Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
ScienceDirect
Research Paper
article info
Measurements indicative of crop development, such as leaf area index, canopy cover or
Article history: biomass are typically performed only a few times throughout the season at irregular time
Received 23 February 2017 intervals. Furthermore, due to the inherent spatial variability that exists in the field,
Received in revised form combining measurements taken at different locations in the field usually leads to large
21 July 2017 uncertainty around the mean value. These factors, together with the fact that crop-soil
Accepted 2 August 2017 models are strongly non-linear, render assimilation of measurements in crop-soil
models non-trivial. This work presents procedures for performing such data assimila-
tion, using the crop model AquaCrop as specific example. The procedures are based on
Keywords: Extended Kalman Filter, with some heuristic adjustments, and enable re-initialisation of
Crop modelling state variables and/or adjustments of selected parameters of the model. The uncertainties
Data assimilation of the measurements are taken into account explicitly in the proposed assimilation
Extended Kalman Filter scheme. The procedures were tested with data obtained from experiments conducted with
Potato potato in Denmark and cotton in Greece. In both cases the data available consisted of
Cotton canopy cover and biomass (average and standard deviation on 5e10 days), and a locally-
FIGARO calibrated AquaCrop model was used as starting point for the assimilation process. The
results demonstrate the soundness of the approach but also emphasise the inherent
limitations associated with data assimilation. In particular, assimilation of easy-to-obtain
canopy cover measurements did not always improve the predictions of biomass.
© 2017 IAgrE. Published by Elsevier Ltd. All rights reserved.
support systems (DSS) (e.g. Amir & Fisher, 2000; Behera &
1. Introduction Panda, 2016; Epperson, Hook, & Mustafa, 1993; Ioslovich,
Borshchevsky, & Gutman, 2012; Jiang et al., 2016; Linker &
There is an urgent need to increase agricultural production Ioslovich, 2014; Shani, Tsur, & Zemel, 2004). In particular,
while preserving water resources, especially in semi-arid and the Food and Agriculture Organization (FAO) has developed
arid regions. Mathematical models that describe the crop-soil- the crop water productivity simulation model AquaCrop (FAO,
atmosphere interactions are central to the investigation of 2009; Steduto, Hsiao, Raes, & Ferengeset, 2009) as a farm-level
irrigation scenarios and the development of improved irriga- tool for simulating various irrigation scenarios. The main
tion strategies. A large number of such scientific models have characteristics of AquaCrop are its simplicity and robustness,
been developed and used to develop irrigation decision
* Corresponding author.
E-mail address: linkerr@tx.technion.ac.il (R. Linker).
http://dx.doi.org/10.1016/j.biosystemseng.2017.08.003
1537-5110/© 2017 IAgrE. Published by Elsevier Ltd. All rights reserved.
58 b i o s y s t e m s e n g i n e e r i n g 1 6 2 ( 2 0 1 7 ) 5 7 e6 6
together with a reasonable accuracy which has been estab- This work presents a general framework and procedures
lished by a number of studies with various crops and at for performing data assimilation in crop-soil models, using
various locations (see https://www.zotero.org/groups/ the crop model AquaCrop as specific example. The procedures
aquacrop_publications for a complete list of publications are based on Extended Kalman Filter, with some heuristic
involving the AquaCrop model). An executable version of this adjustments, and enable re-initialisation of state variables
model can be downloaded freely from the FAO website (http:// and/or adjustments of selected parameters of the model. The
www.fao.org/nr/water/aquacrop.html). Although AquaCrop uncertainties of the measurements are taken into account
has been successfully calibrated for a large number of crops explicitly in the proposed assimilation scheme. The proced-
and locations, the question remains as how to improve model ures were tested using canopy cover and biomass data ob-
predictions by using measurements performed during the tained from experiments conducted with potato in Denmark
season. and cotton in Greece. In both cases a locally-calibrated
Numerous studies on data assimilation have been con- AquaCrop model was used as starting point for the assimila-
ducted with various crop/soil models, most of them focussing tion process.
