Linking Crop Simulation Models To Rs Inputs

The use of remotely sensed information to improve crop model accuracy was proposed as early as two decades ago by Wiegand et al. (1979) and Richardson et al. (1982). They suggested using spectrally derived LAI either as direct input to physiological crop model or as an independent check to model calculation for its re-initialization. The main advantage of using remotely sensed information is that it provides a quantification of the actual state of crop for large area using less labour and material intensive methods than in situ sampling. While crop models provide a continuous estimate of growth over time, remote sensing provides a multispectral assessment of instantaneous crop condition with in a given area (Delecolle et al., 1992).

The different ways to combine a crop model with remote sensing observations (radiometric or satellite data) were initially described by Maas (1988a) and this classification scheme was revised by Delecolle et al. (1992) and by Moulin et al. (1998). Five methods of remote sensing data integration into the models have been identified:

(a) the direct use of a driving variable estimated from RS data in the model;

(b) the updating of a state variable of the model (e.g., LAI) derived from RS

('forcing' strategy);

(c) the re-initialization of the model, i.e., the adjustment of an initial condition to obtain a simulation in agreement with the RS derived observations;

(d) the re-calibration of the model, i.e., the adjustment of model parameters to obtain a simulation in agreement with the remotely-sensed derived observations, also called 're-parameterization' strategy;

(e) the corrective method, i.e., a relationship is developed between error in some intermediate variable as estimated from remotely sensed measurement and error in final yield. This relationship may be applied to a case in which final yield is not known.

Direct use of driving variable

The driving variables of crop simulation models are weather inputs comprising daily observations of maximum and minimum temperature, solar radiation, relative humidity, and wind speed as a minimal subset. In a recent review on this subject, Moulin et al. (1998) cited inadequate availability of RS-derived parameters, due to cloud cover problem and intrinsic properties of sensors and platforms, as a major drawback, for adoption of this approach. However, this is a promising area, given the sparse distribution of weather observational network and recent progress in deriving some of these variables from sensors in space. Rainfall, solar radiation and intercepted/absorbed PAR have received maximum attention.

Maas (1988a) estimated the ratio of daily absorbed PAR (Q) to integrated daily PAR (R) from radiometric NDVI and generated daily values of Q/R by linear interpolation between NDVI measurements for use as driving variable in a simplified maize growth model. The model showed an overestimation of 6.2% in above ground biomass at anthesis. METEOSAT based decadal (10-day) rainfall using cold cloud duration has been used as input to CERES-Millet in Burkina Faso by Thornton etal. (1997b) to forecast provincial millet yields halfway through crop duration to within 15% of their final values.

Forcing strategy

The forcing strategy consists of updating at least one state variable in the model using remote sensing data. LAI has been the most commonly updated state variable. The concept of a simple crop simulation model and its modification for RS-derived LAI forcing is illustrated in Figure 4. Some examples of forcing spectrally derived LAI in crop simulation models are summarized in Table-2. The forcing could either be done only on day of RS observation (Maas, 1988a) or daily LAI profile is generated using some simple parametric model for use (Delecolle and Guerif, 1988).




Figure 4: (a) Simple schematic of a crop simulation model. (b). Modified structure of crop simulation model with RS-based LAI forcing (Delecolle and Guerif, 1988)
Table 2. Selected case studies on use of RS-derived LAI for forcing crop simulation models



