tmp5009 TMP

Improving downscaling: South
America Rainfall
David Mendes
Jos Antonio Marengo
Centro de Cincia do Sistema Terrestre CCST
Instituto Nacional de Pesquisas Espaciais INPE
dmendes@cptec.inpe.br
Taller Iberoamericano de Prediccin Estacional Guayaquil 22-23 de Novembro de 2008
Improving downscaling: South America Rainfall
OBJECTIVE
Data base from observation data (Precipitation) and Model data
(IPCC AR4 models);
Statistical analysis;
Application Artificial Neural Network (ANN) model;
Database of downscaling.
Downscaling
However, the spatial resolution of GCMs remains quite coarse, in the order of 300 x 300 km,
and at scale, the regional and local details of the climate which are influenced by spatial
heterogeneities in the regional physiography are lost. Therefore, there is the need to convert the
GCM outputs into a reliable data set with higher spatial resolution, with daily rainfall and
temperature time series at the scale of the watershed or a region to which the climate impact is
going to be investigated. The methods used to convert GCM outputs into local meteorological
variables required for reliable climate modeling are usually referred to as downscaling
techniques (Hewinton e Crane, 1996).
Two major downscaling approaches namely: dynamical downscaling and statistical
downscaling, are commonly used for climate scenario development at higher resolution.
* Dynamical downscaling generates regional-scale information by developing and using
Regional Climate Model (RCMs) with the coarse GCM data used as boundary conditions;
* Statistical downscaling methods, on the other band, involve developing quantitative
Relationships between large-scale atmospheric variables, the predictors, and local surface
variables, the predictands.
3
Artificial Neural Network (ANN) in Meteorology
Recently, non-linear approaches have been developed (in particular, the

Artificial Neural Network - ANN) and adopted as tools to downscale local and
regional climate variables from large-scale atmospheric circulation variables (e.g.
Crane and Hewitson, 1998; Trigo and Palutikof, 1999).
Reference:
Gardner and Dorling (1998) Review of applications in the atmospheric sciences.
Trigo and Palutikof (1999) Simulation of Temperature for climate change over Portugal.
Sailor et al., (2000) ANN approach to local downscaling of GCMs outputs.
Olsson et al., (2001) Statistical atmospheric downscaling of short-term extreme rainfall.
Bindi and Maselli (2001) ANN technique to topographical and satellite.
Boulanger et al., (2006/2007) Projection of Future climate change in South America.
4
Artificial Neural Network (RNA)

The ANN approach can be viewed as a computer system that is made up of several simple
to the highly interconnected processing elements similar to the neuron architecture of
human brain (McClelland et al., 1986).
Supervised: observed precipitation (CRU);
No-supervised: auto-organization.
In this work, input Nodes is IPCC AR4 models

and observation data.
e.g. by: Moriondo and Bindi (2006).
Multilayer Perceptons
O Multilayer Perceptons - The following diagram illustrates a perceptron network with three
layers:
Input nodes
Hidden nodes
output nodes
Training Multilayer Perceptron Networks

Selecting how many hidden layers to use in the network.
Deciding how many neurons to use in each hidden layer.
Finding a globally optimal solution that avoids local minima.
Converging to an optimal solution in a reasonable period of time.
Validating the neural network to test for overfitting.
Backpropagation or propagation of error

Present a training sample to the neural network.
Compare the network's output to the desired output from that sample. Calculate the error in
each output neuron.
For each neuron, calculate what the output should have been, and a scaling factor, how much
lower or higher the output must be adjusted to match the desired output. This is the local error.
Adjust the weights of each neuron to lower the local error.
Assign "blame" for the local error to neurons at the previous level, giving greater
responsibility to neurons connected by stronger weights.
Repeat from step 3 on the neurons at the previous level, using each one's "blame" as its error.
Method
1 - statistical analysis (simple)
Area
a1
a2
A1
a3
A1
A2
A3
an
aT =
(a + a + a
1
+ ... + a n )
A4
2 - ANN application
Models
WCRP CMIP3 AR4
IPCC AR4 Model
Country
Resolution
Control/scenarios
CCCMA_CGCM3.1
Canad
T47L31
20C3M-A2-A1B-B1
CSIRO-MK3.5
Austrlia
T63L18
20C3M-A2-A1B-B1
MPI-ECHAM5
Alemanha
T63L31
20C3M-A2-A1B-B1
GFDL-CM2.1
EUA
2.5X2.5L24
20C3M-A2-A1B-B1
MIROC_3.2_MEDRES
Japo
T42L20
20C3M-A2-A1B-B1
Period
Control 1961-1990 (20C3M);
Scenarios 2071-2100 (A2 A1B B1).
http://www.pcmdi.llnl.gov/ipcc/model_documentation/ipcc_model_documentation.php
Observation data (Climatic Research Unit)

