
integrate each one of the mentioned sub-indices are
The present work has been divided in three main
analysed.
chapters.
The methods for their obtaining and calculation are
In chapters 2 to 5, a review of methods already
proposed.
developed and applied operationally or semi-
The input variables include the fuel moisture content
operationally for estimating and mapping wildland fire
(live and dead), the risk of ignition associated to man,
risks at Euro-Mediterranean scale is presented.
models of simulation of fire behaviour, vulnerability
The results of two previous European projects
related to population, environmental value of territory,
(Megafires and Spread) and the European Forest Fire
and potential soil erosion.
Information System-Risk Forecast –EFFIS- developed
Considering the state of the art and the available
by the JRC are analysed.
information, it has not been possible to analysis all of
In chapter 6 the basic structure of the wildland fire them with the same detail and precision.
danger rating system is proposed.
Final conclusions about the integration of the
This system is composed by the three most
proposed variables and the distribution of the index are
important components in forest fire risk: ignition,
included in chapter 10.
propagation and vulnerability.
Chapter 11 is dedicated to the numerous references
and the annex (chapter 12) presents a specific work
dealing with mapping the post-fire risk of soil erosion
.

D-08-06.doc
EUFIRELAB

1 SCOPE AND OBJECTIVES

1.1 FIRST PHASE

The main objective of the Wildland Fire Risk and
Hazard Unit of the Eufirelab Project (Unit 8) is the First includes a review of methods already
design of the basic structure and characteristics of a developed and applied operationally or semi-
prototype of a Euro-Mediterranean Wildland Fire operationally for estimating and mapping wildland fire
Danger Rating System. risks at Euro-Mediterranean scale (Megafires and
With that intention, the work of this unit has been Spread projects, European Forest Fire Information
structured in three main sections: System-Risk Forecast - EFFIS -).
- Revision of the variables used in existing fire danger In this section it has been realized a detailed
indices (deliverables D-08-02 and D-08-07), analysis of the proposals, at global scale and for the
- Analysis of methods for mapping the wildland fire Euro-Mediterranean basin, about which we have been
danger (deliverable D-08-05), and able to find information.
- Design and development of a prototype of Euro-
Mediterranean Wildland Fire Danger Rating System
(deliverables D-08-03 and D-08-06), based on the
information analyzed in the previous deliverables.
On the basis of the state of the art (existing
systems, necessary inputs, algorithms...) and the
information availability on the variables of interest, the
consortium is intended to develop a prototype of the
Euro-Mediterranean Wildland Fire Danger Rating
System, designing the basic structure and determining
its characteristics.
This final objective will be developed in two
deliverables.
In the first (D-08-03), a proposal of the basic
structure of the index and a broad selection of the
variables to use is carried out.
In the second one (D-08-06), that we present next,
will be define more precisely the structure and the
variables to include in an index that can be used at
global scale for the Euro-Mediterranean Basin.
According to these objectives, the deliverable D-08-
06 "Towards to Euro-Mediterranean Wildland Fire
Danger Rating System" is carried out in three phases.

D-08-06.doc 1
EUFIRELAB
1.2 SECOND PHASE
In the second phase is proposed the basic structure
of this risk index to integrate the most important
components in the ignition, propagation and
vulnerability of forest fires.
This was carried out taking into account the
proposals and suggestion made in these previous
projects and the analysis performed in preceding
deliverables of this work unit (8), concerning variables
of risk and methods of integration.
In this section it has been determinant the scale of
work (global) and the availability of information (more or
less homogenous) for entire area of study.
These two premises have supposed the proposal of
a relatively simple structure, which could become more
complex and rich in local scales of work as well as in
the case of having available more homogenous
information about variables of interest for whole area.
Even so, as we see in the last section of this
deliverable, it is not at all simple to propose an
operative index for a region so extensive and diverse
(natural and culturally) as Southern Europe.
The structure of this index follows a hierarchic
scheme in which the final index is obtained from three
sub-indices concerning the fire danger associated to
ignition, propagation and vulnerability.
These sub-indices offer independent information
that it can be used individually.
For example, the ignition danger index can be very
useful for the design of the vigilance and monitoring
system, whereas the propagation risk index can be
very interesting for the location of extinction resources.
The vulnerability shows us those zones that could
suppose greater losses in human, ecological and
economic terms.
The combination of these three sub-indices will offer
an overall view about the forest fire danger, considering
all the key elements that need to be valued in an index
of these characteristics.
D-08-06.doc
1.3 THIRD PHASE
In the third phase, after the proposal of this basic
structure, the variables that would have to integrate
each one of the mentioned sub-indices are analyzed. In
some cases the methods for their obtaining and
calculation are proposed.
As it has been previously commented, the selected
variables will depend significantly on the scale of work
and the availability of information.
For this reason, in this section of the deliverable,
study cases at local scale are also presented, with the
intention of transferring that proposal from the local
scale to a global scale.
The input variables in the ignition sub-index include
the fuel moisture content (live and dead) and the risk
associated to man.
In the sub-index of propagation, models of
simulation of fire behavior are included. Finally,
variables associated to population, environmental
value, and soil erosion, are incorporated in the
vulnerability sub-index.
Considering the state of the art and the available
information, it has not been possible to take into
account all of them with the same detail and precision.
Finally, some recommendations about the
integration of the mentioned variables are proposed
and some ideas about the distribution of the index are
included.
2
EUFIRELAB
2 PROPOSED METHODS FOR WILDLAND FIRE DANGER ESTIMATION AT EUROPEAN SCALE
As it was mentioned in the description of work, the
review of the state of the art on fire danger and hazard
assessment and mapping is one of the main objectives
of Unit 08.
The review was divided in two major parts:
- the analysis of the individual environmental and
anthropogenic variables related with forest fire
danger and hazard (accomplished in D-08-02), and
- the description of the models and methods currently
used to combine those variables to estimate fire
danger and hazard and produce maps of their
spatial distribution at different scales (completed in
D-08-05).
A long tradition of research, both European and
worldwide, has been devoted to those issues, specially
at local and regional scales, but few attempts has been
made in order to create an integrate fire risk
assessment system at coarser scales (i.e. Pan-
European).
In the above mentioned reviews missing links in the
integration and spatial assessment of risk information
were observed as most of the works were produced
and tested in specific study areas and they didn´t offer
much information on how to extrapolate the methods to
other areas and, even most important, to other spatial
scales.
The objective of this section is to review existing
methods, systems or products that have been
proposed for fire danger/risk estimation at European
scale.
This review will allow us to analyse and compare the
structure and components of those systems and also to
identify their goodness and, eventually, potential
limitations for being operationally applied to evaluate
fire risk in the framework of national or supranational
decision support systems.
This information is required to cope with the second
objective of Unit 08, which is to propose the structure of
a Euro-Mediterranean Widland Fire Danger Rating
System.
D-08-06.doc
Wildland fire danger may be considered at different
spatial and temporal resolutions: global and local scale
on one hand, and short-term and long-term on the
other.
Both scales of assessments are very important for
fire management.
Considering only spatial scale, global approaches
assume to cover from millions to dozen million square
kilometres, while local studies are focused on hundreds
to few thousands square kilometres.
Global analysis can contribute to establish general
guidelines for fire management at international level,
such as at European Union level, while local scales are
adapted to specific fire prevention resources of small
regions or provinces.
The type of information to be used in fire risk
assessment systems is strongly dependent on the
spatial scale, since global approaches could most
probably not include some critical variables that are
easily available at local or regional scale.
As it was previously mentioned, in this review
section we will include only those methods for wildland
fire danger estimation that had been proposed and
effectively implemented at European scale, that is,
methods that have produced maps of fire risk/danger
for the whole Europe or, at least, an area covering
several European countries.
Systems that were theoretically developed for being
applied at global scales but were never implemented or
only in specific test areas are not included in this
review.
We have found three systems that fulfill the above
mentioned requirements, two of them are the result of
european projects “Megafires” and “Spread” and the
last one is a fire risk forecasting system developed and
implemented by the Joint Research Centre (JRC) in the
framework of the INFOREST Action
(http://inforest.jrc.it/effis/).
3
EUFIRELAB
3 REVIEW OF FIRE DANGER RATING SYSTEM DEVELOPED IN MEGAFIRES PROJECT
3.1 INTRODUCTION
One of the objectives addressed in the MEGAFiReS
project (http://www.geogra.uah.es/megafires/) was to
develop methodologies for fire risk rating and mapping,
at both global and local scales, based on satellite
remote sensing and Geographic Information System
(GIS) techniques.
For the global scale, it was intended to present a
comprehensive view of fire risk assessment at the
European Mediterranean countries.
MEGAFiReS project undertook, for the first time in
Europe, a cross-country evaluation of trends in fuel
moisture content of Mediterranean species using
coarse resolution remote sensing data.
Three different aspects of fire risk assessment were
investigated in MEGAFIRES:
- Short-tem fire risk: foliage moisture estimation from
satellite data
- Fire danger rating using meteorological indices and
satellite data
- Long-term fire risk mapping
Fuel moisture estimation was investigated in
MEGAFIRES using high and low resolution sensors.
This was one of the first attempts to correlate daily-
acquired low resolution satellite data with moisture
content of living foliage on a significant European scale.
Validation of the satellite-derived information was
based on the measurements of foliage moisture
content of different species carried out during 2 years
on 4 networks of test-land plots located in three
different Euro-Mediterranean countries.
No Pan-European maps were produced and,
therefore, we will not describe here the proposed
methodology.
However, the study clearly showed the potential of
low-resolution satellite data such as NOAA-AVHRR to
monitor and map foliage moisture content at global
scales.
On the contrary, the other two aspects of fire risk
that were included in the project, produced methods
and results at Euro-Mediterranean scale and, therefore,
they will be described below.
D-08-06.doc
3.2 USE OF METEOROLOGICAL INDICES AND
SATELLITE DATA
In Europe there is no uniform approach to fire
danger rating and different methods have been applied
and developed in different Mediterranean countries.
Some studies were made in Minerve I and II
European research projects to compare the fire danger
rating capabilities of different meteorological danger
indices, in the attempt to find a common model to be
applied in the European Mediterranean area, but the
final results left some uncertainty about the best model
to use (BOVIO et al., 1994; VIEGAS et al., 1996).
One of the problems faced by the danger indices is
that they are not specifically tailored to assess extreme
fire danger conditions.
In MEGAFIReS a first global attempt was made to
develop a system for rating, from ground weather
measurements, extreme fire conditions in the European
Mediterranean basin.
The system is empirical and is made up of a
combined set of selected existing fire danger indices.
On the other hand, the possibility of using satellite
data to estimate the variables needed to compute the
danger indices specifically tailored for large fires was
investigated.
The objective was to have a more accurate spatial
distribution of the fire danger estimation that could be
less dependent on the location and density of the
weather stations.
A global historical fire and meteorological databases
were addressed for the whole European Mediterranean
basin.
Meteorological variables required to compute fire
danger indices were obtained from the database of the
MARS project implemented at the Joint Research
Center.
In this database, in the time period considered, daily
ground weather data recorded at 12:00 a.m. from about
360 weather stations were interpolated on 1389 grid
cells of 50x50 Km (Figure 1).
Historical meteorological data included maximum
and minimum air temperature, vapour pressure, wind
speed, rainfall, potential evaporation and solar
radiation.
With these weather data, 28 fire danger indices
were computed for each grid cell (daily values for five
years).
The indices computed had different features,
addressing specific components of fire danger (see
table 1).
Due to the high heterogeneity of the fire
environment in the territory, the area of investigation
was restricted and stratified using land cover and
climatic data.
4
EUFIRELAB
Only those grid cells with at least 25% of their area
covered by wildlands (i.e. the sum of grassland,
shrubland and forest land according to the CORINE
Land Cover map legend) were retained for further
investigation, so those grid cells mostly covered by
urban and agricultural lands were masked out.
In addition, only the Köppen climatic types BSh,
BSk, Csa and Csb were considered truly representative
of the investigated fire environment.
The other climatic types, except for some transitory
situations, were in different conditions, often with a
winter-spring fire season and a generally lower level of
fire danger.
In Figure 2 the area selected for the investigation,
stratified in the climatic types is showed.
The restriction of this defined area of interest
resulted in a reduction of the grid cells from 1389 to
499, which multiplied by the number of days considered
(5 fire seasons of 122 days each) give a total of
304 390 statistical units.
The empirical model was build using logistic
regression in order to identify extreme/non extreme fire
danger conditions.
The objective was to identify the set of indices that
together could be used to best estimate the
meteorological conditions for a Danger Day (DD).
It must be pointed out that large fires often last for
more than one day, that is to say that meteorological
danger conditions associated with them remain for a
certain period.
Actually, it often happens that a large fire that burns
for several days spread most of the burned area in few
highly severe days.
In order to consider a reasonable statistical unit an
approximation was introduced defining as Extreme
Danger Day (EDD) a day with at least 1 large fire
burning in the grid cell, excluding multi-day burning
fires and the day when the fire was extinguished, that is
the last day, when is was reasonable to expect a lower
level of fire danger.
The dependent variable in the logistic regression
was a binomial variable that identified an EDD (value 1
as danger day or 0 as non danger day), while the
explanatory variables were a selected set of danger
indices that provided the best fit.
Exploratory data analysis of indices led to identify
and exclude some indices with undesirable distribution
properties.
After test and build models, better results were
obtained when separating the 3 climatic zones i.e.
building different models in each zone and including
indices of different types in each model (short, mid and
long-term moisture content estimators, spread potential
indices and composite indices), showing that when
different components of fire danger are taken into
account the estimates are improved.
All models were globally significant at the 1% level
and so were for individual variable coefficients, with
some exception.
D-08-06.doc
The negative signs found in coefficients were for
indices that directly indicated moisture content values,
while positive signs were for indices that increased with
danger level.
In the typical Mediterranean climatic types (Csa and
Csb classes) the models showed better performances,
while in the semi-arid zone (BS class) predictions were
more uncertain.
Nevertheless, in all zones and, therefore, at global
level, the logistic models developed showed better
performances of each of the 28 individual fire danger
indices considered in the study.
The indices finally selected should not be
considered as being generally better than the others.
Exploratory data analysis showed that many indices
that were not included in the equations had good
discriminating power.
Those selected were a “set” of danger indices that
together improved the rating capabilities of
meteorological severe fire conditions.
To assess the model fitting, observed and predicted
EDD were compared.
This was done grouping the observations according
to different criteria (probability classes, climatic zones
and month) and also applying the performance score
for fire danger indices introduced by (MANDALLAZ and
YE 1997)
The spatial fitting of the model to the data was
qualitatively assessed by means of monthly maps of
DDs plotted over the monthly values of the model
output in each grid cell. One example of these monthly
maps is presented in figure 3.
The agreement between the spatial distribution of
the estimated fire severity and large fires ocurrence,
changed from year to year, but an overall correct
behaviour of the logistic model was recognised.
The logistic regression analysis provided a first
approximation to underline the meteorological indices
better adapted to predict large fire occurrence.
The authors proposed the logistic model as a first
prototype that must undergo further analysis.
Besides, the authors consider the usefulness of
incorporating other refinements such as the integration
of satellite images for improving the spatialization of the
variables.
Within MEGAFIRES project, the potential use of
satellite data for the spatial extrapolation of specific
components of meteorological fire danger was
explored.
The indices considered were the Keetch-Byram
Drought Index (KBDI; KEETCH and BYRAM, 1968) and
the Canadian Drought Code (DC; VAN WAGNER 1987).
The analyses were performed at three different
levels, in a region of Spain, in the Iberian Peninsula
and in the whole European Mediterranean basin.
Both temporal and spatial dimensions were
addressed when analysing correlation of satellite
variables with meteorological danger indices.
5
EUFIRELAB
For Andalusia and the Iberian Peninsula the results
confirmed the correlation between NDVI and KBDI was
quite poor, but the one between NDVI and DC was
interesting.
For the global approach both meteorological danger
indices (DC and KBDI) were computed with records of
the JRC-MARS database for a selection of 56 weather
stations.
The satellite data were the same mosaic of NOAA-
AVHRR images used from the correlation analysis with
the logistic model, with forested area masked out in
circles of 10 km radius around selected weather
stations.
Regarding the temporal analysis, the expected signs
of correlation between NDVI and DC (inverse relation)
were obtained for 44 out of 56 weather stations.
Due to the few dates available, the statistical
significance at the 5% level could be established for
only 9 of them.
Regarding the spatial dimension, the correlation
between NDVI and DC calculated for all available dates
showed the expected trend for 31 dates out of 35.
3.3 FIGURES
Figure 1: Layout of grid cells in the European Mediterranean Basin and location of large summer fires 1991-1995
Figure 2: Area of investigation (study area) stratified in the climatic types
D-08-06.doc
The strongest correlation period, from 19 to 23 July,
with R=-0,74 and 92 records of available data, was
then used to develop a regression equation to derive
DC values as a function of NDVI (figure 4).
The common background of some meteorological
fire danger indices and remote sensing-derived indices
was confirmed by observation at the regional level and
partially at the European Mediterranean basin level,
and this is particularly important for the long-term
components of fire danger.
Furthermore, the DC, better than the KDBI, seemed
to be more related to living vegetation vigour and its
seasonal trend in Mediterranean environment, as it is
monitored by remote sensing techniques.
These consistent similarities, regardless of the local
climatic differences, strongly support the use of satellite
data to estimate temporal trends in some components
of meteorological fire danger, and encourage further
investigation on its potential integration into currently
operational fire danger indices.
6
EUFIRELAB

Figure 3: July 1994 DDs and model values in the study area

Figure 4: Logistic model output estimated from ST in the European Mediterranean basin (July 19-23, 1996).
The mapped area is the one for which the logistic model was developed, i.e. the climatic zones Csa, Csb and
BS, with urban and agricultural 50x50 km2 grid cells masked out

D-08-06.doc 7
EUFIRELAB

3.4 TABLES

Table 1: Indices used in the study

Category N Index Description Variable

1 BEHAVE fine fuel moisture model (ROTHERMEL et al. 1986) BEHAVE

Short-term 2 Canadian Fine Fuel Moisture Code (FFMC) (VAN WAGNER 1987) FFMC
moisture
3 Moisture content estimate in Mark5 McArthur’s Fire Danger Meter (NOBLE et al) Mark5
content
estimators 4 McArthur ‘67 fuel moisture model (MCARTHUR 1967) McArth67
5 USA NFDRS 1-Hour Timelag Fuel Moisture (DEEMING et al. 1978) h1
6 NFDRS 10-Hour Timelag Fuel Moisture (BRADSHAW et al. 1983) h10
7 NFDRS 100-Hour Timelag Fuel Moisture (BRADSHAW et al. 1983) h100
Mid-term and
8 Fosberg’s et al. 1000-Hour Timelag Fuel Moisture (FOSBERG et al. 1981) h1000
Long-term
moisture 9 Canadian Build Up Index (BUI) (VAN WAGNER 1987) BUI
content
10 Canadian Drought Code (DC) (VAN WAGNER 1987) DC
estimators
11 Canadian Duff Moisture Code (DMC) (VAN WAGNER 1987) DMC
12 Keetch-Byram Drought Index (KEETCH and BYRAM 1968)- revision, 1988) KB
13 Canadian Initial Spread Index (ISI) (VAN WAGNER 1987) ISI
Spread 14 Mark3 McArthur’s Fire Danger Meter (NOBLE et al. 1980) Mark3
potential
estimators 15 Mark5 McArthur’s Fire Danger Meter (NOBLE et al. 1980) Mark5F
16 Italian Fire Danger Index (PALMIERI et al. 1992) Impi
17 Canadian Fire Weather Index (FWI) (VAN WAGNER 1987) FWI
18 Portuguese Index (GONÇALVES and LOURENÇO 1990) PortoCif
19 Spanish ICONA Method – probability of ignition (ICONA 1993) Prob
Composite
20 Orieux Index (ORIEUX 1979) Orieux
indices
21 Carrega ’87 Index (CARREGA 1990) Carrega
22 Sol Numerical Risk (SOL 1990) SolRisNum
23 IREPI index (BOVIO et al. 1994) IREPI
24 Anderson et al. adsorption (ANDERSON et al. 1978) EWmax
Equilibrium 25 Anderson et al. desorption (ANDERSON et al. 1978) EDmax
Moisture
26 Van Wagner adsorption (VAN WAGNER 1982) EWmaxW
Content
(EMC) models 27 Van Wagner desorption (VAN WAGNER 1982) EDmaxW
28 Simard’s Equilibrium Moisture Content equations (SIMARD 1968) SEMCmax

D-08-06.doc 8
EUFIRELAB
3.5 ANALYSIS OF LONG-TERM FIRE RISK ON A
EUROPEAN LEVEL
The second MEGAFIRES task regarding risk
assessment at Euro-Mediterranen scale was related
with the definition of an integrated long-term fire risk
index.
The main goal of this analysis (CHUVIECO et al.
1999b) was to identify structural factors of fire risk at
European scale.
As in the case of meteorological indices, fire
occurrence was focused on large events (above 500
hectares).
The basis for the analysis was the compilation of a
database for the whole study area (Greece, Italy, South
of France, Spain and Portugal).
The variables were obtained from National or Global
databases: CORINE, Digital Chart of World, NGDC
Globe project, DMSP data.
Census data were the most difficult to compile, due
to the problem of ensuring consistency among the
different countries.
Census data defined the geographical unit of
reference.
Since most of the human variables were only
available at a provincial level (NUT-3), this division was
used for all the variables.
In the case of geographical layers which cover the
whole territory (such as land cover, elevation and
DMSP data), the average value for each province was
computed to assure consistency among variables.
Within the study area, there are 164 NUT-3
provinces: 48 in Spain, 20 in Italy, 52 in Greece, 18 in
Portugal and 26 in France.
Provincial average sizes vary from one country to
another, Italy having the largest and Greece the
smallest.
3.5.1 Selection of risk variables
The selection of variables relating to long-term
trends of fire occurrence was carried out using previous
technical literature reviews, (CHUVIECO et al. 1997) and
considering the limitations of data among the different
countries.
Three groups of variables were identified: first the
geographical ones, that deal with terrain features
(climate, land cover, roads, rails, etc.); secondly the
demographic variables, which were related to
population characteristics, and finally agricultural
variables, dealing with agricultural structure.
Some variables that could be, at least theoretically,
related to human risk, such as property size,
unemployment, and hunting practices were not
considered, because of the difficulty in either obtaining
data or homogenising them among the different
countries.
Since rural economies have strongly changed in
most Mediterranean countries during the last 30 years,
dynamic variables were also included in the analysis,
by comparing present values with those measured in
1960.
D-08-06.doc
A whole set of 52 variables taken from demographic
and agricultural census were extracted for
164 provinces of the study area.
Regarding the geographical variables, a brief
description of the spatial analysis undertaken follows:
3.5.1.1 Elevation
Elevation was obtained from two sources: National
maps at a 1:1,000,000 scale for Spain and Italy, and
the GLOBE project (compiled by the U.S. Geological
Survey at 1 Km2 resolution). From the elevation data
(Figure 5), mean slope and roughness were computed
using algorithms provided by Idrisi G.I.S. (EASTMAN
1993)
3.5.1.2 Land cover
Land cover, which is one of the layers generated for
the European Environmental Agency (European
Environtmental Agency 1996) was extracted from the
CORINE program.
Since some regions were missing from the
published CD-ROM, the coverage was completed with
data directly provided by the EEA.
The European legend of the CORINE land-cover
program was simplified to six general fuel type
categories: Grasslands, Shrublands, Perennial,
Broadleaf, Agriculture, and Non-Vegetated (Figure 6).
3.5.1.3 Roads and railways density
The density of roads and railways was computed
from national maps and the Digital chart of the World.
3.5.1.4 Urban areas
Urban areas were assumed to be related to fire risk,
since recreational uses of forests are among the main
causes of fire ignition, which can be either accidental or
due to carelessness.
Since urban areas are very dynamic, satellite
information was chosen as a source to map urban land
cover.
Data from the Defence Meteorological Satellite
Program (DMSP) provide a global view of city lights,
since this satellite includes a very sensitive radiometer
operating at night in the visible spectrum (Figure 7).
The National Geophysical Data Center (NGDC) in
Boulder, Colorado, has processed these data to
generate a world database of city lights. (ELVIDGE et al.
1997).
City lights data from Eurasia were extracted from
the Internet server at NDGC.
3.5.1.5 Climatic regions
Climatic regions were generated from the Joint
Research Center climatic database archived at the
MARS unit.
Average values for the last 30 years were used to
classify each cell, according with the Koeppen method.
The original Koeppen system was reduced to a
smaller number of classes and those not presented in
the MEGAFiReS study area were discarded.
9
EUFIRELAB
The final map includes the following climates:
BSh+Csa, BSk, Cfa, Cfb+Cfc, Csb+Csc, Dfb+Dfc.
The proportion of each provincial unit per climate
type was extracted after a cross-tabulation between the
Koeppen system and the provincial limit.
3.5.2 Techniques to estimate large fire occurrence
The number of large fires and burned area in each
province of the EU Mediterranean countries was
compiled from national statistics (Figure 8).
The study period covered was from 1991 to 1995.
Presence of large fires in each provincial unit was
selected as the target variable used to estimate fire
occurrence from the auxiliary variables previously
mentioned.
The final objective was to better understand how
these variables explain patterns of large fire occurrence
in the EU Mediterranean basin.
Two techniques were selected to perform this
analysis: Logistic regression and Artificial Neural
Networks.
Both techniques had been previously used in the
estimation of fire occurrence providing coherent results.
Additional analysis was performed with the number
of large fires per provincial unit.
In this case, linear regression analysis was used
instead of logistic regression, since this latter technique
can only be applied to dicotonomous variables (fire/not
fire).
Neural networks were applied both to the presence
or absence of fires and the number of fires per
provincial unit.
3.5.2.1 Estimation using logistic regression
Different models were tested including different set
of variables.
In order to avoid the effect of multiple correlation
and the noise produced by the large number of
variables, a previous analysis was carried out.
Multiple correlations were computed after grouping
the variables into three categories (geographical,
demographic and agricultural variables).
These variables were input into three different LR,
one for each of the above groups, in order to find out
the most significant variables to explain fire occurrence.
In all cases, the occurrence of large fires (0 not
affected, 1 affected) was used as dependent variable.
LR was performed with the Stepwise Backward
Selection algorithm included in the SPSS statistical
package.
In order to build a global LR model, which includes
the geographical, demographic and agricultural
variables, a new correlation matrix was computed
among them, discarding some variables.
The final model was computed from 161 cases.
Three were discarded because they offered a high
bias.
D-08-06.doc
The variables finally selected were: LDENSI90
(population density), POBACT90 (Active population
90), LALTIMED (mean elevation), LDECIDU (% of
broadleaf forest), and CLIM_BCS (Proportion of area
with Koeppen climates B and CS).
Table 2 provides the LR coefficients for the final
variables and their level of significance.
The signs obtained were logical since the fires are
expected to be higher at greater population densities,
higher elevations, more arid climatic conditions, less
active population and less area covered by broadleaf
trees.
The variables considered in the equation offer
trends in the correlation with the dependent variable
according to what was expected (see column in
table 2).
The significance level varies among them, being
more important for active population, climate and
population density, which should be considered as the
most related to large fire occurrence
Table 3 offers an assessment on the performance of
the model.
A global accuracy of 78.26% was obtained in the
estimation of large fires between 1991 and 1995.
This prediction is quite acceptable, considering the
great diversity of the study area, not only in a
geographical sense but also taking into account the
different national particularities regarding fire
ocurrence.
This performance is even better if only omission
errors are considered, since only 11% of the provinces
where fires occurred were not classified as such.
Commission errors were higher (almost 40% of the
provinces predicted as having a large fire were not
affected), but these errors are less critical than the
omission’s from a fire management point of view.
The LR function derived from the final model makes
it possible to compare the geographical distribution of
expected versus observed occurrence of large fires
(figure 9).
Most provinces were predicted correctly.
To perform a final test on the LR model, a new
equation was generated from a random sample of 60%
of all the provinces.
The other 40% of provinces were used as test
cases.
The LR model was created following a similar
approach to the previous one.
A threshold below 0.5 was also selected to
discriminate between occurrence / not occurrence.
In this case, as could be anticipated, the fitting is
poorer that in the previous equation (table 4).
A global accuracy of 60% was achieved, which
could be considered a good estimation, especially
taking into account the rate of omission errors (37,5%).
The same variables as the previous model were
identified as significant to explain fire occurrence in
these test provinces.
Geographical distribution of the estimations is
included in Figure 10.
10
EUFIRELAB
3.5.2.2 Estimation using linear regression
Linear regression analysis was applied to estimate
the number of large fires in each provincial unit.
Following similar criteria to the logistic regression, a
previous selection of significant variables in each
thematic group (geographical, demographic and
agricultural) was undertaken.
Afterwards, exploratory analysis to the resultant
variables was carried out to assure normal distribution
of the independent variables.
Some were log-transformed.
Criteria for selecting the independent variables were
based on iterative stepwise forward method with
thresholds of significance at 0.05.
The final model was:
NF= 13.467 - 0.273 POBACT90 + 0.161 DIFPOBACT
+ 0.051 BSHCSA - 0.113 DENCAP90 + 0.048
DENOVI90 - 0.060 AGRICULT -0.073 DIFREN,
where
- NF is the number of large fires;
- POBACT90 is the percentage of active population in
each province for 1990;
- DIFPOBACT is the difference in active population
between 1960 and 1990;
- BSHCSA is the proportion of provincial area
covered by KÖPPEN climates Bsh and Csa;
- DENCAP90 is the density of goats for 1990;
- DENOVI90 is the density of sheeps for 1990;
- AGRICULT is the proportion of agricultural cover in
each province and
- DIFREN is the difference in agricultural renters
between 1960 and 1990.
The RMS error of this estimation was 5.4 fires.
3.5.2.3 Estimation using artificial neural networks
The predictive power of artificial neural networks
(ANN) was compared with logistic regression to
estimate large fires ocurrence (CHUVIECO et al. 1998
and 1999, CARVACHO 1998).
The independent variables were the same (human
factors, climatic units, vegetation classes and
topography) used in the regression models.
Also, the same provinces randomly selected to fit
the logistic regression were used to train the ANN.
Then, the trained model was applied to the other
40% of the provinces, studying the predictive capacity
of this technique in comparision with logistic regression.
Figure 10 includes the spatial assessment of both
predictions., where is displayed the estimated versus
observed provinces with fires/no fires.
Table 5 includes the error matrix of ANN analyses.
In the final map (figure 10), red and yellow colors
indicate correct predictions (either estimated or
observed or non estimated and non observed), while
orange and magenta the errors.
The most critical are omission errors (displayed in
magenta) where the models did not predict fires in
provinces that were actually affected
D-08-06.doc
After a comparative analysis, the omission errors
(the most important ones) were higher for the logistic
regression, having also lower global accuracy than
ANN.
The network work better for typical Mediterranean
fires (those occurring at summer time), while it failed in
those provinces affected by winter fires (provinces
located in the North of Spain, Greece and Italy, and
some in South of France).
Errors from logistic regression did not present a
clear spatial pattern, affecting provinces of different fire
characteristics.
In spite of the flexibility and fitting power of the ANN
one of the main drawbacks of this technique refers to
complexity to find out what are the most significant
variables that affect fire occurrence.
In this aspect, the ANN is like a “black box” with few
analytical possibilities to measure the influence of the
independent variables in the estimation.
An indirect method to find out the most critical
variables of the model was undertaken in the Megafires
project.
This method is based on replacing the original
values of each input variable by random values, after
the network is trained.
It was assumed that the increase in the RMS error
produced by such a change should reflect the relative
importance of that variable in the whole fitting.
If the variable was significant, randomising it should
give us a greater RMS with respect to a marginal one.
Repeating this step with each variable showed the
relative importance of all of them.
Randomising Csb-Csc climates, average distance to
the roads, active population in 1990, agricultural area,
differences in population working in the services
between 1960 and 1990 and population density
provided the highest increment in RMS.
Therefore, these should be the most critical
variables in the estimation of fire-not fire.
3.5.2.4 Conclusions
According to the Logistic Regression model the
variables most clearly related to large fire occurrences
was: the proportion of BS-Cs climates, population
density, elevation, unemployment and lack of broadleaf
cover.
The ANN offered a robust estimation of fire
occurrence, but didn't provide as many insights on the
most critical variables.
Estimated fire occurrence maps proved to be a
useful tool for managing global trends of fire risk, by
pointing out those regions, which offer a more
consistent and stable risk of being affected by large fire
events.
Evident difficulties found were related witht the
availability and homogeneity of socio-economic data at
European Scale.
11
EUFIRELAB

3.5.3 Figures

Figure 5: Elevation of theMegafires study area

Figure 6: Fuel type map generated from the Corine Land Cover database

Figure 7: DMSP city lights map of the study area (Source NGDC

D-08-06.doc 12
EUFIRELAB

Figure 8: Occurrence of large fires (above 500 hectares) in the study area 1991-1995).

Figure 9: Actual versus predicted occurrence of large fires by LR techniques 1991-1995)

Figure 10: Observed versus estimated occurrences of large fires in Southern Europe using logistic regression
analysis (top) and neural network analysis (below).
Est-Obs means fires that were both estimated and observed; Nest-Nobs, means both non-estimated and non-
observed; Est-Nobs and Nest-Obs imply commission and omission errors, respectively

D-08-06.doc 13
EUFIRELAB

3.5.4 Tables

Table 2: Coefficients of the final logistic regression model.

Dependent variable occurrence/Not occurrence of large forest fires 1991-1995)
Variable B S.E. Wald df Significance R Exp(B)
LDENSI91 .8349 .2564 10.6024 1 .0011 .1963 2.3045
POBACT90 -.2168 .0500 18.7953 1 .0000 -.2743 .8051
LALTIMED .6631 .2245 8.7219 1 .0031 .1735 1.9408
LDECIDU -.3524 .1572 5.0240 1 .0250 -.1164 .7030
CLIM_BS .0229 .0063 13.1270 1 .0003 .2233 1.0232

Table 3: Observed versus predicted cases

Predicted (cases) % Correct

Observed (cases) 0 1
0 38 25 60.32 %
1 10 88 89.80 %
Overall 78.26 %

Table 4: Error matrix for the estimation of large fire occurrence using logistic regression
(observed versus predicted cases. Test provinces)
Observed
Predicted No Fire Fire Totals Comission error
No Fire 14 15 29 51,72%
Fire 11 25 36 30,56%
Totals 25 40 65
Comission error 44 % 37,5% Global accuracy 60,23%

Table 5: Error matrix for the estimation of large fire occurrence using ANN Observed. Test provinces
Observed
Predicted No Fire Fire Totals Comission error
No Fire 17 12 29 41,39%
Fire 8 28 36 22,22%
Totals 25 40 65
Comission error 32% 30 % Global accuracy 69,23%

D-08-06.doc 14
EUFIRELAB
4 REVIEW OF FIRE RISK ESTIMATION METHOD DEVELOPED FOR THE SPREAD PROJECT
The Spread project (http://www.adai.pt/spread/) is a
European funded project which tried to assess fire risk
conditions at several spatial scales, including the Pan-
European.
Following Spread Deliverable D162 entitled “Fire
risk mapping (II): Validated methods and digital
products covering the whole EU Mediterranean Basin
and selected study sites (local – regional – global
scale)” we present here a summary of the fire risk
integration scheme proposed for the European
Mediterranean area (EUMed).
Acording to previous literature reviews on fire risk
terminology, the Spread project studied the fire risk in
terms of physical probability of fire occurrence
(danger), on one hand, and potential effects caused by
the fire on the other (vulnerability) (CHUVIECO et al.
2003b).
The former refers to the potential that a fire occurs
in a particular area and time on one hand, and to its
propagation capability on the other.
The other component of fire risk is named fire
vulnerability, and concerns the potential effects of a
fire, either on human values and lives and
environmental resources.
The final index should be computed as the product
of the two components
Trying to adopt an approach that may be both
scientifically sound and operationally applicable, the
components that include those two concepts of danger
and vulnerability were adapted within Spread Project to
different spatial and temporal scales, but it was more
focused on the EUMed scale.
D-08-06.doc
The danger component was considered in a broad
perspective, covering the probability of a fuel ignites
(ignition danger) and the potential hazard that this fire
propagates in space and time (propagation danger).
The consideration of both components forms the
wildland fire danger assessment (WFDA) component of
the WFR index defined by the Spread project, as
illustrated in Figure 11.
Therefore, the WFDA is based on the estimation of
two properties: ignition danger and propagation danger.
The former is related to the causal agents of fire as
well as the conditions and properties of the fuel, while
the latter is associated to estimating the behaviour of
the fire.
Fire behaviour is mainly described with reference to
the propagation rate and the intensity of the flame front.
The fire danger assessment system includes an
estimation of fuel moisture content derived from
satellite data (live fuels) and meteorological variables
(dead fuels), as well as an estimation of the historical
patterns of human-caused fires, and the fire
propagation potential, generated from fire-behaviour
simulation programs.
Other factors, such as risk associated to lighting or
flammability, as well as the vulnerability component
could not be derived for the whole EUMed, and was not
addressed in this project.
All the input variables were geographically
referenced and included into a dedicated Geographic
Information System (GIS).
The results were derived for the whole EUMed area
at 1x1 km grid size (with the exception of
meteorological data that were only available at 50x50
km2 grid sizes, and NOAA-AVHRR images with 4.4 x
4.4 km2 pixel size).
The resulting product showed promising potential for
helping fire managers to simulate different danger
scenarios, as well as to obtain a single evaluation of
fire danger conditions for the whole EUMed area size.
A demo was run as a Spread web service during
summer 2004, in which only the fire danger branch
(WFDA) of the framework was implemented and
dynamically updated in quasi-real time during 1 month.
15
EUFIRELAB
4.1 INPUT DATA FOR FIRE RISK MAPPING
Once provided a general explanation of the scheme
adopted in the Spread project, a more detailed
definition of each variable and the data inputs for
generating the index at EUMed scale follows in this
section.
The WFDA is a combination of the following
intermediate maps:
- HDI: Human Danger Ignition (static)
- PD: Propagation Danger (static)
- PI: Probability of Ignition related to fuel moisture
content (dynamic))
Static means that the intermediate map does not
change throughout the season, dynamic means that
the intermediate map will have to be updated every
day. So the final Wildland Fire Danger Assessment
map (WFDA) has to be updated daily.
D-08-06.doc
4.2 HUMAN DANGER IGNITION
The human danger ignition HDI layer was generated
from historical data of fire occurrence.
A probability surface was generated from the
location of ignition points using the kernel density
approach.
Kernel density estimation is based on the estimation
of the density at each intersection of a grid
superimposed on the data, after placing a probability
density (kernel) over each point event (GATRELL et al.
1996, LEVINE 2002).
KOUTSIAS et al. (2002, 2004) introduced it in fire
occurrence for assessing fire occurrence patterns at
landscape level by addressing some of the inherent
positional inaccuracies of the fire ignition locations.
The number of fires observed at community level for
the period 1992-2000 was calculated from national fire
statistics and expressed using the community
centroids.
The adaptive kernel density estimation mode was
chosen due to the non-homogeneous spatial
distribution of community centroids.
The adaptive approach allows for the adjustment of
the bandwidth size in relation to the concentration of
the interpolated points (WORTON 1989).
Locally varying bandwidth size of 10 community
centroids proved to perform best showing a reasonable
variability in the resulting density surfaces avoiding an
under- or over-smoothing.
To avoid over- or under-estimation within and
among countries because of heterogeneities of
different sources the kernel density values within each
country were processed before merging them.
Actually, kernel densities were reclassified to 10
classes based on the equal area criterion within each
country, presupposing equivalence for fire hot spot
areas among the countries
The kernel density interpolation produced
continuous, fire occurrence density surfaces, which in
the Mediterranean context is mainly related to human
fire danger factors, and therefore can be considered as
a static representation of fire danger associated to
human factors thorough a whole fire season.
Figure 12 shows the results of this analysis, which
provides a global view of fire ocurrence distribution,
which is mainly caused by human factors, with higher
occurrences in the NW of Spain and Portugal, SW of
Italy and Greece, the Southern part of the Maritime
Alps, and most territory of Corsica and Sardinia
16
EUFIRELAB
4.3 PROBABILITY OF IGNITION AND FUEL
MOISTURE CONTENT
4.3.1 Fuel Moisture Content of live fuels
The estimation of Fuel moisture content (FMC) of
live fuels was based on an empirical formula derived
from satellite data.
The formula was generated for grasslands and
shrublands based on the analysis of multitemporal
series of NOAA-AVHRR images. (CHUVIECO et al.
2004b)
The index was derived and tested in Central Spain.
For the calibration study, AVHRR images were
acquired daily at the University of Alcala’s HRPT
receiving station, but 8-day composites were created to
avoid cloud and atmospheric contamination.
The compositing criterion was maximizing surface
temperature of the daily series.
The empirical index was based in two different
equations, one for grasslands and the other one for
shrublands (Cistus ladanifer was the species selected
for the calibration phase).
For extending this experience to the whole area
covered by EU Mediterranean countries, it was used
the 10-day NDVI composites of MARS NOAA-AVHRR
images produced by the Joint Research Centre (JRC).
The composites were derived from LAC data
resampled from the original 1.1x1.1 nadir resolution to
4.4x4.4 km2 pixel size.
To maintain the original time sequence used for
developing the empirical formula for FMC estimation, in
the application to the EUMed area the 10-day
composite was updated every 8 days.
In addition auxiliary information on the extents of
grasslands and shrublands in each pixel needed to be
derived.
This information was extracted from the Corine land
cover map of Europe.
Figure 13 shows an example of the FMC generated
from AVHRR images for the whole EUMed area.
4.3.2 Fuel Moisture Content of dead fuels
This map is based on the 10h moisture code of the
USA National Fire Danger Rating System, which
requires a basic set of input meteorological variables
(air temperature and relative humidity).
It was shown that this index could be used to
estimate the average FMC of dead fuels in Central
Spain (AGUADO et al., in preparation), the area used as
calibration site.
Additional testing should be done in other
Mediterranean regions for operational use.
The index was computed every day with weather
observations taken at noon.
Moisture content was expressed as percent of dry
weight.
D-08-06.doc
The meteorological data to compute the index were
extracted from the MARS database maintained at the
JRC.
The spatial resolution of this database is 50x50 km.
The index was computed daily and transferred to
the Internet mapserver designed for the semi-
operational testing of the summer of 2004.
Figure 14 shows the spatial distribution of FMC
estimated from meteorological variables for the same
period as for the figure 13 (in this case, just a single
day, instead of an 8-day period).
As it can be observed, the spatial patterns are
similar in the two figures, in spite of being generated
from two very different sources
4.3.3 Probability of Ignition related to the FMC
This PI index was derived from a combination of the
two previous products.
Both dead and live FMC were converted first to
probability of ignition, based on average values of
moisture of extinction (ME). (CHUVIECO et al. 2004a).
The values of ME are dependent on the fuel
complex, and therefore a fuel type map with a ME
parameter defined for each fuel type is required to
derive this index.
Since no fuel map of Europe with such a
specification is yet available, an estimation of fuel type
distribution was extracted from the CORINE land cover
and a ME value assigned to each fuel type.
Once the FMC of dead and live fuels and their
respective PI were obtained, a global PI for each pixel
was computed following:
PIf = PIlive * Live proportion + PIdeath * Death proportion
Live and Dead proportion expresses the fraction of
live and dead fuel particles in each pixel, which is also
dependent on the fuel type.
As a first approach, this map was again generated
from the CORINE land cover map.
Figure 15 shows an example of the PIf map
computed for the same days as shown in previous
figures.
It should be considered as an integration of the fuel
moisture status of dead and live fuels.
The spatial resolution of the AVHRR data improves
the resolution provided by the MARS meteorological
database, and therefore, geographical patterns are
more evident in the final product.
17
EUFIRELAB
4.4 PROPAGATION DANGER
4.4.1 Average Rate of Spread and Flame Length
In order to obtain a global view of risk associated to
fuel loads, terrain characteristics and wind flows, a
global simulation analysis was performed.
This analysis tried to obtain average values of rate
of spread and flame length, considering different wind
and topographic conditions for the estimated fuel maps
of the whole EUMed area.
This attempt should be considered as a general
overview of average expected fire behaviour at global
scale, in order to rank different danger levels according
to the combination of fuel and terrain spatial patterns.
The estimation of the average RoS was based on
several simulations performed by the Autonomous
University of Barcelona for different fuel types, slope
ranges and wind flows.
As a simulation kernel the wildland simulator Bevins
proposed by COLLINS, which is based on the fireLib
library (COLLINS, 1996) was used.
FireLib is a library that encapsulates the BEHAVE
fire behaviour algorithm (MORGAN et al., 2001).
In particular, this simulator uses a cell automata
approach to evaluate fire spread.
The terrain is divided into square cells and a
neighbourhood relationship is used to evaluate whether
a cell will be burnt and at what time the fire will reach
the burnt cells.
As inputs, this simulator accepts maps of the terrain,
vegetation characteristics, wind and the initial ignition
map.
The output generated by the simulator consists of
two maps of the terrain.
In the first one, each cell is labelled with its ignition
time; in the second one, each cell is labelled with its
flame lenght.
This information must be used to calculate the rate
of spread and an average from among all flame lenght.
To calculate the rate of spread, the distance
between the ignition point and each particular cell in the
terrain is divided by the ignition time of that particular
cell.
This calculation is repeated for each cell in the
terrain to determine the maximum value of the rate of
spread.
This maximum value is used as the rate of spread
for that particular situation.
D-08-06.doc
To provide the propagation danger map, a set of
prototype plots was created, considering all the fuel
models from ROTHERMEL classification and a certain
slope percentage (from 0 to 100%, with a step of 5%).
The total number of plots was 273.
Each plot consists of a grid of cells with 11 columns
x 11 rows (each cell measured 328.083 x 328.083
feet).
The ignition point was located in the middle of the
plot.
For each plot, many input parameter combinations
were used to simulate the wildland fire behaviour and
the average rate of spread and flame lenght were also
calculated.
The parameters considered for variation were: 1-hr
dead fuel moisture, 10-hr dead fuel moisture, 100-hr
dead fuel moisture, live herbaceous moisture and wind
speed and direction.
Average values of rate of spread (RoS, in m/min)
were computed for the different fuel types.
Finally, the RoS values were scaled into a 0-1
range, by normalizing the values between the
maximum and minimum values (figure 16).
Fuel types were derived from the Corine land cover.
The resulting map is considered static, since no
specific conditions are simulated (wind or FMC), but
only general patterns of propagation rates.
Same process as above was performed to compute
the average flame length (FL), measured in metres and
normalized into a 0-1 scale (figure 17).
4.4.2 Propagation danger versus RoS and FL
It derives from the combination of the two
intermediate products previously described: RoS and
FL.
The results of the simulations were mapped at
EUmed scale using CORINE land cover (reclassified
into fuel models) and slope maps
The maps of FL and RoS were then normalized
using linear fitting and multiplied to produce PD:
PD = [(RoSi – RoSmin)/(RoSmax - RoSmin)+0.001] *
[(FLi – FLmin)/(FLmax - FLmin)+0.001]
A small constant (0.001) was added to avoid zero
multiplication in case of minimum values. RoS and FL
were considered in this formula of equal importance,
although this could be tuned up in future improvements,
according to further experience or suggestions.
This map is taken as static, i.e. it will not change
throughout the fire season.
18
EUFIRELAB

4.5 WILDLAND FIRE DANGER ASSESSMENT
The final index of Wildland Fire Danger Assessment
WFDA proposed in SPREAD integrates the three
components previously described, the human danger
index (HDI), the probability of ignition associated to fuel
Status (PIf) and the Propagation Danger (PD), as
follows:
WFDA = (HDI+c) * (PIf+c) * (PD+c)
4.6 FIGURES
- HDI and PD are static, they do not change
throughout the season,
- c is a constant to avoid multiplying by zeros (0.001
is used),
- PIf changes every 8 days with respect to live fuels
and every day with respect to dead fuels.
Therefore the resulting map of WFDA has to be
updated daily. An example is shown in Figure 18,
derived using the maps depicted in Figures 12, 15, 16
and 17 for August 10th 2004
All the EUMed maps mentioned so far were made
available as a demo product in quasi-real time (with 2-3
days delay) within the demo of the Spread Project web
based services.