LAI estimation & interfacing

Evaluation of performance




Ground NDVI-LAI on obs. Dates

AGDM estimation improved



relation: Daily interpolated






Biomass at harvest


[1] Maas, 1988a; [2] Delecolle and Guerif, 1988; [3] Bouman, 1995

[1] Maas, 1988a; [2] Delecolle and Guerif, 1988; [3] Bouman, 1995

Re-initialization strategy

The re-initialization method takes advantage of the dependence of model performance on state variable initial condition. It involves adjustment of initial condition of state variable so as to minimize the difference between a derived state variable or the radiometric signal and its simulation. Maas (1988a) in his simplified maize model adjusted the initial value of LAI (L0) at emergence based on the minimization of an error function between remotely sensed LAI values and simulated LAI values during the course of simulation. Reinitialization using one observation produced results similar to updating (forcing). However, the stability of model estimates obtained through reinitialization increased as more observations were used. The observation at 51 days after emergence, which caused a 42% error using updating, resulted in less than a 3% error using re-initialization. Maas (1988b) demonstrated a similar study for sorghum using satellite observations. The simulation model was developed and verified using 10 fields in Central Texas in 1976 and the re-initialization approach was validated for the 37 fields in South Texas using Landsat MSS data and agronomic observations. Without using the initialization procedure, the average yield for the 37 fields was underestimated by 30%. Use of satellite derived green LAI data to initialize the same simulations resulted in a 2% overestimation of average yield.

Re-calibration/re-parameterization strategy

In this approach it is assumed that model is formally adequate but requires re-calibration. This is achieved by minimizing error between RS-derived state variable and its simulation by the model. This makes such an approach sensitive to errors in deriving state variables from RS data. In this case also, the state variable matched is LAI. However, depending on the model structure, which parameter to tune and number of observations used in analysis is critical.

Maas (1988a) demonstrated the re-calibration for maize model with remotely sensed GLAI observations. A multi-dimensional error function minimization procedure was used which indicated more consistent estimates of LAI and biomass at anthesis as the number of parameters increased in multidimensional re-parameterization.

Delecolle et al. (1992) illustrated the use of re-calibration for rice crop using GRAMI model. Values of one to four parameters in the GRAMI model were re-calibrated to match the simulated LAI profile to observed LAI values. The results showed that improvement in simulated LAI profile by re-calibration depends largely on the number and timing of LAI observations. Clevers and Leeuwen (1996) used ground and airborne radiometric measurements over sugar beet fields to calibrate the SUCROS model. They derived LAI from measurements in optical and microwave wavebands. The adjusted parameters and initial conditions were sowing date, a growth rate, light use efficiency and maximum leaf area. The results showed that re-calibration with both optical and microwave observations estimated yield better than optical data alone. In the absence of optical remote sensing data, radar data yielded a significant improvement in yield estimation with the case of no remotely observed observations. Inoue et al. (1998) related paddy vegetation indices to the fraction of absorbed photosynthetically active radiation (fAPAR) as exponential equations with different parameters for the periods before and after heading. A real time recalibration module based on a simplex algorithm was developed and proved effective in linking remotely sensed fAPAR with a simple rice growth model.

Re-parameterization using Coupled Crop Simulation Models and Canopy-radiation Models

The re-initialization and re-parameterization of crop models can also make direct use of radiometric information instead of deriving canopy parameters from them (Moulin et al., 1998). In this strategy, coupling a radiative transfer reflectance model to the crop production model reproduces the temporal behaviour of canopy surface reflectance, which can be compared with canopy reflectance observed from satellite. Adjusting initial conditions or model parameters carries out the minimization of differences between the simulated and observed reflectance values.

Such an approach has been used by Clevers et al. (1994) and Guerif and Duke (1998) for sugarbeet. The LAI simulated by SUCROS on dates of spectral observations was passed on to PROSPECT-SAIL and SAIL model, respectively, and parameters of SUCROS model adjusted to minimize differences between observed WDVI and simulated WDVI. The re-parameterization of SUCROS reduced yield prediction errors. This approach was extended to microwave RS by Bouman et al. (1999) who simulated radar backscatter of agricultural crops (sugar beet, potato, winter wheat) by integrating LAI and leaf moisture from SUCROS with top soil moisture content by soil water balance model (SAHEL) and radar backscatter model (CLOUD).

