Embedding the Spatial Heterogeneity of the Agricultural Landscape into Crop Model Parameters to Refine Sub-Regional Yield Predictions

Assimilating remote sensing data into crop models has unrivaled potential to enhance the accuracy of yield forecasting systems. The main benefit is the increased adherence of the simulations to observed vegetation dynamics related to farmers management and environmental drivers. Here we propose a new methodology to perform sub-regional yield forecasting, grounded on the derivation of leaf area index (LAI) from remote sensing as input for crop model calibration. The study area comprised 23 NUTS-3 administrative units located in three Italian regions, i.e., Apulia (south), Tuscany (centre) and Veneto (north). Crop maps of barley and maize in 2018 and 2019 were used to derive time series of Normalized Difference Vegetation Index (NDVI), which were tailored in the growing season according to agro-meteorological rules. Multivariate analyses of NDVI allowed identifying local agrotypes and associated LAI time series, which were used to develop parameter sets of the WOFOST_GT model via automatic calibration. Improvements in explaining the variability of official yields were evaluated in hindcast mode (2007-2017) using model parameterizations from literature as the baseline. Crop simulations reflected the latitudinal phenological gradient, leading to the anticipation of 15-30 days of barley and maize flowering and maturing dates from Veneto to Apulia. Simulated LAI dynamics reproduced observations with modelling efficiency always higher than 0.35 in calibration, and average Relative Root Mean Square Error (RRMSE) in explaining yield variability was in the range 15-17% for barley and 5-17% for maize in hindcast mode. The new procedure reduced by 51% and 14% the average RRMSE in yield prediction with respect to the baseline, decreasing by more than 1 Mg ha -1 the absolute errors on both crops. This methodology is scalable to other crops and regions, and the released parameter sets are ready to be used for operational yield forecasting activities