Global Land Precipitation
An historical monthly precipitation data set for global land areas from 1901 to 2000,
gridded at two different resolutions has been constructed and is available for use in
scientific research.
9
Models
CGCM3.1 (T63)
JFM
15
(a)
JJA
JFM
15
(b)
ECHAM5/MPI-OM
15
(e)
15
-15
-15
-15
-15
-30
-30
-30
-30
-45
-45
-45
-45
-60
-90
-75
-60
-30 -60
-45
-80
-65
-50
-35
-60
-90
-75
CSIRO-MK3.5
JFM
(c)
-45
15
15
-30 -60
-90
(f) 15
(g)
(d)
-15
-30
-45
-75
-60
-45
-30-60
-90
15
(h)
-15
-15
-15
-30
-30
-30
-30
-30
-45
-45
-45
-45
-45
-75
-60
-45
-30 -60
-90
-75
-60
-45
-60
-30 -90
-75
-60
-45
-30
-60
-90
-75
-60
-30
(j)
-15
-60
-90
-45
15
-15
-60
JJA
-75
JJA
(I)
GFDL-CM2.1
JFM
JJA
15
-60
MIROC3.2-MEDRES
JFM
JJA
-45
-30
-60
-90
-75
-60
-45
-30
Observation (CRU)
JJA
JFM
15
(a)
15
-15
-15
-30
-30
-45
-45
-60
-90
-75
-60
-45
-30
-60
-90
Control (1961-1990)
(b)
0
-75
-60
-45
-30
10 12 14 1016
Probability - Region 1
Histogram - Region 1
50
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
30
20
95
80
50
5
1
0,0
2,5
5,0
7,5
10,0
Precipitation (mm/day)
12,5
0,01
15,0
Region 1
gfdl echam5 csiro cgcm3
99
20
10
precis miroc
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
99,99
Percent
Frequency
40
0
-2,5
99,9999
15
12
9
6
3
0
15
12
9
6
3
0
15
12
9
6
3
0
15
12
9
6
3
0
6
9
12
15
Annual - Region 1
1961-1990
A1
15
12
9
6
3
0
15
12
9
6
3
0
12
15
obs
11

99,9999
50
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
30
99
95
Percent
Frequency
40
20

precis miroc
50
5
1
-3
3
6
9
12
0,01
15
Region 2
15
12
9
6
3
0
15
12
9
6
3
0
12
9
6
3
0
12
9
6
3
0
15
12
9
6
3
0
80
20
10
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
99,99
6
9
12
15
Annual - Region 2
1961-1990
A2
15
12
9
6
3
0
12
15
obs
12
70
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
60
40
30
80
50
-6
-3
0
3
6
9
12
0,01
15
Region 3
95
10
precis miroc
99
20
20
15
12
9
6
3
0
15
12
9
6
3
0
10
15
20
Annual - Region 3
1961-1990
15
12
9
6
3
0
15
12
9
6
3
0
15
12
9
6
3
0
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
99,99
Percent
Frequency
50
99,9999
A3
15
12
9
6
3
0
12
15
obs
13
50
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
30
20
95
80
50
5
1
-2,5
0,0
2,5
5,0
7,5
10,0
12,5
0,01
15,0
Region 4
precis miroc
99
20
10
15
12
9
6
3
0
15
12
9
6
3
0
6
8
10
12
14
Annual - Region 4
1961-1990
15
12
9
6
3
0
15
12
9
6
3
0
15
12
9
6
3
0
Variable
cgcm3
csiro
echam5
gfdl
miroc
precis
obs
99,99
Percent
Frequency
40
99,9999
15
12
9
6
3
0
A4
0
12
15
obs
14
ANN spatial
Output
15
ANN control
Region 1 - JFM
ANN; RCM vs OBS

JFM
10
10
Variable
ANN
RCM
7
ANN-RCM
8
6
5
4
3
Variable
OBS
ANN
RCM
Obs CRU
ANN output
RCM - precis
1
0
5
OBS
10
ANN; RCM vs OBS

JJA
18
27
36
45
54
Validation months
63
72
81
90
Region 1 - JJA
14
14
Variable
ANN
RCM
13
11
10
9
8
7
6
5
4
3
Variable
OBS
ANN
RCM
12
12
ANN-RCM
10
A1
8
6
4
2
2
1
0
Region 1
7
OBS
10
11
12
13
14
18
27
36
45
54
Validation months
63
72
81
90
Regional Climate Model PRECIS (HadRM3P)- UK Met Office Hadley Centre

16
ANN; RCM vs OBS

JFM
Region 2 - JFM
10
10
Variable
ANN
RCM
7
ANN-RCM
8
6
5
4
3
Variable
OBS
ANN
RCM
Obs CRU
ANN output
RCM - precis
1
0
5
OBS
10
18
27
36
45
54
Validation months
63
72
81
90
Region 2 - JFM
ANN; RCM vs OBS