Intentional
Human
Ignition sources Unintentional

Natural: Lightning
Ignition Danger
Dead
Moisture content
Fuel condition Live
Wildland Fire Flammability
Danger
Assessment

Fuel properties
Propagation
Danger Wind flows

Topography
Figure 11: Structure and components of the Wildland Fire Danger assessment (WFDA).
Red boxes mean components that have not been finally implemented in the project

Figure 12: Fire ignition danger in Southern Europe: an estimation of human fire danger factors spatial distribution

D-08-06.doc 19
EUFIRELAB

Figure 13: FMC of live fuels for the first week of August, 2004

Figure 14: Estimation of FMC for dead fuels. Map from 10th August 2004

Figure 15: Probability of ignition from FMC of 10th August 2004

D-08-06.doc 20
EUFIRELAB

Figure 16: Estimated average Rate of Spread (normalized values from 0 to 1)

Figure 17: Estimated average Flame Length (Normalized values from 0 to 1)

Figure 18: Wildland Fire Danger Assessment for August 10th 2004.

D-08-06.doc 21
EUFIRELAB

5 REVIEW OF THE EUROPEAN FOREST FIRE INFORMATION SYSTEM-RISK FORECAST

5.1 INTRODUCTION
The fire risk module of EFFIS runs on operational
(EFFIS-Risk Forecast) developed BY the JRC (Joint basis 6 meteorological danger indices (Portuguese,
Research Centre) of the European Commision Spanish, Sol Numerical Risk, Italian, Canadian FWI
and Behave fine fuel moisture).
The European Commission DG Joint Research
In addition a Fire Potential Index (FPI) is computed,
Centre set up since 1999 a research group to work
by integrating:
specifically on the development and implementation of
- forecast meteorological data (to estimate fuel
advanced methods for the evaluation of forest fire risk
moisture content),
and mapping of burnt areas at the European scale.
- satellite data (to estimate the relative fraction of live
These activities led to the development of the
fuels from Relative Greenness) and
European Forest Fire Information System (EFFIS).
- a fuel map (to estimate the fuel properties).
Since the year 2003 EFFIS is part of the Regulation
In this case maps with a spatial resolution of 4.4 km
(EC) No 2152/2003 (Forest Focus) of the European
are generated.
Council and Parliament on monitoring of forests and
Daily EU maps are produced from May 1st to
environmental interactions.
October 31st (from 2006 it will be from February 1st)
All the EFFIS activities are coordinated with DG
processing 0-24 hours and 48-72 hours weather
Environment to reach the final users, Civil Protection
forecast data.
and Forest Services, in the Member States
Together they make the core of the currently named
(http://inforest.jrc.it/effis/).
EFFRFS.
EFFIS is aimed to provide relevant information for
Expected forest fire risk level is mapped in 5 classes
the protection of forests against fire in Europe
(very low, low, medium, high, and very high) providing
addressing both pre-fire and post-fire conditions.
1 to 3 days risk forecasts over Europe with an average
On the pre-fire phase, EFFIS is focused both on the spatial resolution of about 40 km.
development of systems to provide forest fire risk
Once the indices are computed, they are distributed
forecast based on existing fire risk indices, and on the
to the civil protection and forest fire services via
development of new integrated forest fire risk indicators
Internet.
(EFFIS - Risk Forecast).
For all indices, maps of past days or averages for a
These indices permit the harmonized assessment of
given historical period can also be generated,
forest fire risk at the European scale.
specifying either a date or a time interval from 1st of
They may be used as tools for the assessment of
May 1st to October 31st
risk situations in cases in which international
(http://inforest.jrc.it/effis/viewer/viewer.html).
cooperation in the field of civil protection is needed.
Currently, the dynamic forest fire risk forecast The long term as well as the advanced integrated
indices are available on the EFFIS web site and sent to approach that will complet the risk assessment system
the Member States Services daily from the 1st of May are currently under further development and are, at the
until the 31st of October. moment, for internal use.
They will be described here as far as the methods,
even at an exploratory research phase, have been
proposed and developed at European scale in several
publications.

D-08-06.doc 22
EUFIRELAB
5.2 METEOROLOGICAL DANGER INDICES
Dynamic indices are short-term indices that assess
the probability of fire ignition and spread.
These factors may be derived directly from
meteorological variables, or indirectly by the effect that
these variables have on vegetation.
Indices that are computed from meteorological
variables are referred to as meteorological fire risk
indices.
On the other hand, the indices that evaluate the
status of the vegetation are the so-called vegetation
stress fire risk indices.
As already mentioned, there is not a consensus on
which meteorological fire risk performs best for the
whole Mediterranean region, although several studies
show that the Canadian Fire Weather Index is well
suited for the estimation of fire risk for the region (SOL,
1999).
Accordingly, several indices operationally used in
southern Europe are computed in the EFFRFS system.
The algorithms in the EFFRFS are those of a
system referred to as EUDIC, (BOVIO and CAMIA 2000)
although the actual computation of the fire risk indices
is now implemented in a distributed geo-database that
enables the on-line integration of all the model inputs.
The indices were initially computed from data
collected from a network of meteorological stations
covering Europe in which data are collected at the
station level, and further interpolated to a 50 km by 50
km grid.
At present, these indices are computed from
forecast data, and fire risk indices are produced as
forecasts for one, two and three days.
Six indices are being computed:
- Behave (Rothermel et al. 1986, VAN WAGNER 1987)
based on estimation of the moisture content of fine
dead fuel.
- Canadian Fire Weather Index (FWI) (VAN WAGNER
1987)made up of six normalized indices that
indicate the daily variation of fuel water content,
initial rate of propagation, and quantity of fuel and
expected intensity of the flame front.
- Portuguese index (GONÇALVES and LOURENÇO
1990) assesses atmospheric conditions in the
proximity of the fuel layer.
- Spanish ICONA method measures probability of
ignition (ICONA 1993)based on litter and fine dead
fuels moisture content.
- Sol’s Numerical Risk (SOL 1990, DROUET and SOL
1993) related to ignition and propagation.
- Italian Fire Danger Index (PALMIERI et al. 1992)
derived from MC ARTHUR's model.
European statistics on forest fire data are used for
the calibration and validation of these indices.
All the indices are computed for southern Europe;
however individual civil protection agencies will be able
to download the indices for its own territory.
D-08-06.doc
5.3 FPI MODEL1.
5.3.1 General presentation
It is a risk index based on the Fire Potential Index
created with BURGAN et al 1998), and integrated
through the combination of diverse types of static and
dynamic variables from different data sources like
meteorological stations, satellite images..
This index has a double structural and dynamic
component.
The structural component constitutes the
information about fuel properties, which is usually used
by indices at local and regional scale, but not at global
scale, due to the difficulty to elaborate cartography of
fuel types at this scale.
The dynamic perspective is offered by the
meteorological variables and the remote sensing data,
that serve to characterize the hydrological state of the
dead fuel and the alive fuel presence.
The main assumption behind the FPI model is that
“fire potential” can be assessed if the moisture content
of live and dead vegetation is reasonably represented
(BURGAN et al 1998).
Therefore, two are the key factors in this model as
can be seen in Figure 19:
1) live and dead fuel loadings,
2) dead fuel moisture content..
In the original FPI model, fuel loadings are obtained
from a fuel model map, while dead fuel moisture
content is estimated using meteorological data.
Satellite data are used to discriminate between
living and “cured” fuel at the pixel level.
Thus, the model requires the following data inputs:
- A fuel map to define the dead and live fuel loads,
and the extinction moisture values for each fuel
type.
- Maximum Value Composites (MVC) of NDVI
(Normalized Difference Vegetation Index) to
calculate the Relative Greenness (RG); the purpose
of the RG in the model is to provide seasonal
adjustment in the proportion of the fuel that is live. In
other words, in the FPI model, the proportion of live
load varies by pixel, as a function of the assigned
fuel model, and of the relative greenness. (BURGAN
and HARTFORD 1993)
- Meteorological data to estimate the moisture content
of the small dead and cured fuels.
Applying the original FPI methodology was
conceptually possible, but neither the input formats nor
the spatial resolution of the data were comparable to
those used in the US.
Consequently, several modifications of the model
were tested.
The main differences between the original FPI and
its European counterpart concern the input data and
the way these data were processed (e.g. formats,
interpolation method, and spatial resolutions).
1
Based on Sebastián et al (2002).
23
EUFIRELAB
Another important difference is connected to the use
of the fuel map, which in the original model is meant to
provide a live-ratio and an extinction moisture value for
every pixel.
To run the model, a fuel map had to be created,
since such a map did not exist at a European scale.
5.3.2 Input data
The temporal resolution of the FPI index is one day.
All the input data have this temporal resolution, with
the exception of the RG, which is computed with 10
days Maximum Value Composites (MVC) of NDVI daily
images.
To avoid the uncertainty of pixel location, a 4.4 km
pixel window was used when deriving the MVC from
which the NDVI is computed (NOAA AVHRR original
spatial resolution is 1.1 km).
Other inputs, in particular, the CORINE Land Cover
database (CLC), in its raster version, have a higher
spatial resolution (0.1 km).
All the interpolated meteorological data have a lower
spatial resolution (50x50 km).
It was decided that the resolution of the NDVI data
would be used as the base resolution of the input data.
This implied a compromise between the higher
resolution of CLC and the lower resolution of
interpolated meteorological data.
The FPI would then be derived with a 4.4 km spatial
resolution. To match
the spatial resolutions, pixel values were replicated
for the case of meteorological data, and a two-step
generalisation was performed in the case of the fuel
map, which was initially derived with the same
resolution as the CLC, i.e. 100 m pixels.
5.3.2.1 Satellite data
At the Joint Research Centre (JRC) a satellite-
tracking antenna receives daily NOAA AVHRR HRPT
data. Within JRC, these images are pre-processed up
to the computation of the NDVI.
Finally, the images are geometrically corrected, and
an overall mosaic of Europe is created at a spatial
resolution of 4.4 x 4.4 km.
Further information on process can be found in
Kerdiles 1997). For the European FPI, historical NOAA
–AVHRR mosaics processed at JRC were used in
several ways:
- First, daily cloud masks were extracted from the
NOAA images and used to derive a daily cloudiness
percentage (%) grid for the study area.
- Second, 10-day Maximum Value Composites (MVC)
of NDVI historical images were used to compute
Relative Greenness (RG) (Burgan and Hartford
1993)
- Finally, the historical NDVI maximum value at a
given location was used along with the overall
historical maximum to parameterise the live-ratio
values, as explained below.
D-08-06.doc
Both the RG and the live-ratio were kept at the 4.4
km spatial resolution of the NDVI images from which
they were derived.
5.3.2.2 Meteorological data
The moisture content of small dead fuels is
considered the major factor in determining fire
ignition.In this model, small dead fuels are represented
by the Ten-Hour Time Lag Fuels.
The moisture content of small dead fuels (Fm10hr)
is determined as a proportion of the equilibrium
moisture content (Emc) of the surrounding atmosphere,
as follows: (FOSBERG and DEEMING 1971)
Fm10hr = 1.28 x Emc
The Emc determines the amount of water vapour
that a given piece of wood can hold. (SIMARD 1968).
In order to compute Fm10hr for a given day it is
necessary to use values for the air temperature,
relative humidity, precipitation, and percentage of
cloudiness for that day.
The two last variables are needed to correct for
solar heating and rainfall respectively.
Except for the percentage of cloudiness – derived
from NDVI images - these meteorological variables
were extracted from the MARS meteorological
database located at the JRC (TERRES 1999).
Once this institution receives data at the station
level, then theses data are spatially interpolated into
50x50 km grids according to the CGMS procedure.
This procedure takes the following characteristics
into account: distance, difference in altitude, difference
in distance to coast, and climatic barrier separation
(VAN DER GOOT 1997).
The 50x50 km pixel size was imposed by the low
density of meteorological stations, particularly in some
areas of southern Europe.
More specifically, to compute the meteorological
inputs of FPI three interpolated variables were
extracted from MARS: maximum temperature of the
day (°C), mean daily vapour pressure (hPa), and total
daily rainfall (mm).
The daily rainfall is the sum of precipitation between
6 UTC on day D and 6 UTC on day D+1.
The interpolated actual vapour pressure and the
interpolated maximum air temperature were used to
compute the relative humidity - which was not directly
available from the database- previous computation of
the saturation vapour pressure.
24
EUFIRELAB
5.3.2.3 Derived data sets
5.3.2.3.1 Fuel types map
In Europe, fuel type maps exist only at the national
or local scales in some countries.
It was thus necessary to derive a European Fuel
types map in order to be able to compute the FPI.
This fact implied the need for a method to link the
ecological aspects of the forest fuels with the fire risk
determinants.
The chosen approach required obtaining, from a
land cover classification, a general framework for
presenting the land cover regions.
Using GIS tools, these land cover regions were
stratified into sub-regions according to phyto-
sociological criteria, which accounted for the floristic
composition and other factors governing the distribution
of the vegetation.
Two databases were thus cross-referenced: the
European classification Corine Land Cover (CLC) and
the Map of Natural Vegetation of Europe (MNVE)
(Conseil de l’Europe/Commission des Communatés
Européennes 1987).
The CLC has a fairly good spatial resolution, with a
minimum mapping unit of 25 ha.
However, thematically the CLC can be considered
relatively poor, since only 44 classes are distinguished.
On the other hand, the MNVE has a lower spatial
resolution (mapping scale of 1:3000.000) but depicts
more than 100 Vegetation associations (V-A).
These associations delimit ecological areas that are
relatively homogeneous, characterised by the
predominance of natural or sub-natural primary
vegetation.
First of all the CLC was used to mask out
agricultural and not vegetated land.
Then the remaining land cover classes were
subdivided into different spatial sub-regions according
to the existing vegetation associations.
Next, the relationship between these Land-cover /
V-A units and the standard National Fire Danger Rating
System (NFDRS) (DEEMING et al. 1978)fuel types was
investigated. Finally, the NFDRS fuel model key was
then used to assign to each identified sub-region a fuel
type depending on the characteristics of the
predominating understory.
In this process there are several issues that are
worth to stress:
- The MNVE identification of the vegetation
associations is based on knowledge of the
ecological factors that govern the vegetation
distribution (mainly climate and soil). This
information is also included in the MNVE data set
along with the description of each V-A, and was also
taken into account when assigning the fuel types
with the help of the fuel model key.
- During this process the CLC is essential, because
the MNVE depicts potential vegetation associations
that sometimes do not coincide with existing ones.
In this case the CLC is given priority and the fuel
type is assigned directly to the land cover class.
D-08-06.doc
Overlay of the two data sets was done at the CLC’s
spatial resolution, resulting in a 100 m spatial resolution
fuel map, which was subsequently generalised to 4.4
km resolution using a two-step procedure.
Notice that this approach has similarities with:
- other global land cover/vegetation mapping
experiences in which ground truth data was not
used, such as the 1-km Global Land Cover
Characteristics database. (LOVELAND et al. 1991),
- the first phases of other experiences in which
ground data was only used in the last phases of the
labelling process.
This is the case, for instance, of the NFDR Fuel
Model Map. (BURGAN et al. 1998)
5.3.2.3.2 Live-ratio map
Following the original FPI methodology, the
maximum Live-ratio was initially obtained from a Fuel
types map.
However, the building of such a map, particularly
when fieldwork is not feasible, entails some subjectivity.
In order to address this problem, it was proposed
(BURGAN, 1999) the use of a maximum live-ratio map
derived through a re-scaling of NDVI historical
maximum values in each pixel.
The maximum NDVI values were computed for a
five-year interval 1994-1998), and the live-ratio was
then derived as:
Live-ratio=0.25 + 0.50 x (NDVImax / NDVIabsolute-max)
Where
- NDVImax represents the maximum NDVI for a given
location,
- NDVI absolute-max is the overall maximum NDVI on
any location in the NDVI mosaic.
The computed live ratio values are scaled between
0.25 and 0.75, in accordance with the live-ratio values
defined for the NFDRS.
5.3.2.3.3 Extinction moisture map
In the FPI model the extinction moisture (ExtM)
value of a pixel is deduced from the fuel type assigned
to that pixel.
However, the ExtM content of broad vegetation
types is normally found in the literature and so it was
reasoned that the direct assignment of ExtM values
was simpler, and could be closer to reality, than their
indirect assignment through the Fuel types map.
Thus it was decided to derive an Extinction moisture
map by assigning an ExtM value to each identified sub-
region.
Table 6 shows the CLC classes considered in the
analysis and the fuel types assigned to them.
The fuel type before the brackets indicates the
general assignation.
Fuel types between brackets correspond to the
different identified sub-regions (therefore they depend
on the predominant vegetation association).
The fourth column shows the initially assigned ExtM
values (those from the NFDRS).
Last column shows the modified extinction moisture
values.
25
EUFIRELAB
To obtain this table, several sources of information
were consulted.
These comprised the extinction moisture values in
the Behave and the NFDRS fuel models.
All the available documentation was consulted and
the conclusions derived discussed with experts until the
ExtM value was agreed.
These conclusions were used to revise the initially
assigned ExtM values, which were only changed in
case of contradiction with the findings of the experts.
Table 6 enables the comparison between the ExtM
values initially used, and new ExtM values assigned
when building the ExtM map.
It can be seen in this table that the use of the ExtM
map simplifies the European FPI model.
5.3.3 Computation of FPI
Once the necessary inputs were derived, following
the original methodology (Burgan et al. 1998), the Fire
Potential Index was computed as:
FPI = 100 x (1-Fm10hcorrected) x (1-Lr)
Where Fm10hcorrected is the moisture content of the
small dead fuels corrected by the solar heating and the
extinction moisture content, and Lr is the Live-ratio
corrected by the Relative Greenness, i.e. the part of the
load that is live.
FPI values range 1 to 100, with 1 meaning “no fire
risk” and 100 meaning “highest risk”.
For operational purposes this range has been
simplified to 5 risk categories from very low to very
high.
The FPI map calculated for the EU is avaliable at
the web site
http://inforest.jrc.it/effis/viewer/viewer.html
D-08-06.doc
5.4 LONG-TERM INDICES
Long-term fire risk indices are those computed from
variables that do not change in a short lapse of time.
Consequently, these maps are updated with a
yearly (or longer) frequency.
Two types of long-term indices could be included in
the EFFRFS.
The first one provides the probability of fire
occurrence, while the second one indicates the
likelihood of damage to a forested area in terms of
economic or environmental losses.
5.4.1 Probability of Fire Occurrence2.
This index intends to evaluate the probability of fire
ignition, rather than the propagation of the fire.
Fire propagation is more related to the dynamic
indices that have been presented in previous sections.
Aspect was included in the model as representative
of the topographic conditions that may influence fire
ignition.
It is related to the type and condition of the available
fuels for ignition.
Since the influence of socio-economic factors
(human factor) is difficult to model, fire recurrence was
introduced in the model as a surrogate of these factors.
Fire statistics for southern Europe show that the
probability of fire occurrence is higher in those areas
that have historically suffer a high rate of fire incidence.
The method developed is based on multi-variant
regression techniques.
Two sources of data of the European Commission
have been used: on the one hand, the database of fires
of the Main directorate of Agriculture; on the other
hand, diverse data bases of the statistical office,
EUROSTAT.
37 independent variables was extracted Of these
two sources of data (see Table 7), including in three
thematic fields: economic, social and environmental
statistics.
All of the variables were chosen because of their
likely relationship with the forest fires phenomena in the
Mediterranean area as described in the literature on
forest.
Some of them, like the fuel types and the
topographic variables, account directly or indirectly for
the hazard.
The others are deemed to somehow affect the
probabilities of ignition.
The third thematic group would integrate the
information relative to fuel types and the fisiographics
variables; the rest would be socio-economic aspects.
In this model the dependent variable is Annual
average of fires, weighted by province (or departement,
in France).
The independent variables were not homogeneous,
neither spatially nor temporally, for all the countries in
the Mediterranean basin, and therefore it was decided
to derive two models.
2
Based on (Sebastián et al. 2001 and Sebastián, 2004)
26
EUFIRELAB

The first model (Model 1) was built for the five EU 5.4.2 Likely Damage.
Mediterranean countries with the set of 32 variables
There are natural areas that are of particular
that ensure the homogeneity in the temporal and
interest.
spatial resolution for the whole study area, i.e.:
This may be due to many different reasons, from the
Portugal, Spain, France, Italy and Greece.
purely economic value of the timber to the unique
Other variables considered important in explaining
environmental qualities of the area.
the fire regime in the Mediterranean, such as surface of
The aim of the likely damage index is to provide a
permanent grasslands, agrarian subsidies, and some
method to highlight areas that should be strongly
economic statistics, were only available (at the required
protected from forest fires.
resolution) for Spain, France and Italy.
Hence, a second model (Model 2) was built for this In the Mediterranean region of Europe the most
reduced region using a total of 37 explicative variables. important quality of most forested areas is their
environmental value.
Concerning the methodology, random sampling
Timber production is usually a secondary asset of
techniques were used to divide the original sample into
these forests.
two subsamples.
An added condition of these forests is the intrinsic
One was used to fit a linear regression model, and
difficulty for regeneration due to the lack of rain and the
the other to validate it.
fragility of soils.
The original sample for Model 1 consisted of
323 observations (i.e. provinces), from which 194 were The likely damage was estimated by assigning to
used to fit the model, and 129 to validate it. each cell a vulnerability degree.
The sample size for Model 2 included
The vulnerability index considers the likely damage
245 observations.
From this sample, 156 observations were used to fit that a fire can cause in a specific area.
the model and 89 to validate the fit. This evaluation can be critical in areas of special
ecological value, in susceptible zones of erosion or
The sampling process was carried out 10 times for prone to the alteration of the hydric balance, and in
both models leading to 10 different and independent areas close to human settlements.
samples to fit the models and 10 corresponding This index considers three factors:
subsets for their validation. - The potential erosion, obtained from the land cover
type, the slope and the regime of rains.
The variable selection was performed by searching
- The protection level of a specific zone, that
for the best intermediate model (sets of explicative
considers its rareness nature, its fragility and its
variables) for each of the obtained sub-samples.
environmental interest.
The information derived from these regressions was
- The distance to human settlements, that considers
then used to fit the final, definitive models.
the human lives and the properties in danger.
The intermediate models were derived using robust
regression, which provides suitable fits even when To derive the mentioned variables, data from the
outliers are present. Eurostat’s database were extracted and processed in a
GIS environment.
An ordered list of the variables that resulted
All the long-term fire risks were normalized between
selected in the final Model 1 (all southern European
the values 0 and 100.
countries) can be found in Table 8; and final Model 2
This range was further divided into five fire risk
(Spain, France and Italy), in Table 9
classes from very low risk to very high risk.
An important consistency was found between the An example of the likely damage index showing the
two models derived. levels of risk and the area for which long-term indices
This supports the thesis of the selected variables as are computed is presented in Figure 21.
the better explicative variables of the forest fire
The two types of long-term indices (probability of fire
phenomena at regional (Mediterranean) scale. The
and likely damage) could be integrated into a single
validation of the models was done through the analysis
index.
of their predictive ability.
This so-called integrated long-term index would
Analysis of the spatial autocorrelation was identify areas that are jointly subject to suffer forest
performed to dextermine whether the selected models damage and high potential losses.
take into account the spatial structure of the dependent The index would help forest fire services locate
variable. those areas to which the highest level of protection
should be given.
6 variables were coincident between the two
models: two types of land cover (presence of shrub and
mixed forest), two variables relative to the agrarian
production (cattle and cereal production), a socio-
economic variable (unemployment) and another
fisiographical one (altitude).
Figure 20 presents an example of a fire probability
map for southern Europe.

D-08-06.doc 27
EUFIRELAB

5.5 FIGURES

Figure 19: Structure of FPI model calculation

Figure 20: Example of a fire probability map for southern Europe (San Miguel-Ayanz et al. 2002)

Figure 21: Example of a likely damage map for southern Europe (San Miguel-Ayanz et al 2002)

D-08-06.doc 28
EUFIRELAB

5.6 TABLES

Table 6: Proposed ExtM values for the considered CLC classes.

CLC Description Fuel type Live ratio Initially assigned Ext. New assigned
assigned Moist. (%) Ext. Moisture
Permanent grassland L 0.67 15 15
Agriculture with natural vegetation T 0.67 15 20
Agroforestry areas L(C) 0.67 15 (20) 15
Broad-leaved forest R ( O,F, T ) 0.40 25 (30,15,15) 30
Coniferous forest T (D,H,C, P,A) 0.67 15 (30, 20, 20, 30, 25, 15) 25
Mixed forest R (O,P,C) 0.40 25 (30, 30, 20) 25
Natural grassland L(A) 0.67 15 (15) 15
Moors and heathland O 0.53 30 30
Sclerophillous vegetation F 0.60 15 15
Transitional woodland-scrub T 0.67 15 15
Sparsely vegetated areas A(S) 0.60 15 (25) 15
Inland marshes N 0.40 25 25
Peat bogs S 0.50 25 25
Salt marhes N 0.40 25 25

Table 7: Independents variables used in the study (Sebastián et al., 2004).

Type
Agrarian production
Economic Variables
Cattle production
Production by economic sectors Secondary sector
Unemployed index
Social Variables
Population variables
Fisiographic variables
Environmental Areas
Fuel types
Variable
Cereal production
Wheat production
Barley production
Corn production
Leguminous production
Agrarian subsidies
Agricultural total production
Permanent grassland area
Agricultural total area
Goat production
Bovine production
Ovine production
Addition goat and ovine production
Primary sector
Tertiary sector
> 25 years old
< 25 years old
Population density
Settlement > 20.000 inhabitants
Protected areas
Road density
Altitude
Slope
Orientation
Land cover type 1 (%)
Land cover type 2 (%)
Land cover type 3 (%)
Land cover type 4 (%)
Land cover type 5 (%)
Land cover type 6 (%)
Land cover type 7 (%)
Land cover type 8 (%)
Land cover type 9 (%)
Land cover type 10 (%)
Land cover type 11 (%)
Land cover type 12 (%)

D-08-06.doc 29
EUFIRELAB

Table 8: Selected variables in Model 1 (Sebastian et al., 2001).

Variables in Final Model 1 Reg. Coeff.
Ln(Percentage of land occupied by Shrubs) 0.1195
Ln(Percent. of land occupied by Mixed Forests) 0.0973
Unemployment 0.0944
Ln(Surface of Permanent grasslands) -0.0936
Altitude -0.037
Ln(Percent. of land occupied by Broad leaves) 0.033
Total Cereal Production -0.0352
Total Cattle Production 0.0255

Table 9: Selected variables in Model 2 (Sebastian et al., 2001).

Variables in Final Model 2 Reg. Coeff.
Ln(Percentage of land occupied by Shrubs) 0.3775
Ln(Percent. of land occupied by Burnt areas) 0.2387
Total Cereal Production -0.2615
Ln(Percent. of land occupied by Mixed Forests) 0.2168
Altitude -0.2115
Total Goats production -0.1525
Ln(Perc. Land occup. by Mixed Forest- Agriculture) 0.1154
Total Cattle production -0.1095
Unemployment 0.1087

D-08-06.doc 30
EUFIRELAB
6 BASIC STRUCTURE AND CHARACTERISTICS OF THE PROPOSED EURO-MEDITERRANEAN
WILDLAND FIRE RISK INDEX
6.1 GENERAL PRESENTATION
Fire risk assessment is routinely performed in most
developed countries that are affected by forest fires.
Estimating forest fire risk involves identifying the
potentially contributing variables and integrating them
into a mathematical expression, i.e. an index.
This index, therefore, quantifies and indicates the
level of risk.
A literature review of forest fire risk methods has
shown how different approaches are used for the
evaluation of fire risk.
The most relevant are producing updating
evaluation of fire danger conditions over the Internet.
Most of these operational systems currently rely
mainly on meteorological data (see section 2.3), which
is considered the most critical variable for fire ignition
and propagation, while other factors of risk are less
considered or they are introduced in an experimental
way.
Fuel loads and fuel water state is commonly
considered, but the difficulties of obtaining accurate
information on both variables are recognised.
Therefore, most commonly fuel loads are roughly
estimated from land cover maps, and fuel state is
associated to weather variation.
Additionally, human factors of fire risk are rarely
considered, in spite of being considered a critical
variable in fire ignition (MARTELL et al., 1987).
In addition to this lack of critical variables in fire
danger estimation, another weakness of current fire risk
assessment systems is related to the concept of risk
itself.
Traditionally fire managers define fire risk
considering the chance of starting or spreading a fire in
a particular time and space.
According to FAO’s terminology (FAO 1986) forest
fire risk is even more restricted, since is defined as “the
chance of a fire starting as determined by the presence
and activity of any causative agent”.
However, other natural hazards evaluate risk
considering not only the potential occurrence of an
event, but also –and more important in many cases-
the potential damage which that event would cause on
persons and values.
From this point of view, the traditional concept of fire
risk, which is still active in most countries and regions,
is not considering a critical component of risk, which
affects the susceptibility to fire of the potentially
affected region.
As it is well known nowadays, fire is a natural factor
in many ecosystems, and therefore, from the
ecosystem point of view, the impact of a given wildfire
could be relatively contained, being within the limits of a
normal natural disturbance.
Fire management should greatly benefit from having
the possibility to consider in the decision making
process the importance of the expected potential
damage.
D-08-06.doc
The assessment of the importance of the expected
damage is closely related to the concept of
vulnerability, which can be defined as the potential
damage caused by the natural/technological disaster
on persons and values (ecological, economic, etc.).
Consequently, an integrated assessment of fire risk
should consider both likely occurrence of fires, as well
as their potential damages (Figure 22).
The Eufirelab unit 8 intends to propose a coherent
and integrated scheme for fire risk assessment, which
would consider a wide range of risk variables, both
associated to fire ignition and propagation, as well as to
fire vulnerability.
Such a proposal will be based on previous
experiences which have shown good potential for risk
assessment.
As previously stated, most operational fire risk
systems are focused on physical factors: weather data
and fuel status mainly.
However, especially in Europe, a growing concern
about the role of human activities in fire ignition and
propagation is widely recognized (KALABOKIDIS et al.,
2002; MARTÍNEZ et al., 2004; VÉLEZ et al., 2002).
Therefore, any risk assessment system should also
include the consideration of socio-economic causes of
fire, as well as the vulnerability associated to human
beings (life, properties and values).
The proposed scheme will include these factors.
Another aspect to be taken into consideration to
design an integrated risk assessment scheme regards
the need of having consistent evaluation of risk at
different temporal and spatial scales.
The great diversity of natural and socio-economic
conditions strongly complicates this task, which is
critical if risk conditions need to be objectively
compared throughout Europe.
Obviously, any risk assessment system based on
physical and human variables need to be adapted to
local conditions, but a hierarchical design may help the
use of a common scheme at different scales, and for a
broad range of social and natural conditions.
On the other hand, Fire risk, as any other natural
hazard, is not static, but needs to be updated regularly.
The temporal updating would strongly depend on
how dynamic the different risk variables are.
Some would require a very short updating period
(such as weather data), while others only need revising
in few months-years (fuel types), or they can be
considered static (elevation, slope).
Based on the indices revised in the preceding
section and considering the availability of spatial
information at the required scale (see previous
deliverables of the unit), Eufirelab unit 8 will aim to
define the scheme for fire risk integration.
It includes the definition of the risk variables, as well
as the integration of those input variables in synthetic
indices for fire risk mapping at European scale.
31
EUFIRELAB
The system will include the assessment of Danger
and Vulnerability (Figure 23).
The former is related to the probability of fire
occurrence and propagation, while the latter is
associated to the potential effects of fire on human
beings and natural values.
The proposed system includes the estimation of two
danger properties: ignition danger and propagation
danger.
The former is related to the causal agents of fire
(excluding lightning because no operational way to
model this factor has been produced) as well as the
conditions and properties of the fuel (live and dead),
while the latter is associated to estimating the
behaviour of the fire.
In this deliverable will be defined the two
components of ignition danger (causal agents and fuel
properties).
The propagation danger is related to fire behaviour
(rate of spread and flame length).
The Danger component requires considering
weather data, fuels and socio-economic data that
should be derived from spatial analysis of point data
(both measured and forecasted), as well as remote
sensing images.
More details can be found in following sections.
D-08-06.doc
As far as vulnerability assessment concerns, the
main potential effects to be analysed have been divided
into ecological and socio-economic factors.
Due to the small number of experiences using this
component in forest fire risk index at global scale, a
preliminary study of the variables to be included in a
vulnerability index will be undertaken in this deliverable.
The socio-economic component of fire vulnerability
will be focused on the potential damages on persons
and properties.
The economic value (in different colour in the
scheme) will not be addressed in this deliverable
because the great difficulty to create this information at
European scale.
The ecological component includes potential soil
erosion after a forest fire and the environmental value
of the territory.
The system will be designed in order to be
implemented at European scale.
The target resolution will be of 1 km².
The supra-national perspective, which will be
proposed for the EUMed countries, should potentially
contribute to the improvement of the European Forest
Fire Information System (EFFIS), currently run at JRC.
32
EUFIRELAB

6.2 FIGURES

Figure 22: Integrated Fire Danger Rating (after ALLGÖWER et al., 2003)

Live fuels (satellite)

Fuel
moisture
Ignition Danger Dead fuels (meteo)
Index

Probability of ignition Human

(human factors) ignition

Propagation
EM - WFRI Danger Index

Population exposition

Vulnerability Environmental value

Index
Economic value

Potencial Soil erosion

Figure 23: Scheme of the Euro-Mediterranean Wildland Fire Risk Index (EM-WFRI)