on assimilation of measurements obtained via remote sensing
(e.g. Aubert, Loumagne, & Oudin, 2003; Dente, Satalino,
Mattia, & Rinaldi, 2008; de Wit and van Diepen, 2007; Gue rif 2. Materials and methods
& Duke, 2000; Ines, Das, Hansen, & Njoku, 2013; Launay &
Guerif, 2005; Olioso et al., 2005; Prevot et al., 2003; 2.1. Experimental data
Vazifedoust, van Dam, Bastiaanssen & Feddes, 2009). Some
of these studies considered only “static” data assimilation, i.e. Two datasets were used to evaluate the performance of the
re-calibration of some model parameters (or initial conditions) proposed procedures. In both cases the measurements con-
using all the data available. For instance, Gue rif and Duke sisted of canopy cover and biomass at various time
(2000) and later Launay and Guerif (2005) developed proced- throughout the season. Individual measurements were not
ures for assimilating remote sensing data into the SUCROS available but only averages and standard deviations (of an
model coupled with the radiative transfer model SAIL. Four unknown number of measurements) were available, so that it
parameters chosen a priori were re-calibrated using least was not possible to calculate the cross-variance matrices even
square optimisation. A similar approach was considered by when the canopy cover and biomass measurements were
Prevot et al. (2003) who used remote sensing data for cali- performed on the same day. At both locations an AquaCrop
brating five parameters of the STICS wheat model. More model calibrated using data from similar experiments con-
recently, Dente et al. (2008) reported assimilation of leaf area ducted in previous years was used as starting point for the
index (LAI) derived from remote sensing into the CERES wheat data assimilation reported below. Details about the calibration
model. Three of the model's parameters were adjusted in of the AquaCrop models can be found in Battilani et al. (2016,
order to minimise a functional that combined deviations be- pages 22e24 and 45e57).
tween observed and estimated LAIs and the estimated and
nominal values of the parameters, taking into account the 2.1.1. Dataset 1: Potato in Denmark
uncertainties associated with the measured LAIs. More com- The experimental data consisted of measurements performed
plex data assimilation methods based on Kalman filter, and its in parallel on four irrigation treatments conducted at Aarhus
extensions Extended Kalman Filter and Ensemble Kalman University-Jyndevad research station in 2014. Details about
Filter, have also been developed. Vazifedoust et al. (2009) the soil, planting pattern and drip irrigation system have been
presented assimilation of satellite data into agro- given by Linker, Ioslovich, Sylaios, Plauborg, and Battilani
hydrological models. After converting the satellite data into (2016) and Zhou, Andersen, and Plauborg (2016). The irriga-
leaf area index and relative evapotranspiration using a land tion treatments consisted of “full irrigation” in which irriga-
surface energy algorithm, assimilation into the distributed tion was applied every two days to replenish soil water to 90%
agro-hydrological model was done with a constant-gain Kal- of field capacity (treatment If in Zhou et al. (2016)); “deficit
man filter. Ines et al. (2013) used an Ensemble Kalman Filter to irrigation” in which crop received 80% of the amount of water
assimilate remote sensing data in the modified DSSAT-CSM- given to If in the tuber initiation stage and 60% of If during the
Maize model. Overall, regardless of the data assimilation tuber bulking stage (treatment Id in Zhou et al. (2016)); “custom
technique implemented, the majority of these studies irrigation” (treatment IFigaroN3 in Battilani et al. (2016), pages
concluded that remote sensing provides useful measure- 22e24); “no irrigation” (treatment I0 in Zhou et al. (2016)). The
ments which can be used to improve yield predictions. irrigation schedules are given in Table 1. Canopy cover (CC)
Despite their increasing popularity, remote sensing mea- and dry matter biomass were measured every nine and five
surements have some limitations in terms of temporal and days, respectively, simultaneously for all treatments. Canopy
spatial resolution. On the other hand, leaf area index, or cover was estimated from digital images taken 1.5 m above
canopy cover, can easily be estimated from downward- the canopy. The CC estimates were obtained by calculating
looking digital colour images which can be acquired by the excess greenness index image (Chen, Zhang, Su, & Guo,
farmers with minimal training (e.g Weiss, Baret, Smith, 2010) and applying thresholding to create a binary image in
Jonckheere, & Coppin, 2004 (software available at https:// which white pixels corresponded to canopy. The threshold
www6.paca.inra.fr/can-eye/Download), Liu & Pattey, 2010, value (25) was determined by trial-and-error and visual in-
Liu, Pattey, & Admiral, 2013 (software available at http:// spection of several images. Biomass was determined using
www.flintbox.com/public/project/5470/)). eight plants, which were dried for 24 h at 80 C in a forced
b i o s y s t e m s e n g i n e e r i n g 1 6 2 ( 2 0 1 7 ) 5 7 e6 6 59
Such an approach has been used for instance by Speetjens, Robust Index of Agreement
Stigter, and van Straten (2009) in a work related to green- P
house climate modelling. Three parameters of AquaCrop 0; ðMk Ok Þ2 s2k
k max
d2 ¼ 1 P 2 (8)
related to canopy development were selected for testing this
k Mk M þ Ok M
approach: CGC, CCX and CDC. CGC is the canopy growth co-
efficient (relevant to early growth stage), CCX is the maximum This index penalises only modelling errors which exceed
canopy cover fraction, and CDC is the daily decrement of the standard deviation of the corresponding measurements.