JJA
10
Variable
ANN
RCM
7
6
5
4
3
2
Variable
OBS
ANN
RCM
5
ANN-RCM
A2
3
2
1
1
0
5
OBS
10
18
27
36
45
54
Validation months
63
72
81
90
Region 2
17
ANN; RCM vs OBS

JFM
Region 3 - JFM
12
12
Variable
ANN
RCM
11
9
8
ANN-RCM
10
10
7
6
5
4
3
Variable
OBS
ANN
RCM
Obs CRU
ANN output
RCM - precis
8
6
4
2
2
1
0
6
OBS
10
11
12
ANN; RCM vs OBS

JJA
18
27
36
45
54
Validation months
63
72
81
90
Region 3 - JJA
10
10
Variable
ANN
RCM
7
6
5
4
3
2
Variable
OBS
ANN
RCM
8
ANN-RCM
A3
1
0
5
OBS
10
18
27
36
45
54
Validation months
63
72
81
90
Region 3
18
ANN; RCM vs OBS

JFM
Region 4 - JFM
10
10
Variable
ANN
RCM
7
ANN-RCM
8
6
5
4
3
Variable
OBS
ANN
RCM
Obs CRU
ANN output
RCM - precis
1
0
5
OBS
10
ANN-RCM
7
6
5
4
3
90
10
Variable
OBS
ANN
RCM
A4
4
3
1
6
81
1
5
OBS
72
63
36
45
54
Validation months
8
27
Variable
ANN
RCM
18
10
10
Region 4 - JJA
ANN; RCM vs OBS

JJA
18
27
36
45
54
Validation months
63
72
81
90
Region 4
19
ERRO
JJA
JFM
95% CI for the Mean
7,5
7,5
5,0
5,0
ERRO (mm/day)
ERRO( mm/day)
95% CI for the Mean
2,5
0,0
-2,5
2,5
0,0
-2,5
-5,0
-5,0
-7,5
-7,5
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
EA EA E A EA RE A REA REA RE A REA RE A REA REA REA REA REA RE A REA RE A REA REA REA REA RE A REA
AR AR AR AR
A A A A
A A A A
A A A A
A A A A
A A A A
c
3
dl
is
iro
m5
ro
gf
ec
cm
cs
mi
ha
pr
c
cg
e
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
AR AR AR AR
A A A A
A A A A
A A A A
A A A A
A A A A
c
3
dl
is
iro
m5
ro
gf
ec
cm
cs
mi
ha
pr
c
cg
e
ANUAL
95% CI for the Mean

2,5
2,0
Mean absolute error
ERRO (mm/day)
1,5
1,0
0,5
0,0
-0,5
-1,0
-1,5
-2,0
SD
-2,5
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
1 2 3 4
A R A R AR A R
A A A A
A A A A
A A A A
A A A A
A A A A
c
3
dl
is
iro
m5
ro
gf
ec
cm
cs
mi
ha
pr
cg
ec
20
PONTUAL
Manaus
Models label
Resolution
Country
CGCM3.1 (T47)
96X48 L32 T47
Canada
CSIRO-MK3.5
192X96 L18
Australia
ECHAM5/MPI-OM
192X96 L31 T63
Germany
GFDL-CM2.1
144X90 L24
USA
MIROC3.2-MEDRES
128X64 L20T42
Japan
Table 1 List of IPCC global coupled climate models analyzed in the present study.
Mean
Standard deviation
JFM
Observation
Multi-models
08.15
06.77 *
09.39
08.39
JJA
Observation
Multi-models
02.81
02.33
04.28
03.87 *
Table 2 Means and standard deviations of precipitation at Manaus, computed from 20 yr of observations and multi-models
21
data for 1970-1989. *Significant differences at the 5%.
Linear Network
Non-linear
10yr calibration
10yr validation
10yr calibration
10yr validation
Validation
Simple
Table 3 Main characteristics for the one types of validation used simple testing.
JFM
SDoutput
RMSE
Error
Linear
6.29
1.21
1.06
1+1
6.30
1.23
1.10
2+1
6.30
1.22
1.08
3+1
6.30
1.22
1.07
4+1
6.30
1.22
1.07
JJA
Linear
3.01
0.58
0.57
1+1
3.06
0.63
0.58
2+1
3.05
0.62
0.59
3+1
3.06
0.62
0.59
4+1
3.05
0.61
0.59
Table 4 Values of explained Standard deviation (SD) for target and output, root mean square error (RMSE) and absolute
error with the linearANN models for JFM and JJA. The non-linear ANNs models are described with the k+1.
22
* Is difference between downscaling and observation (1970-89). ** Is difference between downscaling and multi model (1970-89).
JFM
Observation
Downscaling
Dif *
Dif **
Mean
8.15
8.01
-0.14
1.24
SD
9.39
9.49
0.10
1.26
JJA
Mean
2.81
2.63
-0.18
0.30
SD
4.28
4.21
-0.07
0.34
Table 5 Seasonal statistical of precipitation for observation (1970-89) and downscaling.
Corr. Obs vs Downs = 98.3% - JFM