D-08-06.doc 33
EUFIRELAB
7 COMPONENTS OF THE FIRE RISK SYSTEM: IGNITION DANGER INDEX
7.1 FUEL MOISTURE
7.1.1 Introduction
The moisture content of fuel is a critical parameter in
fire ignition because flammability is closely dependent
on it (DIMITRAKOPOULOS & PAPAIOANNOU, 2001).
The burning capacity of the leaves and plants is
inversely correlated to the moisture quantity.
For fire ignition danger, fuel could be divided in two
categories: dead and live.
Dead fuels lying on the forest floor (fallen branches,
litter, and foliage) are the most dangerous because they
are drier than live fuels and more dependent on rapid
atmospheric changes.
The moisture content of live fuels has a marginal
role in fire ignition, but it is critical in fire propagation
modelling because the amount of water is directly
related to the rate of fire spread (CARLSON & BURGAN,
2003; SNEEUWJAGT & PEET, 1985; VIEGAS, 1998).
It is not easy to measure the water content of the
leaves and plants in an operational way.
Two main directions are generally used:
- water estimation from soil water quantity estimation,
using methods of water reservoir -the water quantity
of soil is computed from rainfall values (input) and
evapotranspiration (output)
- water estimation from remote sensing
measurements of vegetation aspect. This first way is
generally more or less integrated in the
meteorological indices of fire risk.
Commonly, the estimation of dead fuel moisture
content (FMC) is based on meteorological danger
indices, which attempt to account for the absorption–
evaporation relationships in inert materials (SIMARD,
1968).
The application of spatial interpolation techniques is
required because meteorological data are frequently not
available for fire prone areas.
Applying those indices to live fuel moisture trends
could be more complex because live plants are much
less dependent on atmospheric conditions than dead
materials, given their mechanisms to extract water from
the soil reserve and reduce evapotranspiration.
Recent research have obtained strong nonlinear
relationships between live fuels moisture content and
long-term meteorological codes (CASTRO et al., 2003;
VIEGAS et al., 2001), but their results are species
dependant. In this context, spatial interpolation
techniques could introduce additional noise.
Within the context of fire danger estimation, remote
sensing data have shown good correlations with live
fuel moisture content (CHLADIL & NUNEZ, 1995;
PALTRIDGE & BARBER, 1988).
Taking into account the previous comments, a
combination between remote sensing data (related to
live fuel moisture content) and meteorological index
(related to dead fuel moisture content) will be proposed.
D-08-06.doc
7.1.2 Live fuels (satellite information)
Remote sensing data have been frequently used to
estimate the water status of plants, both in agricultural
and ecological research (CARTER, 1991).
In the forest fire danger literature, the water content
of plants is commonly expressed as fuel moisture
content (FMC), defined as the percentage of water
weight over sample dry weight:
FMC=[(Ww-Wd)/Wd)*100]
where Ww is the wet weight and Wd is the dry weight
of the same sample.
This variable is mostly obtained through field
sampling using gravimetric methods (VIEGAS et al.,
1992).
Wet samples are:
- weighed, then
- oven-dried at 60 or 100 oC, and
- weighed again to determine the dry weight.
FMC can be referred to for both live and dead
species.
Within the context of fire danger estimation, good
correlations between live FMC and multitemporal series
of NOAA–AVHRR data have been found for
herbaceous species using normalized difference
vegetation index (NDVI) data (CHLADIL & NUNEZ, 1995;
PALTRIDGE & BARBER, 1988), but problems were found
for shrubs and trees (CHUVIECO et al., 1999a; LEBLON,
2001).
However, in the remote sensing literature, water
content is usually expressed as the equivalent water
thickness (EWT: water content/leaf surface), instead of
FMC, because EWT is directly related to the absorption
depth of the leaf.
Laboratory spectral measurements have been
performed to estimate EWT, showing divergent results
in the visible and near infrared (NIR) depending on
whether they were done at leaf or canopy level,
because of the indirect effects of water content changes
on the whole plant (mainly through the modification of
the leaf area index [LAI]).
However, short wave infrared bands (SWIR: 1.1–2.5
Am) have proven to be the most sensitive to EWT
variations (BOWMAN, 1989; COHEN, 1991; DATT, 1999),
although additional bands are required to reduce the
uncertainty caused by other variables affecting SWIR
reflectance.
Simulation studies based on radiate transfer models
have recently identified a ratio of the near infrared (NIR)
and SWIR band as the most appropriate for retrieving
EWT at leaf and canopy levels (CECCATO et al., 2001,
2002), as previous experimental studies had suggested
(HUNT & ROCK, 1989).
34
EUFIRELAB
In spite of this progress, to estimate EWT from
reflectance measurements, additional efforts need to be
made to derive FMC from satellite data in fire danger
studies because the amount of water per area is not as
critical in fire propagation as the quantity of water per
dry mass.
Assuming that the specific leaf weight (SLW= dry
leaf weight/leaf area) is constant over time for single
species, FMC may be considered a function of EWT
(CHUVIECO et al., 2003a).
Still, when this relation changes significantly over
time, FMC may be indirectly estimated as a result of the
effects of plant drying on the decrease in leaf area
index (LAI) values (mainly in shrub species) and
chlorophyll content (herbaceous species).
Therefore, the estimation of FMC from reflectance
measurements can be undertaken when the estimation
is restricted to single or (physiologically) similar species.
This explains why strong empirical relations between
FMC and satellite variables have been found by several
authors (CECCATO et al., 2003; CHUVIECO et al., 2002;
LEBLON, 2001).
Recent studies have shown that by better estimating
other factors affecting canopy reflectance in the NIR
and SWIR bands, particularly the leaf area index (LAI),
it is possible to apply radiative transfer model (RTM)
inversion techniques to obtain a reliable estimation of
EWT and FMC (ZARCO-TEJADA et al., 2003).
Additionally, plant canopy temperature is affected by
FMC changes because water availability is a critical
parameter in plant evapotranspiration.
Based on this principle, several authors have tested
the use of thermal images to estimate plant water
content, mainly on crops (JACKSON et al., 1981; MORAN
et al., 1994).
Forest and shrub canopies are more complex, but
some workers have shown good relationships between
the differences in air and surface temperature (ST) and
fire danger hazard (VIDAL et al., 1994).
Because these differences are closely dependent on
the density of vegetation cover, the combined use of
surface temperature (ST) and NDVI have shown
statistically stronger relationships with water content
than either of the two variables alone (ALONSO et al.,
1996; CHUVIECO et al., 1999a; PROSPER-LAGET et al.,
1995).
7.1.3 Study case in Central Spain
The following study has been published in Remote
Sensing of Environment (CHUVIECO et al., 2004b).
7.1.3.1 Objectives
The objective of the study case is the assessment of
an empirical approach to estimate FMC of
Mediterranean species based on multitemporal analysis
of NOAA–AVHRR images.
The proposed method is built on statistical fitting of
field collected FMC and satellite data, using a function
of the day of the year to take into account the seasonal
trends of FMC.
D-08-06.doc
The empirical estimation was intended for
operational retrieval of FMC in fire danger
assessments.
Considering the current limitations of meteorological
networks and fuel type maps, it was determined that the
FMC estimation should not require external data sets
other than the information derived from the AVHRR
images and very simple vegetation maps.
The estimation was targeted at grassland and shrub
species, which are the most dangerous in fire
propagation of surface fires.
The empirical fitting was based on a long time series
of field measurements of FMC for the Cabañeros
National Park study site (Central Spain), but collecting
field measurements at other sites with similar species
validated it.
Previous work showed a strong statistical
relationship between AVHRR-derived variables and
FMC for the Cabañeros study site using only summer
images (CHUVIECO et al., 2003a).
In this case, 2 years of field data were used for
calibrating the model and 2 more years for validation in
the same study site.
Additionally, strong relations were also found for
Landsat-TM images (CHUVIECO et al., 2002) and SPOT-
Vegetation images, showing consistent trends among
the three sensors (CHUVIECO et al., in press).
This work follows the same trend towards finding
consistent relations between FMC and satellite-derived
variables, for operational use of satellite data in fire
danger estimation.
In this case, the model is applied to spring and
summer data, uses 4 years for calibration and 2 more
for validation in the same study site, as well as five
additional validation sites, located far away from the
calibration area.
Additionally, it introduces a function of the day of the
year to model the effect of the seasonal trends.
7.1.3.2 Methodology
One of the key elements to obtain a sound empirical
estimation in remote sensing research is the availability
of long time series of field data.
Field measurements must include valid samples of
large areas, on which the satellite images will be
acquired.
For this study, the Cabañeros National Park (located
in Central Spain) was used as a calibration site
because it offered unique opportunities for testing
relations between FMC and satellite-derived variables.
First, the central area of the park is covered by
grassland and shrublands on very gentle slopes.
Because it is a protected area, no agricultural
practices are carried out, and therefore, temporal
changes are associated to vegetation seasonal trends
rather than crop alterations.
Fuel types sampled were grasslands (three plots)
and several shrub species (two plots): Cistus ladanifer,
Erica australis, Phillyrea angustifolia and Rosmarinus
officinalis.
35
EUFIRELAB

Plot sizes were 50 x 50 m, located 3 to 5 km apart, Geometric corrections were based on orbital models
along a range of 20 x 5 km. A complete description of and manual control points and automatic correlation
the fieldwork may be found in CHUVIECO et al. (2003a). improved multitemporal matching.
Daily data were synthesized into 8-day composites
Field measurements were taken from early April to
using maximum NDVI values.
the end of September.
The median value of a 3 x 3 pixel window was
For the model calibration, a time series from 1996 to
extracted from each composite and correlated against
1999 was used.
field measurements.
These samples were collected every 8 days.
When comparing AVHRR images and field
For validation purposes, another field campaign was
measurements, the potential noise caused by the great
carried out in 2001 and 2002.
differences within the area covered may be reduced
In this case, the samples were collected every 16
when using average values of species, instead of single
days because previous analysis did not show shorter
plot averages.
time changes in FMC values.
For instance, average values of grasslands collected
In each plot, three samples per species were
in a length of 10 km (three plots separated linearly 5 km
collected. Average values per plot and period were
each) would be a better representation of what an
computed.
AVHRR pixel is actually measuring than single plot
Long-term trends of FMC values were compared for measurements.
the different grassland and shrubland plots located in
Several authors have discussed the pros and cons
Cabañeros in order to test whether they were showing
of empirical and theoretical models in remote sensing
local or regional differences in FMC trends.
research (STRAHLER et al., 1986).
Considering the large distance between plots (3–5
Theoretical models have two main advantages:
km apart), if significant differences in average values
generalizing power and a better understanding of the
between plots of grasslands or shrublands were not
parameters involved.
found, it could be concluded that FMC temporal
However, they are complex to generate because
changes of these two vegetation types are more
they require many input parameters that are often
significant than those changes caused by their spatial
unavailable and are difficult to validate.
diversity.
Empirical models are commonly based on statistical
Consequently, field measurements of small plots
analysis.
could be considered as representative of the temporal
They are simpler to formulate and provide a
variation of FMC for large plots.
quantitative validation on their exactness, but they are
In this way, the field measurements could be
difficult to generalize, especially when statistical
soundly related to coarse resolution satellite images.
relations are not based on physical properties.
During 2001 and 2002, a different set of field
In the field of water content estimation, a whole
measurements was taken in other regions of Central
range of theoretical models has been proposed in
Spain for validation purposes (Figure 24): a site located
recent years, most of them based on the radiative
in the province of Segovia was covered with
transfer function (BARET & FOURTY, 1997; CECCATO et
grasslands; Atazar–Alberche with grassland and shrubs
al., 2001; CECCATO et al., 2002b; JACQUEMOUD et al.,
(a mixture of C. ladanifer and R. officinalis); Ibérica and
1996; ZARCO-TEJADA et al., 2003).
Pre-pirineo with shrubs (R. officinalis and other shrub
They are solid approaches but require further
species), and Cádiz with C. ladanifer.
assessment and must demonstrate their operational
These plots are 200 to 500 km apart from the
application with field campaigns.
Cabañeros site and have different elevations but
These models estimate the EWT, which is the
include similar species, parts of the Mediterranean
variable directly associated to leaf water absorption.
ecosystem except of the Pre-pirineo site.
FMC is equal to EWT divided by SLW. EWT can be
The plots were selected so as to include
estimated using a radiative transfer function, but dry
homogeneous plant coverage on as gentle slopes as
matter content cannot be directly retrieved because the
possible.
water is masking its effect on reflectance (JACQUEMOUD
However, the shrub species were frequently mixed,
et al., 2000).
even with some trees.
For this reason, ZARCO-TEJADA et al. (2003) use a
Among the different mixtures, only those plots with a
simplified inversion model to obtain dry matter for FMC
significant coverage of C. ladanifer or R. officinalis
estimation, after deriving EWT from a radiative transfer
(more than 60%) were selected for validation purposes.
model.
To assure consistency in the results, the field
protocol of these assessment sites was the same as Several authors (ALONSO et al., 1996; CHLADIL &
that of Cabañeros. NUNEZ, 1995; CHUVIECO et al., 1999a, 2002, 2003a;
HARDY & BURGAN, 1999; PALTRIDGE & BARBER, 1988)
AVHRR images were acquired by the University of
proposed empirical fittings for estimating FMC from
Alcala’s HRPT receiving station. Raw digital to
satellite data.
reflectance conversion was based on NOAA
coefficients (including degradation rates), and surface
temperature (ST) was based on methods proposed by
Coll & Caselles 1997).

D-08-06.doc 36
EUFIRELAB
Most commonly, these studies were based on
AVHRR images, although there are also some
examples using Landsat-TM images (CHUVIECO et al.,
2002).
The empirical model was based on linear regression
analysis, where FMC was the dependent variable and
the independent variables were AVHRR variables,
NDVI and ST, and a function of the day of the year.
Two models were generated, one for grasslands and
one for shrubs.
The major physiological differences between these
two communities made it advisable to split the fittings.
Cistus ladanifer was selected as a representative of
Mediterranean shrub species because it is widely
represented in Spain.
Other species of the same family (Cistus sp.) are
also broadly distributed across the Mediterranean
basin.
The statistical model was built from 88 periods (22
periods of 8 days during 4 years: 1996 to 1999),
covering spring and summer conditions of the
Cabañeros National Park.
The time series include a wide range of rain
patterns, with some dry years 1999 and 1997 with
precipitation close to 200 mm in 6 months), and more
humid ones 1996 and 1998), with 250 and 230 mm,
respectively.
The equation was validated using data from the
same study area (Cabañeros), as well as the other
study sites previously described during the 2001 and
2002 spring and summer seasons.
Satellite variables considered in the linear
regression were NDVI and ST.
The former would be positively related to FMC
because the drying of the plant reduces chlorophyll
activity in grasslands, as well as leaf area index in
shrub species.
On the contrary, ST would be expected to be
negatively related to FMC because the cooling effect of
evapotranspiration is reduced when plants get dry and
introduce mechanisms to reduce water loss.
Before obtaining the estimations of FMC from linear
regression models, an analysis of trends between FMC
and these two variables (NDVI and ST) was
undertaken.
In a similar way to other study areas (ALONSO et al.,
1996; KALLURI et al., 1998; MORAN et al., 1994;
PROSPER-LAGET et al., 1995), NDVI showed a negative
correlation with ST in both grasslands and shrublands
for the spring and summer seasons.
This trend must be related to the physiological
reaction of plants to higher temperatures and lower
moisture contents, which, depending on the plants, may
change leaf colour, deteriorate leaf structure, modify
leaf angle distribution by leaf curling or reduce LAI by
leaf loss, and/or decrease evapotranspiration.
Based on these relationships, some authors have
proposed a regression model of NDVI and ST to
estimate plant evapotranspiration (KALLURI et al., 1998)
and fire hazard levels (PROSPER-LAGET et al., 1995).
D-08-06.doc
A scatterplot of NDVI against ST for different FMC
values observed in the Caban˜eros site showed the
trend towards the appearance of low values of FMC
when low values of NDVI and high ST values occur,
both for grasslands and shrublands.
The trends are more evident for grasslands because
they present a wider range of both FMC and NDVI
values.
Following the logic of Verstraete and Pinty 1996),
the design of an optimal index for discriminating
different FMC values should be based on lines
perpendicular to the main axis of NDVI and ST
variation, which show potential sensitivity for
discriminating FMC values.
Additionally, a temporal variable based on the day of
the year (from 1 to 365) was included in the empirical
fitting to take into account seasonal trends in FMC,
following a logic already tested in Mediterranean
conditions (CASTRO et al., 2003; CHUVIECO et al., in
press).
Considering that these temporal trends are more
contrasted for grasslands than shrublands, two different
functions were computed:
FDg = (sin(1.5 x π x (Dy + Dy1/3)/365))4) x 1.3
FDc = (sin(1.5 x π x Dy/365))2+ 1) x 0.35
where FDg and FDc are the functions of the day of
the year (Dy) for grasslands and C. ladanifer,
respectively, and the sine angle is computed in radians.
This function was derived by fitting a periodical
function to the temporal average of FMC values of
grassland and C. ladanifer for 6 years of measurements
in Cabañeros 1996–2001).
The function has a wider variation for grasslands
than C. ladanifer, which agrees with the stronger
contrast in the water content of herbaceous species.
The constant terms were used just to scale the
functions in a similar range among them.
7.1.3.3 Results
For the 6 years of field data, the t tests applied to the
temporal differences of the three grassland plots in the
Cabañeros site, separated between 3 and 5 km, did not
show significant differences among them.
Similarly, the average temporal trends of the two
shrub plots located 3 km apart were not significantly
different.
Therefore, it could be concluded that the average
FMC values of small plots (50 x 50 m) are
representative of large areas (several kilometres apart),
at least in the Cabañeros site, and consequently,
temporal data extracted from those small plots can be
assumed representative of the large plots observable in
AVHRR coarse-pixel size images.
As mentioned above, the equations to estimate FMC
from AVHRR data were derived from multiple linear
regression analysis, using NDVI, ST and FD for 4 years
of Cabañeros field data 1996 to 1999).
37
EUFIRELAB
The resultant equations were:
FMCg = -57.103 + 284.808 x NDVI – 0.089 x ST +
136.75 x FDg
FMCc = 70.195 + 53.520 x NDVI – 1.435 x ST +
122.087 x FDc
where:
- FMCg and FMCc are the estimated FMC values of
grasslands and C. ladanifer, respectively;
- NDVI is the normalized difference vegetation index
(range -1 to + 1);
- ST is the surface temperature (in Celsius degrees);
- FD, the function of the Day of the Year.
The determination coefficients (r2) obtained were
0.737 (p < 0.001) for grasslands and 0.672 (p < 0.001)
for C. ladanifer.
Significance values of independent variables were
lower than 0.01 for NDVI and FDg in the case of
grasslands, and ST and FDc in the case of C. ladanifer.
ST was not significant for grasslands (p>0.5)
because most of its discrimination power is included in
the FDg variable.
However, it was incorporated to improve spatial
estimations, given that FDg does not change spatially.
For the same reason, NDVI was kept in the case of
C. ladanifer, in spite of its low significance (p = 0.187).
The contribution of NDVI to FMC estimation is more
significant for grasslands because the drying process in
herbaceous species is commonly followed by a loss of
chlorophyll activity and a LAI decrease.
In both cases, the factor accounting for the seasonal
trends (FD) is very significant (p < 0.001).
As expected, NDVI and FMC show a positive
correlation, whereas for ST it is negative, confirming the
physiological assumptions previously stated.
The assessment of these equations was carried out
on other time periods (2001–2002) in the same study
site (Cabañeros), as well as on other study sites, and
good results were obtained in all cases.
Relations are very coherent at all sites, in spite of
being at a distance of over 200 km for Segovia and
Atazar–Alberche sites, and more than 500 km for the
Ibérica site, and with different altitude ranges (up to 500
m of height increase in the case of Segovia).
For the Cádiz site, only four observations were
available, but r2 values were also very high (0.96).
The Pre-pirineo site had no samples or neither
grasslands nor C. ladanifer and was not included in this
analysis.
The temporal trends are very well estimated, and
deviations of actual versus predicted FMC values are
low and have no consistent bias.
The worst estimation was observed for grasslands in
late spring (early June) and in the middle of the summer
(August).
The former is related to the sudden decrease in
FMC, which in both years, changes from over 170% to
just 30–35% in 16 days.
D-08-06.doc
This severe decrease is reflected in the reduction of
NDVI and increasing ST, but not as steep as the field
FMC values show.
As a result, overestimations in this period reach up
to 60% of FMC.
Additionally, negative FMC values were estimated in
August of the second year, caused by very low NDVI
values (below 0.1 for this period).
However, negative estimations are not a major
obstacle for operational purposes because a simple
filter could be applied to the empirical model to avoid
them.
Additionally, FMC values of grasslands during most
of July and August are below 30%, which may be
considered the limit for live species.
Therefore, for practical purposes, grasslands may
be considered as dead fuels for the central part of the
summer.
The validation of the C. ladanifer showed an even
better fitting than grasslands, with very close
estimations both in spring and summer in the
Cabañeros site.
The highest deviations from the observed FMC
values never reached 20%, and for most periods, they
are under 10% of FMC.
The other two validation sites for grasslands (Ávila–
Segovia and Alberche–Atazar) also showed very good
fittings, with r2 values of 0.881 and 0.905 (Figure 25).
There is a slight tendency towards overestimation in
Alberche–Atazar and underestimation in Ávila–Segovia,
but the relation in both cases is close to a 1:1.
The scattergram also shows a nonlinear estimation
trend, especially in Alberche–Atazar, which may be
related to the saturation of NDVI in the upper part of the
range (BARET & GUYOT, 1991).
In fact, polynomial equations between observed and
estimated FMC values provide r2 values higher than for
lineal trends, with 0.96 for Ávila–Segovia and 0.95 for
Alberche–Atazar.
Nonlinear relationships should also be explored at
the calibration stage in the future.
FMC estimation for C. ladanifer shows good fittings
in all assessment sites: Alberche–Atazar, Cabañeros
and Cádiz, although for the latter, only 4 observations in
the summer of 2001 were available (Figure 26).
The empirical model has a slight tendency towards
overestimation, especially for lower values of FMC.
The lower values offer a better fitting between
estimated and observed FMC with differences lower
than 10% of FMC during mid-summer.
Considering certain physiological similarities
between C. ladanifer and another widespread
Mediterranean shrub, R. officinalis, the empirical
function was also applied to other study sites where this
shrub species had been sampled in the field.
The results were very positive for all sites
(Cabañeros, Atazar–Alberche and Ibérica), with r2 over
0.85.
The temporal trends also show good fittings, with
nonsignificant biases (Figure 27).
38
EUFIRELAB
An underestimation was observed for the spring
season, but the fittings improved in the summer, when
fire danger is higher, and therefore, the need for
accurate estimations is more demanding.
The absolute errors were higher for R. officinalis
than for C. ladanifer, especially in the Ibérica site, which
may also be caused by the mixture of the field plots,
where several shrub species grow in the same area.
7.1.4 Probability of ignition related to live fuel
moisture content
The empirical model generated in the previous
section using NDVI, ST and function of Day of the Year
showed a consistent predictive power to estimate FMC
of grasslands and C. ladanifer, a typical Mediterranean
shrub species.
The model was tested in plots located several
hundreds of kilometers apart and with different altitude
ranges.
Therefore, this model may be tested on operational
scenarios in Mediterranean conditions, applying both
spring and summer data.
The model only requires two basic satellite variables
(NDVI and ST), the day of the year and a regional map
of vegetation types, which distinguishes grasslands
from shrublands.
This variable could also be derived from the
multitemporal classification of AVHRR data, following
any of the methods applied to derive global land cover
maps (DEFRIES & TOWNSHEND, 1994).
Considering the spatial and temporal resolution of
AVHRR images, the empirical index may be used for
short-term estimations of fuel moisture content (Figure
28).
After estimating the FMC of live fuels through
satellite information it is essential to transform these
values to probability of ignition in order to be able to
assess the probability of ignition in live fuels and to
combine this probability with probability of ignition
based on other factors (i.e. human factors etc).
To do so, CHUVIECO et al. (2004a) developed a
simple method to convert FMC values to danger ratings
based on computing ignition potential from thresholds of
moisture of extinction adapted to each fuel.
Fire danger is restricted to the likelihood of fire
occurrence, given a particular fuel moisture content.
This likelihood will be defined in terms of probability
of ignition associated with fuel moisture content status.
These values can then be integrated with other
variables associated with ignition sources (e.g.,
lightning, human), which could also be expressed in
terms of ignition probability.
The method proposed from Chuvieco et al. (2004a),
to convert FMC values to ignition potential associated
with fuel status (IPf ranging from 0 to 1) is based on the
concept of moisture of extinction (ME).
This is defined as the threshold moisture content
above which a fire cannot be sustained (ROTHERMEL,
1972).
D-08-06.doc
The ME of living fuels varies between 12% for some
grass fuels and 200% for the needles of some conifer
species.
For most live vegetative fuels the ME is in the range
of 120%-160%, while for dead fuels it is in the range of
12%-40%.
This method assumes that ME values act as relative
thresholds to ignition for each fuel, above which the IP
dramatically decreases.
Although, the IPf for FMC values higher than ME
would be zero, a conservative approach is
recommended here, assuming that a marginal IPf exists
even at high values of FMC.
For this reason, it is proposed to assign a maximum
IPf value of 0.2 to the FMC that equals the ME value of
each fuel.
FMC values lower than ME would have IPf values in
the range of 0.2–1, the IPf being linearly inversely
proportional to FMC values.
For FMC values greater than the ME, IPf values
would range from 0.2 to 0.
Null IP (IPf = 0) was assigned to the maximum FMC
value recorded in the historical series of FMC field
measurements.
Schematically this method is based on the following
algorithm (Figure 29):
If FMC > ME, then
IPf = {1– [(FMC – ME)/(FMCmax – ME)]} × 0.2
else
IPf = 0.2 + [(ME – FMC)/(ME – FMCmin)] × 0.8
where FMCmax and FMCmin are the maximum and
minimum FMC values of each fuel type derived from
field FMC samplings.
While these samplings are site specific, they might
nevertheless be assumed to apply to fairly large regions
with similar environmental conditions.
ME values are fuel-type specific.
For dead fuels, ME values were taken from the
BEHAVE fire behaviour prediction system (Burgan and
Rothermel 1984), ranging from 12% to 40%.
Grasslands with FMC values lower than 30% were
treated as dead fuels and assigned ME values of 12%
(model 1) and 15% (model 2) following values proposed
for the BEHAVE model (BURGAN and ROTHERMEL
1984).
For live fuels, threshold values were taken from
specialized literature.
For shrubs, an average value of ME was selected
(105%), following the experimental results of
DIMITRAKOPOULOS and PAPAIOANNOU (2001) and
BURGAN (1979).
For annual grasslands, the ME threshold was fixed
at 40% (ALBINI 1976).
According to ANDERSON (1982), the ME of
grasslands is directly related to fuel depth.
This explains why short grasses have low values
(12%–15%) and tall grasses (more than 70 cm high)
have higher values (25%–30%).
39
EUFIRELAB
The use of linear functions to convert FMC to IPf is
based on the experience of several authors (such as
DIMITRAKOPOULOS and PAPAIOANNOU, 2001) who have
shown similar trends between FMC and ignition delay
for a wide range of Mediterranean species.
The slope coefficient of the linear regressions was
related in this case with the flammability of the species.
DELICHATSIOS et al. (1991) proposed a similar
solution, they termed this regression coefficient
“flammability delay increment.”
The use of a linear function instead of an
exponential one (as suggested by other authors, such
as TRABAUD 1976 and ALBINI and REINHARDT 1995) is
justified by the fact that recent experimental results with
numerous species of the Mediterranean Basin clearly
show that there is a gradual, linear relation between
flammability (as measured by ignition delay time) and
moisture content (DIMITRAKOPOULOS and PAPAIOANNOU
2001).
Older studies used a limited number of species and
not the ISO 5657-1986(E) methodology for measuring
ignition time of materials (TRABAUD 1976; VALETTE
1992).
In a highly exponential (nonlinear) relationship, small
variations of the FMC in the upper scale of the curve
(i.e., near the threshold of the moisture of extinction)
may result in relatively big differences in the
flammability of the fuels.
However, Mediterranean species seem to respond
with gradual fluctuations in their ignition delay time to
changes in the FMC status (DIMITRAKOPOULOS and
PAPAIOANNOU 2001).
This is especially noticeable in live fuels and highly
flammable species (i.e., species with relatively low
ignition delay time).
Efforts are underway to improve the highly
exponential relationship between the moisture and fuel
flammability in the Canadian Forest Fire Danger Rating
System (MCALPINE, 1995; in BEMMERZOUK, A.M., 1997)
7.1.5 Probability of Ignition in Dead Fuels
(Meteorological Index)
7.1.5.1 Meteorological Danger Indices
Dead fuel includes a wide range of materials
(senescent grasses, dry leaves, small twigs, and
organic material in the topsoil).
Water content of dead fuels is the most important
factor in determining fire danger potential.
On one hand, the water content of the fuel is
inversely related to the probability of ignition, due to the
fact that part of the energy necessary to start a fire is
used up in the process of evaporation right before the
fire starts to burn (CHANDLER et al. 1983).
On the other hand, water content also affects fire
propagation since the source of the flames is reduced
with humid materials, therefore reducing flammability
(ROTHERMEL 1972).
For this reason, the estimation of water content in
these fuels is a critical variable both in determining fire
ignition in a specific time and place, as well as in
predicting the behaviour of the fire itself.
D-08-06.doc
The water content of dead fuel is constantly
changing, depending mainly on atmospheric conditions
(SIMARD 1968).
Loss or gain of water content will vary depending on
the physical and chemical characteristics of the fuel and
the presence of varying atmospheric activity (rain,
condensation, etc.).
The water content of dead fuels is determined by
various methods.
The most precise one is direct sampling by
gravimetric methods.
Following this approach, water content is computed
from the difference of wet and dry weights of the
samples.
Most commonly, the water content is expressed as a
percentage of the dry weight (BLACKMAR and FLANNER
1968; DESBOIS et al. 1997).
Direct sampling provides exact measurements, but it
is costly and labour intensive, especially when wide
area estimations are required.
Additionally, this method does not provide an
instantaneous measurement, as the samples must be
oven dried during a certain number of hours (24 or 48
have been commonly suggested).
Other methods are based on the use of previously
calibrated wooden sticks that are assumed to be good
representatives of certain types of fuel (SIMARD 1968).
In the USA, standard 10-hour fuel sticks with an
oven-dry weight of 100 grams are commonly used.
The sticks are continuously weighed.
Finally, meteorological danger indices (MDI) have a
long tradition in fire danger estimation, because they
comprise different critical variables related to fire
ignition and fire propagation.
The meteorological danger indices vary in
complexity and in the number of variables to be
considered, from those that only require temperature
and relative humidity to those which are based on
complicated numerical models (for a summary of the
models most used see VINEY 1991).
These indices rely on current and past weather
conditions, since they also try to estimate the degree of
dryness of different forest fuels.
Since most of the countries have a relative dense
network of weather stations for different purposes
(agriculture yield prediction, disaster prevention, traffic
regulation, etc.), MDI values can be operationally
computed for extended territories.
Additionally, with the growing availability of
automatic weather stations, these indices may be
computed very frequently and measured in real time.
However, often the location of weather stations is
not very appropriate for fire danger estimation, since
they are located in urban or agricultural areas.
Therefore, spatial interpolation techniques to
estimate weather variables at forested areas are
required.
These interpolation methods always introduce a
certain estimation error, which is added to the actual
estimation of fuel moisture content (FMC).
40
EUFIRELAB
Even so, this method is the most widely used to
determine the FMC of dead fuel.
There are a lot of fire danger meteorological indices
which use in different ways air or/and ground
temperature, air or/and wood sticks humidity (relative or
absolute), wind speed, and soil water reserve as an
expression of the possibility for the plant to get a water
satisfaction rate.
From all these meteorological indexes found in fire
prevention literature, the Fine Fuel Moisture Content
(FFMC) of the Canadian CFFWI (Canadian Forest Fire
Weather Index) and the 10-h moisture code of the
National Forest Fire Danger Rating System (NFDRS)
provide a sound estimation approach for short-term
changes in dead fuels FMC.
Both the CFFWI and the NFDRS have several
components and their usefulness had been validated
for fire prevention in several Mediterranean contexts
(RODRÍGUEZ Y SILVA 2002; SEBASTIÁN 2002).
The FFMC estimates the water content of the top
layer on the ground (L-layer) from measurements of
temperature, relative humidity, wind velocity, and
precipitation registered in the last 24 hours (VAN
WAGNER 1987).
This water content index has an empirical character
derived from the relationship between these
meteorological variables and the water content of a
standard fuel (jack pine and lodgepole pine type), at the
same time integrating the accumulative effect of
atmospheric conditions in the hours previous to the
measurements.
This index has a timelag of approximately 0.66 days.
On the other hand, the index 10-h, which is part of
the NFDRS, estimates the water content of
combustibles with a width of 1.2 to 2.5 cm (BRADSHAW
et al. 1983).
This index uses the concept of Equilibrium Moisture
Content (EMC).
The EMC of a fuel element under given
environmental conditions is the moisture content that
the element will attain if left for sufficient time in those
(constant) conditions.
The EMC is a function of the temperature and the
relative humidity (SIMARD 1968), as well as the
atmospheric conditions present at the moment of
measuring the samples.
A study undertaken in the Department of Geography
at University of Alcalá has compared the usefulness of
two meteorological danger indexes in fire prevention
used in Canada (FFMC) and the United States (10-h) in
order to evaluate the FMC of dead fuels in a
Mediterranean territory.
Taking measurements in the field over a period of
6 years compared the codes.
The results did not show a significant difference
between the two; therefore it is recommended that the
index with less meteorological variables (10-h) be used.
This index allows an estimation of FMC of
flammable dead fuels with an RMS around 5% (AGUADO
et al, in preparation).
D-08-06.doc
As noted before, one of the main challenges of using
these meteorological fire danger indexes is extending
the measurement obtained in the specific weather
stations to cover the rest of the territory (FUJIOKA 1987).
This problem is compounded more and more
because the Forest Services use the GIS technology in
forest fire prevention when they integrate spatial
variations in the data.
As a result, once an index with a better predictive
capacity has been selected, it is necessary to get a
spatial distribution of this index.
The spatial dimension of meteorological indices can
be obtained using a mapping methodology based on
logical relationship between the value of an index (or of
its components) and the environmental parameter
values, as it has been explained in a previous
deliverable (D-08-05).
This methodology improves the spatial information:
from a map built with some points (the index value of
which is indicated) and many empty cells, we get a map
with all the cells filled with the index value for each cell.
Next section shows an example of this methodology
at local scale.
7.1.5.2 Spatial interpolation techniques: the case of
Alpes-Maritimes (France)
Mapping the meteorological indices values is in fact
one of the most important steps of the risk
determination.
As explained in previous deliverable D-08-03,
meteorological indices are always computed from
stations network data.
Every data is representative only of the sensor that
measured it, and then there are two concepts
concerning the use of meteorological data measured in
a given place, to elaborate a meteorological risk index:
- Either a region is covered for instance by a 20
stations network, and the index is computed for each
station. That means the risk is known only for 20
points, and is unknown for all the places, which exist
outside these points. The users, as foresters or
firemen have to fill themselves, with their
imagination, the numerous places where the value is
not computed, and in fact they interpolate in their
mind the absent values. This risk concept is a
punctual one and the problem is the needed
information is absent on large surfaces.
41
EUFIRELAB
- Or, in the same region covered by 20 stations, it is
considered that there are sub-regions constituted by
different areas, which are relatively homogeneous.
For example, in France, MeteoFrance decided they
are 7 different « fire regions » in the department of
Alpes-Maritimes (Nice). For every « fire region »
(their surfaces are different from the others), the
meteorological risk index is computed only from the
stations, which are inside the fire region, generally 2
or 3 stations. In fact, each fire region is «virtual»
because each meteorological variable which is used
and contribute to compute the index of each fire
region is an average value computed from the
stations. Thus the index is obtained from data
characterizing a non-existent place which results
from computed values (means, barycenter from 2 or
3 stations).
In this case we can say that each fire region (from
200 to 1500 km2) index is supposed to represent an
area which is homogeneous enough, so that in details,
the risk is he same in each point of the region.
Obviously, especially in mountainous relief, as very
often in Mediterranean regions, this homogeneity does
not exist…
We propose here another possibility, that can be
used everywhere, if a stations network exists: it consists
to obtain with scientific arguments the most probable
meteorological risk index value for a very small area, a
pixel, the surface of which is for instance from 2500 m2
(50 m side) to 1 km2 (instead of some hundred or
thousands kilometres…).
The principle is to interpolate data from a station to
elaborate the risk value of each pixel.
In the previous deliverables D-08-03 and D-08-05,
we explained the « environmental » regression
methodology: a statistical law is found, establishing a
link between the meteorological data of a place (a pixel)
and the environment of this place, especially the
topographic and relief environment (altitude, relative
altitude, aspect, slope, sea distance).
Realizing index risk maps induce to answer to 2
sorts of problems and to make choices:
- To decide the width of each pixel, in a raster GIS;
- To decide which methodology will be used:
interpolation of data from the network stations and
computing the index for each pixel, or computing the
index for each station (from its data) and
interpolation of the index for each pixel.
The meteorological index we have used in this work
is I85/90 CARREGA index for 3 reasons:
- We know it very well as one of us is the author!
- This index is simple because composed with only
3 parameters (wind, air humidity, soil water reserve),
that allows to understand more easily how maps are
changing from a situation to another, and how
spatial interpolation works;
D-08-06.doc
- This index has been used by French firemen and
foresters of the Alpes-Maritimes during more than 10
years 1990-2000) inside of an Expert System
elaborated by Ecole Nationale Supérieure des Mines
de Paris, and has proved to be very efficient,
according to the users.
7.1.5.2.1 The width of each pixel: spatial resolution.
The smaller a pixel is, the more accurate the risk
definition will be, and that may be very interesting,
especially in complex topography.
But such a precision needs computer resources.
Working on a large area as Mediterranean Europe
for instance, at a given moment, does not necessitate
having a very high spatial resolution.
But, a very high resolution is needed for operational
work, i.e. by foresters or firemen.
In this case, the spatial scale used has to be very
accurate.
In this work we used a DTM of Alpes-Maritimes
(4400 km2) with 50 m spatial resolution, (degraded to
1 km2 if needed).
The map of figure 30 gives an idea of the
topography of the region where this methodology was
checked (department of Alpes-Maritimes).
In fact, we decided to work on the south part of this
region because excepted one, all the available
automatic stations are located in this south part,
allowing thus to obtain a higher measurement network
density for meteorological data.
As explained in previous deliverables, different
environmental parameters can be obtained from the
DTM, as:
- Altitude of each pixel;
- Relative altitude, indicating if a given pixel is at a
lower position (bottom of a valley) or dominating;
- Aspect, with a value varying from 1° (North) to 180°
(South) that is the maximum, East and West being
the same;
- Slope, above the given pixel, in degrees;
- Sea distance.
One example of maps is given that allow comparing
two parameters as aspect, for 50m pixel (fig. 31) and
1km pixel (fig. 32).
7.1.5.2.2 Two different logics: direct interpolation of
the index, or interpolation of meteorological
data in order to compute the index.
Direct interpolation of the index, 50 m definition
It is in fact the simpler way to obtain a
meteorological risk index map.
The methodology needs a station network and an
environmental regression model.
We decided to use always the same model, even if
according to the date and the corresponding
meteorological situation, the score of the regression
analysis is not the same.
The environmental parameters used are: altitude
(m), relative altitude (m), aspect (°), slope (°), distance
from the sea (km).
42
EUFIRELAB

Maps of figure 33 and figure 34, at 06h and 12h Thus the interpolation can be made by complex
UTC are risk index interpolation maps using a DTM 50 models or by spatial interpolation as kriging or simpler,
m resolution for each pixel. as IDW method (weight of each point is related to the
The risk index value of each 50 m pixel was distance between other points).
interpolated directly by environmental regressions, from
Direct interpolation of the index, 1 km definition
the risk index value computed for each meteorological
station. The methodology is the same as section i), but the
geographical scale changes: each pixel is 1 km instead
On 28 August at 6:00, the influence of altitude is
of 50m.
obvious at the end of the night: risk is higher on slopes
Comparing both maps of fig. 33-34 and fig. 37-38
facing to South but especially on higher altitudes.
shows the differences concerning the spatial accuracy,
High risk values (red) are not numerous, and the
which is less operational and usable when pixel width is
dominant situation is a low risk value, because we are
1km.
just at the beginning of the day, with generally low wind
But the advantage of this 1 km dimension is that less
speed, and high air humidity.
computer capacity and memory storage is used.
The high value of the index in altitude is due to wind
speed. Maps of figure 37 and figure 38, at 06h and 12h
At 12:00, the difference is obvious, due to UTC are risk index interpolation maps using a DTM 1
decreasing air humidity (temperature has increased) km resolution for each pixel.
and increasing wind speed (due to air heating and The risk index value of each 1 km pixel was
turbulence). interpolated directly by environmental regressions, from
the risk index value computed for each meteorological
Interpolation of the meteorological variables which
station.
constitute the risk index value, 50 m definition
Figures 39 to 43 are 1 km definition risk index maps
In this case the interpolated values are not the risk
(index interpolated) for August 31st 2003.
index initially computed for each station.
The sequence of maps allows to evaluate the the
The interpolation concerns the meteorological
risk level which is changing very quickly in time, and
variables.
which is different at the same moment from a place to
These ones are interpolated from the station
another one.
network for each pixel of the map, taking into account
During this day a Foehn wind effect occurred at the
its environmental parameter.
end of the morning, with high temperature, very low air
Obtaining its meteorological values allows
humidity and increased (but moderate) wind speed, with
computing for each pixel its risk index value.
a soil water reserve, which was very low.
The environmental parameters used to interpolate A “mixed fire” (forest, crops and houses) started
meteorological data are also: altitude (m), relative about 12:00 UTC very close to the sea, burnt more than
altitude (m), aspect (°), slope (°), distance from the sea 200 ha despite of about 800 firemen and destructed 10
(km). houses.
This fire was studied and explained in D-08-02.
Maps of figure 35 and figure 36, at 06h and 12h
It is interesting to notice that the meteorological risk
UTC are risk index maps resulting from meteorological
index was very high in the place where the fire began
variables interpolation (using a DTM 50 m resolution for
(in dark at 12:00).
each pixel).
The risk index value of each 50 m pixel was 7.1.5.2.3 Conclusion
computed directly from its meteorological values.
The possibility to obtain a map with a spatial
These values were interpolated by environmental
repartition of risk indices, and especially of a
regressions, from meteorological values of each
meteorological index, seems to be very interesting,
meteorological station.
compared with the simple knowledge of some points
Comparing the two maps (interpolation of the risk only, or to large areas in which the risk value is
index or interpolation of meteorological components of supposed to be the same.
the index, and then computing the index) allows
The methodology used influences the results as the
noticing differences of behavior: mapping directly the
comparison between the two sorts of maps shows.
index tends to smooth the local differences and
But the most satisfactory way to obtain a map
prevents to know what is the meteorological variable
(interpolation of meteorological variables) from an
which dominates the others.
intellectual point of view is also the longer.
On the opposite, the risk map resulting from
Thus, it is easier to interpolate directly the value of
interpolation of meteorological variables shows easily
the risk index.
some “spots” the risk of which is obviously different
In fact the validity of the result depends upon the
(blue spots of fig 35 and 36).
performance of the coefficient of correlation.
They are due to the interpolation method of wind
This one is generally better for water soil reserve
speed which is not an environmental one because the
than for humidity, and better for 50 m definition than for
logic of spatial variation of wind does not obey
1 km.
principally to topographic conditions.

D-08-06.doc 43
EUFIRELAB

7.1.6 Figures

Coruña
Bilbao

6
Zaragoza
Barcelona

5
Madrid
4 2
Madrid
3 Grasslands
Valencia Shrublands
1
Both

Name (Elevation)
1 Cabañeros (700m)
2 Atazar (900m)
3 Alberche (800m)
Sevilla
4 Segovia (1200m)
5 Ibérica (900m)
7 6 Pirineo (1000m)
7 Cádiz (400m)

Figure 24: Location of the sampling plots

Grasslands Alberche-Ataz ar y = 0.841x + 6.908

2
R = 0.8815
350

300

250
E stimated

200

150

100

50

0
0 50 100 150 200 250 300 350
Observed

y = 0.9525x - 3.8232
2
Grasslands Ávila-S egovia R = 0.9053

350

300

250
Estimate d

200

150

100

50

0
0 50 100 150 200 250 300 350
Ob se rv e d

Figure 25: Observed and estimated FMC values for grasslands in Alberche-Atazar and Ávila-Segovia
Validation sites (2001 and 200)

D-08-06.doc 44
EUFIRELAB

y = 0.7376x + 29.912
Cistus all sites
R2 = 0.7241

180
160
140
120

Estimated
100
80
60
40
20
0
0 20 40 60 80 100 120 140 160 180
Observed

Cabañeros Cádiz Alberche-Atazar

Figure 26: Observed and estimated FMC values for C. ladanifer in Alberche-Atazar, Cabañeros and Cádiz.
Validation sites (2001 and 2002)

Rosemary Cabañeros (2001-2002)

200
180
160
140
120
FMC

100
80
60
40
20
0
3-4 29-5 16-7 2-9 9-5 27-6 13-8 30-9

Observed Estimated

Figure 27: Observed and estimated FMC values for R. officinalis in Cabañeros.Validation sites (2001 and 2002)

Figure 28: Live fuel moisture content map from Euro-Mediterranean countries (August, 10th; 2004).
Based on CHUVIECO et al., 2004a, processed by A. CAMIA, JRC

D-08-06.doc 45
EUFIRELAB

FMC PIf
max
0
150 0.2
ME 0.4
100 0.6
0.8
50 1

Figure 29. Scheme to convert fuel moisture content (FMC) to ignition potential (IP) (example of dead fuels)

Figure 30: The topography of the Department of Alpes-Maritimes: altitude.DTM with Pixel width: 50 m.