Since canopy-radiation models such as SAIL have parameters in addition to LAI, their uncertainty could affect the results from this approach. Moulin and Guerif (1999) concluded that the error in canopy reflectance estimates as a result of omitting data on canopy leaf angle and soil reflectance, two parameters in SAIL model, is so large that direct use of simulated canopy reflectance in simulation models for yield prediction is severely affected. However, the use of vegetation indices drastically reduced the errors linked to crop structure (NDVI) and to soil reflectance (TSAVI).

Corrective approach

Sehgal et al. (2001b) used this strategy for generating the wheat yield maps for farmers' fields during rabi 1998-99 in Alipur block (Delhi). The RS inputs as estimated LAI were linked to wheat simulation model WTGROWS for yield mapping and results were validated with yield observations on farmers' fields. Biometric relation of grain yield and leaf area index (LAI) is derived from simulation model by running model for a combination of input resources, management practices and soil types occurring in the area. Then this biometric relationship is applied to all the crop fields of the study area for which the LAI is computed from remote sensing data. The WTGROWS simulated grain yield for the combination of inputs showed yields varying between 1.1 and 4.9 t ha-1. The corresponding range of simulated LAI on 27 Jan 1999 was 0.6 to 4.2. The regression equation fitted between simulated LAI on 27th Julian day (i.e. 27-January-99) and simulated grain yield showed saturating logarithmic nature with a R2 value of 0.81. The relationship is given below:

This empirical biometric relation was applied to the LAI map of the wheat pixels and grain yield map for farmer's fields of Alipur block, Delhi, was generated. The predicted yields ranged from 2.1 to 4.8 tha-1. The comparison of predicted grain yield and observed yield for the 22 farmers' fields showed high correlation coefficient of 0.8 and a root mean square error (RMSE) of 597 kg ha-1 which was 17 per cent of the observed mean yield (Figure 5).

Figure 5: Comparison of predicted grain yield by modified corrective approach and observed values for 22 farmers' fields. The 1:1 line and its ±15 per cent band lines are also shown (Sehgal etal., 2001b)

Development of a RS-based CGMS for wheat in India

Sehgal et al. (2001a) reported the development of a prototype Crop Growth Monitoring System (CGMS) for wheat using WTGROWS simulation model on a 5'x5' grid in GIS environment for generating daily crop growth maps and predicting district-wise grain yield. The inputs used were RS based wheat distribution map, daily weather surfaces, soil properties map and crop management input databases in a GIS environment and analysis for wheat season of 1996-97 was carried out. The inputs, their processing in GIS and framework for CGMS are summarized in Figure 6. The grid-wise final simulated grain yields are shown in Figure 7. The figure clearly indicates spatial patterns in yield variability. The high grain yields in Kurukshetra and Karnal and low yields in Bhiwani, Rohtak, Yamunanagar and Ambala are brought out clearly. The comparison of simulated grain yields aggregated at district level and estimates by the State Department of Agriculture is shown in Figure 8. In general, the model simulated yields were higher than observed. This could be due to a number of yield reducing factors such as pest, weed, soil constraints, which operate in field but are not considered by the model. The model predicted yields were within ±10% of reported yields in 12 out 16 districts. The RMSE of 335.4 kg ha , which is less than 10 per cent of the State mean yield, was obtained. Only in two districts, Mahendragarh and Bhiwani, the simulated district yields were lower than observed yields while for Kaithal, Karnal, Ambala and Yamunanagar, the simulated yields were higher than observed yields by more than 10 per cent.

-----[INPUTS]------ --------[GIS]--------- [CENTRAL RDBMS] [SIMULATION

-----[INPUTS]------ --------[GIS]--------- [CENTRAL RDBMS] [SIMULATION

PTF —Pedo Transfer Functions

* Selection by SHELL as per run parameters


PTF —Pedo Transfer Functions

* Selection by SHELL as per run parameters

Figure 6. Schematic diagram of a crop growth monitoring system showing the linkages between inputs, spatial layers in GIS, and relational database to WTGROWS simulation model (Sehgal etal., 2001a)