Corr. Obs vs Downs = 93.6% - JJA
23
JFM - Manaus
Probability
JFM - Manaus - Linear
120
99,99
Variable
Observ ation
Multi-models
100
95
Percent
Multi-models
99
80
60
40
80
50
20
5
20
1
0
20
40
60
80
100
120
Observation
(a)
10
20
Data
30
40
(a)
JJA - Manaus
Probability
JFM - Manaus - Non-linear - 1+1
60
99,99
Variable
Observ ation
Multi-models
50
95
Percent
Multi-models
99
40
30
20
80
50
20
5
10
1
0
10
20
30
40
50
60
Observation
(b)
Figure 2 Quantiles plots of observed and

multi-models from the nearest grid point over
Manaus (1970-1989) for precipitation in JFM
(a) and JJA (b).
10
20
Data
30
40
(b)
Figure 3 Comparison between the modeled

and observed probability frequency for the
linear (a-c) and non-linear [(a-d) 1+1]
models.
24

JFM
40
Variable
Observ ed
Downscaled
30
20
10
36
72
108
144
180
216
252
288
324
360
Julian day
1+1 - JFM
(a)
Figure 4- Twenty daily means of downscaled (multi-models)

and observed output from nearest grid point to Manaus.
40
Variable
Observ ed
Downscaled
30
20
10
36
72
108
144
180
216
252
288
324
360
Julian days
(b)
25
CONCLUSION
In conclusion, when applied to precipitation, the ANN approach,
using a MLP scheme for optimization, makes it possible to
compute the optimal set of weights for a linear combination of
the models (using in this paper), and a penalizing function or
probability that such a change occurred, based on the presentclimate model biases and their projection dispersion.
Before concluding, it is important to highlight that climate
models present significant errors in simulating the patterns and
amplitude of present-day climate conditions;
26
Future
Application ANN for scenarios (A2-A1B-B1) (2071-2100);

Application other method of statistical downscaling;
(e.g. Statistical Downscaling model - SDSM);
GRACIAS PELA ATENCION
27

tmp5009 TMP

Caricato da

Informazioni sul documento

Descrizione originale:

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

tmp5009 TMP

Caricato da

Copyright:

Formati disponibili

Improving downscaling: South

Taller Iberoamericano de Prediccin Estacional Guayaquil 22-23 de Novembro de 2008

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

Artificial Neural Network (ANN) in Meteorology

Recently, non-linear approaches have been developed (in particular, the

Improving downscaling: South America Rainfall

Artificial Neural Network (RNA)

In this work, input Nodes is IPCC AR4 models

e.g. by: Moriondo and Bindi (2006).

Improving downscaling: South America Rainfall

Training Multilayer Perceptron Networks

Improving downscaling: South America Rainfall

Backpropagation or propagation of error

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

Observation data (Climatic Research Unit)

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

gfdl echam5 csiro cgcm3

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

gfdl echam5 csiro cgcm3

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

ANN; RCM vs OBS

ANN; RCM vs OBS

Regional Climate Model PRECIS (HadRM3P)- UK Met Office Hadley Centre

Improving downscaling: South America Rainfall

ANN; RCM vs OBS

ANN; RCM vs OBS

Improving downscaling: South America Rainfall

ANN; RCM vs OBS

ANN; RCM vs OBS

Improving downscaling: South America Rainfall

ANN; RCM vs OBS

ANN; RCM vs OBS

Improving downscaling: South America Rainfall

95% CI for the Mean

95% CI for the Mean

95% CI for the Mean

Mean absolute error

Improving downscaling: South America Rainfall

96X48 L32 T47

192X96 L31 T63

Improving downscaling: South America Rainfall

Improving downscaling: South America Rainfall

Table 5 Seasonal statistical of precipitation for observation (1970-89) and downscaling.

Corr. Obs vs Downs = 98.3% - JFM

JFM - Manaus - Linear

JFM - Manaus - Non-linear - 1+1

Figure 2 Quantiles plots of observed and

Figure 3 Comparison between the modeled

Improving downscaling: South America Rainfall

Figure 4- Twenty daily means of downscaled (multi-models)

Application ANN for scenarios (A2-A1B-B1) (2071-2100);

GRACIAS PELA ATENCION

Potrebbero piacerti anche