D-08-06.doc 46
EUFIRELAB

Figures 31 and 32: Aspect of each 50 m pixel (left) and 1 km (right)

Figures 33 and 34: High definition (50m) Meteorological risk index map on 28 August 2003 at 06 and 12 UTC.

Figures 35 and 36: High definition (50m) Meteorological risk index map on 28 August 2003 at 06 and 12 UTC.

D-08-06.doc 47
EUFIRELAB

Figures 37 and 38: Low definition (1 km) Meteorological risk index map on 28 August 2003 at 06 and 12 UTC.

Figures 39 and 40: 1km definition risk index maps for 31st August 2003 at 03:00 and 06:00 UTC

Figures 41, 42 and 43: 1km definition risk index maps for 31st August 2003 at 09:00, 12:00 and 15:00UTC

D-08-06.doc 48
EUFIRELAB
7.2 PROBABILITY OF IGNITION (HUMAN
FACTORS)
7.2.1 Introduction
Human activity is the main factor of forest fires
ignition worldwide, having a special impact on
Mediterranean countries of Europe.
From different statistical sources we know that most
of the wildfires in Mediterranean Europe (above 90%)
occur as a consequence of human activities that can
directly act as fire ignition sources or indirectly create
the conditions that favours fire ignition and/or fire
propagation.
For example, according to data published by the
Spanish Forest Service (DGCN) for the period 1988-
1999, 96% of the fires in Spain were produced, either
directly or indirectly, by human causes, which indicate
the close link between fires and human activities.
The only natural cause of fires is lightning, which in
Spain accounts for a small percentage of only 4%
according to statistics of the past 15 years.
This is probably due to the fact that storms are not
as important in Spanish forested areas as in other parts
of the planet.
In fact, this may well be the only meteorological
phenomena in this climatic area that does not favour
fires ignition.
Curiously enough, this distribution of fires according
to its cause is quite different to other typical
Mediterranean ecosystems such as the Californian
one, where lightning may cause up to 40% of fires.
There are other non Mediterranean parts of the
planet where lightning is of high importance such as in
Scandinavia and Russia (20 to 30%), Western North
America (40 to 60%) and in the Australian region of
Victoria (10 to 30%).
Consequently, it is clear the importance of including
human factors in any comprehensive fire danger index.
Nevertheless most current operational indices of fire
danger rely on physical variables (mainly weather
data), and human factors are not ordinarily considered
(Chuvieco et al., 2003b).
Modelling human factors associated to fire ignition is
very difficult, since economic and recreational activities
linked to fire are very disperse in time and space.
An alternative to model human factors of fire danger
is the estimation based on indirect indicators that relate
fire occurrence with spatial variables linked to human
activities.
Human ignition danger may be defined as the
probability of a fire occurring as a result of the presence
and activity, either directly or indirectly, of human
beings.
The evaluation of human activity as an agent of fire
ignition is a complex task, since data on human
activities on forested areas are rarely available (VEGA
GARCÍA et al. 1995).
D-08-06.doc
In fact, the temporal and spatial data required to
evaluate the human factors in fire danger rating simply
do not exist (MARTELL, OTUKOL, and STOCKS 1987),
while other variables of fire danger such as
temperature or relative humidity are routinely
generated.
Moreover, human activities are very dynamic on
time and space, which difficult the estimation of specific
spatial patterns, or rather these are more difficult to
determine, as in the case of pyromania or specific
deliberate motivations and attitudes.
However, for other causes, such as those relating to
recreational activities or agricultural burnings are easier
to model using spatial variables.
The complexity of dealing with the human factors of
fire danger have frequently led to investigators and fire
managers to either leave them out of their risk models,
or consider them only marginally.
Most commonly, when those factors are considered,
the estimations are based on indirect assessments of
human risk activity, namely indicators of activities which
are the usual cause of fire and which are generally
structural in nature, i.e. population density or urban-
wildland interface.
The first studies on human factors of fire danger
were based on indirect variables, obtained mainly from
censuses and survey sources (ALTOBELLIS 1983; BAIRD
1965; BERTRAND and BAIRD 1975; CHRISTIANSEN and
FOLKMAN 1971; COLE and KAUFMAN 1963; DOOLITTLE
1979; DOOLITTLE and WELCH 1974; FOLKMAN 1965,
1973; JONES, TAYLOR, and BERTRAND 1965).
Later on, in the eighties the human factor is
analysed spatially, and it is frequently based on
cartographic aspects, that considered some human
variables along with natural features, such as slope or
fuel types (Aerial Information Systems Inc. 1981;
CHUVIECO and CONGALTON 1989; DONOGHUE and MAIN
1985; DONOGHUE, SIMARD, and MAIN 1987; LYNHAM and
MARTELL 1985; PHILLIPS and NICKEY 1978; MARTELL,
BEVILACQUA, and STOCKS 1989; YOOL et al. 1985).
The number of such studies has increased in the
last years due to the greater availability of digital
cartographic and statistical information managed
through Geographical Information Systems (GIS)
(ABHINEET et al. 1996; CHOU 1992; CHOU, MINNICH, and
CHASE 1993; CARDILLE, VENTURA, and TURNER 2001;
CHUVIECO and Salas 1996; CHUVIECO et al. 1999b;
LEONE et al. 2003; SALAS and CHUVIECO 1994;
THOMPSON 2000; VASCONCELOS et al. 2001; VEGA-
GARCÍA et al. 1995; VEGA GARCÍA et al. 1996).
These recent studies commonly use variables
related to recreational activities in forested areas,
proximity to roads and trails, population density,
distance to human settlements, forest property types,
etc.
These variables are easier to spatialize, but they
may have marginal importance in some areas (VEGA-
GARCIA et al., 1995).
Despite the importance of these empirical works,
there are still many aspects of the fire-human
relationships that require further research.
49
EUFIRELAB
For instance, we do not have yet a clear
understanding on the spatial consistency of those
relationships between fire and human factors.
Some variables may be critical in some regions,
while marginal in others, depending on the socio-
economic and environmental structure of the territory.
Additionally, the human factors require a
comprehensive assessment, which includes not only
spatial-variables (distances, fragmentation, interfaces)
but also socio-economic ones (unemployment, rural
population age, population density…).
Recent conferences in forest fire research have
emphasised the importance of this holistic approach
(Viegas 1998, 2002).
In the following sections two examples of human
ignition danger assessment are showed; one case at
regional scale and another one at European Scale.
We are aware that both aproaches need to be
improved and adapted in order to integrate them
properly in a euro-mediterranean risk index using the
previously specified resolution of 1 km2.
Finally a fire occurrence map is presented as a
possible alternative
7.2.2 Estimation of human ignition danger at
regional scale: the case of Alpes-Maritimes
(France)
Natural ignition due to lightning in the Alpes-
Maritimes (SE France) is relatively rare since most
storms are usually accompanied by rain due to the
proximity of the sea which induces conditions of high
atmospheric vapour.
Approximately 90% of fire ignitions between 1991
and 2003 in the Alpes-Maritimes were of human origin,
and this figure is probably an underestimate of long
term averages since the data include the fires of 2003
which was a particularly hot and dry summer with more
“dry” storms than usual during the month of August.
It is therefore reasonable to attempt to estimate the
probability of fire ignition as a function of the location
with respect to human settlements since it is in
proximity to human activities that most fires find their
origin.
7.2.2.1 Description of study area and database
Site location: the department of the Alpes-Maritimes
was chosen since it includes both low altitude climatic
conditions and Mediterranean vegetation with a
dominance of summer fires and a high altitude
mountain area where fires tend to occur at the end of
winter.
The surface area of the department is about 4300
km2.
Data period: a 13 year period, from 1991 to 2003
represents the time during which data on fire ignition
were systematically collected and recorded.
Description of the data: the French National Forest
Office (“Office National des Forets”, or ONF) provided
the bulk of the fire data for the period defined above.
D-08-06.doc
This included information on the location and likely
cause of fire ignition, as well as surface area burned
(not used in this study).
The “Prométhée” database provides less
comprehensive information for fires going back to 1973.
Fire ignition data were complemented by a Digital
Elevation Model (50 m resolution), digitized road
network from the “BD Carto” data base of the IGN
(National Geographic Institute), and vegetation and
land-use maps at a 1:25,000 scale from the Corinne
Land Cover data base.
Elevation, vegetation, land-use, and road networks
maps for the Alpes-Maritimes are shown in Figures 44
to 47.
In summary, the Alpes-Maritimes department is
particularly mountainous in the North of the department
and hilly in the South along the coast.
The Mediterranean climate has a pronounced dry
period in the summer, which is accompanied by dry
periods in the mountains during the winter.
Within a 10-15 km swath along the coast, the
vegetation is typically Mediterranean, especially West
of Nice.
The population of about 1 million inhabitants is
concentrated in the coastal area with a succession of
large cities (Cannes, Antibes, Nice, Monte-Carlo...).
The area of Mediterranean vegetation is therefore
concentrated in the suburban fringe just North of the
urban coastal area where houses are mainly individual
villas surrounded by forests
Forest fire ignition risk can be considered high in
this sector due to abundant dry vegetation in the
summer and a relatively dense road network and a
concentration of human activities.
Forest fire ignition: During the 1991-2003 period, a
total of 550 fires were recorded in the Alpes-Maritimes.
A precise ignition location is available for only 362
(about 66%) of these, and a known can be defined for
only 225 fires, or 50% of the total number of fires
(Figure 48).
The spatial distribution of fire ignitions in the
department according to known cause is shown in
Figure 49, identifying 8 categories of fire ignition.
The remaining fires were classed as cause unknown
7.2.2.2 Modelling fire ignition risk : a first « intuitive »
approach.
An initial approach to estimating fire ignition risk is to
look at individual aspects related to human activities
and vegetation characteristics separately.
In this case, the impact of urban development, road
network density, and vegetation type were analysed
individually.
7.2.2.2.1 Urban development:
According to a study conducted by KALABOKIDIS
(2002) in Greece, urban areas can strongly influence
fire ignition up to a distance of about 300 m, and
beyond that distance, the impact decreases sharply.
50
EUFIRELAB

We therefore quantified the number of fires Summing the variables creates a raster layer where
occurring within that distance for the different types of the cell values can range from -1 to 16 and where
urban setting shown in Figure 46. greater values indicate greater ignition risk: <0 = no
risk, 0 = very low risk, 0≤4 = low risk, 4≤8 =
For each urban setting, a coefficient was attributed
intermediate risk, 8≤12 = high risk, ≥12 = very high risk.
according to the number of fires recorded within the
300 m limit; coefficient values ranged from -1 to 3. 7.2.2.2.5 Discussion
The value of -1 was integrated was used in the very
The fire ignition map in Figure 50 shows the
high density urban areas where the risk of a fire is
greatest risk (orange and red) in the urban fringe of the
considered zero, even though 2 fires were identified in
coastal and pre-coastal areas.
that sector.
Table 10 summarizes the number of fires occurring The highly urbanised coastal areas are classed as
within the 300 m limit of each urban type and for each no risk (green) due to the negative coefficient attributed
fire ignition cause. to this sector where forest fires are extremely unlikely.
The bottom of Table 10 shows the coefficients that
were attributed to these types The fire ignition risk distribution can be analysed
according to three types of human occupation.
7.2.2.2.2 The road network. In the South, the dense urban area reduces the fire
ignition risk to zero.
The distance of 300 m from a house applied above
In the extreme North, there is a forested zone that is
in Table 10 was maintained for fires occurring in
too far from human activities to present much risk of a
proximity to a road as shown in Table 11, which shows
fire ignition that can be attributed to anthropogenic
that there are far more fires occurring near municipal
causes.
roads than any other type, and it is for this reason that
The low fire ignition risk is therefore associated with
they have been attributed a coefficient value of 3.
the low density of human activities.
A coefficient of -1 was attributed to autoroutes
Fire ignitions in this zone were due to lightning
because they tend to have large paved shoulders and
which strikes randomly.
few stopping areas in the Alpes-Maritimes, making
Finally, there is an intermediate zone where fire
them less likely to be sources of fire ignition.
ignition risk is at a maximum.
Fire ignition was observed to occur near an
autoroute at Aix-en-Provence,but this remains a rare In this sector, human activities are sufficiently dense
event and was ignored in this study. to create a high probability of fire ignition and yet low
enough for the surface not to be largely non-vegetated
7.2.2.2.3 Vegetation risk.
and artificial, as in the urban centre.
As for the two preceding factors, a risk coefficient
There is sufficient vegetation to provide a serious
was attributed to the different types of vegetation
fire risk and the human activities are neither dense
according to the relative weight of fire ignitions in each
enough to create largely artificial surfaces
type.
The number of fires and coefficients are described Despite the logical distribution of forest fire ignition
in Table 12. risk described above, the model has some faults.
Summing the coefficients of the three variables
The “moor” vegetation class had no impact on fire
should give a maximum value of 9 (3 variables times a
ignition and was eliminated from the analysis.
maximum coefficient value of 3), but the greatest value
On the other hand, bushes (“maquis”) were strongly
obtained is 16.
associated with fire ignitions and were attributed the
This indicates that some cells were counted more
greatest coefficient.
than once. In the dense road network of the urban
Shrub species favour fire ignition whereas tress
fringe, cells located less 300 m from two or more roads
species play an important role in fire propagation under
would be counted once for each road.
certain vegetation humidity conditions.
The fire ignition risk is therefore overestimated in
7.2.2.2.4 The “structural” risk associated with fire this sector.
ignition.
In addition, the choice of a 300 m limit is somewhat
The structural risk map shown below (Figure 50) arbitrary since this value, observed in Greece, may not
was elaborated working in raster mode with a cell size be the most appropriate in the Alpes-Maritimes region.
of 250 m, the resolution of the Corinne Land Cover.
Finally, the coefficient values varying from -1 to 3
The IGN BD Carto data base was in vector format
were selected intuitively and may not reflect the actual
and had to be rasterized.
weighted importance they represent.
Each raster cell was attributed the value of the
coefficient attributed to it above. 7.2.2.3 The second modelling approach: statistical
model A
The final ignition risk map was obtained by summing
the individual coefficients of the variables defined After the initial intuitive approach, a statistical
above. method was used to estimate fire ignition risk.

D-08-06.doc 51
EUFIRELAB
This second model is based on the number of fires
occurring with varying distances of the two main
factors, the road network and urban land-use.
In this approach, the number of fires occurring with
successive bands of 25 m was calculated to estimate
the range of impact of these factors on fire ignition.
The vegetation variable was not taken into account
since its spatial distribution was not adapted to the
approach.
In addition, no data were available on the density
and height of the different vegetation layers (of the
shrubs in particular) which could have been useful
descriptors of the contribution of the vegetation to
ignition risk.
7.2.2.3.1 Analysis of the factors at the departmental
scale
For the entire Alpes-Maritimes department, the
number of fires per band of 25 m from urban areas was
counted (Figure 51).
Unlike the distance from urban settlements, Figure
52 shows a correlation between the number of fire
ignitions and distance from the road network.
The number of fires decreases exponentially with
increasing distance with a coefficient of determination
2 created by applying the regression equations shown
(r ) greater than 0.95. Distance from roadways is
therefore considered a good predictor of fire ignitions.
In order to improve the model, the Alpes-Maritimes
department was separated into two sectors of different
urban density.
7.2.2.3.2 Separating the department into two sectors.
Before separating the department into two zones,
the dense urban coastal area was removed from the
study since the risk of a forest fire ignition there is
negligible.
The northern part of the department is mountainous
and rural with low population and road densities as can
be seen in Figures 53 and 54.
Between the northern part and the dense urban
coastal area, an urban fringe can be found..
7.2.2.3.3 The urban fringe.
- the impact of urban development : in this sector, the
number of fires decreases exponentially with the
distance from human structures (Figure 55). The
limits of the sectors appear justified since there is
now a good relationship between the number of fire
ignitions and distance from human settlements-
- the road network : fire ignitions decrease
exponentially with distance from a road in the urban
fringe.
7.2.2.3.4 The mountainous and rural northern sector
of the department
- the impact of urban development: in the northern
sector, there is no apparent relationship between
the number of fire ignitions and distance from
human settlements as can be seen in Figure 57.
Human activities are therefore not a good predictor
of fire ignitions in this area.
D-08-06.doc
- the road network: as it was the case in the urban
fringe, the number of fire ignitions decreases
exponentially with distance from a roadway (Figure
58). Distance from the road network is therefore a
good predictor of fire ignitions
For both the urban fringe and northern mountainous
areas, the distance from a roadway is a good predictor
of fire ignitions for a distance of up to 250 m from a
road.
Urban development is a good predictor only in the
urban fringe for a distance of up to 500 m.
In the northern sector, there is no apparent
relationship between human settlements and fire
ignitions, so this factor was eliminated in this sector in
the following analyses.
7.2.2.3.5 The fire ignition risk map developed using
statistical method A.
The fire ignition risk map using this method was
elaborated in the following way.
A distance grid is created for each of the factors for
the urban fringe and mountainous area.
It is important to note that these two geographical
entities are treated separately.
Once that has been done, a new raster data layer is
above.
This results in three new raster layers: two for the
urban fringe where both urban density and road
network are significant factors, and one for the
mountainous area where only road network is taken
into consideration.
Each layer shows the potential number of fires
estimated by the regression equations.
Classes of fire ignition risk are determined from the
sum of the factors in the urban fringe and the fires
predicted by the road network alone in the mountainous
area.
The results of Model 2, method A are shown in
Figures 59 (risk map alone) and 60 (risk map and
actual fire ignitions)
7.2.2.3.6 Discussion.
The relationship between fire ignition risk and actual
fires shown in Figure 60 appears satisfactory in the
mountainous area since most of the fire ignitions occur
with the high risk zone.
However, in the urban fringe several of the high risk
zones have few or no actual fire ignitions.
The validity of the model in this sector is therefore
put into question.
One of the reasons for the poor correspondence is
perhaps because the road network and urban density
are not independent since increasing urban density
inevitably increases road density and vice versa.
This covariance is not taken into account in the
urban fringe where the results of the predictive
equations are summed.
In the mountainous zone, much of the surface area
is in a low class risk, and some fires have occurred far
from both houses and roads.
52
EUFIRELAB
These can be attributed to either natural fires started
by lightning or to humanly caused fires associated with
pastoralism.
Shepherds sometimes ignite fires to maintain grass
quality and occasionally these fires escape their
control.
This hypothesis could be tested by determining the
location of these fire ignitions with respect to the sheep
grazing areas, but this information is not available
through local agricultural authorities.
Since the statistical approach used in the second
model proved deficient in the urban fringe, another
statistical approach was tested.
7.2.3 The third model: statistical approach B.
This third model is also based on the distance from
urban activities and roadways but it uses logistic
regression which takes into account all of the cells.
This type of regression uses binary type data:
absence or presence of a fire ignition between 1991
and 2003.
7.2.3.1 The method.
To optimize the calculations, the grid size was
increased to 1 km2 squares.
The dependent variable in this regression is the
number of fire ignitions between 1991 and 2003 in each
1 km2.
For each cell the mean distance from the centre of
the cell to a road or settlement is calculated.
Table 13 shows that the tendencies observed above
are still valid, but the relationship is weaker since cells
without a fire ignition are also included in the
regression.
In logistic regression, dependent variable is normally
binary (0 or 1), but since some cells may have been
affected by several fires (>1), the Y variable is
transformed into a series of repetitions which each
correspond to a Bernoulli law with the same
P parameter, so that Y follows a binomial law with
parameters n, and P, as shown in the equation below.
1 − P ( X 1, X 2)
LogitP ( X 1, X 2) =
P ( X 1, X 2)
= Exp-(a0 + a1X1 + a2X2)
The intensity of the risk is related to the
exp-(a0+a1*X1+a2*X2) term.
7.2.3.1.1 Results and risk maps.
The solution to the above equation gives the results
summarized in Table 14.
The results of the logistic regression show that the
values of the coefficients are statistically significant.
The values of their critical probabilities are always <
0.05.
D-08-06.doc
In addition, the coefficients enable us to estimate
the probability of fire ignition as a function of distance
from urban activities or settlements (X1) and roadways
(X2).
In order to obtain the erosion risk map, the equation
is applied to the database.
We then obtain P values which range from 0.26 to
0.008.
These values are then subdivided into four fire
ignition risk classes; added to these classes are the two
classes which include the cells not taken into account
(zero and very low).
Figures 61 and 62 show the spatial distribution of
fire ignition risk based on this method..
Nearly half the surface area is in the low risk class,
yet fires have occurred in these zones in the past. Most
of these fires can be attributed to either lightning, which
strikes randomly, or pastoralism, which tends to be
located far from roads and houses.
The remaining area has a variable risk according
the distance from roads and urban dwellings as
predicted by the logistic regression.
Much of the urban fringe is in a high risk class, as it
was for the first model.
Pressure from human activities in this zone makes it
particularly vulnerable to fire ignitions.
The coefficient of determination (r2) for the logistic
regression model is only 0.063, but this does not
invalidate the model since the r2 value is considered
less meaningful in logistic regression than in classical
linear regression.
On the other hand, the probability associated with
Pearson’s Khi2 value is pertinent and the risk of
erroneously attributing explicative value to the road and
urban variables is less than 0.1%.
7.2.4 Model comparison and conclusion.
From a simple visual evaluation of the risk maps,
the three models appear to give similar results apart
from differences in the spatial resolution.
In all three cases, the coastal area along the sea
has no risk due to its high urban density, a
mountainous zone in the North where the risk is low,
and finally an urban fringe zone between the two others
where the risk is particularly great.
Despite the similarities, some differences between
the models persist.
The first two models have spatial resolutions of 250
m whereas the third has a resolution of 1 km.
The loss in spatial detail is compensated for by a
more objective approach and valid statistical
parameters.
Models 1 and 3 are the preferred options: the first
for its simplicity and ease of application, the third for its
rigour.
53
EUFIRELAB
The three models agree with Figure 63 which shows
the general risk behaviour associated with human
pressure on the environment: in the first segment
(mountainous), the risk is low due to low human
pressure, the main cause of fire ignition.
In the third segment (coastal), the risk is also low,
but in this case it is due to the scarcity of the vegetation
in this densely urbanised environment.
In the second segment (urban fringe), the risk is at a
maximum due to an abundant supply of combustible
vegetation and the greater likelihood of fire ignition due
to a high concentration of human activities.
Improving fire ignition risk modelling can only be
done with better data acquisition techniques.
Both fire ignition location and cause must be
investigated and recorded systematically in order to
provide a fuller and more reliable data base.
In Spain, for example, forest fires causes are
defined for the vast majority of fires, leaving very few
unexplained.
Example of human ignition danger estimation at
european scale
Continuing the work developed in Megafires Project
(see section 2.1) a geo-statistical analysis to model
wildland fire occurrence was carried out at provincial
level in Southern Europe from socio-economic and
demographic indicators of 1960 and 1990 together with
variables that describe land cover and agricultural
statistics.
This work is has been presented in Koutsias et al
(2005) and is extended and inproved in a new paper
actually in revision (KOUTSIAS et al. 2006).
Here is presented main results obtained in the first
one.
In this study the classical OLS linear regression
modeling has been used along with geographically
weighted regression modeling to explain long-term
wildland fire occurrence patterns at Southern Europe.
By applying the global OLS linear regression we
identified critical underlying causal factors for wildland
fire occurrence from a set of socio-economic,
demographic, land cover and agricultural statistics.
Traditional regression modeling assumes that the
relationship between the dependent variable and the
explanatory variables is constant regardless of their
geographical location.
This assumption, which is known as stationarity, is
often violated in real world situations.
However, this constraint is overcome by another
approach, known as geographically weighted
regression (GWR), which considers that such
relationships vary in space according to their location
and that allows local variations to be taken into
account.
In wildland fire occurrence modeling, especially
when the geographical extent of the study area is large
(i.e. the whole Southern Europe), it would be more
reasonable to find varied rather than constant
relationships.
D-08-06.doc
In this study, non-stationarity is the main assumption
made for modeling the driving factors behind wildland
fire occurrence.
To overcome the constraints of traditional
regression modeling which assumes stationary
processes we applied geographically weighted
regression (GWR) analysis.
GWR considers that relationships vary in space
according to their location and allows local variations to
be taken into account.
7.2.5 Study Area and Wildland Fire Database
The geographical extent of our study area includes
Portugal, Spain, Southern France, South-Eastern
Switzerland, Italy and Greece (see Figure 64).
Within the study area census data at provincial level
were acquired from national statistics.
In total, 153 geographical units were identified, 18 in
Portugal, 48 in Spain, 14 in South France, 20 in Italy, 2
in South-Eastern Switzerland and 51 in Greece.
For most of the countries these geographical units
correspond to NUTS-3 level (Nomenclature of
Territorial Units for Statistics) except for Italy where
they correspond to NUTS-2 level and for Portugal
where they correspond to their national district level.
Unfortunately, this not fully homogenized division
among the countries, which resulted from limitations on
data availability, could eventually influence negatively
the regression analysis.
The basic data used in our study come from the
MEGAFiReS project 1999) with an update for including
the Southern part of Switzerland together with the
number of wildland fires occurred within the period
1992-2000.
The database is composed by socio-economic and
demographic indicators together with variables that
describe land cover and agricultural activities.
In total, 77 variables were established for 1960 and
1990.
Differences of some variables (mainly for population
characteristics) between the two date sets were also
computed showing changes in population
characteristics associated to urbanization and land
abandonment.
These selected variables were considered of being
potential underlying causal factors for explaining long-
term fire occurrence patterns.
Besides the census data, the mean annual number
of wildland fires, having occurred between 1992-2000
at provincial level, was also computed from national
forest fires statistics.
This variable is used as depedent variable.
For Portugal, Spain, Southern France and Italy
wildland fire observations were provided using the
community centroids for the period 1992-2000.
For Switzerland the x and y coordinates of the
ignition points were used having occurred in the same
period.
Finally, for Greece, wildland fire observations were
provided using the x and y coordinates of the ignition
points for the period of 1985 to 1995.
54
EUFIRELAB
Out of this original database, forest fire data were
aggregated at provincial level so that to much the
aggregation scheme used for the census statistics.
Wildland fire occurrence was expressed using the
mean annual number of fires to overcome the
inconsistence observed between the time period of
Greek data and the rest of the countries.
A qualitative visualization of the spatial distibution of
the dependente variable reveals critical regions for fire
occurrence.
In future studies the mean annual number of fires is
going to be estimated for two different fire size classes
(burned area above 1 ha and above 100 ha).
7.2.6 Methodology: Geographically Weighted
Regression (GWR), the linear and logistic
case
GWR tries to capture the spatial variation by
calibrating a multiple regression model so that at
different points in space different relationships between
variables can be found (ZHANG and SHI 2002).
A regression model is fitted at each data point by
weighting all observations from that point as a fuction of
distance.
Consequently, the neighbors sampled around the
point influence the regression coefficients more
strongly than the observations farther away.
GWR estimates the parameters at each point in the
study area which then can be mapped using for
instance geographic information systems (GIS) to
investigate local spatial variation in the regression
relationships (FOTHERINGHAM et al. 2002).
In geographically weighted regression the
relationship between y and x can be expressed as:
p was found on the % of forested area, the % of the
y i = β 0 (u i ,ν i ) + ∑ x ij β j (u i ,ν i ) + ε i
j =1
where
- β0(ui, νi), and βj(ui, νi) are estimated coefficients as
a function of location, and
- ε is a random error term.
Besides the linear regression model we developed
also a global and GWR logistic model.
In addition to linear regression, where the outcome
variable is supposed to be continuous, logistic
regression presupposes the existence of a binary or
dichotomous dependent variable.
This main difference between the two regression
models renders logistic regression very popular,
especially when the experimental question can be
expressed in a dichotomous way.
Besides this, the independent variables in logistic
regression modelling can be a mixture of continuous
and categorical variables.
Consequently, the assumption of multivariate
normality is not presupposed in logistic regression.
Previous use of LR in fire occurrence at provincial
level in Southern Europe can be found in Chuvieco et
al. 1999b).
D-08-06.doc
7.2.7 Results and discussion
7.2.7.1 Global versus GWR linear model
A stepwise ordinary least squares (OLS) regression
was carried out for developing a model with the most
significant variables.
Eight of the original variables have been chosen by
the model which explains 53% of the variation of the
dependent variable.
Among them the density of livestock (i.e., sheep) in
1990, the density of agricultural employees in 1990, the
percentage of forested area, and the difference of the
youth index between 1960 and 1990 are the most
important variables based on the criterion of
standardized coefficients of the model.
The GWR provided significantly better results than
the global regression model, since the variance of the
dependent variable explained increased to 68.65 %.
This is caused by the consideration of the non-
constant relationships between the dependent and the
explanatory variables, which change throughout the
EUMed basin.
The residuals in GWR (Figure 64) are significantly
less than those of OLS regression, indicating the better
performance of the former model.
Based also on an ANOVA test, statistical significant
improvement of the GWR model over the OLS linear
regression model was found.
A Monte Carlo significance test for the parameters
of the model found significant spatial variability on the
intercept and on the coefficients of the difference for
the youth index between 1960 and 1990, the number of
agricultural employees in 1990, and the density of
sheep in 1990.
Based on the same test, non-significant variability
difference of the active population between 1960 and
1990, the density of agricultural employees in 1990, the
% of difference in size of agriculture exploitations
between 1960 and 1990, and the number of sheep in
1990.
7.2.7.2 Global versus GWR logistic model
Besides the OLS linear regression model we
developed also a global and GWR logistic model. As
mentioned, logistic regression presupposes a binary
dependent variable that takes the value 1 in case of an
event and 0 otherwise.
Thus we created a new variable which is based on
the reclassification of the kernel density surfaces (two
classes based on the equal area criterion for each
country).
Then, the median value (1 or 0) found inside each
polygon (= provinces) was attached as the dependent
variable (Figure 65). We used the same independent
variables as in the case of linear regression.
55
EUFIRELAB

Three out of the four most significant variables of

7.2.8 Conclusion
linear regression (the density of agricultural employees
in 1990, the percentage of forested area, and the In conclusion, GWR seems to be a valuable
difference of the youth index between 1960 and 1990) approach for exploring and modeling non-stationary
were considered also by logistic regression as the most relationships between the response and explanatory
significant ones based on the Wald statistic. The GWR variables and thus to better understand the spatial
logistic model improved the deviance (-2LL) from processes in wildland fire occurrence.
178.35 to 116.54. The reduction of Moran’s I spatial autocorrelation of
Also, the GWR logistic model classified 84.97 % of the residuals indicates that GWR models were more
the observations successfully, which is significantly appropriate than the classical global ones.
higher than 66 % of the ordinary global logistic model.
The residuals in GWR logistic model, as shown in
Figure 65, are less than those of ordinary logistic
regression indicating the better performance of the
former model.
7.2.7.3 Spatial Autocorrelation of Residuals

To further explore the advantages of the GWR over

the classical global approach we calculated the
correlograms of the residuals which show how they are
distributed in space (Figure 66).
A reduction of Moran’s I spatial autocorrelation is
evident in lag distances beyond 100 Km especially in
the GWR linear model.
This reduction indicates that the model is
appropriate.
The existence of spatial autocorrelation in OLS
regression violates the assumption of independent
errors (LICHSTEIN et al. 2002).

D-08-06.doc 56
EUFIRELAB

7.2.9 Figures

Figures 44 and 45: Topographic map (Source: BD Carto IGN) and Vegetation map (Source: Corinne LC)

Figure 46 and 47: Landuse types map (left) (Source Corinne LC) and road network map (right) of Alpes-Maritimes.
(Source: BD Carto IGN)

D-08-06.doc 57
EUFIRELAB

Figure 48: Recorded fires between 1991 and 2003 in the Alpes-Maritimes

Left: Figure 49: Spatial distribution of fires occurring during the 1991-2003 period according to cause

Right: Figure 50: Map showing the structural forest fire ignition risk
The black points show actual fire ignitions, and the colours indicate the calculated risk

Figures 51 and 52: Number of fires versus distance to habitats (left) and to road network (right)

D-08-06.doc 58
EUFIRELAB

Figures 53 and 54: Urban density (left) and road network density (righ) in the Alpes Maritimes deparment

Figures 55 and 56: Number of fires versus distance from homes (left) and from a roadway (right) in the seashore

Figures 57 and 58: Number of fires versus distance to settlements (left), and to a roadway (right) in northern sector

D-08-06.doc 59
EUFIRELAB

R= ∑ F(Xi)
ie Rlito = 45,545*Exp(-0,0126*X1) + 23,65*Exp(-0,0053*X3) (1)
Ra-pays = 38,10*Exp(-0,0101*X2) (2)
X
1: Littoral road network X2 : Northern sector road network X3 : Littoral distance to habitations

Figures 59 and 60: Fire ignition risk map, model 2 (left) with 1991 - 2003 ignitions shown in white (right).

Figures 61 and 62: Fire ignition risk map obtained using the logistic regression model (left)
with fire departures between 1991-2003 (right)

D-08-06.doc 60
EUFIRELAB

7.2.10 Tables

Table 10: Number of fire ignitions according to urban density and ignition cause
for fires less than 300 m from a residence.
Discontinuous Agricultural with Natural areas with
Origin Dense urban Single dwellings
urban isolated dwellings isolated dwellings
Unknown 2 11 17 23 22
Voluntary 0 1 3 7 2
Other 1 0 2 10 6
BBQ 0 0 0 3 0
Gain 0 0 0 0 1
Electrical 0 0 0 0 0
Pastoralism 0 0 0 0 0
Lightning 0 0 0 0 1
Construction 3 0 4 12 9
% 5 10 18 34 32
Coefficients -1 1 2 3 3

Table 11: Number of fire ignitions according within 300 m of a road for each ignition cause.
Origin Autoroute National route Departmental Municipal
Unknown 3 13 47 112
Voluntary 0 2 4 21
Other 0 5 0 5
BBQ 0 6 10 21
Gain 0 0 2 0
Electrical 0 0 0 0
Pastoralism 0 0 1 15
Lightning 0 0 1 0
Construction 0 4 9 41
% 0.08 8 20 62
Coefficients -1 1 2 3

D-08-06.doc 61
EUFIRELAB

Table 12: Number of fire ignitions according to vegetation type and ignition cause.
Origin Coniferous Broadleaf Mixed Shrubs Grass Bush
Unknown 37 43 3 13 0 53
Voluntary 4 3 0 3 0 10
Other 1 1 0 0 0 0
BBQ 10 8 3 0 0 10
Gain 0 3 0 2 0 3
Electrical 1 0 0 0 0 1
Pastoralism 5 1 0 1 0 6
Lightning 5 5 0 0 0 1
Construction 11 8 1 1 0 7
% 27 27 3 7 0 35
Coefficients 2 2 1 1 0 3

Table 13: Correlation matrix

Fires Urban distance Road distance
Fires 1 -0,060 -0,109
Urban distance -0,060 1 0,119
Road distance -0,109 0,119 1
Bold font indicates a significant corrrelation at α=0.05 (bilateral test)

Table 14: Results of the logistic regression and equation giving the fire ignition probability
as a function of two factors, X1 and X2.
Estimated model parameter values (maximum likelihood) :
Parameter Estimated value Standard deviation Khi² Pr. > Khi²
Constant -3,515 0,107 1073,719 < 0,0001
Urban distance (X1) 0,000 0,000 10,999 0,001
Road distance(X2) -0,005 0,001 50,738 < 0,0001

P(X1, X2) = 1 / (1+ Exp-(-3,515 - 0,000158*X1 - 0,005*X2))

D-08-06.doc 62
EUFIRELAB
7.3 HUMAN IGNITION DANGER IN SOUTHERN
EUROPE BASED ON FIRE OCCURRENCE
MAPS
The objective of this and next Eufirelab deliverables
is to elaborate a fire risk index at European scale.
In order to do it, and taking into account the
resolution of satellite images and the size of the study
area, it was decided that the ideal resolution of the pixel
for the GIS layers must be 1 square kilometre.
In the case of the human ignition factors it is very
difficult to get and have access to available and reliable
data at this resolution or scale, for the whole or at least
big part of the Euro-mediterranean area.
The main problem we find with the spatial variables
(distances, fragmentation, interfaces) derived from
cartographic layers like roads, urban areas, forest and
land use covers, is that this original layer are very
generalist at this global scale, with very poor detail. In
the cases of socio-economic factors (unemployment,
rural population age, population density…) estimated
from census data, the main inconvenience is that they
are only available in these moments for provinces.
This data unavailability at the desired scale
(municipality level) for the whole area forced us to
discard the socio-economic variables because we don’t
consider appropriate to rasterize at 1 km2 the
information derived from provinces.
The same problem regarding vector layers, like
roads, with scales larger than 1: 1.000.000.
Because of these present difficulties to have an
operational human danger ignition layer, one possibility
can be generating it from historical data of fire
occurrence.
From the location of ignition points and using the
kernel density approach, that is a well known and
explored interpolation method, we can generate a
probability surface, as is showed in section 2.2 about
Spread Project review.
Occurrence maps provides a global view of fires,
which are mainly caused by human factors in Europe.
Visual exploration of spatial distribution of point
observations is popular in point pattern analysis and
takes place prior to any further analysis (BAILEY and
GATRELL 1995, Fotheringham et al. 2000).
By visually examining point observations, one may
acquire a prior idea about their spatial distribution,
suggest hypotheses about explanatory factors, and
decide for further statistical analysis and modeling
(SADAHIRO 2003).
In the following section is explained deeply the
generation of this map, according with the information
of Deliverable D143 “Fire Risk and Human Factors (III)”
from Spread Project.
7.3.1 Fire Occurrence Hot Spot Areas
For the analysis of large-scale fire occurrence
patterns, we often need to transform fire ignition data
that come in different formats and accuracies (e.g. x
and y locations or number of fires per area unit) into
continuous surfaces.
D-08-06.doc
As a non-parametric smoothing technique, the
kernel density estimation (SILVERMAN 1986) is well
suited for converting fire ignition point data into
continuous “fire occurrence distribution surfaces”.
Kernel density estimation is based on the estimation
of the density at each intersection of a grid
superimposed on the data, after placing a probability
density (kernel) over each point event.
Depending on whether constant or multiple adaptive
values are used for the smoothing parameter, kernel
density estimation is distinguished into the fixed and
adaptive method, respectively. (SILVERMAN 1986,
Levine 2002).
KOUTSIAS et al. (2004) applied the kernel approach
to assess fire occurrence patterns at landscape level by
addressing some of the inherent positional inaccuracies
of the fire ignition locations.
DE LA RIVA et al. (2004) used also this approach to
express fire occurrence patterns at municipality level by
using fire ignition observations.
The current study proposes an adaptive kernel
density interpolation approach applied to community
centroids where the number of fires per community is
used as the intensity variable for the kernel density
estimation.
The final aim is to define large-scale fire occurrence
patterns and to identify ‘fire occurrence hot spot areas’
in southern Europe, using fire ignition observations
aggregated at community level. In our study the
number of fires form the period 1992-2000 have been
estimated from national fire statistics and expressed at
community level using the community centroids (Figure
67).
The geographical extent of the study covers the
European Mediterranean countries including Portugal,
Spain, Southern France, Italy, Greece, as well as
South-Eastern Switzerland (Canton of Grison and
Ticino) from central Europe (Figure 67).
Motivated by the non-homogeneous spatial
distribution of community centroids (Figure 67) we
decided to choose the adaptive kernel density
estimation mode instead of the fixed one.
The adaptive mode allows for the adjustment of the
bandwidth size in relation to the concentration of the
community centroids (WORTON 1989).
For the adaptive kernel approach different
bandwidth sizes have been tested (from 1 to 40 at
various steps) that are indicative of typical resulting
patterns of under- to over-smoothing.
Here we present four representative steps (2, 5, 10,
20).
For a homogeneous process with no spatial
dependence, the expected number of events within a
distance d of a randomly chosen event equals ?pd2
where ? is the density (GATRELL et al. 1996).
Thus, if the expected number of events corresponds
to the number of centroids used in the adaptive kernel,
the estimated distance d can be used for defining the
bandwidth size in the fixed mode.
The corresponding bandwidth sizes in the fixed
mode are 3622, 5726, 8099 and 11453.
63
EUFIRELAB

We applied both approaches, the fixed and the Thus kernel densities have been reclassified to 10
adaptive kernel density interpolation, at national level in classes based on the equal area criterion within each
Portugal in order to decide about the most proper country (Figure 69), presupposing equivalence for fire
approach and the most appropriate bandwidth size hot spot areas among the countries.
(Figure 68).
The effect of small bandwidth size to kernel density
surfaces is evident in the corresponding images in
Figure 68.
Low bandwidth sizes for both, the adaptive and the
fixed kernel density, (e.g. comprising 2 and 5
community centroids or 3622 and 6726 meters),
created surfaces in which the local variation of the
density surfaces is enhanced around the interpolated
points.
This effect is more evident in the fixed kernel density
approach because the bandwidth size remains
constant throughout the whole extent of the study area
while the concentration of the interpolated points
(community centroids) varies significantly.
This variation in the concentration of community
centroids is taken into account by the adaptive kernel
mode since the bandwidth size varies locally so each
time the same number of community centroids to be
used.
Based on visual inspection, among the different
bandwidth sizes used the one that corresponds to 10
community centroids is considered of being the most
appropriate to avoid under- or over smoothing.
To evaluate the density surfaces three grids of 2, 5
and 10 Km resolution were overlaid over the density
surfaces.
Subsequently the number of fires and the sum of
density kernel values were estimated.
The correlation coefficients (Table 15) were slightly
higher between the number of fires and the adaptive
kernel density surfaces.
This is another indication that the adaptive kernel
density interpolation performs better compared to the
fixed kernel density as the former allows for locally
varying bandwidth sizes.
The same methodology described previously and
applied at national level in Portugal has been applied to
the rest of the countries.
The absolute values of kernel densities for each
country depend on the number of fire ignition points as
well as on the spatial distribution of community
centroids and wildland fire observations.
The total number of wildland fire ignition
observations varies greatly between the different
countries and consequently kernel density values are
also very diverse.
The consequence of this variability found among the
countries results in an underestimation when the kernel
density surfaces are joined together under a common
classification scheme as shown in Figure 69.
This problem can be overcome if a reclassification
of kernel density surfaces is applied before merging all
data.

D-08-06.doc 64
EUFIRELAB

7.3.2 Figures

Figure 63: Fire ignition risk evolution according to vegetaion and human activity characteristics. (Carrega, 1992)

Figure 64: Observed and predicted values of OLS and GWR linear models and the distribution of their residuals.

Figure 65: Observed and predicted values of ordinary and GWR logistic models and their residual distribution.

D-08-06.doc 65
EUFIRELAB

Figure 66: Correlograms of the residuals

Figure 67: Community Centroid

Figure 68: Kernel density estimates using the fixed (upper images) and
the adaptive kernel approach (lower images).