74°30'E 75°E 75°30'E 76° 76°30'E 77°E 77°30'E

74°30'E 75°E 75°30'E 76° 76°30'E 77°E 77°30'E

3.51 3.76 4.0 4.25 4.50 4.75


Figure 7: Grid-wise simulated wheat yields by WTGROWS simulation model for 1996-97 season in Haryana (Sehgal et al., 2001a)

Sehgal et al. (2002) demonstrated a technique for estimating date of sowing (DOS) using RS-derived spectral-temporal crop growth profiles and CGMS simulation capability and evaluated the capability of CGMS for spatial yield mapping and district level yield prediction for Haryana State during 2000-01 crop season. The technique for estimating district-wise DOS matched the RS-derived date of peak NDVI (from multi-date WiFS sensor aboard IRS-1D satellite) to date of peak LAI simulated in CGMS for a range of plausible dates of sowing (Figure 9). The peak date of NDVI was computed by fitting Badhwar model to the multi-date NDVI values. The CGMS performance was evaluated by incorporating RS-derived date of sowing in predicting district level wheat yields with and without use of district-wise N fertilizer application


3000 3500 4000 4500 5000

Observed Yield (kg/ha)

3000 3500 4000 4500 5000

Observed Yield (kg/ha)

Figure 8: Comparison of simulated district wheat yields by CGMS with observed values reported by the State Department of Agriculture, Haryana , for 1996-97 season. The 1:1 line and its ±10 per cent band lines are also shown (Sehgal et al., 2001a)

rate computed from district-wise fertilizer consumption statistics. The correlation between district yield simulated by CGMS and official State Department of Agriculture (SDA) estimates was only 0.163 when constant median/mean inputs of DOS, N fertilizer and irrigation application were specified for all the districts. The correlation increased to 0.52 when RS-CGMS-derived district-wise DOS was used as input and further increased to 0.74 when information from consumption statistics of N fertilizer use was additionally specified.

It is clear from the above studies that the potential of integrating crop simulation model, RS inputs and GIS has been well proven in a number of case studies. While techniques for geophysical and crop biophysical parameter retrieval are becoming available and producing products of required accuracy, the available crop simulation models need to be provided with GIS integration and iterative run options to benefit from this integration.




7 / /

* /

' i f




s s


y /


04 f


12 g

DAY Of VEAfl(Z001)

Figure 9: An illustration for obtaining date of sowing (DOS) for Karnal district by matching date of simulated LAI peak with date of fitted NDVI profile peak (Tmax). Different simulated LAI curves correspond to various dates of sowing (320 to 350) in year 2000 (Sehgal et al, 2002)


Remote sensing data provide a complete and spatially dense observation of crop growth. This complements the information on daily weather parameters that influence crop growth. RS-crop simulation model linkage is a convenient vehicle to capture our understanding of crop management and weather with GIS providing a framework to process the diverse geographically linked data. Currently RS data can regularly provide information on regional crop distribution, crop phenology and leaf area index. This can be coupled to crop simulation models in a number of ways. These include, (a) direct use of RS inputs as forcing variable, (b) re-initializing or re-calibrating CSM so that its outputs of LAI match RS observations, and (c) using simulation model to estimate impact of variation in a state variable (e.g. LAI) and final yield and using CSM-RS differences to modeling yield predictions. These approaches have been demonstrated through case studies on wheat in India at different spatial scales (village, grid and district). CSM-RS linkage has a number of applications in regional crop forecasting, agro-ecological zonation, crop suitability and yield gap analysis and in precision agriculture.

In future the RS-CSM linkage will be broadened due to improvements in sensor capabilities (spatial resolution, hyper-spectral data) as well as retrieval of additional crop parameters like chlorophyll, leaf N and canopy water status. Thermal remote sensing can provide canopy temperatures and microwave data, the soil moisture. The improved characterization of crop and its growing environment would provide additional ways to modulate crop simulation towards capturing the spatial and temporal dimensions of crop growth variability.

0 0

Post a comment