Assimilation of Earth Observation leaf area index into the STICS crop model using a formal Markov Chain Monte Carlo algorithm

English Abstract:

Assimilation of Earth observation (EO) derived crop biophysical descriptors into crop models is proven efficient to predict yield in complex agricultural ecosystems. However, the performance of crop models is affected by the accuracy of key input parameters related to soil properties and management practices, which often are not readily available at the regional scale. In this case, data assimilation could improve model performance to predict crop yield by re-initializing key model parameters. The objective of this study is to evaluate the performance of two optimization algorithms to re-initialize seeding date and density, and soil moisture at field capacity, all of which affect the yield predictions of the STICS crop model. Simulations were performed over a small region near Ottawa (ON, Canada) where soybean, corn, and spring wheat were grown. Images from multispectral satellite data (Landsat TM, SPOT and Formosat-2) and compact airborne spectrographic imager (CASI) were acquired over several growing seasons. The LAI was calculated with the modified transformed vegetation index (MTVI2). The two optimization techniques evaluated in this study were 1) the Nelder-Mead Simplex method and 2) an adaptive formal Markov Chain Monte Carlo named Differential Evolution Adaptive Metropolis (DREAM). The DREAM algorithm is expected to offer great advantages compared to the Simplex method for regional-scale predictions because: 1) its outcome is expected to be less dependent on the initial values of the optimized parameters, 2) it provides correlation coefficients between optimized model parameters, and 3) it provides uncertainties associated with model parameters and predictions. This was shown for hydrological simulations but not for yield predictions. Our preliminary results indicated that the DREAM method lead to a root mean square error of 0.7 t ha^-1 regardless of how the initial values for the parameters were assigned, while the Simplex method showed a high level of sensitivity (root mean square error ranging from 0.5 to 2.3 t ha^-1) to initial values when they were taken from properly assigned pre-defined ranges.