canopy cover toward the end of the season (for more detailed We propose this modification of the index of agreement d1 in
information about these parameters, see AquaCrop docu- order to estimate how far beyond the margins defined by the
mentation available at http://www.fao.org/nr/water/ standard deviations of the measurements is the model tra-
aquacrop.html). Since all three parameters are related to jectory. This so-called Robust Index of Agreement is always
canopy cover and not biomass, only the canopy cover mea- larger than d1 and should be close to one when the model fits
surements were used when testing this procedure. The com- well to the measurements.
ponents of the partial derivatives matrix Hk were calculated
numerically at each instance of canopy cover measurement. 2.5. AquaCrop implementation
The (unknown) diagonal components of the matrix Pk were
calculated from the assumption that the corresponding Figure 1 shows the flowchart of the implementation of the
standard deviation of each of the chosen parameters is equal procedures detailed in Sections 2.2 and 2.3. This imple-
to 20% of the current value of corresponding parameter. To mentation was performed in Matlab 2016a using custom
prevent extreme variations of the parameter values, the in- functions that read/wrote the appropriate text files and called
crease or decrease of the value of each parameter at each the EXE AquaCrop file (plug-in Version 5.0).
adjustment was limited to 5%.
2.5.1. Assimilation procedure
2.4. Performance indices One of the features included in AquaCrop Version 5.0 is the
ability to perform “Hot Starts”: When a simulation ends within
The performance of the assimilation process was quantified the growing cycle, all the results are stored in a temporary file
via four indices: (<AquaCropRootFolder>ySIMULyHotStartData.SIM). If the
first day of the subsequent run corresponds to the next day of the
Normalised Root Mean Square Error growing cycle, AquaCrop recognises this situation as a ‘hot start’
and loads the saved results as initial conditions for the next run.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Although this file is currently undocumented, Prof. Raes kindly
u
1u XðMk Ok Þ2 provided us with the format of this file, which enabled us to write
NRMSE ¼ t (5)
M k
N custom procedures for reading and overwriting the entries cor-
responding to the state variables “canopy cover” and “biomass”.
where M denotes the mean value of the measurements, and N
is the number of days on which measurements are available. 2.5.2. Model recalibration procedure
Good agreement between model and measurements corre- As mentioned in Section 2.3, adjustment of the values of the
sponds to low values of this index. parameters required numerical calculation of the partial de-
rivatives matrix Hk. Each partial derivative was approximated
Weighted Least Squares via the average of the change observed for a one-day step after
changing the parameter value by ±10%:
XðMk Ok Þ2
WLS ¼ (6)
s2k CCþ CC0 CC CC0
k DER ¼ average ;
D D
In the weighted least squares, the weights of the squared
residuals are inversely proportional to the corresponding where DER denotes the approximation of the derivative, CC0
variances: measurements with low variance are given higher denotes the canopy cover value with the nominal parameter,
weights and measurements with higher variance are given and CCþ and CC denote the canopy cover values after
lower weights. Good agreement between model and mea- increasing and decreasing the parameter by D%. The changes in
surements correspond to low values of this index. the parameter values were implemented via a custom procedure
which overwrote the corresponding entries in the CRO file con-
Index of Agreement taining the crop parameter values. Note that since the runs
performed for these calculations affected the Hot Start file Hot-
P 2 StartData.SIM, an unaltered version of this file was saved before
k ðMk Ok Þ
d1 ¼ 1 P 2 (7) starting the calculations and restored after their completion.
k Mk M þ Ok M
0.979
0.990
0.982
0.988
0.999
1.000
1.000
0.999
were used only to reset the state variables as described in
d2
Section 2.2. In the second analysis, both data assimilation
0.926
0.927
0.916
0.933
0.994
0.997
0.999
0.995
adjustments, Section 2.3) were performed.
d1
Three scenarios were evaluated:
39.389
12.054
15.376
15.376
- Only canopy cover measurements used
WLS
1.792
0.770
9.670
9.110
- Only biomass measurements used (data assimilation only)
- Both canopy cover and biomass measurements used
NRMSE
0.141
0.134
0.070
0.039
0.103
0.149
0.167
0.106
Simulations with the nominal (locally calibrated) AquaCrop
models, without any data assimilation, were used as reference.
