Using Leaf Area Index, retrieved from optical imagery, in the STICS crop model for predicting yield and biomass of field crops

Using remote sensing (RS) data to re-initialize input parameters that are not readily available is considered an effective way to improve regional-scale crop modelling. Our objective is to evaluate the conditions of application of the re-initialization (e.g., number of image acquisitions, spatial resolution of the homogeneous field subunits, etc.). Although several aspects of the coupling between crop models and RS data have been investigated, extensive verification using multiple years is required prior to implementation of the approach at the regional scale, in order to evaluate its robustness and consistency in response to climatic variations. We evaluated the performance of STICS, a functional crop model, in predicting yield and biomass by using Leaf Area Index (LAI) retrieved from RS data to re-initialize selected input parameters over ten growing seasons (1999–2008) for rainfed corn, soybean and spring wheat fields cultivated near Ottawa (Ontario, Canada). The impact of the number of RS images and of the definition of the homogeneous spatial units required to re-initialize input parameters was also evaluated. Green LAI was estimated with the modified transformed vegetation index (MTVI2) using either airborne hyperspectral data (Compact Airborne Spectrographic Imager, CASI) or more readily available multispectral satellite data (Landsat TM and SPOT). Re-initialization of the model parameters involves using the simplex optimization algorithm to minimize the weighted sum-of-squared differences between LAI retrieved from RS and LAI estimated by the crop model. Yield and biomass predictions were greatly improved through the re-initialization of seeding date, seeding density and field capacity. Almost no bias was observed, and the relative root mean square error (RMSE%) was 13% for yield and 23% for biomass (versus 22% and 44% without optimization). The improvement in model predictions was particularly noticeable in the case of water-stress conditions or a deficit of growing degree-days, indicating that the method is sensitive to climate variability. The results were very close to the yield and biomass predictions (i.e., RMSE% of 11% for yield and 17% for biomass) obtained with actual management and soil properties. Most of the improved predictions were associated with re-initialization of the seeding date. When only two LAI values were used to re-initialize the seeding date, the RMSE% values for yield and biomass predictions were 15% and 27%, respectively. Finally, we showed that overlaying field boundaries onto soil texture was sufficient to accurately predict yields. The addition of a third layer, based on LAI-homogeneous zones, did not improve yield predictions because the model was not able to capture some of the small within-field yield variations (<0.5tha⁻¹).