D-08-06.doc 66
EUFIRELAB

Figure 69: Large-scale fire occurrence patterns at European level by kernel density estimation

7.3.3 Tables

Table 15: Correlation coefficients between number of fires and kernel densities at 2, 5 and 10 km grid resolutions

D-08-06.doc 67
EUFIRELAB
8 PROPAGATION DANGER INDEX
The propagation danger component hast to cover
the potential hazard that a fire propagates in space and
time.
This index is associated to estimating the behaviour
of the fire.
Fire behaviour is mainly described with reference to
the propagation rate and the intensity of the flame front.
Not too many studies include this type of index. At
global scale, a case applied to Euro-Mediterranean
basin has been included in the Spread project,
described in chapter 4.
The propagation danger index (PD) is composed by
two subindex: the average rate of spread (RoS), and
the flame length (FL).
Both variables are calculated using BEHAVE
algorithm.
The input data is obtained from CORINE and a
global Digital Terrain Model.
The results is a map taken as static, it will not
change throughout the fire season.
Taking into account the characteristics of the index
described in Spread project (input variables, spatial and
temporal resolutions), we propose this index as the
“Propagation Danger Index” ot he Euro-Mediterranean
Wildland Fire Risk Index (EM-WFRI) defined in this
deliverable.
The definition and estimation of this index are also
described in the same chapter 4.
D-08-06.doc
8.1 AVERAGE RATE OF SPREAD AND FLAME
LENGTH
In order to obtain a global view of risk associated to
fuel loads, terrain characteristics and wind flows, a
global simulation analysis was performed.
This analysis tried to obtain average values of rate
of spread and flame length, considering different wind
and topographic conditions for the estimated fuel maps
of the whole EUMed area.
This attempt should be considered as a general
overview of average expected fire behaviour at global
scale, in order to rank different danger levels according
to the combination of fuel and terrain spatial patterns.
The estimation of the average RoS was based on
several simulations performed by the Autonomous
University of Barcelona for different fuel types, slope
ranges and wind flows.
As a simulation kernel the wildland simulator
proposed by COLLINS D. BEVINS, which is based on the
fireLib library (COLLINS, 1996) were used.
FireLib is a library that encapsulates the BEHAVE
fire behaviour algorithm (MORGAN et al., 2001).
In particular, this simulator uses a cell automata
approach to evaluate fire spread.
The terrain is divided into square cells and a
neighbourhood relationship is used to evaluate whether
a cell will be burnt and at what time the fire will reach
the burnt cells.
As inputs, this simulator accepts maps of the terrain,
vegetation characteristics, wind and the initial ignition
map.
The output generated by the simulator consists of
two maps of the terrain.
In the first one, each cell is labelled with its ignition
time; in the second one, each cell is labelled with its
flame lenght.
This information must be used to calculate the rate
of spread and an average from among all flame lenght.
To calculate the rate of spread, the distance
between the ignition point and each particular cell in the
terrain is divided by the ignition time of that particular
cell.
This calculation is repeated for each cell in the
terrain to determine the maximum value of the rate of
spread.
This maximum value is used as the rate of spread
for that particular situation.
To provide the propagation danger map, a set of
prototype plots was created, considering all the fuel
models from Rothermel classification and a certain
slope percentage (from 0 to 100%, with a step of 5%).
The total number of plots was 273. Each plot
consists of a grid of cells with 11 columns x 11 rows
(each cell measured 328.083 x 328.083 feet).
The ignition point was located in the middle of the
plot.
68
EUFIRELAB

For each plot, many input parameter combinations
were used to simulate the wildland fire behaviour and
the average rate of spread and flame lenght were also
calculated.
The parameters considered for variation were: 1-hr
dead fuel moisture, 10-hr dead fuel moisture, 100-hr
dead fuel moisture, live herbaceous moisture and wind
speed and direction.
Average values of rate of spread (RoS, in m/min)
were computed for the different fuel types.
Finally, the RoS values were scaled into a 0-1
range, by normalizing the values between the
maximum and minimum values (figure 70).
Fuel types were derived from the Corine land cover.
The resulting map is considered static, since no
specific conditions are simulated (wind or FMC), but
only general patterns of propagation rates.
Same process as above was performed to compute
the average flame length (FL), measured in metres and
normalized into a 0-1 scale (figure 71).
8.2 PROPAGATION DANGER (PD)
It derives from the combination of the two
intermediate products previously described: RoS and
FL.
The results of the simulations were mapped at
EUmed scale using CORINE land cover (reclassified
into fuel models) and slope maps
The maps of FL and RoS were then normalized
using linear fitting and multiplied to produce PD:
PD = [(RoSi – RoSmin)/(RoSmax - RoSmin)+0.001] *
[(FLi – FLmin)/(FLmax - FLmin)+0.001]
A small constant (0.001) was added to avoid zero
multiplication in case of minimum values.
RoS and FL were considered in this formula of
equal importance, although this could be tuned up in
future improvements, according to further experience or
suggestions.
This map is taken as static, i.e. it will not change
throughout the fire season.

8.3 FIGURES

Figure 70:Estimated average Rate of Spread (Normalised values from 0 to 1)

Figure 71: Estimated average Flame Length (Normalised values from 0 to 1)

D-08-06.doc 69
EUFIRELAB
9 VULNERABILITY INDEX
9.1 POPULATION VULNERABILITY
9.1.1 Introduction
A vulnerability analysis is realized to obtain
information about the susceptibility of individuals,
property, and the environment to the adverse effects
of a given hazard in order to develop appropriate
prevention strategies.
The analysis of this information helps to determine
who is most likely to be affected, what is most likely
to be destroyed or damaged, and what capacities
exist to cope with the generated effects.
To perform a proper and detailed vulnerability
analysis it should be necessary:
- to identify that part of population which is
vulnerable because of proximity or exposure
factors;
- to assess vulnerability of buildings, infrastructure
and other aspects of the physical environment
due to factors such as site, materials used,
construction technique, and maintenance;
- to evaluate transportation systems,
communications systems, public utilities (water,
sewage, power), and critical facilities (i.e.,
hospitals) for weaknesses; to identify most
vulnerable population (e.g., older adults, children,
single-parent families, the economically
disadvantaged, the disabled);
- to estimate the level of poverty, jobs that are at
risk, and the availability of local institutions that
may provide social support;
- to determine the community’s potential for
economic loss and recovery following a disaster;
- to identify existing measures and resources that
can reduce the impact of a given hazard.
The degree to which populations are vulnerable to
hazards is not solely dependent upon proximity to the
source of the threat or the physical nature of the
hazard, social factors also play a significant role in
determining vulnerability (CUTTER et al., 2000).
The primary social effects of wildfires are physical:
houses and other kinds of human structures will burn;
people may be injured or may die.
However, it should be noted that death and injury
depend upon unique sets of circumstances
surrounding each fire event: these cannot be
modelled here.
For instance, as noted by CASE et al. (2000), in a
geographic information system, we can only describe
the total number of people at potential risk, and the
number of people having special characteristics
which place them in jeopardy.
Some individuals belong to special populations
because they have characteristics, which lead them
to suffer effects disproportionate to the general
population.
Special populations include families with very
young children, the elderly, and households whose
members have incomes below the poverty line.
D-08-06.doc
Risks to these populations, as pointed out by CASE et
al. (2000), is expressed as a greater difficulty of
successful evacuation, greater susceptibility to health
impairment, loss of possessions (which could be
irreplaceable for people in two of these special groups),
and loss of jobs and income.
In brief, houses which are dependent upon domestic
water supplies and fuel-wood, which lie in remote areas,
and which have difficult routes of ingress and egress are
at greater risk of damage or destruction in the event of
wildfire.
People who are very young or very old are more likely
to experience adverse health consequences in the event
of wildfire, and people who have very low incomes are
more likely to suffer irreversible economic and social
consequences. When these groups of people suffer from
wildfires, they have less individual means of recovery,
and require more help from the communities they live
within (CASE et al., 2000).
To identify those sections of a community most likely
to be affected by a particular hazard and to determine
areas that require strengthening to prevent or mitigate the
effects of the hazard, data on several variables must be
collected (among which, information on the size, density,
location, and socio-economic status of the at risk
community).
Vulnerability assessment is one of the least
investigated tasks, due partly to the lack of relevant
detailed socio-economic data and the difficulty of their
effective spatial representation for integration with
physical environmental data on hazards (CHEN et al.,
2003).
Vulnerability assessment is however an important task
in risk assessment and has social significance for a
hazard-prone vulnerable community.
Assessing vulnerability in spatial terms requires a
wide range of physical and socio-economic knowledge
and expertise.
In this context, GIS can be used for database
establishment, analytical modelling, and decision support
in a decision-making process.
GIS can play an important and integral role in
lessening the adverse impacts of natural hazards on
society.
Having a wide range of spatial analysis techniques
and tools, they are helpful for identifying, measuring, and
assessing many aspects of natural hazards and their
consequences (CHEN et al., 2003).
GIS spatial analysis in particular, applying various
methods and techniques, has the ability to employ
physical environmental and socio-economic data for risk
and vulnerability analysis.
High-quality GIS databases support subsequent risk
assessment and rational decision-making in a spatial and
temporal context, which can help risk managers and the
public understand how complex hazards and their
consequences will affect vulnerable communities.
70
EUFIRELAB

Spatial information system can be used to Additionally, some crucial social effects depend very
estimate the physical risk to people and houses, and much on the exact nature of the circumstances
to draw useful inferences about the nature of surrounding each fire or upon problematic factors that
secondary, less tangible, and less predictable social cannot be adequately described in any known model
effects resulting from wildfire (CASE et al., 2000). (CASE et al., 2000).
To reach these tasks geo-referenced data layers Some general and basic census data can be utilized
must be collected from a variety of sources. to develop indicators of relative risk or vulnerability to
human populations.
When we have to assess vulnerability to wildfire,
The elements chosen for this analysis can be the total
and in general in all the natural hazards context, the
population and population density within each unit (for
required data can be divided into three categories -
instance a 1 km2 grid) and populations that might be
physical environmental data, socio-economic
placed at additional risk due to their circumstances
environmental data, and management-related data
(children, elderly, low-income).
(CHEN et al., 2003).
These factors can be chosen as examples, not as an
Socio-economic environmental data are in exhaustive listing of relevant social information; however
particular used to assess community vulnerability and they can give some information about the ways in which
include population and housing census data and data people experience wildfire, and how they may be affected
on utilities and access. by it, providing decision makers with useful information in
making judgments about prioritising mitigation and
Because of the increasing emphasis on protection programs (CASE et al., 2000).
community vulnerability assessment, collecting socio-
economic data is essential. We could construct a density map of these special
According to Granger 1998) detailed information populations, reasoning that proportionately greater risk
on setting, shelter, sustenance, security, and society occurs where people and houses with these particular
is required. characteristics occur in greater numbers (CASE et al.,
For example, data on shelter include construction 2000).
materials of the walls, roofs, and floors, and the ages
Another crucial data set is represented by the number
of buildings.
of ignitions per 1 km2 grid over a certain historical period.
Data on utilities (e.g., water, electricity,
On the assumption that areas experiencing the most
telecommunication, gas), data on security facilities
ignitions in the past will continue to experience the
(e.g., hospitals, police stations, emergency
highest relative ignition rates, areas most likely to
operations centres), and data on access (e.g., roads,
experience future wildfire ignitions can be identified
bridges, tunnels, railways) are also required,
(NEUENSCHWANDER et al. 2000).
wherever possible.
The importance of assessing population vulnerability
The concepts previously expressed can be
is particularly relevant when we have to face the problem
usually applicable to a local scale, as a consequence
of wildland-urban interface (WUI) areas.
of the great amount of required input data and their
very refined spatial resolution. Wildland-urban interface in the European environment
and more specifically in the Mediterranean area is a very
The possibility to obtain a refined result is typically
complex spatial context with many interrelated social,
dependent on the chosen working scale and is strictly
natural resource and wildfire issues.
related to the availability and the resolution of data.
The problem of wildfires in the wildland-urban
The scale and resolution of the data that can be
interface has recently become quite relevant because of
used (typically available at the state scale) limit the
the increasing number of dwellings near to or inside
detail that can be produced.
natural areas and of the increasing number of wildfires
This choice can be conflicting with the desire of involving these sites.
scientists to be precise, and the needs of land
As previously underlined, Geographic Information
managers to identify specific project activities and
Systems (GIS) can be useful tools for WUI management,
locations (SAMPSON and NEUENSCHWANDER, 2000),
through their capability of handling in an integrated
but is consistent with the objective of producing a
environment multi source and multi resolution spatial data
strategic assessment, at least concerning population
(BURROUGH and MCDONNEL, 1998).
vulnerability.
At the 1:1,000,000 scale, corresponding to the
one selected to develop the Euro-Mediterranean
Wildland Fire Danger Rating System, the social data
sets above all are necessarily not detailed enough to
allow us to quantify the complete suite of social
effects we know can occur.

D-08-06.doc 71
EUFIRELAB
9.1.2 Example of population vulnerability
mapping in Piemonte Region (Italy)
As an example of GIS analyses that can be
performed in WUI areas to assess wildfire risk and
vulnerability see CAMIA et al. (2004).
The analyses described in this work were
performed in the framework of WARM Project.
They were firstly aimed at identifying and typifying
settlements, providing a first description of the
surrounding environment and of fire suppression
preparedness.
As previously mentioned, the study area is
characterized by small urban settlements, presenting
both situations of classical interface and intermixes.
Actually, an entire set of conditions, ranging from
the densely inhabited village with boundaries facing
wildland and forests until the isolated houses are
represented.
The authors wanted to identify and classify the
different situations according to the degree of
urbanization within the wildland area, and therefore
the general wildland-urban mixing conditions for the
different settlements, being clear the usefulness of
such a characterization from a fire management point
of view.
This task was approached using density analysis
techniques to the house layer of the database, and
applying thresholds to the derived house density
layer and identify in this way settlement boundaries.
Settlement types were then characterized by
computing the number of houses in each polygon
discriminated, performing a spatial join between the
settlements and the houses layers.
The presence of specific fuel types, as well as the
topographic arrangement of the area, provides useful
indications about the fire hazard and risk conditions
around the settlements, but also about the possibility
and strategies for fire fighting resources.
In addition to such key features, from a fire
suppression point of view, it is important to know the
accessibility of the settlements to be protected, as
well as the relative location of fighting resources (fire
stations) and water supplies.
Although a road density map can provide useful
information on the accessibility of the sites, much
more informative is the map derived from the network
analysis that allows to classify each settlement
according to the time needed to reach it from the
nearest fire station or the time from the nearest water
supply.
Working at a global scale within a 1-km² grid, an
example of what can be obtained with simple and
necessarily coarse data, is reported in Camia et al.
(2002).
In this work, a methodology was developed to
analyse the problem of WUI wildfires and to
characterize its spatial distribution at the scale of the
Italian Regions.
D-08-06.doc
Italy is divided into 20 Regions, characterized by deep
environmental differences; the study was applied in
Piemonte, a Region located in the north-western part of
Italy, with a surface of about 2,5 million hectares.
It has a quite heterogeneous territory, characterized
by plains mainly interested by agriculture, hilly areas and
mountains among the highest of Europe.
There are 1.209 municipalities, whose population
ranges from about 1.000.000 inhabitants to less than 100
inhabitants.
In the country areas the secondary houses are quite
frequent and their number has been strongly increasing in
the last years.
The fire season is typically a winter-early spring one,
which is mainly characterized by surface fires.
Crown fires are mostly due to the presence of
evergreen coniferous stands, while most forests are
made up by broadleaved species.
Most wildfires break out and spread in the mountain
belt from 200 m to 1000 m a.s.l..
Data about wildfires are collected and recorded with a
spatial resolution of 1-km2 (i.e. on a UTM kilometric grid
with square cells of 1000 m size).
For each wildfire, the cell where the fire started is
recorded, while the adjacent cells that could have been
interested by the fire spread are not reported.
Only recently, information on fire perimeter was
introduced among data collected following a fire event.
Among the available data there is no information
concerning possible infrastructures damaged or
threatened by fire.
WUI wildfires are therefore not identifiable from
historical series.
Considering these constraints due to data deficiency
or resolution, a methodology that analyses past events
examining and assessing their territorial background was
proposed.
The geographical units of the analysis correspond
necessarily to 1-km² cells.
To locate WUI fire prone areas in Piemonte territorial
elements that contribute to create the WUI environment
were identified.
Areas where natural vegetation and infrastructures are
both present and intermingle were looked for using and
overlaying layers containing urban areas and woods, with
predefined criteria.
Thus, working mainly with GIS facilities, proper layers
derived from wildfires, urban areas and forest databases
were overlaid applying thresholds and neighboring
constraints in order to extract those cells that are
expected to contain WUI environment prone to wildfires.
Data processing that eventually led to the selection of
territorial contexts where WUI areas are potentially
threatened by wildfires, allowed to assess the size of the
problem at regional scale, its spatial distribution and to
find the most affected areas.
72
EUFIRELAB
The first GIS layer analysed is represented by
wildfires data.
An historical series of 20 years, from 1980 to
1999, was considered.
From these data the number of wildfires occurred
in this period within each grid cell was obtained.
With this respect, it is important to underline once
again that it is possible to know only the cell in which
the wildfire started and not the neighboring cells
involved, if any.
To include also those cells that, although not
directly recorded as sites where fire actually started,
were expected to be affected by fires started in
neighbouring cells, some further criteria were
introduced.
The second input layer is represented by the
territory urbanization.
Urban areas have been extracted from the land
use map (1:100,000) of Piemonte Region
cartography.
These areas were intersected with the cells
containing wildfires in order to select those cells
characterized both by houses and by wildfires.
In this step all cells containing exclusively urban
area and not burned wildland were excluded.
In this way, starting from all wildfires occurred in
Piemonte from 1980 to 1999, cells with wildfires
occurred next to urbanized areas were identified.
As mentioned, wildfire starting within cells with no
urban area but adjacent to cells with urban area
presents a potential menace.
Therefore, we added to the previous selection,
cells less than 1,000 m away from other cells
characterized by the presence of urban areas and
with more than 3 wildfires or more than 10 ha of
burned surface in the 20 considered years.
This step allows to include not only directly
affected areas, but also those urban areas
considered potentially prone to WUI wildfires.
The third input layer is related with forests.
The forested surface of each cell was calculated
from the land use map, selecting between the
previously extracted cells only those typified by a
forestry cover higher than 5%.
Considering starting data and analysis scale, it
isn’t possible to assert these wildfires directly
interfered with human dwellings, but just that they
occurred in environments characterized by the
presence of WUI areas.
To characterize areas where WUI wildfires are
particularly critical, all cells between those previously
selected, in which events characterized by a burned
surface higher than 30 ha and 15 ha occurred, have
been identified.
D-08-06.doc
These two values of burned surface were chosen
since 30 ha correspond to the burned surface threshold
that typifies a critical fire according to the Regional Fire
Management Plan (BOVIO et al., 1999), while 15 ha have
been prudentially added as an in-between attention
threshold for WUI fires in Piemonte.
On the base of WUI wildfires frequency, a zoning of
municipalities has been realized.
Piemonte municipalities were classified, identifying
5 frequency classes, according to the number of wildfires
potentially interesting WUI areas between 1980 and
1999.
9.1.3 Estimation of population vulnerability at Euro-
Mediterranean scale
Vulnerability is not directly measurable.
However, it is possible to induce vulnerability using
information about the characteristics of the geographical
zone considered.
As previously underlined, to realize a Wildland Fire
Danger Rating System operative at an Euro-
Mediterranean scale, we must face the problem of using
appropriate data at relevant scale and precision for the
expected aim, currently available for all the countries
involved.
Thus the question is: What do we have and how
reliable is it?
The difficulty when trying to approach the risk at global
scale is to find relevant and available indicators allowing
a comparison between all countries.
Simple indices relying on good data and with stated
limitation and subjectivity might be much more efficient
than complex ones that cannot be computed because of
the lack of (reliable) data.
Extrapolations from local researches to global scale
are rarely applicable as data may not be of comparable
formats or simply not available.
If a model requests a large amount of inputs, the
chances that such model will never be used, by lack of
data or by too fuzzy data, are significant.
On the other hand, a model based on too few
parameters will lead to large gap between observed facts
and expected figures.
Since no more detailed and precise data are currently
available and/or easily accessible for all the Euro-
Mediterranean countries, it was decided to work with the
CLC2000 geographic data layer (EC JRC-IES, 2005).
CORINE land cover (CLC) is a geographic land
cover/land use database encompassing most of the
countries of the European Community and the majority of
the Central and East European countries (Figure 72).
CLC 2000 is the year 2000 update of the first CLC
database, which was finalised, in the early 1990s as part
of the European Commission programme to COoRdinate
INformation on the Environment (Corine).
It also provides consistent information on land cover
changes during the past decade across Europe.
The CLC2000 database currently covers 32 countries.
73
EUFIRELAB
CLC describes land cover (and partly land use)
according to a nomenclature of 44 classes organized
hierarchically in 3 different levels (Figure 73).
The first level (5 classes) corresponds to the main
categories of the land cover/land use (artificial areas,
agricultural land, forests and semi-natural areas,
wetlands, water surfaces).
The second level (15 classes) covers physical and
physiognomic entities at a higher level of detail (e.g.
urban zones, forests, lakes), while level 3 is
characterized by 44 more detailed classes.
The nomenclature has been developed in order to
map the whole Community territory, including the
foreseen extension to other eligible countries.
The use of the CLC nomenclature with 44 classes
at three hierarchical levels is mandatory.
Additional national levels can be mapped but
should be aggregated to level 3 for the European
data integration.
No unclassified areas should appear in the final
version of the data set.
CLC was elaborated based on the visual
interpretation of satellite images.
Ancillary data (aerial photographs, topographic
and vegetation maps, statistics, local knowledge)
were used to refine interpretation and the assignment
of the territory into the categories of the CORINE land
cover nomenclature.
The smallest surfaces mapped (mapping units)
correspond to 25 hectares.
Only area elements (polygons) are identified.
Areas smaller than 25 ha are allowed in the
national land cover database as additional thematic
layers, but should be aggregated/generalized in the
European database.
Linear features less than 100 m in width are not
considered
The scale of the output product was fixed at
1:100.000.
Thus the location precision of the CLC database
is 100 m.
From the CLC2000 layer, information on both
potentially combustible land covers and populated
ones could be derived, providing a common base for
all the involved countries, from which to indirectly
obtain information on population vulnerability to
wildfires.
An index to assess population vulnerability was
defined, made up by three main sub-indices.
These three sub-indices are calculated on a
1-km2 grid and their values within each cell are
summarized to give the final index.
D-08-06.doc
9.1.3.1 SUB-INDEX 1
This sub-index qualifies each grid cell according to the
presence within the cell itself of urban land covers and
potentially combustible ones. The presence of these land
covers is quantified with this sub-index independently
from their spatial location and topology within the cell.
Third level CLC2000 classes were used.
The surface occupied inside the cell by the following
land cover classes is calculated in a GIS environment:
- Areas occupied by forests and semi-natural areas,
shrubs and herbaceous plants; recently burned areas
are also included (CLC2000 codes: 2.4.4; 3.1.x; 3.2.2,
3.2.3, 3.2.4, 3.3.4 ; see Table 16);
- Areas characterized by a major presence of houses
and consequently population (CLC2000 codes: 1.1.x;
1.2.x; see Table 16);
This sub-index is meant to take into account those
situations characterized by a significant presence of both
combustible vegetation and population potentially at risk.
The total surface occupied within each cell by the two
land cover categories (a and b) is then classified in ten
percentage classes.
The value of the sub-index, going from 0 to 5, is
obtained through a bi-dimensional matrix (Figure 74),
combining the classes of the two land cover categories
present in a cell.
In the matrix higher values are assigned to those cells
in which there is a major presence of both urban land
covers and potentially combustible ones, in order to
identify those situations characterized by a higher
probability to have wildland-urban interface areas or by a
higher probability to have a widespread human presence,
independently from the precise location of settlements.
9.1.3.2 SUB-INDEX 2
This sub-index is based on the assessment of the
spatial configuration of specific land covers in the cell.
The linear development of those zones in which urban
land covers are in contact with potentially combustible
land covers is measured.
Considering the scale and precision of CLC2000 data,
this contact is more probably to be considered as a
neighbourhood relationship between the two categories
of land cover.
The land covers selected for this analysis are:
- Areas occupied by forests and semi-natural areas,
shrubs and herbaceous plants; recently burned areas
are also included (CLC2000 codes: 2.4.4; 3.1.x; 3.2.2,
3.2.3, 3.2.4, 3.3.4 ; see Table 16);
- Areas characterized by urban fabric, continuous or
discontinuous (CLC2000 codes: 1.1.x; see Table 16);
The main aim of this sub-index is to identify those
situations in which, given a certain surface occupied
within a cell by the two categories of urban and potentially
combustible land uses, as assessed by sub-index 1,
there is a high probability that people, buildings or
infrastructure are in proximity of fuels susceptible to burn.
74
EUFIRELAB
Please note that the susceptibility and fire-prone
nature of these fuels should be assessed though
other components of the Euro-Mediterranean
Wildland Fire Risk Index proposed in this deliverable.
As a consequence, the effective population
exposition to fire risk comes from the integration of
the Vulnerability index here discussed and the
Ignition and Propagation Danger indices previously
discussed.
The sub-index value is assigned according to the
linear development of probable WUI areas measured
in the cells and classified following Table 17.
Only those cells having at least one of the two
above described sub-indices different from 0 are
selected to proceed with the analyses.
This means that cells with no or little presence of
urban and potentially combustible land uses and cells
with no or little linear development of WUI areas are
considered as not vulnerable areas, given the low
detail of initial data and the working scale.
D-08-06.doc
9.1.3.3 SUB-INDEX 3
Cells with a certain level of possible vulnerability, as
resulting from the previous steps, are then further
differentiated according to the presence of a particular
urban land cover, i.e. the “discontinuous urban fabric”
(CLC2000 code 1.1.2). This land cover identifies areas
where most of the land (between 30 to 80 %) is covered
by structures; building, roads and artificially surfaced
areas are however associated with vegetated areas and
bare soil, which occupy discontinuous but significant
surfaces. This land cover, frequently corresponding to the
external part of settlements, is considered a more
vulnerable urban cover because of the concurrent
presence of structures and artificial green areas or
wildlands, often associated in a complex spatial pattern.
The surface occupied by discontinuous urban fabric
inside cells previously selected is then calculated,
assigning to each cell a value corresponding to a surface
class as defined in Table 18.
The final vulnerability index was obtained as the sum
of the three sub-indices in the cell, ranging from a
minimum of 1 to a maximum of 14.
Three vulnerability classes were identified, as reported
in Table 19.
An example of each sub-index and of the final
population vulnerability index, calculated according to the
methodology above described, is reported in Figure 75.
The described methodology was applied to the Italian
province of Torino (North-Western Italy, Piemonte
Region), corresponding to a NUTS 3 level. The CLC2000
data layer of the province and the DTM are shown in
Figure 76.
Resulting values of sub-index 1, sub-index 2, sub-
index 3 and the final vulnerability index are instead
reported in Figure 77.
75
EUFIRELAB

9.1.4 Figures

Figure 72: Example of CLC2000 geographic data layer. Yellow lines correspond to the boundaries of CLC classes
(see figure 73 for CLC nomenclature); red lines correspond to 1-km2 grid cells.

Figure 73: CLC2000 nomenclature (from EC JRC-IES, 2005)

D-08-06.doc 76
EUFIRELAB

urban areas
1 2 3 4 5 6 7 8 9 10

91 - 100
11 - 20

21 - 30

31 - 40

41 - 50

51 - 60

61 - 70

71 - 80

81 - 90
0 - 10
surface %

0 - 10 0 0 0 0 0 0 0 0 0 0

1
11 - 20 0 1 1 1 2 2 2 1 1

2
21 - 30 0 1 2 2 2 3 3 3

3
combustible areas
31 - 40 0 2 2 3 3 4 4

4
41 - 50 0 2 2 3 4 5

5
51 - 60 0 2 3 4 5

6
61 - 70 0 2 3 4

7
71 - 80 0 3 4
8
81 - 90 0 3
9
10

91 - 100 0

Figure 74: Bi-dimensional matrix to determine sub-index 1 values within each 1-km² grid cell.

Figure 75: Example of each sub-index and of the final population vulnerability index calculated according to the
methodology here described

D-08-06.doc 77
EUFIRELAB

Figure 76: DTM and CLC2000 data layer of the Italian province of Torino (North-Western Italy), where the
population vulnerability index here proposed was tested

D-08-06.doc 78
EUFIRELAB

Figure 23: Sub-index 1, sub-index 2, sub-index 3 and the final population vulnerability index computed for the
province of Torino (Italy)

D-08-06.doc 79
EUFIRELAB

9.1.5 Tables

Table 16a: Corine land cover nomenclature of the classes selected for the analyses (Populated Land Covers)
CLC code CLC description
Areas mainly occupied by dwellings and buildings used by administrative/public utilities or
authorities, including their connected areas (associated lands, approach road network,
parking-lots).
1.1.1 Continuous urban fabric: Most of the land is covered by structures and the transport
1.1 Urban fabric network. Building, roads and artificially surfaced areas cover more than 80 % of the total
surface. Non-linear areas of vegetation and bare soil are exceptional.
1.1.2 Discontinuous urban fabric: Most of the land (Between 30 to 80 %) is covered by
structures. Building, roads and artificially surfaced areas associated with vegetated areas
and bare soil, which occupy discontinuous but significant surfaces.
Areas mainly occupied by industrial activities of transformation and manufacturing, trade,
financial activities and services, transport infrastructures for road traffic and rail networks,
airport installations, river and sea port installations, including their associated lands and
access infrastructures. Includes industrial livestock rearing facilities.
1.2.1 Industrial or commercial units
Artificially surfaced areas (with concrete, asphalt, tarmacadam, or stabilised, e.g. beaten
earth) without vegetation occupy most of the area, which also contains buildings and/or
vegetation.
1.2 Industrial,
commercial and Particularity of class 1.2.1: Agricultural farms: Areas of other than housing buildings, in-
transport units door spaces, stables, garages, workshops, lay-by and storing areas, often also bad land
with ruderal vegetation, part of farms. The farms are often located in outskirts or close to
rural settlements with agricultural function. Concentration of agricultural buildings in areas
of various sizes was associated with collectivisation of agriculture. The quoted areas
smaller than 25 ha are included in class 1.1.2.
1.2.3 Port areas: Infrastructure of port areas
including quays, dockyards and marinas.
1.2.4 Airports
Airports installations: runways, buildings and associated land.
1.4.1 Green urban areas:
1.4 Artificial, non-
Areas with vegetation within urban fabric, includes parks and cemeteries with vegetation,
agricultural
and mansions and their grounds
vegetated
areas 1.4.2 Sport and leisure facilities: Camping grounds, sports grounds, leisure parks, golf
courses, racecourses, etc. Includes formal parks not surrounded by urban areas.
2.4.2 Complex cultivation patterns: Juxtaposition of small parcels of diverse annual crops,
pasture and/or permanent crops.
2.4 Heterogeneous Particularity of class 2.4.2: Complex cultivation patterns with scattered houses Alternation
agricultural areas of small plots (smaller than 25 ha) of arable land with annual or permanent crops with
scattered garden huts or scattered houses. They are usually situated in proximity of rural or
urban settlements and are used for growing agricultural crops, fruit, and vegetable for the
particular households.
Areas occupied by forests and woodlands with a vegetation pattern composed of native or
exotic coniferous and/or deciduous trees and which can be used for the production of
timber or other forest products. The forest trees are under normal climatic conditions higher
than 5 m with a canopy closure of 30% at least. In case of young plantation, the minimum
cut-off-point is 500 subjects by ha.
3.1 Forests 3.1.1 Broad-leaved forest: Vegetation formation composed principally of trees, including
shrub and bush understoreys, where broad-leaved species predominate
3.1.2 Coniferous forest: Vegetation formation composed principaly of trees, including shrub
and bush understoreys, where coniferous species predominate.
3.1.3 Mixed forest: Vegetation formation composed principally of trees, including shrub and
bush understoreys, where neither broad-leaved nor coniferous species predominate.

D-08-06.doc 80
EUFIRELAB

Table 16b: Corine land cover nomenclature of the classes selected for the analyses (Combustible Land Covers)
CLC code CLC description
3.2.2 Moors and heathland: Vegetation with low and closed cover, dominated by bushes,
shrubs and herbaceous plants (heather, briars, broom, gorse, laburnum, etc.)
3.2 Shrubs and/or 3.2.3 Sclerophylous vegetation: Bushy sclerophyllous vegetation, includes maquis and
herbaceous garrigue. In case of shrub vegetation areas composed of sclerophyllous species such as
vegetation Juniperus oxycedrus and heathland species such as Buxus spp. or Ostrya carpinifolia with
associations no visible dominance (each species occupy about 50% of the area), priority will be given to
sclerophyllous vegetation and the whole area will be assigned class 3.2.3.
3.2.4 Transitional woodland/shrub: Bushy or herbaceous vegetation with scattered trees.
Can represent either woodland degradation or forest regeneration/recolonisation.
Natural areas covered with little or no vegetation, including open thermophile formations of
sandy or rocky grounds distributed on calcareous or siliceous soils frequently disturbed by
erosion, steppic grasslands, perennial steppe-like grasslands, meso- and thermo-
3.3 Open spaces
Mediterranean xerophile, mostly open, short-grass perennial grasslands, alpha steppes,
with little or no
vegetated or sparsely vegetated areas of stones on steep slopes, screes, cliffs, rock fares,
vegetation
limestone pavements with plant communities colonising their tracks, perpetual snow and
ice, in land sand-dune, coastal sand-dunes and burnt areas.
3.3.4 Burnt areas: areas affected by recent fires, still mainly black.

Table 17: Sub-index 2 values according to the linear development of probable WUI areas (m). Natural breaks
(Jenks) were selected to identify the six classes corresponding to the index values.
Sub-index 2 Linear development of probable WUI areas (m)
0 0
1 1-500
2 501-1000
3 1001-1700
4 1701-2700
5 2700-4700

Table 18: Sub-index 3 values according to the surface occupied in the cell by CLC2000 code 1.1.2.
Natural breaks (Jenks) were selected to identify the classes corresponding to the index values.
Sub-index 3 CLC 112 surface (m2)
0 0
1 1-62.000
2 62.001 -200.000
3 200.001 - 380.000
4 380.001 - 650.000
5 > 650.000

Table 19: Final vulnerability index values and corresponding vulnerability classes.
Classes Vulnerability index
Low 1-5
Medium 6-10
High 11-14

D-08-06.doc 81
EUFIRELAB
9.2 VULNERABILITY RELATED TO
ENVIRONMENTAL VALUE
9.2.1 Presentation
Ecological and landscape assessments are
essential in the design of an integrated fire risk index
which takes into account aspects related to the
vulnerability of forested areas.
Such assessments are of great value in sustainable
land management and help managers in decision
making processes within a multipurpose scenario,
ranging from natural environment conservation policies
and biodiversity promotion strategies to urban and
regional planning.
For instance, given the fact that the budget available
for preventive silviculture is usually limited or
insufficient, is it reasonable to apply these tasks to
areas with the highest ecological and landscape value.
These areas are, therefore, the most vulnerable to
this line of events.
In addition, when a forest fire is spreading, it would
also seem reasonable for a fire manager to allocate fire
suppression resources according to a certain scale of
values.
This scale would bear in mind the presence of
population, property (buildings, etc.) and also
outstanding natural environments, such as natural
parks, reserves, special protection zones, preserved
landscapes, etc.
In some previous studies, this problem has been
addressed by using GIS tools.
However, in some cases (Comunidad de Madrid,
2000) the only parameter which is taken into account to
measure priorities in forest fire fighting, is the ecological
value of the area according to the quality and the
degree of protection of the forested ecosystems.
There is no assessment of the landscape and no
landscape ecology indices are applied. In other cases
(GULINCK et al. 2001), the method used to asses a
given landscape is based on land use data, which does
not include information from other supplementary data
sources.
In this context, Landscape Ecology provides the
appropriate conceptual framework that meets the
demand for information which regional planners
require.
It is well known that landscape ecology is a
multidisciplinary science which mainly aims at solving
the problems to do with land management and
development on a local and regional scale (NAVEH and
LIEBERMAN, 1994).
This new science facilitates territorial analysis by
attempting to understand and compare different spatial
patterns by means of patches with different shapes,
quantities, classes, etc. (HONG et al. 2000; RIITTERS et
al. 1995).
Landscape is under the influence of ecological and
human processes, both on a regional and local scale,
which imprint changes in its structure and composition.
D-08-06.doc
The study of landscape structure/composition by
means of spatial statistics helps us characterize the
territory and understand the spatial-temporal
relationship among the different elements that compose
a landscape.
Landscape ecology has developed many indices
that measure spatial textures and shapes, as well as
the spatial structure of the landscape (MCGARIGAL et al.
2002; RIITTERS et al. 1995; MCGARIGAL y MARKS, 1995):
patch density, size, compactness, fractal dimension,
dispersion, diversity, etc.
Metric assessment and analysis of landscapes,
using a GIS (MCGARIGAL et al. 2002; BERRY, et al.
1999; IGIS, 1997; MCGARIGAL y MARKS, 1995), allows
to characterise the structures and changes in forested
land occupation and land use within a territory, and
also helps us find the environmental involvement of its
activity (MALDENOFF y BAKER, 2000).
The studies carried out by Martinez et al., 2006
propose a simple method to map the landscape and
ecological value of an area, by integrating several
indicators in an exclusively ecological approach,
without considering visual or aesthetic aspects.
The results show that the methodology can be
usefully applied on a regional/global scale.
In addition, it is a valuable source of information for
fire fighting authorities in decision making processes
related to preventive forestry and to the distribution of
extinction resources and manpower when a forest fire
is spreading and threatening the areas which are most
vulnerable from an ecological point of view.
We consider it would interesting to include the
landscape assessment methodology proposed by
these authors in the section concerning the vulnerability
of an integrated risk index on a European level.
The methodology starts by integrating different
indicators, under an exclusively ecological approach,
and seeks to work out the intrinsic ecological value of a
territory.
Figure 78 shows the methodological flow chart
proposed by Martínez et al. (2006):
The main data source used to obtain the landscape
value is the land cover map from the CORINE-Land
Cover project from the year 2000 (CLC2000) available
in digital format for all of Europe.
Two criteria groups are considered.
i) Vegetation and land use
The first group of indicators is based on the intrinsic
characteristics of vegetation, such as the degree of
proximity to climax vegetation.
In addition, it analyses the importance of vegetation
types in its regional context (rarity) and in its global
context (representativity).
ii) Landscape Ecology Indicators
As well as the variables from the previous group,
other landscape ecology indicators are used, especially
indices related to diversity, connectance and
juxtaposition or interspersion, which take into account
the spatial distribution of the patches within a territory.
82
EUFIRELAB

MARTINEZ et al. (2006) propose a selection of
indices (Simpson's Diversity Index, Interspersion
Juxtaposition Index and the Connectance Index) with
which to get to know the diversity, homogeneity or
fragmentation of a given landscape.
Simpson’s Diversity Index (SIDI) assesses the
number of different land cover types and the
distribution of the area proportional the types of land
covers.
Values near 0 indicate that there is only one patch
or land cover (non diversity) whereas values close to 1
(high diversity) indicate that the different types of land
covers as well as the proportional distribution of their
areas is more balanced.
The Interspersion Juxtaposition Index (IJI) allows us
to understand the spatial configuration of the patches
as well as its contiguity and its degree of interspersion.
Low values (0) represent landscapes where patches
are distributed randomly, whereas high values (100)
correspond to landscapes which are distributed in an
equal contiguity.
This index shows whether patches are grouped or
distributed homogeneously in space.
The Connectance Index (CI) is defined by the
number of functional unions among all the patches
within the same type.
It show where each pair of patches are connected
according to a 500 meter analysis window.
9.2.2 Figures
Values range from 0 (consisting in just one match or
when none of the matches are connected) to 100,
when all the patches in the analysis window are
connected.
The Landscape Integrated Value (LIV) is the simple
average of the combination of the six components
mentioned.
An example of this product can be seen in figure 79.
Regarding the ecological value, the authors suggest
considering whether the territory analysed is inside a
preserved area of some kind: protected natural
environments, special protection areas for birds, sites
and habitats of community interest, preserved
woodlands and public woodlands, etc.
This information can be easily accessed on a
national/European scale.
In view of the results obtained in this work and
although the workload capacity of GIS could include
other elements for improving landscape and ecological
assessment, we believe that the methodology
proposed by the authors is straightforward and easy to
understand.
In addition, it has basic data requirements, available
on a European scale, that guarantee the applicability of
the method and its integration, together with other
factors, into a synthetic fire risk index.

