Improving crop yield prediction accuracy by embedding phenological heterogeneity into model parameter sets

The assimilation of Remote Sensing (RS) data into crop models improves the accuracy of yield predictions by considering crop growth dynamics and their spatial heterogeneity due to the different management practices and environmental conditions. This study proposes a new method for performing sub-regional yield predictions (Nomenclature of territorial units for statistics, NUTS-3 level) using RS time series data and crop models. The primary objective was to release a procedure whereby the heterogeneity of the agricultural landscape (i. e. the agrophenotypes) observed from RS is used to drive crop model calibration. Fine-resolution barley and maize distribution maps (100 m) and related crop calendars have been used to derive the “where” and “when” of the 8-day MODIS NDVI time series data, located in Apulia, Tuscany, and Veneto in 2018–2019. Principal Component Analysis and Hierarchical Clustering have been applied to NDVI seasonal profiles to identify the agrophenotypes, which have been used to derive crop growth and leaf area index dynamics. These data served as the reference to optimize the most relevant parameters of the WOFOSTGT model, using gridded weather data as input (0. 25° resolution, Copernicus ERA5). Parameter distributions from multiple automatic calibrations have been sampled to characterize the agricultural heterogeneity within 22 NUTS-3 administrative units. Yield statistics from the Italian National Institute of Statistics have been used as reference data to test the accuracy of the yield simulation by the crop model WOFOSTGT. The agrophenotypes reflected the wide north-south latitudinal gradient experienced in Italy by the two crops, leading to a gap of 15–30 days in barley and maize flowering and harvest dates across the study area. Average Nast-Sutcliffe modelling efficiency in reproducing LAI dynamics from RS (0. 6) and Relative Root Mean Square Error (RRMSE) in predicting yield data (12. 1% for barley, 3. 7% for maize) in hindcast simulations demonstrated the effectiveness of our approach. The average RRMSE was reduced by 32. 7% (barley) and 8. 5% (maize) compared to baseline, decreasing by 0. 7–1 Mg ha−1 absolute yield errors on both crops. The inclusion of local agrophenotypes in the yield prediction workflow reduced errors in yield prediction compared to unsupervised simulations at NUTS-3 level. The script for agrophenotypes extraction and the model parameter sets are released to the scientific community, to foster improvements and further applications to other crops, ecoclimatic regions, satellite sensors and spatial scales.