0.979
0.965
0.996
0.998
1.000
1.000
1.000
0.997
d2
3.1.1. Data assimilation
Typical results are shown graphically in Fig. 2. For this specific
0.924
0.860
0.926
0.946
0.995
0.991
0.997
0.993
d1
treatment, assimilation of the canopy cover (CC) measure-
Biomass
ments (red dashed line) led to two major adjustments, on days
43 and 56. It should be noticed that while the first adjustment
12.054
42.118
18.862
18.862
WLS
3.720
8.208
1.271
12.34
led to better prediction of CC on day 49, it led to large over-
estimation on day 56, which was only partially corrected by
the second adjustment. Figure 2b shows that the first adjust-
NRMSE
0.157
0.134
0.099
0.149
0.195
0.131
0.074
0.119
ment led to increased biomass development while the second
adjustment had virtually no influence on biomass.
0.979
0.990
0.982
0.988
1.000
1.000
1.000
1.000
d2
0.926
0.927
0.913
0.933
0.997
0.995
0.999
0.990
d1
Canopy cover
15.327
39.389
18.604
13.217
20.848
WLS
9.110
2.699
0.770
NRMSE
0.147
0.141
0.039
0.170
0.148
0.173
0.083
0.127
The best results according to NRMSE and WLS are indicated in bold.
Table 3 e Results of the assimilation procedure for potato.
0.978
0.965
0.996
0.998
1.000
0.995
1.000
0.991
d2
0.924
0.858
0.926
0.946
0.994
0.981
0.997
0.984
d1
None
12.629
41.986
19.484
12.076
15.173
32.174
WLS
4.856
1.268
NRMSE
0.157
0.134
0.149
0.197
0.101
0.168
0.074
0.162
IFigaroN3
Data used
Treatment
Id
I0
I0
If
Assimilation of biomass measurements (green dash-dotted According to NRMSE and WLS indices, assimilation of CC
line) led to one major adjustment (on day 80), which caused measurements alone improved CC predictions only in two
overestimation of the final biomass. Assimilation of both CC of the four treatments. Biomass predictions were improved
and biomass measurements (black dotted line) led to “super- in three of the treatments.
position” of the previous results: Canopy development According to NRMSE and WLS indices, assimilation of
adjusted according to CC measurements and biomass devel- biomass measurements alone improved biomass pre-
opment adjusted according to biomass measurements. dictions in three of the four treatments.
Similar trends were observed for all four treatments, as can According to NRMSE and WLS indices, assimilation of both
be inferred from Table 3 which summarises the results. The CC and biomass measurements led to the best results in
following observations can be made from Table 3: terms of CC and biomass.
The best results according to NRMSE and WLS are indicated in bold.
Fig. 3 e Results for cotton (left frames: Plot #1, right frames: Plot #2). Symbols and colours as in Fig. 2. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)
64 b i o s y s t e m s e n g i n e e r i n g 1 6 2 ( 2 0 1 7 ) 5 7 e6 6
0.961
0.997
1.000
1.000
the value of the index d1, which reflected the overall good
d2
agreement between model and measurements.
0.928
0.958
0.990
0.994
predictions were always very close (or within) the margins
d1
defined by the standard deviation of the measurements,
even when no data assimilation was performed.
104.79
21.326
24.176
WLS
9.945
3.1.2. Data assimilation and model recalibration
Table 4 summarises the results obtained when using the
NRMSE
0.364
0.151
0.105
0.449
measurement for simultaneous adjustments of the model
states and the parameters related to CC development. It
can be observed that, as in Table 3, the indices d1 and d2 are
not informative. According to the NRMSE index, this pro-
0.885
0.999
1.000
1.000
d2
cedure always led to improved predictions of CC and
biomass. However, this was not always reflected in the
WLS index.
0.831
0.968
0.991
0.994
d1
Biomass
3.2. Results for cotton
22.485
281.40
22.264
WLS
7.770
3.2.1. Data assimilation
The results are shown in Fig. 3. First, it should be noticed that,
despite the fact that the same irrigation treatment was
NRMSE
0.092
0.232
0.232
0.365
applied to both plots, the observed crop development differed
significantly between the two plots. The differences observed
in canopy development can be attributed mainly to differ-
ences in soil composition, soil initial water content and soil
0.962
1.000
1.000
1.000
d2
compaction, which all strongly influenced root growth and
overall crop development. As a result of these measurement
differences, data assimilation led to contrasting results in the
0.928
0.958
0.998
0.993
d1
Canopy cover
22.565
26.650
These adjustments had only a very small impact on biomass
WLS
9.910
0.143
0.105
0.276
0.371
1.000
1.000
0.998
0.993
24.244
23.902
WLS
7.765
biomass.
The summary of the results presented in Table 5 shows
that assimilation of CC and/or biomass measurements did not
NRMSE
0.269
0.376
biomass development.
Results for canopy cover
Plot 1
Plot 2
The best results according to NRMSE and WLS are indicated in bold.