Figure 78: Landscape assessment methodological flow chart

D-08-06.doc 83
EUFIRELAB

Figure 79: Landscape integrated value map for the Madrid region (MARTÍNEZ et al., 2006)

D-08-06.doc 84
EUFIRELAB

9.3 POTENTIAL SOIL EROSION Research indicates that fires are likely to increase
runoff rates and sediment yields relative to undisturbed
9.3.1 Introduction
forested land (AHLGREN and AHLGREN, 1960; LAIRD and
Soil erosion is a widespread threat to European HARVEY, 1986; May, 1990; SCOTT and VAN WYK, 1990;
soils. SOTO et al., 1991; SOLER and SALA, 1992; SWANSON,
The main consequences of erosion are not only on- 1981).
site soil degradation, declining fertility, desertification
A great number of pan-European soil erosion risk
and reduced infiltration and water storage capacities,
assessment efforts have been made with a variety of
but there are also off-site impacts that include
approaches such as GLASOD approach (Global
eutrophication of rivers and lakes, destruction of wildlife
Assessment of the Current Status of Human-Induced
habitats, siltation of dams, reservoirs, rivers, as well as
Soil Degradation), INRA (Institut National de la
infrastructure and property damage by muddy floods
Recherche Agronomique, 1988) approach, HOT-
(RUBIO et al., 2006).
SPOTS, IMAGE/RIVM, CORINE, USLE/ESB,
Some authors have highlighted its impact on global
PESERA.
food security (CROSSON, 1997; LOMBORG, 2001).
Widely used models include RUSLE (Renard et al.,
In addition, soil erosion results in emission of soil
1997) and MUSLE (SMITH et al., 1984).
organic carbon to the atmosphere in the form of CO2
Another model, Soil Erosion Model for
and CH4, thereby enhancing global warming (LAL,
Mediterranean regions (SEMMED, DE JONG, 1994a)
2004).
was developed with the objective of using satellite data,
Global warming in turn is expected to increase
similarly to the Thornes model (THORNES, 1985),
erosion rates (NEARING et al., 2004).
Agricultural Nonpoint Source Pollution model (AGNPS:
European Research Programmes initiated by the
YOUNG et al., 1989) and the Areal Nonpoint Source
European Commission demonstrate that although the
Watershed Environment Response Simulation model
Mediterranean region is historically the most severely
(ANSWERS: BEASLEY et al., 1980).
affected by erosion, there is growing evidence of
Furthermore, important European policies and
significant erosion occurring in other parts of Europe
directives, such as the Water Framework Directive (EC,
(e.g. Austria, Czech Republic and the loess belt of
2000), the European Commission Soil Thematic
Northern France and Belgium).
Strategy (COM, 2002), and instruments of the Common
Soil erosion can therefore be considered, with
Agricultural Policy, such as agro-environmental
different levels of severity, an EU-wide problem
measures (EC, 1999), address the issues of soil
(MONTANARELLA et al., 2003).
erosion.
Soil erosion occurs under diverse conditions and is
driven by interactions of many factors such as land use, Although there are many approaches in relation to
climate, soil conditions, and topography that are difficult soil erosion risk assessment, only a few can be found
to quantify. in the literature that take into account the effects of
A proper assessment of erosion problems is greatly wildfires on vegetation and soil conditions and are
dependent on their spatial, economic, environmental, suitable for burned woodland areas.
and cultural context (WARREN, 2002). Most models and strategies address soil erosion
As a result, a comprehensive understanding of soil problems in agricultural areas only.
erosion is still very difficult.
9.3.2 Objectives
Forest fires are a further factor that can lead to
In accordance with the aim of this deliverable, a
excessively high rates of erosion and contribute to
vulnerability index of soil erosion will be created for
environmental degradation.
Mediterranean countries that suffer high intensity and
Every year there are more than 50,000 forest fires in
frequency of forest fires.
Europe, affecting over 500,000 ha of forest and other
The questions and objectives that were set
woodlands, the majority of which occurs in the
regarding potential soil erosion in the previous
Mediterranean Region (European Commission, 2002).
deliverable (D-08-03) are the following:
The continuous impacts of fires, with progressively
- To what extent can planners prepare for the
shorter recovery periods, together with torrential rains
consequences of a forest fire before it occurs in
characteristic of the Mediterranean climate, favour the
order to respond more quickly to the potential runoff
intensification of erosive processes (GIOVANNINI et al.,
and erosion risks
1990).
- To elaborate a pre-fire strategy that would shorten
Wildfires and prescribed fires affect the vegetation,
the post-fire reaction time
soils, wildlife, and water resources of watersheds.
- To evaluate the use of the model quoted in D-08-03
They impose a wide range of effects which depend
(ANNEX I) at different spatial scales and to answer
on the mosaic of fire severity and post-fire hydrologic
how model scale affects the objectives and the
events (NEARY, 2004).
operational use of the model.

D-08-06.doc 85
EUFIRELAB
From the above questions, two approaches can be
put forward.
In the first, a “what if” logic underlies the model in
order to define potential erosion risk areas.
The underlying question is therefore “what would be
the post-fire soil erosion risk if a forest fire were to take
place? (sub-chapter 9.3.3)”.
In this case, scenarios can be tested at the
European scale to define potential soil erosion risk.
The focus here is on gaining a spatial perspective of
the distribution of high-risk zones.
In a second approach, an operational model is
elaborated to be used in a Mediterranean context in the
case of an actual fire.
Here, the question is “how can Mediterranean
countries best respond to a fire in order to deal
efficiently with the erosive consequences? (sub-chapter
9.3.4)”.
9.3.3 A European scale model to predict potential
erosion risk
The aim of this approach is to perform a pre-fire
prediction of post-fire soil erosion risk.
Post fire soil erosion risk is strongly related to the
effects of fire on vegetation and soil conditions.
In other words, before a fire actually occurs, we
must firstly predict the kind and intensity of fire effects
on soil properties and vegetation; secondly, we must be
able to estimate the consequences of these changes
on soil erosion rates.
One approach to addressing this problem is to
modify the method quoted in D-08-03 deliverable
(ANNEX I) in order to be able to acquire information
about the post-fire erosion risk that may arise before a
Wildfire takes place.
This ‘a priori’ knowledge would give European level
planners a vision of the spatial distribution of post-fire
erosion risk and help in the preparation of strategies at
the European program scale to deal with this potential
danger.
9.3.3.1 Discussion and suggestions about the
method described in ANNEX I
The method quoted in D-08-03 deliverable takes
into account four factors to produce a map of soil
erosion risk at the catchments scale (FOX et al., 2006).
These factors are:
- Slope
- Pre-fire vegetation
- Fire severity
- Soil erodibility
These factors were estimated after a wildfire in the
Massif des Maures area near St Tropez in southern
France which means that this method poses temporal
difficulties in the typical Mediterranean climatic context
where there is a limited time between a summer or
autumn fire and the following rainy winter period.
D-08-06.doc
As a result, this method has to be modified in order
to meet the following needs:
- to be applied in pre-fire time and thus help us to
acquire an ‘a priori’ knowledge of the most soil
erosion prone areas
- to provide European scale information of the
potential erosion risk.
The way the four erosion factors were estimated
and how they can now be addressed in order to predict
potential soil erosion at the European scale in pre-fire
time is discussed below.
9.3.3.2 Slope
This factor can be considered static since it does not
change over time.
A 25-m DEM was used for slope estimation, and a
50-m resolution was considered too coarse to be
operational.
At the European scale, 30 to 50-m DEMs are
available and are considered sufficient of the objectives
of this approach.
Slope needs to be estimated at the European scale
only once since it does not change significantly over
time.
FOX et al. (2006)formed five categories, according to
slope magnitude and coefficients were ascribed to each
category as follows:
9.3.3.3 Pre-fire Vegetation density
This factor was estimated from the standing charred
trunks during a field survey and three categories were
included.
Coefficients were ascribed to each factor level as
follows table 20:
It is obvious that at the European scale, vegetation
data cannot be collected manually.
However, CORINE land cover or remote sensing
data could be used to estimate pre-fire vegetation.
Vegetation cover provides protection of the soil
against erosion processes.
This happens not only before fire (INBAR et al.,
1997), but also after fire because in low intensity and
medium intensity fires the surface organic matter is
increased providing protection from soil erosion (fallen
needles, twigs, etc).
In addition to the mulch effect provided by fallen
pine needles (SHAKESBY et al., 1993; SHAKESBY et al.,
1994), standing dead vegetation may favour infiltration
(near stems and in burned roots) and slow runoff
velocity, so the net expected effect is lower erosion
where vegetation stands are denser.
ANDREU et al. 1996 demonstrated that organic
matter content strongly decreases immediately after
medium to intense fires, with a subsequent increase
with time.
However, in zones that suffered a low to medium
fire intensity its value increases.
This could be due to the accumulation of plant
residues, not completely burnt, according to the studies
of CHRISTENSEN 1987.
86
EUFIRELAB
It was found from the literature that for land cover
estimation, the following methods were applied:
- in the IMAGE/RIVM approach, Olson’s land cover
database and statistical information from FAO was
used
- in the CORINE approach, the CORINE landcover
database was used
- in the USLE/ESB approach, an exponential scaling
function was linked to the Normalized Difference
Vegetation Index (NDVI) extracted from NOAA
images
- in the PESERA approach, a plant growth model was
used initially, but in monitoring mode it was replaced
by NDVI from Earth Observation Sensors NOAA-
AVHRR or SPOT-VGT (also applied by Zhang et al.,
2004)
- an NDVI (TUCKER et al., 1985) was also extracted
from Landsat TM data (Bayaramin et al., 2006) in
the semi-arid area of Beypazari in Ankara in order to
create a layer of vegetation cover which was used in
a study of soil erosion risk assessment. NDVI was
grouped as “fully protected” or “not fully protected”
based on ground truth information with the global
positioning system. The map created can be seen
further down (Figure 80). The pixel size from NOAA-
AVHRR is of 1 km spatial resolution so in order to
improve the spatial resolution, LANDSAT remote
sensing data could be used in local scale to
calculate Vegetation Indices (VIEDMA et al., 1997:
Santos et al., 2000: DELGADO et al., 2001) in which
pixel size is 30 m. This can easily be used with a
30-m DEM.
Resolution of 30 by 30-m pixel size can be
considered appropriate because of the great fluctuation
in vegetation density that can be found inside a 1 by 1
km pixel size of NOAA-AVHRR.
Although LANDSAT remote sensing data could be
used for NDVI calculation, it is not convenient for
European scale projects because of its swath width
(185 km).
The Disaster Monitoring Constellation (DMC) can be
used for NDVI calculation.
This system records reflected electromagnetic
radiation over the same spectral range as Landsat-
ETM 2, 3 and 4 bands, it offers 32 m spatial resolution,
daily coverage and a very wide swath of 600-km. NDVI
data can be used in combination with thresholds in
order to classify the pixels into the three
aforementioned categories of land cover density.
In addition, a fourth category, high vegetation
density, can be added to the table above.
This class was not encountered in the Massif des
Maures catchment described in deliverable D-08-03,
but would be found at the European scale.
Finally, it is worth mentioning that the DMC is the
first earth observation constellation of five low cost
satellites providing daily images for applications
including global disaster monitoring.
D-08-06.doc
9.3.3.4 Fire severity
It must be defined as potential fire severity.
Fire severity was defined as either variable or
intense and coefficients were ascribed to each level as
following: Table 22
“Variable severity” were sites where either the fire
had passed quickly and charred the trunks and
underbrush, but left the crown more or less intact, or
the intense combustion of the standing vegetation was
confined to a small area (dozens of m2).
The “intense severity” category consisted of sites
where the fire had consumed the underbrush and
crown entirely over an extensive area.
According to this classification, this factor could also
be defined as fire type because the former category
corresponds more or less to the description of a
surface fire, whereas the latter to the description of a
crown fire. In order to estimate fire severity, the main
factors contributing to the outbreak of a crown fire
should be taken into account.
The primary factors influencing crown fuel ignition
are:
- Height of the under-story vegetation and
- The vertical distance between the ground/surface
fuel strata and the lower boundary of the crown fuel
layer (Miguel et al., 2006).
These two factors are the most significant for
predicting danger of crown fire ignition.
Although, they cannot be estimated easily at the
European scale from remote sensing data, because
methods that can be applied for their estimation, are
usually, if not always, by means of LIDAR data
(ANDERSEN et al., 2005: DRAKEA et al., 2002:
MORSDORF et al., 2004: RIANO et al., 2003a).
This system is airborne and an inventory effort at
such a scale would be practically and financially
unfeasible.
Instead of this method, it is proposed to initially
reclassify vegetation types into the two aforementioned
categories (variable or intense fire severity) based on
the following:
- vegetation types from the CORINE database in
relation with the potential existence of a dense
understory below specific dominant tree species
- the flammability of species (Dimitrakopoulos and
Papaioannou, 2001)
It is widely known that some dominant tree species
are associated with specific densities and structures of
understory vegetation, as is the case for Pinus
halepensis where we find very dense underbrush in
contrast to Fagus species under which there is limited
understory vegetation.
Flammability of species on the other hand can help
us to predict the intensity and severity of a forest fire
before it occurs.
87
EUFIRELAB
The periodicity of estimation for this factor could
follow that of the Propagation Danger Index described
elsewhere in this deliverable since they are highly
correlated.
Other factors that can lead to high severity wildfires
are the high density and favourable horizontal structure
of vegetation cover which may result in extensive
spreading, low live and dead fuel moisture contents,
intense winds, high temperatures and low relative
humidity.
The vegetation cover density could be assessed as
stated earlier with NDVI calculation whereas
information about the live and dead fuel moisture
content from the corresponding indices developed in
other sections of this deliverable.
Wind speed, temperature, and relative humidity can
be estimated either from historical meteorological data
or from meteorological stations across Europe.
9.3.3.5 Soil erodibility
Three classes of soil conditions were defined and
coefficients were ascribed to each class as following:
Table 23
Shallow stony soils are assumed to generate more
runoff but surface material is less mobile due to the
high stoniness, and in many cases, the bare bedrock.
In contrast, deep soils with fewer stones may have
higher infiltration rates, but they are more erodible,
particularly with respect to the weak aggregate stability
and possible presence of hydrophobicity.
Therefore, it was decided to increase the value of
the coefficient with decreasing stone content.
In order to estimate this factor, without altering the
logic upon which it was built and its functionality, it is
proposed to use the method applied in the CORINE
approach: erodibility is estimated from soil texture,
depth and stoniness, extracted from the soil map of the
European Communities (CEC, 1985).
The CORINE model, which has been applied to
many countries in the European Community, is
constituted by combining 4 parameters: soil erodibility,
erosivity, topography and vegetation cover (DENGIZ and
AKGUL, 2005).
The following diagram (Figure 81) shows the three
parameters used to estimate erodibility and how they
can be integrated into four classes.
Figure 82 (taken from KIRKBY, 2001) shows the
overall flow chart of CORINE model for risk
assessment of soil erosion
In the IMAGE/RIVM approach, terrain erodibility was
estimated based on soil type and landform.
Landform was classified into types by using the
difference between minimum and maximum altitudes
for each grid cell whereas soil type:
- was derived from the FAO Soil Map of the World
and
- is composed of soil depth, soil texture, and bulk
density (GOBIN et al., 2003).
D-08-06.doc
The disadvantage of this approach is that the
information derived is at a 50-km spatial resolution,
which renders it difficult to interpret at sub-national
scales (GOBIN et al., 2003).
Soil type can be considered a very important factor
affecting erodibility since some soil types present
higher susceptibility under the same conditions.
For example, PAPAMIXOS 1996 states that terrarosa
soils present higher soil erosion risk in comparison to
other soil types, especially for high slope values.
Studies of GIOVANNINI et al. 1990 indicate that,
depending on the fire intensity; CaCO3 could suffer a
break-up and change to CaO or Ca+2 (ANDREU et al.,
1996).
This new constitution is far more erodible and can
be washed away by water more easily.
GOFAS (2001) refers also to.
It has been mentioned in the D-08-03 deliverable
that high surface temperatures volatise organic
materials and create gases that move downward in
response to a temperature gradient and then condense
on soil particles causing them to become water
repellent (LETEY, 2001).
This causes an increase in surface runoff and
therefore soil erosion rate.
Many researchers focus on the changes in soil
properties as the essential element affecting erosion
rates (IMESON et al., 1992; KUTIEL and INBAR, 1993;
SEVINK et al., 1989).
GIOVANNINI and LUCCHESI 1991 stress the
importance of the soil conditions as the main factor in
the erosion processes.
In soils with a high percentage of organic matter,
intense fire may lead to complete destruction of the
organic layer and the exposure of a hydrophobic
mineral soil layer, which acts as a repellent mantle
(DEBANO, 1981).
This increases runoff and erosion rates (SEVINK et
al., 1989).
IMESON et al. 1992 state that for fires of medium and
high intensity, the degradation of the vegetation cover
and soil organic matter produce the surface
accumulation of hydrophobic substances that reduce
infiltration and increase runoff (ANDREU et al., 1996).
This situation causes increases in runoff production
and soil removal, which favour the loss of nutrients.
The severity of water repellency depends on the
combined interactions of soil properties and the soil-
heating regime developing during a fire.
The longevity of fire-induced water repellency
depends on some of the same factors that affect its
formation.
Water repellency produced by low to moderate
severity fires is usually of shorter duration than that
produced by high severity fires (DEBANO, 2006).
The results of preliminary field observations suggest
that water repellency might well be an important factor
responsible for the accelerated erosion experienced
during the first few years following wildfires (KRAMMES
and DEBANO, 1965).
88
EUFIRELAB
An initial laboratory study showed that water
repellency could be intensified by heating a soil -
organic matter mixture in a muffle furnace at different
temperatures for different lengths of time (DEBANO and
KRAMMES, 1966).
It was hypothesized that a more efficient coating of
mineral soil particles occurred at lower temperatures
and for shorter periods of heating than in the case of
longer periods of heating at higher temperatures, which
destroyed the organic substances responsible for the
water repellency.
High intensity fires, of 1,200 °C or more, do not
always result in high severity impacts on the soil if their
duration is short, but low intensity fires of just 300 °C
that smoulder for a long time in roots or organic matter
can produce large changes in the nearby soil (NEARY,
2004).
The variability of erosion is associated with fire
severity. Fire severity has been reported (INBAR et al.,
1998; ROBICHAUD, 2000) as being one of the main
reasons for causing erosion variability.
Fire effects on soil depend mainly on fuel type, fire
severity/intensity, topography and soil conditions
(PAPAMIXOS, 1996).
Nevertheless, the way soil erodibility is addressed in
this method can be considered as simplified since
some of the factors affecting susceptibility (soil type
and wildfire effects on soil properties) are not taken
fully into account.
However, extracting this type of data before an
actual event, and even after an event without direct
measurements, is unfeasible at this time.
9.3.3.6 Conclusions
In the Mediterranean, heavy rains follow the fire
season and this means that intervention measures
have to be undertaken upon urgently.
Up until today there is a number of factors, which
are related to soil erosion risk assessment that are
difficult to estimate.
The proposed methodology is based on the use of
factors that are easy to estimate either directly or
indirectly.
It constitutes an approach for “fast and coarse” risk
assessment of potential soil erosion after a fire event.
It should be mentioned that Soil Erosion is closely
related to the effects of fire on vegetation and soil.
The proposed methodology still needs to be
evaluated by conduction of extensive field surveys after
a fire event.
D-08-06.doc
9.3.4 An operational model for the Mediterranean
context
The post-fire effects of a wildfire are usually
temporally and operationally problematic because the
first winter following a summer forest fire is the most
vulnerable period, (SOTO and DÍAZ-FIERROS, 1998;
Vacca et al., 2000), and runoff and erosion
management strategies have to be implemented in the
first weeks after a fire.
The techniques used and their spatial location must
often be decided upon urgently and in an administrative
context where the sources of funding, funding channels
and agency responsibilities are all evolving over a
period of days to months (FOX et al., 2006).
In deliverable D-08-03, an operational model was
presented to respond to post-fire runoff and erosion
risks.
A full description of the model can be found in
Appendix 1.
Initial testing of the model proved satisfactory, but
an additional procedure has been introduced to
improve it further still.
The model itself will not be revised here since it has
been the subject of discussion in the preceding section
and can be consulted in full in the appendix.
The following section deals only with the
improvement to the model and it resides in its capacity
to estimate the evolution in erosion rate over the first
years following a fire.
The change in erosion rate depends principally on
vegetation re-growth, and this in turn is closely related
to soil properties.
In the Mediterranean environment, both soil
properties and vegetation characteristics are strongly
influenced by topography.
The objective was therefore to predict temporal
changes in erosion rate over a 6 year period based on
the impact of topography on soil properties and
vegetation growth.
The proposed addition to the model was developed
from a detailed study of the relationships between
topography, landuse, soil properties, and post-fire
vegetation re-growth in the same catchment used to
test the initial model.
Although the relationships from this single
catchment cannot be extrapolated to the entire
Mediterranean region, they are typical of many
Mediterranean environments.
89
EUFIRELAB
9.3.4.1 Landuse versus topography
In Mediterranean regions, there is a distinct
relationship between landuse and topography.
The climate is conducive to the production of olives,
vines and other crops.
In Mediterranean coastal areas of high population
density, suitable land, possessing both relatively low
slopes and fertile soils, are shared between agricultural
production and human occupation.
This is clearly illustrated in Figure 83, where the
valley bottoms are occupied almost entirely by
vineyards or urban development and the steeper
upland areas are forested with some individual
housing.
The spatial distribution in landuse has implications
for post-fire soil erosion risks.
In Mediterranean France and throughout much of
the Mediterranean region, forest fires occur
systematically on the steepest slopes of the catchment.
The post-fire context is therefore naturally highly
erosive where slopes tend to be both steeper and
longer (Figure 84): slope length in the field is
determined essentially by field length, whereas in
forested areas it is not uncommon to find slopes that
extend from the crest down into valley bottoms where
slope inclination levels off.
The impacts of slope inclination and length on
erosion rates are well known and are not elaborated
upon here.
In addition to the impact of slope inclination and
length on erosion rate, the steeper slopes also
influence sediment deposition.
Sediments detached from slopes have a low
probability of being deposited before reaching a
tributary, and, once in the tributary, there is little
likelihood of deposition before they reach the main
channel.
As can be seen in Figure 85, the area that could
serve as a potential zone of deposition is much more
restricted in the forested zone.
This is quite different from an agricultural context
where fields along river channels can serve as sinks for
sediments eroded upslope and where channel
deposition can be important.
Therefore, the relationship between landuse and
topography affects not only erosion rate, but it also has
an impact on sediment redistribution.
9.3.4.2 Topography versus soil characteristics
The relationships between topography and soil
characteristics were determined by an analysis of the
impacts of topographic units on soil properties.
Slopes were divided into three categories: top-slope
convexities, North-facing slopes and South-facing
slopes. In many Mediterranean environments, water is
a limiting factor and soils tend to be deeper and better
developed on North facing slopes where evapo-
transpiration rates are lower.
D-08-06.doc
In this study, two soil properties were taken into
consideration: soil depth, measured using a hand-held
soil corer, and grain size distribution.
Textural analyses were carried out on surface (top 5
cm) samples for each topographic section and these
were compared to sediments accumulated in sediment
traps put in place after the 2003 fire described in the
previous deliverable.
Thirty sites were selected, ten on each slope
section.
Figure 86 shows the distribution of soil depths
according to slope section.
North facing slopes clearly have deeper soils
(α=0.05, r2=0.74 in an analysis of variance).
Soil depths on these soils since measurements
were stopped at a depth of 1 m due to the difficulties in
manipulating the hand held corer at greater depths.
Differences in soil depth arise essentially from long-
term soil forming processes.
Better moister conditions on North facing slopes
increase the rate and duration of pedogenic processes
and favour vegetation growth (described below).
Greater rooting depths and vegetation densities, in
turn, contribute to more active soil formation.
In addition to differences in soil depth, we can also
note changes in surface grain size distributions.
The stacked histogram of Figure 87a shows that the
fine fraction (< 0.2 cm) content on the South facing and
convex slopes is less than half the content on the North
facing slopes.
The sediments trapped in sediment traps (Figure
87b) positioned in a stream channel draining a burned
slope (see FOX et al., in press) can explain this
tendency.
Log Debris Dams (LDD) are located in the channel
and tend to trap coarser sediments.
Sedimentation basin (Basin) sediments represent
the fraction of sediments moved through the channel
after initial settling behind the LDDs.
Figure 87b clearly shows selective erosion where
the finer fraction is removed preferentially from burned
slopes.
The results suggest that post-fire erosion processes
occur longer and/or more frequently on South facing
and convex slopes.
Finally, it should be noted that the gravel size
fraction (0.2-0.5 cm) represents a greater fraction of the
surface texture on convexities than on South facing
slopes.
In this case, runoff on the low inclination convexities
has sufficient velocity remove the finest fraction, but not
the coarse (> 0.2 cm) sediments, as is the case on the
South facing slopes.
90
EUFIRELAB
9.3.4.3 Topography versus vegetation
For each of the 3 plots identified, vegetation
characteristics were measured in two ways.
Vegetation height was measured on a random
sample of 20 plants located in a 2 m by 2 m perimeter.
This was carried out by hand using a tape measure.
Vegetation density was quantified by photographing
a 50 cm by 50 cm grid (Figure 88) and then by noting
the percentage of points occupied by vegetation in a
random sample of 50 points.
For each x,y coordinate in the 50-points sample, the
surface was noted as either vegetated (1) or non-
vegetated (0).
This value was then converted to a percentage.
Three photos were taken at each of the 30 plots for
a total sample size of 90-grid pictures.
As expected, both vegetation height and density
were greater on North facing slopes than South facing
slopes and top-slope convexities (Figure 89).
Furthermore, soil depth, vegetation height and
surface cover were all positively correlated.
Slope was significant in analyses of variance
(α=0.05) for both vegetation height (r2=0.74) and
surface cover (r2=0.63).
The greater vegetation growth on North-facing
slopes supports the hypothesis evoked above that
selective erosion is more intense on South facing
slopes than North facing slopes since the latter take
more time to recover after a fire.
In addition, South-facing slopes have always had
more human occupation, as is shown by old terraces
and newer housing, so the likelihood of fire ignition has
always been greater on South facing slopes.
This, combined with drier conditions, suggests that
South-facing slopes may burn more often than North
facing slopes.
9.3.4.4 Implications of topographic effects for erosion
modelling
The model described in Annex gives an
instantaneous spatial distribution of post-fire soil
erosion risk.
However, post-fire erosion, though at its maximum
during the first year, can continue over a period of
several years.
Faster vegetation re-growth on North facing slopes
suggests that erosion will be decrease more rapidly
here than on South-facing slopes or top-slope
convexities.
This is supported by both the vegetation data
gathered in-situ and the soil and textural results.
To complete the model, it is therefore necessary to
consider this temporal evolution.
Figure 90 shows the initial calculation of the index of
soil erosion risk.
The initial model re-classed these values into
categories according to erosion risk.
It should be noted that these values represent a
post-fire structural risk.
D-08-06.doc
Actual erosion rates depend on random rainfall
effects: for the same erosion index, erosion rates can
be significantly different according to the total rainfall
and number of high rainfall intensity events.
Hence, an erosion index of 90 (maximum value in
Figure 90) can correspond to an actual erosion rate of
from a few T ha-1 to dozens of T ha-1 according to post-
fire rainfall patterns.
In addition to the impact of rainfall described above,
several processes contribute to determine the evolution
in post-fire erosion.
At least three can be described here:
- The initial rates of erosion depend on the magnitude
of the first post-fire rainfall events as described
above.
- The initial decrease in erosion rates can be
attributed to the loss of the most easily mobilised
sediments. As the finer fraction is eroded, the
increasing stone cover at the surface progressively
decreases sediment detachability and erosion rates.
- Erosion rates decrease further due to the
progressive establishment of a vegetation cover on
the slopes. This occurs faster on North facing slopes
than South-facing slopes.
The trends cited above are shown in figure 91.
The decrease in the erosion index trend is different
for North-facing and South-facing slopes since
vegetation dynamics proceed at different rates.
The change in erosion index over time was
estimated using non-linear decaying exponential curves
for each of the slope types (Table 24).
The two extremes, North- and South-facing slopes,
were defined so that the index reached approximately
zero within three years for the North-facing slope and
within six years for the South-facing slope.
The 3-year limit for the North-facing slope was
determined in the field by observations of the filling of a
sedimentation basin.
The 6-year limit for the South-facing slope is an
estimation based on the current vegetation cover.
Convex slopes were treated with the same equation
as South-facing slopes and West- and East-facing
slopes were given decay values intermediate to the two
extremes.
Slope orientation was defined for each cell in
Figure 90.
Top-slope convexities were defined as slopes above
150-m with an inclination of less than 10%.
This proved to be a simple but effective method of
identifying these topographic features.
Once the orientation was simplified into the five
categories in table 24, the appropriate regression
equation was applied to each of the cells, year by year
for 6 years, in order to predict its evolution over time.
For each year, the erosion index value was
converted to an erosion risk class, as in the initial
method (see Annex).
The result is a dynamic erosion risk spatial
distribution showing the potential erosion risk for each
year during the 6-year period (Figures 92a to 92f).
91
EUFIRELAB
The map for the sixth and final year is not shown
since the entire surface is in the low class category.
In the initial year after the fire, the North facing
slopes initially in the very high-risk class have dropped
to the high and intermediate classes.
By the third year, all of the North-facing slopes are
in the low risk class.
Many of the South-facing slopes persist in the very
high-risk class after the first year, despite the
exponential decrease in erosion risk, and many of
these slopes remain in the intermediate class even
three years after the fire.
Showing the temporal evolution in erosion risk is an
added factor in determining where erosion control
measures should be put into place after a fire.
9.3.5 Figures
Figure 80: Map of NDVI produced by Bayaramin et al. (2006) for their study area.
D-08-06.doc
9.3.4.5 Conclusions
The Mediterranean environment is subject to regular
forest fires and it is possible that the number and
intensity of fires will increase in coming years.
Reacting quickly to post-fire erosion risk is therefore
an issue that must be dealt with by European countries
in particular.
The above deliverable demonstrated two
approaches to dealing with post-fire erosion risk.
In the first, a method for determining pre-fire
potential risk at the European scale is described.
In this context, the issue addressed is how to
acquire apriori knowledge about post-fire erosion risk
before an actual fire occurs.
The second approach refines a post-fire erosion
mapping method that was presented in deliverable D-
08-03.
The major improvement concerns the integration of
post-fire erosion risk evolution over a period of 6 years
after the fire.
In this approach, it was demonstrated that
topography plays a major role in determining post-fire
vegetation recovery and soil depth, and these, in turn,
have a major influence on post-fire erosion.
Field data support the approach and it has
considerable potential for testing in a number of
Mediterranean post-fire contexts.
92
EUFIRELAB

Soil texture (ST) Soil Depth (SD) Stoniness (SS)

0 for Bare rock 1 for > 75 cm 1 for >10 %

1 for C, SC, SiC 2 for 25-75 cm 2 for <10 %
2 for SCL, CL, SiCL, LS, S 3 for 25 cm
3 for L, SiL, Si, SL

Soil erodibility(K)

0 for ST *SD *SS=0

1 for 0 < ST *SD *SS < 3
2 for 3< ST *SD *SS < 6
3 for ST *SD *SS > 6

Figure 81: Methodology for CORINE soil erodibility assessment (CORINE, 1992).

Soil Texture, ST
0 for Bare rock
1 for C, SaC, SiC
2 for SsCL,CL,SiCL,Lsa
3 for SaL, L, SiL, Si

Erodibility, K
Soil Depth, SD 0 for ST.SD .SS =0
1 for >75 cm 1 for 0< ST.SD .SS <3
2 for 25 – 75 cm 2 for 3< ST.SD .SS <6
3 for <25 cm 3 for ST.SD .SS >6

Soil Stoniness, SS
1 for >10%
2 for <10%

Fournier Index, F Erosivity, R

1 for Σpi2/Σp <60 1 for F.B<4
2 for 60<Σpi2/Σp<90 2 for 4<F.B<8
3 for 91<Σpi2/Σp<120 3 for F.B>8
4 for Σpi2/Σp>120
Potential Soil Erosion
Bagnouls-Gaussen Risk, EP
Aridity Index, B 0 for K.R.S = 0
1 for Σ(2Ti-pi) =0 1 for 0 < K.R.S < 5
2 for 0<Σ(2Ti -pi)<50 2 for 5 < K.R.S < 11
3 for 50<Σ(2Ti-pi)<130 3 for K.R.S > 11
4 for Σ(2Ti-pi)>130

Slope angle, S
1 for <5%
2 for 5-15% Actual Soil Erosion
3 for 15-30% Risk, E A
4 for >30% 0 for EP.V = 0
1 for EP.V = 1-2
2 for EP.V = 3-4
Land Cover, V
3 for EP.V >=5
1 for fully protected
2 for not fully protected

Figure 82: Methodology for CORINE Soil Erosion assessment

D-08-06.doc 93
EUFIRELAB

Hilly metamorphic
uplands Alluvial plain

Topography
and
Landuse

Forest
Vineya rds
Ot her
Urban

Figure 83: Steeper uplands in the coastal area of the Massif des Maures (France) are occupied by forest, the
lowland plains are occupied by vineyards and urban development.

a) Slope inclination b) Slope length

Mean (%) Median (%) Mean (m) Median (m)

Forest: 23.3 25.5 Forest: 357.1 400.0
Non-forest: 10.1 6.6 Non-forest: 162.7 162.0
35 400

) 28
%( 300
e
p
y )
T 21 m(
e ht
s g 200
u n
d e
n 14 L
a Forest e
p
L ol
r Non-forest
e
p S
100
a 7
er
A
0 0
0 20 40 60 80 100 Forest Vineyard
Slope Inclination (%)

Figure 84: Highly erosive topographic conditions with few possibilities for sediment deposition in the fire zone

D-08-06.doc 94
EUFIRELAB

50
100

40
80

Soil Depth (cm)

30 60
Area (%)

20 40

10 20

0 0
Forest Non-forest North South Convex

Left Figure 85: Percentage area in the forest and non-forest zones with a slope inclination less than 4%.
Right Figure 86: The North-facing slopes have a greater depth than the South-facing and convex slopes.

a) Slope grainsize b) Sediment trap

characteristics grainsize characteristics
100 100
0.2-2 cm

80 ) 80
) %(
%( n
oi
n t
oi 60 u 60
t bi
u rt
bi si
rt
si D
D e
z
e
zi
40 si 40
ni
s ar
ni
ar G
l
G 20 ai 20
ti
nI

0 0
North South Convex BASIN LDD
< 0.2 cm < 2.0 mm
0.2-0.5 cm • Selective erosion of the finer fractions (< 2 mm) from slopes.
2-20 mm
0.5-1 cm > 20 mm
1-2 cm • Increasing stone cover with erosion – South and Convex slopes.
>2 cm
• Intermediate fraction (0.2-0.5 cm) detached from South facing
slopes but not convexities

Figure 87: North-facing slopes have finer textures than South facing and convex slopes (a)
This results from selective erosion of the finer fraction (measured in sediment traps - b)).

D-08-06.doc 95
EUFIRELAB

Figure 88: Example of a grid photograph used to measure vegetation density.

a) Undergrowth vegetation height b) Undergrowth vegetation cover

120 100

80 R2=0.63
) R2=0.74
m
c( )
80 %(
t r 60
h
g e
i v
o
e c
h n
n oti
40
oi at
t 40 e
at g
e
g
e
V
e 20
V

0 0
North South Convex North South Convex

Correlations: Vegetation height/area: r=0.82; Soil depth / Vegetation height:

r=0.87; Soil depth / Vegetation area: r=0.75

Figure 89: Vegetation height (a) and density (b) were both greater on North facing slopes than on South facing
slopes and top-slope convexities.

D-08-06.doc 96
EUFIRELAB

Theoretical range: 1-135; Actual range: 1-90

Figure 90: Raw values for soil erosion risk calculated from the initial soil erosion model.

90
Peak erosion = f(rainfall events)

Initial decrease = f(sediment

availability, increased stoniness
through selective erosion)
Initial Erosion Index

60

Greatest decrease = f(vegetation

growth)

30
South:

North:

0
0 1 2 3 4 5 6 7
Time (years)

Figure 91: Theoretical curves for soil erosion index trends over time: rates decrease and the greatest decrease
arises due to differences in vegetation growth, which are significantly different on North and South facing slopes

D-08-06.doc 97
EUFIRELAB

Figure 92a and b: Initial soil erosion (left) and one year after the fire risk maps

Figure 92c and d: Erosion risk maps two (left) and three (right) years after the fire

Figure 92e: Erosion risk maps four (left) and five (right) years after the fire

D-08-06.doc 98
EUFIRELAB

9.3.6 Tables

Table 20: Coefficients assigned to each slope class

Slope (%) Coefficient
0-5 1
5-10 2
10-20 3
20-30 4
>30 5

Table 21: Coefficients assigned to each vegetation density

Pre-fire vegetation density Coefficient
Medium 1
Low 2
Bare 3

Table 22: Coefficients assigned to each fire severity level.

Fire severity Coefficient
Variable 1
Intense 3

Table 23: Coefficients assigned to each soil condition class

Soil erodibility Coefficient
Shallow soil with high stone content (depth < 20 cm) 1
Intermediate depth and stoniness 2
Deep soil with few stones (depth > 50 cm) 3

Table 24: Non-linear equations used to predict the evolution in the erosion index
Slope orientation Non-linear equation

North EI = IV × e −2 yr
South EI = IV × e −0.8 yr
Convex EI = IV × e −0.8 yr
East EI = IV × e −1.2 yr
West EI = IV × e −1.6 yr

D-08-06.doc 99
EUFIRELAB

10 EURO-MEDITERRANEAN WILDLAND FIRE RISK INDEX

10.1 GENERAL PRESENTATION In any case, several straightforward solutions can

be raised to make this conversion, using for instance
The basic tools for the application of the proposed
linear functions from the maximum and minimum
EM-WFRI for the regional wildfire management are
values adjustment.
available once the basic structure has been defined
The inclusion of scaled sub-indices to probability
and the different components of this index have been
units will make easier the use of the different
identified.
integration methods.
Next, the methodology to integrate these
components and finally the most suitable procedure to The hierarchical structure of the index, as well as
spread the information should be defined. the probability units format of the different sub-indices,
Concerning this, the multiple suggested possibilities will allow the inclusion of new variables or the
in the D-08-05 must be remarked. substitution or removal of some of the proposed ones,
Given the diversity of available methods, and the in function of the available information of the territory in
variety of potential applications and users, whether of which will be used.
the total of a part of the identified risk components, a Moreover, it will allow introducing modifications in
unique integration methodology has not been function of the work scale, as given some of the
considered to propose. included variables have been presented in this
Instead of this, the index structure and the obtained deliverable (climate, population, etc.)
input variables are offered to the users, leaving them to
In the definition of the risk index structure, the
determine the most suitable integration method
proposed input variables have a spatial nature and a
according to the index destination and the availability of
digital format, as far as possible according to the
combination of its components with additional
available information at the European Union level.
information that can be useful for the management.
Remote sensing information (with a very suitable
In fact, a survey that was performed in Spain to
spatial and temporal component for this kind of studies)
potential users of this index (forest managers and
was considered, as well as point data (meteorological
technicians) in the framework of the FIREMAP
variables, census data, etc.) that would be spatialized
research Project (http://www.geogra.uah.es/firemap)
by means of any spatial interpolation algorithm, and
reveals the preference of managers and technicians to
cartographic information available in digital format
have the basic components of the index (variables and
(Corine Land Cover maps, digital terrain models,
sub-indices), instead of the final index value.
national cartographic databases, etc.).
Thus it enables the use of the information and its
All this information leads to a suitable series of
integration in the prevention systems available for each
spatialized and geo-referenced information layers in
institution.
digital format for using in a Geographical Information
In order to make easier the integration of variables System (GIS)
according to the user interests, the final risk index is
In this context, GIS is a one of the most suitable
generated from three sub-indices: the ignition danger
tools for obtaining a risk wildfire index with the
index, the propagation danger index, and the
proposed characteristics in this deliverable, allowing
vulnerability index.
the management, manipulation, analysis, modelling
These indices are obtained separately, in this way
and representation of the input data as well as the
each one offers specific information about each need in
output indices.
terms of wildfire prevention (ignition, propagation, and
Moreover, they make considerably easier the
wealth), and can be used independently for each
updating processes of the information.
management task.
In this updating process, the existing connection
The three indices combination will offer the global
between the GIS and the remotely sensed images
risk value, and depending on the user, the indices
processing should be considered, thus this is the
weighting can be varied according to the importance of
source of an important part of the dynamic information
each factor in the territory or in the considered
used in the risk index (mainly fuel moisture content).
management task.
On the other hand, the disposal of this information in
It would be recommended for the integration of the a GIS means that it can be used in forest management
variables that all indices that are taken into account for for different facets other than risk indices.
the global risk -not only the three intermediates, but
However, the use of this tool, that considerably
also those that have been used to generate the three
facilitates the process, does not guarantees the
ones- were scaled in probability units (0-1 or 0-100%).
goodness of the obtained index, which reliability will
The way to achieve the conversion from the original
depend on the selection and accuracy of the input
index value to probability units has been included in
variables, as well as the success in the weighting risk
this deliverable for some of the indices (i.e, fuel
assigned to each variable in the model.
moisture content).
Undoubtedly, GIS offer a wide range of possibilities
to define the fire risk integration model:

D-08-06.doc 100
EUFIRELAB
- qualitative methods, where arbitrary weights are
based on the judgment of an expert;
- quantitative indices, based on multi-criteria
evaluation or other survey system;
- regression techniques, where statistical estimation
methods are applied to explain fire occurrence;
- neural network, similar conceptually to the
regression models, although with different fitting
procedures. All these methods were analysed in
detail in the deliverable D-08-05. Its use allows, with
a good enough effective cost and speed, the
creation of new models, the model application at
different scenarios or the modification of risk classes
limits, among others. Because of it, it seems
unquestionable the use of this tool for the
generation of the risk index proposed in this
deliverable.
In these last years an important number of GIS
applications have appeared in the web, with few use
requisites, that make considerably easier the access to
this application from everywhere.
The simplest ones only allow the visualization or
download of a variable of interest, as the more complex
ones allow the user to interact with the proposed
models.
Several examples of both situations are found in the
field of wildfire risk assessment.
For instance, in the United States we can obtain in
real time the National Fire Danger Rating System
estimation or any of its sub-indices that is composed
(USDA Forest Service; http://www.fs.fed.us/land/wfas).
In the European Union, several risk indices by
means of the Risk Forecast System can be obtained in
the European Forest Fire Information System (EFFIS;
http://effis.jrc.it/wmi/viewer.html).
In contrast to this simpler situation, in which the user
only checks the generated information by the different
institutions, situations that are more complex may
appear, where the user can interact with the index
generation.
One of the most interesting examples with this
option is the Fire-Climate-Society Strategic Model
(FCS-1) developed by the Arizona University
(http://walter.arizona.edu),
10.2 FIGURES
Figure 24: Analytic Hierarchy Process (AHP) weights to the Fire-Climate-Society Strategic Model
(Wildfire Alternatives, University of Arizona).
D-08-06.doc
The Fire-Climate-Society Strategic Model is part of
the Model Wildfire Alternatives (WALTER), located at
the University of Arizona.
WALTER is an interdisciplinary research initiative
aimed at improving our understanding of the processes
and consequences of interactions among wildfire,
climate and society.
FCS-1 Fire-Climate-Society (FCS-1) is an online,
spatially explicit strategic wildfire-planning model with
an embedded multi-criteria decision process that
facilitates the construction of user-designed risk
assessment maps under alternative climate scenarios
and varying perspectives of fire probability and values
at risk.
The model is generated by integrating the following
variables or sub-indices: fuel moisture stress index, fire
return interval departure, large fire ignition probability,
lightning probability, human factors of fire ignition,
recreational value, species habitat richness, property
value, personal landscape value.
One of the main advantages of this system is that it
can be used with a standard model, where the weights
of the different variables are assigned according to the
model defined by the experts who created the system,
or the user can decide the risk weights of each input
variables, as shown in the figure 93.
This system seeks to capitalize on advances in geo-
spatial, analytical, and web delivery technology to
provide access to scientific and management activities.
Bearing in mind the required structure of the Euro-
Mediterranean Wildland Fire Risk Index EM-WFRI:
- the nature of variables that compose it,
- the possibility of modifying and updating these
variables,
- the integrating of different methods,
- its use at different spatial scales, etc.,
it is obvious:
- to recommend generating this risk index by means
of GIS and,
- to integrate the use of this tool in a web service.
In this way a very flexible tool would be created, and
could be used for wildfire management in very different
facets of prevention and at different scales (local,
regional, national, supranational), allowing the user to
define the most suitable scenario at the time of use.
101
EUFIRELAB
11 REFERENCES
ABHINEET J; RAVA SA; SINGH RK; DAS KK; ROY PS
1996. Forest fire risk modelling using remote
sensing and geographic information system. Current
Science (70 10):928-933.
AERIAL INFORMATION SYSTEMS Inc. 1981). Final report.
Spatial risk method for predicting wildland fire
occurrence. Crestline, California: USDA, Pacific
Southwest Forest and Range Experiment Station.
AGUADO I; CHUVIECO E; CAMARASA A; MARTÍN MP;
CAMIA A 1998). Estimation of Meteorological Fire
Danger Indices from Multitemporal Series of NOAA-
AVHRR data. Pages 1131-1147 in D. X. Viegas,
editor. III International Conference on Forest Fire
Research - 14th Conference on Fire and Forest
Meteorology. ADAI, Coimbra.
AHLGREN IF; AHLGREN CE 1960. Ecological effects of
forest fires. Bot. Rev. 26, 483–533.
ALBINI FA 1976. Estimating wildfire behavior and
effects. USDA For. Serv. Gen. Tech. Rep. INT-30.
ALBINI FA; REINHARDT ED 1995. Modeling ignition and
burning rate of large woody natural fuels. Int. J.
Wildland Fire, 5: 81–91.
ALEXANDRIAN D; GOUIRAN M 1990. “Les causes
d'incendies : levons le voile.” Revue forestière
française - Espaces forestiers et incendies Numéro
spécial 1990.
ALONSO M; CAMARASA A; CHUVIECO E; COCERO D;
KYUN I; MARTÍN MP; SALAS FJ 1996. Estimating
temporal dynamics of fuel moisture content of
Mediterranean species from NOAA–AVHRR data.
EARSEL Advances in Remote Sensing, 4(4), 9 –
24.
ALTOBELLIS A 1983. A survey of rural population density
and forest fire occurrence in the South, 1956-1970.
New Orleans: USDA Forest Service, Southern
Forest Experimental Station. Report SO-294.
ANDERSEN HE; MCGAUGHLEY RJ; REUTEBUCH SE 2005.
Estimating forest canopy fuel parameters using
LIDAR data. Remote Sensing of Environment, 94,
441-449.
ANDERSON HE; SCHUETTE RD; MUTCH RW 1978.
Timelag and equilibrium moisture content of
ponderosa pine needles. Research Paper SE-74,
USDA Forest Service, Intermountain Research
Station, Ogden, Utah.
ANDREU V; RUBIO JL; FORTEZA J; CERNI R 1996.
Postfire Effects on Soil Properties and Nutrient
Losses. International Journal of Wildland Fire 6 (2),
pp. 53-58.
ANDREU V; IMESON AC; RUBIO JL 2001. Temporal
changes in soil aggregates and water erosion after a
wildfire in a Mediterranean pine forest. Catena 44,
69-84.
D-08-06.doc
BAILEY TC; GATRELL AC 1995. Interactive Spatial Data
Analysis. Longman Scientific and Technical, Burnt
Mill, Essex, England.
BAIRD AW. 1965. Attitudes and Characteristics of forest
residents in three Mississippi counties., 48:
Mississippi State University, Social Science
Research Center Report No. 29.
BARBERIS A; DETTORI S; FILIGHEDDU MR. 2003.
Management problems in Mediterranean cork oak
forests: post-fire recovery. Journal of Arid
Environments 54, 565-569.
BARET F.; FOURTY T. 1997. Estimation of leaf water
content and specific leaf weight from reflectance
and transmittance measurements. Agronomie, 17,
444- 455.
BARET F; GUYOT G. 1991. Potentials and limits of
vegetation indices for LAI and APAR assessment.
Remote Sensing of Environment, 35, 161-173.
BAYRAMIN I; ERPUL G; ERDOGAN EH. 2006. Use of
CORINE methodology to Assess Soil Erosion Risk
in the Semi-Arid Area of Beypazari, Ankara. Turk J
Agric For, 30, 81-100.
BEMMERZOUK AM. 1997. Evaluation of the CFFDRS
and the KBDI in the Mediterranean-type
environment of Crete (Greece. M. Sc. Thesis,
Chania, Greece.
BERRY J; BUCKLEY DJ; MCGARIGAL K. 1999. Integrating
Landscape Structure Programs with GIS.
www.innovativegis.com
BERTRAND AL; BAIRD AW. 1975. Incendiarims in
Southern Forest: a decade of sociological research:
Report. USDA Forest Service, Southern Forest
Experimental Station and Social Science Research
Center at Mississipi University.
BLACKMARR WH; FLANNER WB. 1968. Seasonal and
diurnal variation in moisture content of six species of
Pocosin shrubs. Research Paper: SE-33. pp. 11.
USDA Forest Service, Southeastern Forest
Experiment Station. Asheville, NC.
BLANCHI R ; JAPPIOT M et al. 2002. Forest fire
assessment and cartography - a methodological
approach. Forest fire research and wild land fire
safety, Coimbra.
BOSSARD M; FERANEC J; OTAHEL J. 2000. CORINE land
cover technical guide. Addendum 2000. European
Environment Agency - Technical report n° 40.
BOVIO G; CAMIA A. 1998. An analysis of large forest
fires danger conditions in Europe. Pages 975-994 in
DX VIEGAS, editor. 3rd International Conference on
Forest Fire Research, Luso-Coimbra, Portugal.
BOVIO G; CAMIA A. 2000. Description Eudic software for
fire danger indices calculation.
102
EUFIRELAB
BOVIO G; SOL B; VIEGAS DX. 1994. Synthese des
Travaux sur l'intercomparison des indices de danger
meteorologique d'incendie. Final Report of Minerve I
Project EC, Environment and Climate Programme.
BOVIO G; CAMIA A.; GOTTERO F. 1999. Piano Regionale
per la difesa del patrimonio boschivo dagli incendi
1999-2001. Regione Piemonte, Assessorato
Economia Montana e Foreste.
BOWMAN WD 1989. The relationship between leaf water
status, gas exchange, and spectral reflectance in
cotton leaves. Remote Sensing of Environment, 30,
249– 255.
BRADSHAW LJ; DEEMING JE; BURGAN RE; COHEN JD
1983. The 1978 National Fire-Danger Rating
System. Technical Documentation: GTR INT-169.
pp. 44. USDA Forest Service. (Ogden, Utah.
BURGAN RE. 1979. Estimating live fuel moisture for the
1978. National Fire Danger Rating System. USDA
For. Serv. Res. Pap. INT-226.
BURGAN RE; ROTHERMEL RC. 1984. BEHAVE: fire
behaviour prediction and fuel modeling system. Fuel
subsystem. USDA For. Serv. Gen. Tech. Rep. INT-
167.
BURGAN RE; HARTFORD RA. 1993. Monitoring
Vegetation Greenness with Satellite Data. GTR INT-
297, USDA Forest Service, Ogden, Utah.
BURGAN, R. E., KLAVER, R. W. and KLAVER, J. M. 1998.
Fuel models and fire potential from satellite and
surface observations. International Journal of
Wildland Fire 8:159-170.
BURGAN RE. 1999. Personal Communication. Not
published.
CAMIA A; BOVIO G; AGUADO I; STACH N. 1999.
Meteorological fire danger indices and remote
sensing. Pages 39-59 in E. Chuvieco, editor.
Remote Sensing of Large Wildfires in the European
Mediterranean Basin. Springer-Verlag, Berlin.
CAMIA A ; BOVIO G ; MARZANO R. 2002. Preliminary
analysis of wildland-urban interface fire prone areas
in Italy. In: Viegas DX (ed.), Proc. IV Int. Conf. on
Forest Fire Research and 2002 Wildland Fire Safety
Summit, Luso (Portugal), Millpress Science Pub.,
Rotterdam, pp.9; ISBN 90-77017-72-0.
CAMIA A; VARELA V; MARZANO R; EFTICHIDIS G. 2004.
Spatial analysis in European Wildland-Urban
Interface environments using GIS. In: Xanthopoulos
G. (ed.), Proc. II International Workshop “Forest Fire
in the Wildland-Urban Interface and Rural Areas in
Europe: an integral planning and management
challenge”; 15–16 May 2003, Athens (Greece.
Mediterranean Agronomic Institute of Chania
(MAICh), Graphic Art and Publishing Unit, Chania,
Crete, 185 – 196.
CARDILLE JA; VENTURA SJ; TURNER MG. 2001.
Environmental and social factors influencing
wildfires in the Upper Midwest, United States.
Ecological Applications 11 (1):111-127.
D-08-06.doc
CARLSON JD; BURGAN RE. 2003. Review of users’
needs in operational fire danger estimation: The
Oklahoma example. International Journal of Remote
Sensing, 24(8), 1601– 1620.
CARREGA P. 1990. Climatology and index of forest fire
hazard in Mediterranean France. Pages B.05-01/11
in International Conference on Forest Fire
Research, Coimbra, Portugal.
CARREGA P. 1994. Topoclimatologie et habitat. U.F.R.
Espaces et Cultures. Nice, Université de Nice -
Sophia Antipolis: 408.
CARTER GA. 1991. Primary and secondary effects of
water content on the spectral reflectance of leaves.
American Journal of Botany, 78, 916– 924.
CARVACHO L. 1998. Evaluación de la estimación de
grandes incendios forestales en la cuenca
mediterránea europea por redes neuronales y
regresión logística. Serie Geográfica 7:73-85.
CASE P; BANKS B; BUTLER E; GOSNELL R. 2000.
Chapter 7. Assessing Potential Wildfire Effects on
People. Co-published simultaneously in Journal of
Sustainable Forestry, 11(1/2), 159-175; and:
Mapping Wildfire Hazards and Risks (ed: Neil
Sampson, R. Dwight Atkinson, and Joe W. Lewis),
Food Product Press, an imprint of The Haworth
Press, Inc., 2000, , 159-175.
CASTRO FX; TUDELA A; SEBASTIÁ MA. 2003. Modeling
moisture content in shrubs to predict fire risk in
Catalonia (Spain. Agricultural and Forest
Meteorology, 116, 49– 59.
CEC 1985. Soil Map of the European Communities,
1:1,000,000. CEC Luxembourg, 124pp., 7 maps.
CECCATO P; FLASSE S ; TARANTOLA S; JACQUEMOUD S;
GRÉGOIRE JM. 2001. Detecting vegetation leaf water
content using reflectance in the optical domain.
Remote Sensing of Environment, 77, 22–33.
CECCATO P; GOBRON N; FLASSE S; PINTY B;
TARANTOLA S. 2002. Designing a spectral index to
estimate vegetation water content from remote
sensing data: Part 1. Theoretical approach. Remote
Sensing of Environment, 82, 188– 197.
CECCATO P; LEBLON B; CHUVIECO E; FLASSE S;
CARLSON JD. 2003. Estimation of live fuel moisture
content. In E. Chuvieco (Ed.), Wildland fire danger
estimation and mapping. The role of remote sensing
data (pp. 63–90. Singapore: World Scientific
Publishing.
CERDÀ A; IMESON AC; CALVO A. 1995. Fire and aspect
induced differences on the erodibility and hydrology
of soils at La Costera, Valencia, south-east Spain.
Catena 24, 289-304.
CHANDLER CP; CHENEY P; THOMAS P; TRABAUD L;
WILLIAMS D. 1983. Fire in Forestry. Forest Fire
behavior and effects. John Wiley and Sons: New
York.
103
EUFIRELAB
CHEN K; BLONG R; JAKOBSON C. 2003.Towards an
Integrated Approach to Natural Hazards. Risk
Assessment Using GIS: With Reference to
Bushfires. Environmental Management, 31(4), 546-
560.
CHLADIL MA; NUNEZ M. 1995. Assessing grassland
moisture and biomass in Tasmania. The application
of remote sensing and empirical models for a cloudy
environment. International Journal of Wildland Fire,
5, 165– 171.
CHOU YH. 1992. Spatial autocorrelation and weighting
functions in the distribution of wildland fires.
International Journal of Wildland Fire 2 (24):169-
176.
CHOU YH; MINNICH RA; CHASE RA. 1993. Mapping
probability of fire occurrence in San Jacinto
Mountains, California, USA. Environmental
Management 17 (1):129-140.
CHRISTENSEN NL. 1987. The biogeochemical
consequences of fire and their effects on the
vegetation of the Coastal Plain of the southeastern
United States. In: The role of fire in ecological
systems, (L. Trabaud Ed.. SPB Academic
Publishing 1- 19.
CHRISTIANSEN J; FOLKMAN W. 1971. Characteristics of
people who start fires...some preliminary findings.
Berkeley, Cal.: USDA, Forest Service. Pacific
Southwest Forest and Range Experiment Station.
Research Note PSW-251.
CHUVIECO E; CONGALTON RG. 1989. Application of
remote sensing and Geographic Information
Systems to Forest fire hazard mapping. Remote
Sensing of Environment 29:147-159.
CHUVIECO E; SALAS FJ. 1996. Mapping the spatial
distribution of forest fire danger using GIS.
International Journal of Geographical Information
Systems 10 (3):333-345.
CHUVIECO E; SALAS FJ; VEGA C. 1997. Remote sensing
and GIS for long-term fire risk mapping. Pages 91-
107 in E. Chuvieco, editor. A review of remote
sensing methods for the study of large wildland
fires. Departamento de Geografía, Universidad de
Alcalá, Alcalá de Henares.
CHUVIECO E; DESHAYES M; STACH N; COCERO D;
RIAÑO D. 1999a. Short-term fire risk: Foliage
moisture content estimation from satellite data. In E.
Chuvieco (Ed.), Remote sensing of large wildfires in
the European Mediterranean Basin (pp. 17– 38.
Berlin: Springer-Verlag.
CHUVIECO E; SALAS FJ; CARVACHO L; RODRÍGUEZ-
SILVA F. 1999b. Integrated fire risk mapping. In
Remote Sensing of Large Wildfires in the European
Mediterranean Basin, Ed. E CHUVIECO. pp 61-84.
Berlin: Springer-Verlag.
D-08-06.doc
CHUVIECO E; RIAÑO D; AGUADO I; COCERO D. 2002.
Estimation of fuel moisture content from
multitemporal analysis of Landsat Thematic Mapper
reflectance data: Applications in fire danger
assessment. International Journal of Remote
Sensing, 23(11), 2145–2162.
CHUVIECO E; AGUADO I; COCERO D; RIAÑO D. 2003a.
Design of an empirical index to estimate fuel
moisture content from NOAA–AVHRR analysis in
forest fire danger studies. International Journal of
Remote Sensing, 24(8), 1621– 1637.
CHUVIECO E; ALLGÖWER B; SALAS FJ. 2003b.
Integration of physical and human factors in fire
danger assessment. In Wildland Fire Danger
Estimation and Mapping. The Role of Remote
Sensing Data, ed. E. CHUVIECO, pp: 197-218.
Singapore: World Scientific Publishing.
CHUVIECO E; AGUADO I; DIMITRAKOPOULOS A. 2004a.
Conversion of fuel moisture content values to
ignition potential for integrated fire danger
assessment. Canadian Journal of Forest Research-
Revue Canadienne De Recherche Forestiere 34
(11): 2284-2293.
CHUVIECO E; COCERO D; RIAÑO D; MARTÍN P; MARTÍNEZ-
VEGA J., RIVA, J., and PÉREZ, F. 2004b. Combining
NDVI and surface temperature for estimation of live
fuel moisture content in forest fire danger rating.
Remote Sensing of Environment, 92, 322-331.
CHUVIECO E; COCERO D; AGUADO I; PALACIOS-
ORUETA A; Prado E. (in press. Improving burning
efficiency estimates through satellite assessment of
fuel moisture content. Journal of Geophysical
Research - Atmospheres
COHEN WB. 1991. Response of vegetation indices to
changes in three measures of leaf water stress.
Photogrammetric Engineering and Remote Sensing,
57(2), 195– 202.
COLE LW; KAUFMAN HF. 1963. Socio-economic factors
and forest fire in Mississippi counties. Mississippi.
USA: Mississippi State University Social Sciences
Research Centre. Preliminary Report 14.
COLL C; CASELLES V. 1997. A global split-window
algorithm for land surface temperature from AVHRR
data: Validation and algorithm comparison. Journal
of Geophysical Research, 102B14, 16697–16713.
COLLINS DB. 1996. “FireLib User Manual and Technical
Reference”. http://www.fire.org.
COMUNIDAD DE MADRID 2000): Plan de Protección Civil
de emergencia por incendios forestales en la
Comunidad de Madrid (INFOMA), BOCM nº 138 pp.
4-28.
CONSEIL DE L’EUROPE / COMMISSION DES COMMUNATES
EUROPEENNES. 1987. Carte de la Végétation
Naturelle des Etats membres des Communatés
Européennes et du Conseil de l’Europe. Office for
Official Publications of the European Communities,
Luxembourg.
104
EUFIRELAB

CORINE 1992. Soil Erosion Risk and Important Land DELGADO XP 2001. Spatial patterns of forest fires in
Resources in the Southern Regions of the European Catalonia (NE of Spain) along the period 1975-
Community. EUR 13233, Luxembourg. 1995. Analysis of vegetation recovery after fire.
Forest Ecology and Management, 147, pp.67-74 .
CROSSON P 1997. Will erosion threaten agricultural
productivity? Environment 39 (8), 4– 12. DELICHATSIOS MA, PANAGIOTOU T, KILEY F. 1991. The
use of time-to-ignition data for characterizing the
CRUZ GM; BUTLER WB; ALEXANDER EM 2006.
thermal inertia and the minimum (critical) heat flux
Predicting the ignition of crown fuels above a
for ignition or pyrolysis. Combust. Flame, 84: 323–
spreading fire. Part II: model evaluation.
332.
International Journal of Wildland Fire, Vol. 15, pp.
61-72. DENGIZ O, AKGUL S, 2005. Soil Erosion Risk
Assessment of the Golbasi Environmental
CUTTER SL, MITCHELL JT, SCOTT MS 2000. Revealing
Protection Area and Its Vicinity Using the CORINE
the Vulnerability of People and Places: A Case
Model. Turk J Agric For, 29, 439-448.
Study of Georgetown County, South Carolina.
Annals of the Association of American Geographers, DESBOIS N, DESHAYES M, BEUDOIN A. 1997. Protocol for
90(4), 713-737. fuel moisture content measurements. In 'A review of
remote sensing methods for the study of large
DATT B 1999. Remote sensing of water content in
wildland fires. (Ed E. Chuvieco) pp. 61-72.
eucalyptus leaves. Australian Journal of Botany, 47,
Departamento de Geografía, Universidad de Alcalá.
909–923.
(Alcalá de Henares, Madrid.
DE LA RIVA J , PEREZ F , RENAULT NL , KOUTSIAS N.
DIMITRAKOPOULOS A, PAPAIOANNOU KK 2001.
2004. Mapping forest fire occurrence at a regional
Flammability assessment of Mediterranean forest
scale. Remote Sensing of Environment 92(3):363-
fuels. Fire Technology, 37, 143– 152.
369
DISASTER MONITORING CONSTELLATION
DE LUIS M, GONZALEZ-HIDALGO JC, RAVENTOS J. 2003.
http://directory.eoportal.org/pres_DMCDisasterMonit
Effects of fire and torrential rainfall on erosion in a
oringConstellationAlSAT1BILSAT1NigeriaSat1UKD
Mediterranean gorse community. Land Degradation
MCBeijing1.html
and Development 14, 203-213.
DOERR SH, SHAKESBY RA, WALSH RPD. 1998. Spatial
DEBANO LF, KRAMMES JS 1966. Water repellent soils
variability of soil hydrophobicity in fire-prone
and their relation to wildfire temperatures.
eucalyptus and pine forests, Portugal. Soil Science
International Association of Hydrological Science
163, 313-323.
Bulletin 2, 14-19.
DOERR SH, SHAKESBY RA, WALSH RPD. 2000. Soil
DEBANO LF 1981. Water repellent soils: a state-of-the-
water repellency: its causes, characteristics and
art. USDA Forest Service General Technical Report
hydro-geomorphological significance. Earth-Science
PS W-46, Berkeley, CA.
Reviews 51, 33-65.
DEBANO LF. 2000a. Water repellency in soils: a
DONOGHUE LR, MAIN WA. 1985. Some Factors
historical overview. Journal of Hydrology 231-231,
Influencing Wildfire Occurrence and Measurement
4-32.
of Fire Prevention Effectiveness. Journal of
DEBANO LF. 2000b. The role of fire and soil heating on Environmental Management 20 (1):87-96.
water repellency in wildland environments: a review.
DONOGHUE LR, SIMARD AJ, MAIN WA. 1987.
Journal of Hydrology 231-231, 195-206.
Determining the Economic Relationship Betwen Law
DEEMING JE, BURGAN RE, COHEN JD. 1978. The Enforcement Activities and Arson Wildfires: a
National Fire-Danger Rating System - 1978. GTR Feasibility Study in Arkansas. Journal of
INT-39, USDA Forest Service, Ogden, UT. Environmental Management 25:377-393.
DEFRIES RS, TOWNSHEND JRG 1994. Global land- DOOLITTLE ML. 1979. Southern woods-burners: a
cover: Comparison of ground-based data sets to descriptive analysis: USDA Forest Service.
classifications with AVHRR Data. In G.M. Foody, Southern Forest Experiment Station. Research
and P.J. Curran (Eds.), Environmental remote Paper SO-151
sensing from regional to global scales ( pp. 84– 110.
DOOLITTLE ML, WELCH GD. 1974. Fire prevention in the
Chichester: John Wiley and Sons.
Deep South: personal contact pays off. Journal of
DE JONG SM, 1994. Applications of reflective remote Forestry 72:488-490.
sensing for land degradation Studies in a
DOUGUEDROIT A, JOLY C. 2002. Spatial distribution of
Mediterranean Environment. PhD thesis, Utrecht
the forest fires in lower Provence : the influence of
University, The Netherlands, 237 pp.
Marseille. Forest Fire Research and Wildland Fire
DELETIC A. 2001. Modelling of water and sediment safety, Coimbra, Milpress.
transport over grassed areas. Journal of Hydrology
248, 168-182.

D-08-06.doc 105
EUFIRELAB
DRAKEA JB, DUBAYAHA RO, CLARK DB, KNOXD RG,
BLAIRD BJ, HOFTONA MA, CHAZDONE RL,
WEISHAMPELF JF, PRINCEA SD. 2002. Estimation of
tropical forest structural characteristics using large-
footprint lidar. Remote Sensing of Environment, 79,
305-319.
DROUET JC, SOL B. 1993. Mise au point d'un indice
numerique de risque météorologique d'incendies de
forêts. Forêt Mediterranéenne 14:155-162.
EASTMAN JR 1993. IDRISI. Clark University, Worcester.
ELVIDGE CD, BAUGH KB, KIHN E, KROEHL HW, DAVIS ER
1997. Mapping city lights with nightime data from
the DMSP Operational Linescan System.
Photogrammetric Engineering and Remote Sensing
63:727-734.
ESPELTA JM, RETANA J, HABROUK A 2003. An economic
and ecological multi-criteria evaluation of
reforestation methods to recover burned Pinus nigra
forests in NE Spain. Forest Ecology and
Management 180, 185-198.
EUROPEAN COMMISSION 2002. Forest Fires in Europe –
2001 fire campaign – Official Publication of the
European Commission, S.P.I.02.72 EN.
EUROPEAN COMMISSION, JOINT RESEARCH CENTRE (DG
JRC), Institute for Environment and Sustainability
(IES), 2005) - CORINE Land Cover updating for the
year 2000. IMAGE2000 and CLC2000. Products
and Methods. Maria Vanda NUNES DE LIMA (Ed..
EUR 21757 EN, ISBN 92-894-9862-5.
EUROPEAN ENVIRONTMENTAL AGENCY 1996. Natural
Resources, CD-Rom. European Environtmental
Agency, Copenhagen.
EUROSTAT, 1999. Statistical references. OYce for
oYcial publications of the European Communities,
No.2/1999, Luxembourg.
FAO 1986. Wildland fire managemente terminology.
Report nº 70, FAO Forestry Paper, Roma. M-99.
ISBN 92-5-0024207.
F.A.O. 1988. Soil map of the world 1:500.000 scale.
F.A.O. Rome.
FERREIRA AJD, COELHO COA, SHAKESBY RA,
WALSH RPD 1997. Sediment and solute yield in
forest ecosystems affected by fire and rip-ploughing
techniques, central Portugal: a plot and catchment
analysis approach. Physical Chemistry of the Earth
22, 309-314.
FOLKMAN WS. 1965. Forest fires as accidents: an
epidemiological approach to fire prevention
research, 15. Berkeley, California: Pacific Southwest
Forest and Range Experiment Station. USDA.
FOLKMAN WS. 1973. Fire prevention in Butte County,
California: their knowledge and attitudes regarding
forest fire prevention., 32. Berkeley, California:
USDA, Forest Service Pacific Southwest Forest
Range Experiment Station. PSW-98.
D-08-06.doc
FOSBERG MA, JE DEEMING. 1971. Derivation of the 1-
and 10-hour timelag fuel moisture calculations for
fire danger rating. Research Note RM-207, USDA,
Forest Service, Fort Collins, Colorado.
FOSBERG MA, ROTHERMEL RC, ANDREWS PL. 1981.
Moisture content calculations for 1000-Hour Timelag
Fuels. Forest Science 27:19-26.
FOTHERINGHAM AS, BRUNSDON C, CHARLTON M. 2000.
Quantitative Geography: Perspectives on Spatial
Data Analysis. SAGE Publications, London.
FOTHERINGHAM AS, BRUNSDON C, CHARLTON M. 2002.
Geographically weighted regression: The analysis of
spatially varying relationships. John Wiley and Sons,
New York, NY. 269 p.
FOX D, BEROLO W, CARREGA P, DARBOUX F 2006.
Mapping erosion risk and selecting sites for simple
erosion control measures after a forest fire in
Mediterranean France. Earth Surface Processes
and Landforms, Vol 31, pp. 606-621.
FUJIOKA FM. 1987. Fire weather network design: an
update. In '9th Conference on fire and forest
meteorology'. American Meteorological Society. pp.
165-169. (San Diego, California.
GAGARI P, LYRINTZIS G, BALOUTSOS G,
XANTHOPOULOS G 2001. Contribution of log erosion
barriers to soil protection and vegetation recovery
after wildfire in a Pinus halepensis forest, in Greece.
In Proceedings of the International Conference,
Forest Research: a Challenge for an Integrated
European Approach, Thessaloniki, Greece,
Radoglou K. (ed.), 1:311-316.
GATRELL AC; BAILEY TC; DIGGLE PJ; ROWLINGSONT BS
1996. Spatial point pattern analysis and its
application in geographical epidemiology.
Transactions of the Institute of British Geographers
21:256-274.
GIOVANNINI G; S LUCHESI; M GIACHETTI; 1990.
Beneficial and detrimental effects of heating on soil
properties. In: Fire in ecosystem dynamics, (J.G.
Goldamrner and M.J. Jenkins Eds. SPB Academic
Publishing 95-102.
GIOVANNINI G, LUCCHESI S 1991. Is the vegetative cover
the primary factor controlling erosion in burnt soils?
In: Sala, M., Rubio, J.L. Eds. , Soil Erosion and
Degradation as a Consequence of Forest Fires.
Barcelona, p. 16.
GIOVANNINI G, VALLEJO R, LUCCHESI S, BAUTISTA S,
CIOMPI S, LLOVET J. 2001. Effects of land use and
eventual fire and soil erodibility in dry Mediterranean
conditions. Forest Ecology and Management 147,
15-23.
GOBIN A, KIRKBY C, GOVERS G 2003. Pan-European
soil erosion risk assessment. European Commision
funded fifth framework project.
GOFAS A 2001. Εγχειρίδιο Δασοπροστασίας. Εκδόσεις
Γιαχούδη-Γιαπούλη.
106
EUFIRELAB
GONÇALVES MZ, LOURENÇO L. 1990. Meteorological
index of forest fires risk in the Portuguese mainland
territory. Pages B.07-01 - B.07-14 in I International
Conference on Forest Fire Research., Coimbra.
GRANGER K. 1998. Developing an understanding of
urban geohazard risk. Australian Journal of
Emergency Management 13(4):13–17.
GULINCK H, MÚGICA M, DE LUCIO JV, ATAURI JA. 2001. A
framework for comparative landscape analysis and
evaluation based on land cover data, with an
application in the Madrid region (Spain), Landscape
and Urban Planning, 55 pp. 257-270.
HARDY CC. 2005. Wildland hazard and risk: problems,
definitions and contexte. Forest ecology and
management.
HARDY CC, BURGAN RE. 1999. Evaluation of NDVI for
monitoring live moisture in three vegetation types of
the Western U.S. Photogrammetric Engineering and
Remote Sensing, 65, 603– 610.
HAYES JC, BARFIELD BJ, BARNHISEL RI 1984.
Performance of grass filters under laboratory and
field conditions. Transactions of the American
Society of Agricultural Engineers 27, 1321-1331.
HONG S, BARRY H, DE ZONIA, MALDENOFF DJ. 2000. An
aggregation index (AI) to quantify spatial patterns of
landscape, Landscape Ecology, 15 pp. 591-601.
http://perso.wanadoo.fr/04as/carte/cartes/pacageol.htm
http://www.meteofrance.com/FR/climat/dpt_tempsdumo
is.jsp?LIEUID=DEPT06
http://www.prim.net/citoyen/definition_risque_majeur/int
rofeux.
http://www.promethee.com/prom/home.do (base de
données Prométhée)
HUFFMAN EL, MCDONALD LH, STEDNICK JD. 2001.
Strength and persistence of fire-induced soil
hydrophobicity under ponderosa and lodgepole
pine, Colorado front range. Hydrological Processes
15, 2877-2892.
ICONA. 1993. Manual de operaciones contra incendios
forestales. 5.1/65, ICONA, Madrid.
IGIS 1997. FRAGSTATS*ARC User’s Manual
www.innovativegis.com/products /fragstatsarc
ILLERA P, FERNANDEZ A, DELGADO JA. 1996. Temporal
evolution of the NDVI as an indicator of forest fire
danger. International Journal of Remote Sensing
17:1093-1105.
IMESON AC, VERSTRATEN JM, VAN MULLIGEN EJ,
SEVINK J. 1992. The effects of fire and water
repellency on infiltration and runoff under
Mediterranean type forest. Catena 19, 345-361.
INBAR M, WITTENBERG L, TAMIR M. 1997. Soil Erosion
and Forestry Management After Wildfire in a
Mediterranean Woodland, Mt. Carmel, Israel.
International Journal of Wildland Fire 7 (4), 285-294.
D-08-06.doc
INBAR M, TAMIR M, WITTENBERG L. 1998. Runoff and
erosion processes after a forest fire in Mount
Carmel, a Mediterranean area. Geomorphology 24,
17-33.
JACKSON RD, IDSO SB, REGINATO RJ, PINTER PJ. 1981.
Canopy temperature as a crop water stress
indicator. Water Resources Research, 17, 1133–
1138.
JACQUEMOUD S, USTIN SL, VERDEBOUT J, SCHMUCK G,
ANDREOLI G, HOSGOOD B. 1996. Estimating leaf
biochemistry using the PROSPECT leaf optical
properties model. Remote Sensing of Environment,
56, 194–202.
JACQUEMOUD S, BACOUR C, POILVE H, FRANGI JP. 2000.
Comparison of four radiative transfer models to
simulate plant canopies reflectance: Direct and
inverse mode. Remote Sensing of Environment, 74,
471– 481.
JOLY C, SUSINI F et al. 1999. Analysis of the occurrence
of fire patterns in a limited area Forest fires: Needs
and Innovations, Athens, Greece.
JONES AR, TAYLOR ML, BERTRAND AL. 1965. Some
human factors in woods burning, 28. Baton Rouge,
La.: La. State University Agricultural Experiment
Station Bulletin Nº. 601.
KALABOKIDIS KD, KONSTANTINIDIS P, VASILAKOS C.
2002. GIS analysis of physical and impact on
wildfire patterns. In Proceedings of IV International
Conference on Forest fire research and wild land
fire safety, Coimbra. (D.X. Viegas, Ed.)
KALLURI SNV, TOWNSHEND JRG, DORAISWAMY P. 1998.
A simple single layer model to estimate transpiration
from vegetation using multispectral and
meteorological data. International Journal of Remote
Sensing, 19(6), 1037– 1053.
KEETCH J, BYRAM G. 1968. A Drought Index for forest
fire control. Research Paper. SE-38, USDA Forest
Service, Southeastern Forest Experiment Station,
Asheville, NC.
KERDILES H. 1997. Software for processing AVHRR
data for the Communities of Europe (SAPCE):
Algorithms used in SPACE version 2. Part 1:
Calibration, atmospheric correction and cloud
detection. Internal document at ARIS Unit, SAI, JRC
Ispra.
KIRKBY M. 2001. MEDRAP Athens December 2001
Physical Aspects of Soil Erosion Control. page 11.
KOUTSIAS N, KALABOKIDIS K, ALLGÖWER B. 2002. Fire
occurrence patterns at landscape level: beyond
positional ac-curacy of ignition points with kernel
density estimation methods. Poster presentation in
2002 World Conference on Natural Resources
Modeling, “Modeling Natural and Biotic Resources
in a Changing Planet”, 23-26 June 2002, Sigri,
Lesbos.
107
EUFIRELAB
KOUTSIAS N, KALABOKIDIS K, ALLGÖWER B. 2004. Fire
occurrence patterns at landscape level: beyond
positional accuracy of ignition points with kernel
density estimation methods. Natural Resource
Modeling 17(4): 359-375.
KOUTSIAS N, J MARTÍNEZ, E CHUVIECO, B ALLGÖWER.
2005. Modelling Wildland Fire Occurrence in
Southern Europe by a Geographically Weighted
Regression Approach. Paper presented at 5th
International Workshop on Remote Sensing and
GIS Applications to Forest Fire Management: Fire
Effects Assessment, 16-18 June 2005, at Zaragoza,
Spain.
KOUTSIAS N, MARTÍNEZ J, ALLGÖWER B. 2006. Global
(OLS) and Local (Geographically--Weighted
Regression) Models for Explaining Long-term
Wildland Fire Occurrence Patterns in Southern
Europe. Journal of Geophysical Research (in press)
KRAMMES JS, DEBANO LF. 1965. Soil Wettability: a
neglected factor in watershed management. Water
Resources Research 1, 283–286.
KUTIEL P, INBAR M. 1993. Fire impact on soil nutrients
and soil erosion in a Mediterranean pine forest
plantation. Catena 20, 129–139.
LAIRD JR, HARVEY MD. Complex response of a
chaparral drainage basin to fire. In: Hadley, R.F. Ed.
, Drainage Basin Sediment Delivery, Vol. 159. IAHS
Publ., pp. 165–183.
LAL R, 2004. Soil carbon sequestration impacts on
global climate change and food security. Science
304 (5677), 1623– 1627.
LE BISSONNAIS, Y. 1996. Aggregate stability and
assessment of soil crustability and erodibility : I.
Theory and methodology 47, 425-437.
LEBLON, B. 2001. Forest wildfire hazard monitoring
using remote sensing: A review. Remote Sensing
Reviews, 20(1), 1 –57.
LETEY, J. 2001. Causes and consequences of fire-
induced soil water repellency. Hydrological
Processes 15, 2867-2875.
LEVINE, N. 2002. CrimeStat II: A Spatial Statistics
Program for the Analysis of Crime Incident
Locations (version 2.0), Ned Levine and Associates
Annandale, VA and The National Institute of Justice
Washington, DC.
LICHSTEIN, J.W., SIMONS, T.R., SHRINER, S.A,
FRANZREB, K.E. 2002. Spatial autocorrelation and
autoregressive models in ecology. Ecological
Monographs 72 (3): 445-463.
LIGDI, E., AND MORGAN, R.P.C. 1995. Contour grass
strips: a laboratory simulation of their role in soil
erosion control. Soil Technology 8, 109-117.
LOMBORG, B., 2001. The Skeptical Environmentalist:
Measuring the Real State of the World. Cambridge
University Press.
D-08-06.doc
LOVELAND, T.R., MERCHANT, J.W., OHLEN, D.O. AND
BROWN. J.F. 1991. Development of a Land-Cover
Characteristics Database for the Conterminous U.S.
Photogrammetric Engineering and Remote Sensing
57:1453-1463.
LYNHAM, T.J., AND MARTELL., D.L. 1985. An analysis of
annual people-caused forest fire ocurrence in
Ontario. Paper read at 8th National Conference on
Fire and Forest Meteorology, April 29-May 2, at
Detroit, Michigan.
MALDENOFF, D. J. AND BAKER, W.L. 2000): Spatial
Modelling of Forest Landscape Change, Cambridge
University Press.
MANDALLAZ, D., AND YE. R. 1997. Prediction of forest
fires with Poisson models. Canadian Journal of
Forest Research 27:1685-1694.
Manuel d’utilisation de SYSTAT 8.0 1998. Régression
logistique : p 517 - 584
MARQUES, M.A., AND MORA E. 1998. Effects on erosion
of two post-fire management practices: clear-cutting
versus non-intervention. Soil and Tillage Research
45, 433-439.
MARTELL, D.L., OTUKOL, S. AND STOCKS, B.J. 1987. A
logistic model for predicting daily people-caused
forest fire occurrence in Ontario. Canadian Journal
of Forest Research 17:394-401.
MARTELL, D.L., BEVILACQUA, E. AND STOCKS, B.J. 1989.
Modelling seasonal variation in daily people-caused
forest fire occurrence. Canadian Journal of Forest
Research 19 (12):1555-1563.
MARTIN, C., CHEVALIER, Y., GIMENEZ, H., AND QUILLARD,
J. 1998. Variations de la composition chimique des
eaux de drainage et de ruissellement et
modifications des caractères agronomiques des sols
après incendie de forêt dans le bassin versant du
Rimbaud (Var, France. Géomorphologie 3, 215-232.
MARTINEZ, J., CHUVIECO, E., MARTIN, M.P. 2004.
Estimating human risk factors in Wildland fires in
Spain using logistic regression. II International
Symposium on Fire Economics, Planning and
Policy: A Global Vision. Córdoba, Spain. 19-22 de
Abril, 2004. 15 pp.
MARTINEZ, J, ROMERO, R. AND ECHEVARRIA, P. 2006):
Valoración paisajística y ecológica de la Comunidad
de Madrid: su integración en un índice sintético de
riesgo de incendios forestales. XII Congreso
Nacional de Tecnologías de la Información
Geográfica. Granada. CDRom.
MATAIX-SOLERA, J., AND DOERR, S.H. 2004.
Hydrophobicity and aggregate stability in calcareous
topsoils from fire-affected pine forests in south-
eastern Spain. Geoderma 118, 77-88.
MCARTHUR, A.G. 1967. Fire behaviour in eucalypt
forests. Leaflet 107, Commonwealth of Australia
Forest and Timber Bureau, Canberra, Australian
Capital Territory.
108
EUFIRELAB
MCGARIGAL, K. AND MARKS, B.J. 1995): FRAGSTATS:
spatial pattern analysis program for quantifying
landscape structure. Portland, Department of
Agriculture, Forest Service.
MCGARIGAL, K., MARKS, B., HOLMES, C. AND ENE, E.
2002): Fragstats 3.3. Spatial Pattern Analysis
Program for Quantifying Landscape Structure,
Department of Natural Resources Conservation,
University of Massachusetts.
MONTANARELLA, L., 2003. Soil erosion risk in Europe.
European Commision, Joint Research Centre,
Institute for Environment and Sustainability. T.P.
280, 21020 Ispra (VA) ITALY.
http://ies.jrc.cec.eu.int/Projects/ESB/.
MORAN, M.S., CLARKE, T.R., INOUE, Y., AND VIDAL, A.
1994. Estimating crop water deficit using the relation
between surface-air temperature and spectral
vegetation index. Remote Sensing of Environment,
49, 246– 263.
MORGAN, P., HARDY, C.C., SWETNAM, T.W., ROLLINS,
M.G., AND LONG, D.G. 2001. Mapping fire regimes
across time and space: Understanding coarse and
fine-scale fire patterns. International Journal of
Wildland Fire, Vol. 10: 329-342.
MORSDORF, F., MEIER, E., KOTZ, B., ITTEN, K. I.,
DOBBERTIN, M. ALLGOWER, B. 2004. LIDAR-based
geometric reconstruction of boreal type forest
stands at single tree level for forest and wildland fire
management. Remote Sensing of Environment, 92,
353-362.
MUNOZ-CARPENA, R., PARSONS, J.E., AND GILLIAM, J.W.
1999. Modeling hydrology and sediment transport in
vegetative filter strips. Journal of Hydrology 214,
111-129.
NAVEH, Z. AND LIEBERMAN, A. 1994): Landscape
Ecology. Theory and Application, Nueva York,
Springer-Verlag.
NEARING, M.A., PRUSKI, F.F., O’NEAL, M.R., 2004.
Expected climate change impacts on soil erosion
rates: a review. Journal of Soil and Water
Conservation 59 (1), 43– 50.
NEARY, D.G., KLOPATEK, C.C., DEBANO, L.F., AND
FOLLIOTT, P.F. 1999. Fire effects on belowground
sustainability: a review and synthesis. Forest
Ecology and Management 122, 51-71.
NEARY, G. D., 2004. An Overview of Fire Effects on
Soils. Southwest Hydrology, September/October
2004.
NEUENSCHWANDER, L.F., MENAKIS J.P., MILLER, M.,
SAMPSON, R.N., HARDY, C., AVERILL, B., MASK, R.,
2000. Chapter 3. Indexing Colorado Watersheds to
Risk of Wildfire. Co-published simultaneously in
Journal of Sustainable Forestry, 11(1/2), 159-175;
and: Mapping Wildfire Hazards and Risks (ed: Neil
Sampson, R. Dwight Atkinson, and Joe W. Lewis),
Food Product Press, an imprint of The Haworth
Press, Inc., 2000, , 35-55.
D-08-06.doc
NOBLE, I.R., BARY, G.A.V. AND GILL. A.M. 1980.
McArthur's fire-danger meters expressed as
equations. Australian Journal of Ecology 5:201-203.
ORIEUX, A. 1979. Conditions météorologiques et
incendies di forêts en région mèditerranéenne. Note
Technique du Service Météorologie Métropolitain
Ministère des Transports, Direction de la
Météorologie.
PALMIERI, S., INGHILESI, R., SIANI, A.M. AND
MARTELLACCI., C. 1992. Un indice meteorologico di
rischio per incendi boschivi. Bollettino Geofisico
15:49-62.
PALTRIDGE, G.W., AND BARBER, J. 1988. Monitoring
grassland dryness and fire potential in Australia with
NOAA/AVHRR data. Remote Sensing of
Environment, 25, 381– 394.
PAPAMIXOS, T., N., 1996. Δασικά Εδάφη, Σχηματισμός,
Ιδιότητες, Συμπεριφορά. Aristotle University of
Thessaloniki, Υπηρεσία Δημοσιευμάτων.
PAUSAS, J.G. 1997. Resprouting of Quercus suber in
NE Spain after fire. Journal of Vegetation Science 8,
703-706.
PAUSAS, J.G. 2004. Changes in fire and climate in the
eastern Iberian peninsula (Mediterranean basin.
Climatic Change 63, 337-350.
PAUSAS, J.G, AND VALLEJO, V.R. 1999. The role of fire
in European Mediterranean Ecosystems. In Remote
Sensing of Large Wildfires in the European
Mediterranean Basin, Chuvieco E. (ed.), Springer-
Verlag, 3-16.
PEW, K.L. AND LARSEN, C. 2001. “GIS analysis of spatial
and temporal patterns of human-caused wildfires in
the temperate rain forest of Vancouver island,
Canada.” Forest ecology and management 140: 1-
18
PHILLIPS, C.B., AND NICKEY. 1978. The concept of
"spatial risk" and its application to fire prevention.
Fire Management Notes 39 (4):7-8-19.
PINAYA, I, SOTO B, ARIAS M, AND DÍAZ-FIERROS F. 2000.
Revegetation of burnt areas: relative effectiveness
of native and commercial seed mixtures. Land
Degradation and Development 11, 93-98.
PROSPER-LAGET, V., DOUGUÉDROIT, A., AND GUINOT,
J.P. 1995. Mapping the risk of forest fire occurrence
using NOAA satellite information. EARSEL
Advances in Remote Sensing, 4(3), 30– 38.
PROSSER IP, AND WILLIAMS L. 1998. The effect of
wildfire on runoff and erosion in native Eucalyptus
forest. Hydrological Processes 12, 251-265.
RENARD, K.G., FOSTER, G.R., WEESIES, G.A., MCCOOL,
D.K., YODER, D.C., 1997. Predicting soil erosion by
water: a guide to conservation planning with the
Revised Universal Soil Loss Equation. Agricultural
Handbook, vol. 703. U.S. Department of Agriculture
404 pp.
109
EUFIRELAB
RIAÑO, D., MEIER, E., ALLGÖWER, B., CHUVIECO, E. AND
USTIN, S. L. 2003. Modelling airborne laser
scanning data for the spatial generation of critical
forest parameters in fire behavior modelling.
Remote Sensing of Environment, 86, 177– 186.
RICHARD, J.-F., AND BEROUTCHACHVILI, N. 1996. “Vers
l'élaboration d'un système d'information sur les
paysages du monde.” Cahiers des sciences
humaines 32: 823-842.
RIITTERS, K.H; O’NEILLS, R.V; HUNSAKER, C.T;
WICKHAM, J.D; YANKEE, D.H. AND TIMMINS, S.P.
1995): A factor analysis of landscape pattern and
structure metrics, Landscape Ecology, 10, pp. 23-
39.
RIVA, J.D.L., PEREZ-CABELLO, F. et al. 2004. “Mapping
Wildfire occurrence at regional scale.” Remote
sensing of environnement 92: 363-369.
ROBICHAUD, PR, HUNGERFORD RD. 2000. Water
repellency by laboratory burning of four northern
Rocky Mountain forest soils. Journal of Hydrology
231–232: 207–219.
ROBICHAUD, P. 2004. Post-fire rehabilitation treatments:
are we learning what works ? Southwest hydrology,
Watersheds on Fire 3, 20-21.
RODRIGUEZ Y SILVA, F. 2002. Visual Forest Fire
Weather Index (vffwi) a mathematical model for the
prediction of forest fires weather danger in
mediterranean ecosystems. In 'IV Conference on
Forest Fire Research'. ADAI, (Luso,Portugal.
ROTHERMEL, R.C. 1972. A Mathematical Model for
Predicting Fire Spread in Wildland Fuels. Res. Pap.
INT-115. pp. 40. USDA Forest Service. (Ogden,
Utah.
ROTHERMEL, R.C., WILSON, R.A. MORRIS, G. A., AND
SACKETT., S.S. 1986. Modelling moisture content of
fine dead wildland fuels: input to BEHAVE fire
prediction system. Research Paper INT-359, USDA
Forest Service Intermountain Research Station,
Odgen, Utah.
RUBIO, J.L, FORTEZA, J., ANDREU, V., AND CERNI, R.
1997. Soil profile characteristics influencing runoff
and soil erosion after forest fire: a case study
(Valencia, Spain. Soil Technology 11, 67-78.
RUBIO, J. L., BOARDMAN J., HELMING K., 2006. Soil
erosion across Europe: Research approaches and
perspectives. Catena, Article in Press.
SADAHIRO, Y. 2003. Stability of the surface generated
from distributed points of uncertain location.
International Journal of Geographical Information
Science 17:139-156.
SALAS, F.J., AND CHUVIECO, E. 1994. GIS applications
to forest fire risk mapping. Wildfire 3:7-13.
D-08-06.doc
SAMPSON, R.N., NEUENSCHWANDER L.F., 2000. Chapter
2. Characteristics of the Study Area, Data Utilized,
and Modeling Approach. Co-published
simultaneously in Journal of Sustainable Forestry,
11(1/2), 159-175; and: Mapping Wildfire Hazards
and Risks (ed: Neil Sampson, R. Dwight Atkinson,
and Joe W. Lewis), Food Product Press, an imprint
of The Haworth Press, Inc., 2000, , 15-33.
SAN MIGUEL-AYANZ, J., ANNONI, A. AND SCHMUCK. G.
1998. The use of satellite imagery for retrieval of
information on wildfire damage in Mediterranean
landscapes. in ERIM’98, International Symposium
on Remote Sensing of Environment: Information for
sustainability, Tromso.
SAN MIGUEL-AYANZ, J., BARBOSA, P., SCHMUCK, G., AND
LIBERTA, G. 2002. Towards a coherent forest fire
information system in Europe: The European Forest
Fire Information System (EFFIS. in D. X. Viegas,
editor. IV International Conference on Forest Fire
Research and 2002 Wildland Fire Safety Summit.
ADAI, University of Coimbra, Coimbra, Portugal.
SANTOS, MÁRIO R. CAETANO, PAULO M. BARBOSA AND
JOSÉ U. PAÚL, 2000. A comparative study of
vegetation indices to assess land cover change after
forest fires. Remote Sensing for Earth Science,
Ocean, and Sea Ice Applications, 3868, pp.232-240.
SCOTT, D.F., VAN WYK, D.B., 1990. The effects of
wildfire on soil wettability and hydrological behaviour
of an afforested catchment. Journal of Hydrology
121, 239–256.
SEBASTIÁN LÓPEZ, A. 2004. Modelización a escala
europea del riesgo de incendio forestal, desde la
perspectiva dinámica y estructural. Universidad de
Valladolid, Valladolid.
SEBASTIÁN LÓPEZ, A., SALVADOR, R., AND SAN-MIGUEL-
AYANZ, J. 2001. Two models for a structural forest
fire risk index at a regional scale. in 21st EARSeL
Workshop Remote sensing and GIS applications to
forest fire management: New Methods and Sensor,
Paris. France.
SEBASTIÁN LÓPEZ, A., SAN-MIGUEL-AYANZ, J., AND
BURGAN, R.E. 2002. Integration of satellite sensor
data, fuel type maps and meteorological observation
for evaluation of forest fire risk at the pan-European
scale. International Journal of Remote Sensing
23(13): 2713-2719.
SEVINK, J., IMESON, A.C., VERSTRATEN, J.M., 1989.
Humus form development and hillslope runoff, and
the effects of fire and management under
mediterranean forest in NE-Spain. Catena 16, 461–
475.
SHAKESBY, R.A., COELHO, C.O.A., FERREIRA, A.J.D.,
TERRY, J.P., and WALSH, R.P.D. 1993. Wildfire
impacts on soil erosion and hydrology in wet
Mediterranean forest, Portugal. International Journal
of Wildland Fire 3, 95-110.
110
EUFIRELAB

SHAKESBY, R.A., BOAKES, D.J., COELHO, C.O.A., SOTO, B, BASANTA R, PEREZ R, AND DIAZ-FIERROS F.
GONÇALVES, A.J.B., AND WALSH, R.P.D. 1994. 1995. An experimental study of the influence of
Limiting soil loss after forest fire in Portugal: the traditional slash-and-burn practices on soil erosion.
influence of different post-fire timber clearing Catena 24, 13-23.
practices. In Proc. 2nd Int. Conf. Forest Fire
STRAHLER, A.H., WOODCOCK, C.E., AND SMITH, J.A.
Research II, 1161-1170.
1986. On the nature of models in remote sensing.
SHAKESBY, R.A., BOAKES, D.J., COELHO, C.O.A., Remote Sensing of Environment, 20, 121–140.
GONÇALVES, A.J.B., AND WALSH, R.P.D. 1996.
SWANSON, F.J., 1981. Fires and geomorphic processes.
Limiting the soil degradational impacts of wildfire in
In: Proceedings, fire regimes and ecosystems
pine and eucalyptus forests in Portugal. Applied
conference, Honolulu. Gen. Tech. Rep. WO-26
Geography 16, 337-355.
USDA, Washington, DC, pp. 401–420.
SHAKESBY, R.A, DOERR, S.H, AND WALSH, R.P.D. 2000.
TADESSE, L.D. and MORGAN, R.P.C., 1996. Contour
The erosional impact of soil hydrophobicity: current
grass strips: a laboratory simulation of their role in
problems and future research directions. Journal of
erosion control using live grasses. Soil Technology
Hydrology 231-232, 178-191.
9, 83-89.
SILVA, F.C.O.R. 2002. "Visual forest fire weather index
TERRES, J.M. 1999. The Crop Growth Monitoring
(VFFWI)" a mathematical model for the prediction of
System implemented by JRC/ARIS unit for the
forest fire weather danger in Mediterranean
information needs of the EC DG VI - Agriculture.
ecosystems. Forest fire research and wild land fire
Proceedings of the international symposium
safety, Coimbra.
Modelling Cropping Systems - European Society for
SILVERMAN, B.W. 1986. Density Estimation for Statistics Agronomy. EX 18746/C. ORA/PRO 60194R
and Data Analysis. Chapman and Hall, London,
THOMAS, A.D., WALSH, R.P.D, AND SHAKESBY, R.A.
England.
1999. Nutrient losses in eroded sediment after fire in
SIMARD, A.J. 1968. The moisture content of forest fuels. eucalyptus and pine forests in the wet
A review of the basic concepts. Information Report: Mediterranean environment of northern Portugal.
FF-X-14. pp. 47. Forest Fire Research Institute. Catena 36, 283-302.
(Ottawa, Ontario.
THOMAS, A.D., WALSH, R.P.D., AND SHAKESBY, R.A.
SMITH, S.J., WILLIAMS, J.R., MENZEL, R.G., COLEMAN, 2000. Solutes in overland flow following fire in
G.A., 1984. Prediction of sediment yield from eucalyptus and pine forests, northern Portugal.
Southern Plains grasslands with the Modified Hydrological Processes 14, 971-985.
Universal Soil Loss Equation. Journal of Range
THOMPSON, W.A. 2000. Using forest fire hazard
Management 37 (4), 295– 297.
modelling in multiple use forest management
SNEEUWJAGT, R.J., AND PEET, G.B. 1985. Forest fire planning. Forest Ecology and Management
behaviour tables for Western Australia. Perth: 134:163-176.
Western Australian Department of Conservation and
THORNES, J.B., 1985. The ecology of erosion.
Land Management.
Geography 70, 222–234.
SOL, B. 1990. Estimation du Risque Météorologique
TRABAUD, L. 1976. Inflamabilité et combustibilité des
D'incendies de Forêts dans le Sud-Est de la France.
principales espèces des garrigues de la region
Revue Forestière Française 17:263-271.
mediterranéenne. Oecol. Plant. 11: 117–136.
SOL, B. 1999. A European program improves the
TUCKER, C.J., J.R. TOWNSHEND AND T.E. GOFF. 1985.
operational fight against forest fires: MINERVE
African land-cover classification using satellite data.
Program and meteorological fire danger in
Science, 227: 367-374.
southeastern France, in DELFI Proceedings,
Athens, pp. 185-192. VACCA, A., LODDO, S., OLLESCH, G., PUDDU, R., SERRA,
G., Tomasi, D. and Aru, A. 2000. Measurement of
SOLER, M., SALA, M., 1992. Effects of fire and of
runoff and soil erosion in three areas under different
clearing in a Mediterranean Quercus ilex woodland:
land use in Sardinia (Italy. Catena 40, 69-92.
an experimental approach. Catena 19 3r4 , 321–
332. VALETTE, J.C. 1990. “Inflammabilité des espèce
forestières méditerranéennes: conséquences sur la
SOTO, B., BENITO, E., BASANTA, R., DIAZ-FIERROS, F.,
combustibilité des formations forestières.” Revue
1991. Influence of antecedent soil moisture on
forestière française - espaces forestiers et incendies
pedological effects of fire. In: Sala, M., Rubio, J.L.
numéro spécial 1990.
Eds., Soil Erosion and Degradation as a
Consequence of Forest Fires. Barcelona.
SOTO, B. and DÍAZ-FIERROS, F. 1998. Runoff and soil
erosion from areas of burnt scrub: comparison of
experimental results with those predicted by the
WEPP model. Catena 31, 257-270.

D-08-06.doc 111
EUFIRELAB

VALETTE, J.C. 1992. Inflammabilities of Mediterranean VIEDMA, O., MELIA, A, J., SEGARRA, D., GARCÍA-HARO,
species. In Forest Fire Risk and Management: J., 1997. Modelling rates of ecosystem recovery
European School of Climatology and Natural after fires by using Landsat TM data. Remote
Hazards Course. Halkidiki, Greece, 27 May – 4 Sensing of Environment, 61(3), pp. 383-398.
June 1992. Edited by P. Balabanis, G. Eftichidis,
VIEGAS, D.X., ed. 1998. Proceedings of the III
and R. Fantechi. European Commission, Halkidiki,
International Conference on Forest Fire Research
Greece. pp. 51–64.
and 14th Conference on Fire and Forest
VALLEJO, R, AND ALLOZA J.A. 1998. The restoration of Meteorology. Held in Luso, Coimbra, Portugal, on
burned lands: the case of eastern Spain. Large November 16-20th: ADAI - Associaçao para o
Forest Fires, 91-108. Desenvolvimento da aerodinâmica Industrial - and
University of Coimbra.
VAN DER GOOT, E, 1997. Technical description of
interpolation and processing of meteorological data VIEGAS, D.X. 1998b. Fuel moisture evaluation for fire
in CGMS. Internal document at SAI, JRC Ispra behaviour assessment. In G. Eftichidis, P. Balabaris,
and A. Ghazi (Eds.), Advanced study course on
VAN WAGNER, C.E. 1982. Initial moisture content and
wildfire management. Final Report, Marathon (pp.
the exponential drying process. Canadian Journal of
81– 92. Athens: European Union.
Forest Research 12:90-92.
VIEGAS, D.X., ed. 2002. Proceedings of the IV
VAN WAGNER, C.E. 1987. Development and structure of
International Conference on Forest Fire Research
the Canadian Forest Fire Weather Index System.
and 2002 Wildland Fire Safety Summit. Held in
Technical Report 35. pp. 37. Canadian Forestry
Coimbra, Portugal, on 18-23 November: ADAI -
Service. (Ottawa, Ontario.
Associaçao para o Desenvolvimento da
VASCONCELOS, M.J.P., SILVA, S., TOMÉ, M., ALVIM, M., aerodinâmica Industrial - and University of Coimbra,
and J.M.C. PEREIRA. 2001. Spatial prediction of fire Portugal.
ignition probabilities: comparing logistic regression
VIEGAS, D.X., VIEGAS, T.P., AND FERREIRA, A.D. 1992.
and neural networks. Photogrammetric Engineering
Moisture content of fine forest fuels and fire
and Remote Sensing 67 (1):73-83.
occurrence in central Portugal. International Journal
VEGA GARCÍA, C., WOODARD, P.M. TITUS, S.J. of Wildland Fire, 2(2), 69–85.
ADAMOWICZ, W.L. AND LEE, B.S. 1995. A logit model
VIEGAS, D.X., BOVIO, G., CAMIA, A., FERREIRA, A., AND
for predicting the daily occurrence of human caused
SOL. B. 1996. Testing Meteorological Fire Danger
forest fires. International Journal of Wildland Fire 5
Methods in Southern Europe. Pages 571-589 in
(2):101-111.
13th Conference on Fire and Forest Meteorology,
VEGA GARCÍA, C., LEE, B.S., WOODARD, P.M., AND Lorne - Australia.
TITUS. 1996. Applying neural network technology to
VIEGAS, D.X., PIÑOL, J., VIEGAS, M.T., AND OGAYA, R.
human-caused wildfire occurrence prediction. AI
2001. Estimating live fine fuels moisture content
Applications 10 (3):9-18.
using meteorologically-based indices. International
VELEZ, R et al. 2000 La defensa contra incendios Journal of Wildland Fire, 10, 223–240.
forestales: fundamentos y experiencias. McGraw
VINEY, N.R. 1991. A review of fine fuel moisture
Hill, Madrid
modelling. International Journal of Wildland Fire
VELEZ, R. AND MÉRIDA, J.C. 2002. Forest fire 1(4): 215-234.
management in Spain: Some examples of
VOUGHT, L.B.M, PINAY, G., FUGLSANG, A., AND
systematic analysis of a comprehensive database to
RUFFINONI, C. 1995. Structure and function of buffer
improve effectiveness and efficiency. Forest fire
strips from a water quality perspective in agricultural
research and Wild land fire safety, Coimbra.
landscapes. Landscape and Urban Planning 31,
VELEZ, R., REANO, P., et al. 2002. Forest fire prevention 323-331.
with a target: the rural people. Forest fire research
WALLIS, M.G., AND HORNE, D.J. 1992. Soil water
and wildland fire safety, Coimbra.
repellency. In Advances in Soil Science, Vol. 20,
VERSTRAETE, M.M. AND PINTY, B. 1996. Designing Stewart BA (ed., Springer: New York, 91-146.
optimal spectral indexes for remote sensing
WALSH, R.P.D., COELHO, C.O.A., SHAKESBY, R.A., AND
applications. IEEE Transactions on Geoscience and
TERRY, J.P. 1992. Effects of land use management
Remote Sensing, 34(5), 1254– 1265.
practices and fire on soil erosion and water quality in
VIDAL, A., PINGLO, F., DURAND, H., DEVAUX-ROS, C., the Agueda river basin, Portugal. Geoökoplus III,
AND MAILLET, A. 1994. Evaluation of a temporal fire 15-36.
risk index in Mediterranean forest from NOAA
thermal IR. Remote Sensing of Environment, 49,
296– 303.

D-08-06.doc 112
EUFIRELAB

WALSH, R.P.D., BOAKES, D.J., COELHO, C.O.A.,

YOUNG, R.A., ONSTAD, C.A., BOSCH, D.D., ANDERSON,
FERREIRA, A.J.D., SHAKESBY, R.A., AND THOMAS,
W., 1989. AGNPS: a nonpoint-source pollution
A.D. 1995. Post-fire land use and management and
model for evaluating agricultural watersheds.
runoff responses to rainstorms in northern Portugal.
Journal of Soil and Water Conservation 44 (2), 168–
In Geomorphology and Land Management in a
173.
Changing Environment, McGregor DFM, and
Thompson DA (eds), Wiley,Chichester, 283-308. ZARCO-TEJADA, P.J., RUEDA, C.A., AND USTIN, S.L.
2003. Water content estimation in vegetation with
WARREN, A., 2002. Land degradation is contextual.
MODIS reflectance data and model inversion
Land Degradation & Development 13 (6), 449– 459.
methods. Remote Sensing of Environment, 85, 109–
WOHLGEMUTH, P.M., HUBBERT, K.R., AND ROBICHAUD, 124.
P.R. 2001. The effects of log erosion barriers on
ZHANG, L.J., SHI, H.J. 2004. Local modeling of tree
post-fire hydrologic response and sediment yield in
growth by geographically weighted regression.
small forested watersheds, southern California.
Forest Science 50 (2): 225-244.
Hydrological Processes 15, 3053-3066.
ZHANG QUANFA, GORAN PAVLIC, WENJUN CHEN, RASIM
WORTON, B.J. 1989. Kernel methods for estimating the
LATIFOVIC, ROBERT FRASER, JOSEF CIHLAR, 2004.
utilization distribution in homerange studies. Ecology
Deriving stand age distribution in boreal forests
70:164-168.
using SPOT VEGETATION and NOAA AVHRR
YOOL, S.R., ECKHARDT, D.W., ESTES, J.E., AND imagery. Remote Sensing of Environment, 91, pp.
COSENTINO., M.J. 1985. Describing the brushfire 405–418.
hazard in southern California. Annals of the
Association of American Geographers 75 (3):417-
430.

D-08-06.doc 113
EUFIRELAB
12 ANNEX: MAPPING POST-FIRE SOIL EROSION RISK
FOX D, BEROLO W, CARREGA P
UMR 6012 Espace CNRS, Equipe GVE, Department of Geography, University of Nice Sophia Antipolis
BP 3209 06204 Nice Cedex 3, France.
12.1 INTRODUCTION
The Mediterranean climate is particularly conducive
to large-scale forest fires.
Mean annual precipitation is sufficient to support a
dense mixture of oak, pine, and other forest species
and the warm dry summer season makes both the live
and accumulated biomass highly inflammable for up to
at least two months of the year (VALLEJO and ALLOZA,
1998; PAUSAS and VALLEJO, 1999.
The abandonment of traditional underbrush clearing
practices combined with an increase in accidental and
intentional fire ignitions has led to more frequent forest
fires throughout much of the Mediterranean region
(MORENO and VALLEJO, 1999, in FERNANDEZ et al.,
2003).
Changing climatic conditions are likely to increase
the number of fires and area burned in coming years
(PAUSAS, 2004).
The impacts of a major forest fire on runoff and
erosion are well-known, and these include greater peak
flows and soil loss until the return of a vegetation cover.
The first winter following a summer forest fire is the
most vulnerable period and runoff and erosion rates
generally decrease rapidly afterwards (SOTO and DÍAZ-
FIERROS, 1998; VACCA et al., 2000).
It is therefore essential that runoff and erosion
management strategies be implemented in the first
weeks after a fire.
Several natural factors determine the impact of a fire
on post-fire runoff and erosion rates, and these include
the following: pre-fire vegetation, topography, slope
aspect, fire severity, changes in soil properties, and
post-fire rainfall (WALSH et al., 1992; RUBIO et al., 1997;
INBAR et al., 1998; SOTO and DÍAZ-FIERROS, 1998;
THOMAS et al., 2000; DELUIS et al., 2003).
Although many of these factors (such as slope or
rainfall intensity) are common to all soil erosion
contexts and need little explanation, the changes in soil
erodibility brought about by the intense heat are
particular to forest fire conditions.
For this reason, these will be reviewed briefly before
presenting the erosion control strategy.
D-08-06.doc
12.2 THE IMPACT OF FOREST FIRES ON SOIL
ERODIBILITY
The increases in runoff and soil erosion rates
observed after a forest fire are due primarily to the
destruction of the vegetation cover (PROSSER and
WILLIAMS, 1998; WOHLGEMUTH et al., 2001).
However, the combustion of the standing vegetation
and litter layer provokes several changes in soil
properties.
The study of low severity fires or traditional slash
and burn methods has confirmed some of the short-
term benefits of forest fires on soil fertility (pH, nutrient
status) and their influence on post-fire regrowth (SOTO
et al., 1995; NEARY et al., 1999).
However, the effects are not all beneficial since the
combustion of the soil organic matter can also lead to
lower cation exchange capacities, and high severity
fires can lead to significant nutrient losses through
volatilisation and accelerated erosion (MARTIN et al.,
1998; NEARY et al., 1999; THOMAS et al., 1999).
Two further changes in soil characteristics, the
development or enhancement of hydrophobicity and a
decrease in aggregate stability, have a direct impact on
soil erodibility (ANDREU et al., 2001.
The formation of soil hydrophobicity, or water
repellency, has been described extensively elsewhere
(e.g. WALLIS and HORNE, 1992; DEBANO, 2000; DOERR
et al., 2000; DEBANO, 2000b; HUFFMAN et al., 2001), so
only a few aspects related directly to runoff generation
and soil erosion will be considered here.
Water repellency develops naturally in most
Mediterranean forests (DOERR et al., 1998; MATAIX-
SOLERA and DOERR, 2004).
During a forest fire, heat from the combustion
vaporizes organic substances, some of which migrate
downwards into the soil where they condense at cooler
temperatures and coat mineral particles (DEBANO,
2000)
After a forest fire, it is therefore common to find a
thin layer of hydrophobic soil at the surface or within a
depth of a few centimetres (DEBANO, 2000; MATAIX-
SOLERA and DOERR, 2004).
Although the effect of hydrophobicity on runoff and
erosion has clearly been demonstrated at the plot
scale, it has been difficult to demonstrate its importance
for a catchment (SHAKESBY et al., 1993; SHAKESBY et
al., 2000).
One reason for this is the high spatial variability in
hydrophobicity: at the catchment scale, preferential flow
in decayed root channels, cracks, rodent burrows, and
hydrophilic patches may account for a large proportion
of the infiltrated water (IMESON et al., 1992; FERREIRA et
al., 1997; SHAKESBY et al., 2000).
114
EUFIRELAB

The impact of forest fires on soil aggregate stability 12.3 SITE DESCRIPTION
is more ambiguous than for hydrophobicity.
The study area is located in south-east France in
On the one hand, forest fires tend to produce
the Massif des Maures near St Tropez; the
greater water repellency in most forest soils, and this
approximate centre of the burnt area is 43°16’ N, 6° 28’
tends to reduce slaking, thereby increasing aggregate
E.
stability (MATAIX-SOLERA and DOERR, 2004).
The Giscle catchment has a surface area of about
On the other hand, aggregate stability is positively
234 km2 and is composed of two major topographical
correlated with organic matter content, so the
features.
destruction of the organic compounds present in the
The lower portion of the catchment (roughly 25% of
soil tends to diminish aggregate stability (CERDÀ et al.,
the total area) is an alluvial plain occupied by vineyards
1995; GIOVANNINI et al., 2001).
and urban development.
The net effect probably depends on initial organic
The upper portion (about 75%) of the catchment has
matter content, soil texture, fire intensity, and the size
forested hilly terrain dominated by a mixture of
range of aggregates being considered.
Mediterranean oak and pine species.
(MATAIX-SOLERA and DOERR 2004) showed that finer
A major forest fire which occurred at the end of
sieve fractions (< 0.25 mm) are more hydrophobic than
August, 2003, burned more than 2000 ha located at the
larger ones (> 0.25 mm), and that both hydrophobicity
head of the Giscle river and a few of its tributaries.
and aggregate stability are correlated with soil organic
Local authorities were therefore concerned about
matter for different sieve fractions, but this question
the impact of increased runoff on the urbanised areas
requires further research.
downstream as well as the possible increase in
Soils on south-facing slopes tend to have both lower sediment load entering the port.
aggregate stability values and greater erosion rates
Annual precipitation is about 950 mm with rainy
than soils on north-facing slopes (CERDÀ et al., 1995;
seasons in the autumn and spring.
MARQUÉS and MORA, 1998; ANDREU et al., 2001).
The hilly forested zone is underlain by metamorphic
Several reasons can explain this: south-facing gneiss and schists, and the soils are classified as
slopes in the Mediterranean climate tend to have Rankers in the FAO classification.
greater potential evapo-transpiration rates leading to Soil textures for the < 2 mm fraction are typically
sparser vegetation cover and therefore thinner soils about 75% sand, 10% silt, and 15% clay. Mean Weight
and lower soil organic matter contents. Diameter (MWD), measured using the method of Le
Bissonnais 1996), for soils in the valley bottom is 3.28
In addition, the drier conditions on south-facing
mm (std. dev. = 0.07), classifying these soils as highly
slopes may make them more susceptible to forest fires
erodible.
with more frequent burns and exposure to soil erosion
processes; finally, post-fire recovery on south-facing
slopes is slower than on north-facing slopes (CERDÀ et
al., 1995; PAUSAS and VALLEJO, 1999), so erosion
processes continue for longer.

D-08-06.doc 115
EUFIRELAB
12.4 METHODS
12.4.1 Mapping soil erosion risk
Since undisturbed forests typically have very low
soil erosion rates, most soil erosion models were
developed for agricultural contexts and were not
suitable for the burned forest site.
Measuring soil erosion and elaborating a regression
model were also out of the question due to the lack of
time.
After an initial survey of the digital data available,
the dominant factors affecting soil erosion (described
below) were mapped and combined to create an
erosion map.
The multiplicative method is described below after a
presentation of the factors.
In the field, each layer was mapped on a 1:25,000
topographic map; some areas of the burned upper
catchment were inaccessible except via footpaths,
much of the burned catchment was accessible only via
unmade roads (fire prevention access roads and
private lanes), and only the lower portion of the
catchment could be accessed by car.
Therefore, cycling throughout the catchment and
combining direct on-site observations with panoramic
views from viewpoints using binoculars for remote
sites, carry out most of the mapping.
Roughly 2,000 ha had to be mapped in a few days,
so each factor was attributed a number of qualitative
classes that could be estimated quickly without time-
consuming measurements.
The methods used are outlined below after a
description of the soil erosion factors.
12.4.2 Soil erosion factors
Slope: the importance of slope inclination for soil
erosion is well known, and a Digital Elevation Model
(DEM) of the catchment with a 50 m grid was already
available, but the resolution quickly revealed itself to be
insufficient, so a 25 m DEM was ordered.
Pre-fire vegetation: pre-fire density was estimated
from the standing charred trunks, and three categories
were included – bare, low density, and medium density.
Rock outcrops were devoid of any vegetation and
the distinction between low and medium density was
based mainly upon the ease of walking through the
charred forest.
In addition to the mulch effect provided by fallen
pine needles described above (SHAKESBY et al., 1993;
SHAKESBY et al., 1994), standing dead vegetation may
favour infiltration (near stems and in burned roots) and
slow runoff velocity, so the net expected effect is lower
erosion where vegetation stands are more dense.
Fire severity: this is a major factor influencing runoff
and soil erosion rates (ROBICHAUD and WALDROP, 1994,
cited in LETEY, 2001; RUBIO et al., 1997; PROSSER and
WILLIAMS, 1998).
Two severity categories were included: variable and
intense.
Variable severity sites were designated as areas
where:
D-08-06.doc
- either the fire had passed quickly and charred the
trunks and underbrush but left the crown more or
less intact, or
- where the intense combustion of the standing
vegetation was confined to a small area (dozens of
square meters)
In intense severity zones, the fire had consumed the
underbrush and crown entirely over an extensive area.
Soil erodibility: soil depth and stoniness distribution
within the catchment are controlled by slope inclination,
position within the catena, and aspect.
Thin stony soils were generally found on steep
slopes, convexities, and south-facing slopes, while
deeper less stony soils were found on gentler slopes,
concavities, and north-facing slopes.
Three classes were defined:
- thin and stony,
- of intermediate depth and stoniness, and
- deep with few stones.
A soil corer was used along a dozen slope transects
and thin soils were designated as those that had
depths of < 20 cm.
A soil was considered deep if it was > 50 cm in
depth.
Ascribing coefficients and combining the soil erosion
factors.
Mapping the soil erosion risk was based on the
multiplication of coefficients (equ. 1) ascribed to the soil
erosion factors summarized in Table 1.
Erosion Index = Slope x Vegetation density x Fire
severity x Soil erodibility (1)
For each factor, a raster layer with a 25 m cell size
was created using geographic information software.
Each cell within the layer was assigned a coefficient
(Table 1) and the layers multiplied (equ. 1) to provide a
theoretical range of values of from 1 to 135.
In the absence of a suitable, functional, and readily
available soil erosion model, values for the coefficients
were attributed arbitrarily according to personal
experience and field observations.
Estimating coefficients for slope, vegetation and fire
severity was relatively straightforward.
The soil erodibility coefficients were more
problematic: shallow stony soils generate more runoff
but surface material is less mobile due to the high
stoniness, and in many cases, the bare bedrock.
Deep soils with fewer stones may have higher
infiltration rates, but they are more erodible, particularly
with respect to the weak aggregate stability and
possible presence of hydrophobicity.
Therefore, we decided to increase the value of the
coefficient with decreasing stone content, as can be
seen in Table 1.
116
EUFIRELAB
12.4.3 Partial validation of the soil erosion risk
map
Once the soil erosion risk map was completed, the
catchment was divided into sub-catchments and the
percentage area in the high and very high risk
categories (Table 2) was calculated for each sub-
catchment.
On 2 December, 2003, a rainfall event generated
significant runoff in the catchment: peak rainfall
intensity values were not available but roughly 35 mm
fell in the morning, and about 35 mm fell in the
preceding five days (measured at about five km from
the fire site.
Runoff samples from most of the major streams
were obtained manually by plunging a 0.5 L bottle into
the stream during the storm.
Replicate suspended sediment samples from the
main channel and selected tributaries were collected
within a period of about 10 minutes so rainfall/runoff
conditions were considered comparable.
The amount of suspended sediment transported
through the channel depends on both sediment
concentration and discharge, but the timing imposed by
sampling simultaneously and the difficult field
conditions made it impossible to measure discharge
accurately.
Therefore, we only measured suspended sediment
concentrations.
An order of magnitude based on visual estimates
would give discharges of about 10 m3 s-1 for the main
channel and about 20-30% of that value for the
tributaries.
It was therefore possible to compare the stream
suspended sediment load to the percentage of the
catchment in the high and very high-risk categories.
For comparison, we sampled a stream draining a
vineyard sub-catchment and an unburned forest area.
D-08-06.doc
12.5 RESULTS
12.5.1 Distribution of the soil erosion factors
Figures 1 to 4 show the distribution of the four
factors of soil erosion in the catchment.
More than 30% of the slopes in the study area have
an inclination greater than 20% (Figure 1), showing the
natural vulnerability of the zone to erosion processes.
In addition, many of the steeper slopes were those
most affected by the fire (Figure 2).
Typically, pre-fire vegetation densities (Figure 3)
were low on convexities or south-facing steep slopes
with shallow stony soils (Figure 4).
Conversely, deeper soils with a denser pre-fire
vegetation cover tend to be concentrated in the
southeast corner of the map where slopes are gentle.
12.5.2 The soil erosion risk map
Values resulting from the multiplication of the factors
ranged from 1 to 90, and were subdivided into soil
erosion classes according to breaks between peaks in
the histogram distribution.
Table 2 presents the values for each class, and
Figure 5 shows the results.
Each class tends to correspond to a typical
environment (Table 2.
As stated above, the erosion model used is
extremely simple and the subjective definition of the
coefficients is open to some debate.
But the objective was to provide a quick spatial
representation of general trends and not a quantitative
estimate of soil loss.
So, it is unlikely that a more time-consuming
modelling approach would have significantly altered the
spatial distribution of the erosion classes identified.
The authors had been working in the catchment for
three years prior to the fire, and soil erosion in the
vineyards had been mapped using both the Revised
Universal Soil Loss Equation (RUSLE) and a
regression equation based on measurements carried
out in a reference area.
Neither of these models, designed for agricultural
contexts, could be applied to burned forest sites, and
the data requirements for more complex deterministic
models were unrealistic within the time frame allowed
for the post-fire management proposals.
117
EUFIRELAB
12.5.3 Partial model validation
Significant runoff occurred during the 2 December,
2003, storm (Figs 6, 7, 8. Stream discharge samples (2
replicates) were collected for three of the four main
sub-catchments.
Sampling locations are shown in Figure 5: Tourré
- measured just West of the “Pont de Bois” before
entering the main Giscle channel
- the subcatchment surrounding the “Val de Gilly”
sampled just before entering the Giscle channel
- the main channel sampled at the “Pont Gué”
(sampled upstream of other tributaries. The Périer in
the southern part of the catchment could not be
sampled in similar rainfall/runoff conditions, so it
was not analysed. In addition, samples for the
burned north-facing slope
- draining into the Giscle main channel just South of
the “Pont Gué”; samples of runoff from a vineyard
and an unburned forest were collected for
comparison purposes.
For each of the three sub-catchments sampled and
the burned north-facing slope, stream suspended
sediment concentration was plotted against the sum of
the percentage area in the high and very high-risk
erosion classes (Figure 9).
The numbers in Figure 9 correspond to sampling
locations described in Figure 5. Sediment concentration
values represent the means of 2 samples.
Differences in concentration between samples were
never greater than 10% except for the Gilly, for which it
was 31%.
Although there are only four points, there is a
general trend for the sediment concentration to
increase with increasing area in the High and Very High
erosion classes.
The exception is the Giscle (sample 3) which is
located immediately downstream of a long straight
section with natural sediment deposition.
Therefore, the low suspended sediment
concentration of the Giscle was attributed to natural
deposition in the stream bed upstream of the sampling
area.
Sediment concentrations for the vineyard (5) and
unburned forest were 1.64 g l-1 and 0 g l-1, respectively.
D-08-06.doc
12.6 DISCUSSION
Identifying critical sites for post-fire erosion control
methods depends on an accurate spatial estimation of
high-risk zones.
Several factors contribute to make this task
somewhat more difficult than in a typical agricultural
setting.
These include the following:
- The short intervention time between the forest fire
and the installation of post-fire erosion control
methods. Identifying critical areas must be done
within a period of weeks after the fire for the
methods to be efficient during the first winter rains.
- The lack of spatial data on soil properties that
contribute to enhance or reduce the erosion risk
(soil depth, stoniness, hydrophobicity, aggregate
stability…. Data for soil properties in agricultural
areas tend to be more abundant than for forests
since soil productivity depends on these properties.
Data for forested areas are scarce.
- In the Mediterranean region, forested slopes tend to
be particularly steep since flatter areas are occupied
by agricultural activities, including vineyards, fruit
and olive groves, and cereal crop. After a major
forest fire, slope angle is perhaps the single most
important factor determining erosion rates. DEM
spatial resolutions should be in the order of 25 m to
accurately represent the terrain.
- Pre-fire vegetation density and fire intensity are
spatially variable and along with slope angle they
play a major role in determining soil erosion risk.
The case study presented above describes an
operational strategy to map erosion risk at the scale of
a large catchment.
Work is currently underway to investigate two future
directions of research.
The first is to elaborate a pre-fire strategy that would
shorten the post-fire reaction time.
To what extent can planners prepare for the
consequences of a forest fire before it occurs in order
to respond more quickly to the potential erosion risk?
The second is to determine how the spatial scale of
the data and model affect output.
Soil erosion risk maps have been produced to
identify potential high risk areas at the European scale
for essentially arable land.
The question now asked is “Can a similar approach
be used for post-fire soil erosion risk?”.
118
EUFIRELAB

12.7 CONCLUSIONS In the future, soil erosion modelling should be

accompanied by the development of soil conservation
Major forest fires in the Mediterranean environment
models where high risk sites are evaluated for
occur regularly and might become more frequent in the
suitability to different erosion control methods.
future as the climate changes.
Increasing the efficiency of the approach and the
A method for choosing appropriate sites for the
effects of varying spatial scale are two further research
erosion control measures is based firstly on an
directions that are currently being pursued.
assessment of the spatial distribution of erosion risk.
It should be noted that although the soil erosion risk
map produced in the first stage of the method is useful
in identifying high risk areas, it does not represent in
itself a sufficient tool for developing a soil conservation
strategy.

D-08-06.doc 119
EUFIRELAB

12.8 FIGURES

Figures 1 and 2: Slope classes (left) and Fire Severity (right)

Figures 3 and 4: Estimated pre-fire vegetation density (left) and distribution of soil characteristics (right)

Figure 5: Soil erosion risk map showing the erosion risk categories.

D-08-06.doc 120
EUFIRELAB

Figure 6: Runoff observed in the burnt forest during the 2 December, 2003, storm

Figures 7 and 8: Runoff flowing over a low bridge (Pont Gué) during the 2 December, 2003, storm (left)
and the same bridge (Pont Gué) at normal discharge (right)

5
4
Sediment Concentration (g L )
-1

4 2

2
1

1 3 1 TOURRE
2 GILLY
3 GISCLE
0 4 GULLEY
20 40 60 80
Area (%) in High & Very High Erosion

Figure 9: Relationship between sediment concentrations in storm runoff and proportion of area (%) defined as high
or very high erosion risk (numbers refer to site locations described in Fig. 5)

D-08-06.doc 121
EUFIRELAB

12.9 TABLES

Table 1: Coefficients ascribed to the four soil erosion factors used to estimate erosion risk: slope, pre-fire
vegetation density, fire severity, and soil erodibility
Slope (%) Pre-Fire Fire Severity Soil Erodibility
Vegetation Density
0-5 1 Medium 1 Variable 1 Shallow soil with high stone content 1
5-10 2 Low 2 Intense 3 Intermediate depth and stoniness 2
10-20 3 Bare 3 Deep soil with few stones 3
20-30 4
>30 5

Table 2: A brief description of the typical environments observed in the catchment for a range of erosion risk values
Erosion Area
Class Summary Description
Value km2 (%)
Valley bottoms with gentle slopes, low severity fire, dense pre-fire
Low ≤5 3.3 (16) vegetation and deep soils. Included are also some flat topslope sections
where soils are stonier and shallower.
Convexities and concavities upslope and downslope, respectively, of
major linear slope sections. These areas have gentler slopes than the
Intermediate 6 - 15 8.3 (40)
linear sections and are frequently found along the limits of the pre-fire
vegetation and fire serverity classes.
Steep slopes with bare or low pre-fire vegetation densities in the
severely burned area dominate. One significant exception is the north-
facing slope located south of the main Giscle channel. In this area, pre-
High 16 - 26 6.5 (31)
fire vegetation density and soil depth are greater than elsewhere in the
high erosion risk class, but the steep slopes make it vulnerable to
erosion.
Severe fire conditions in low density pre-fire vegetation (with patches of
Very high ≥ 27 2.7 (13)
bare slope) on very steep slopes.

D-08-06.doc 122