Pixel-scale maize yield prediction integrating process-based crop modelling, geostatistics, remote sensing, and machine learning

The reliability of pixel-scale crop yield models depends on how they are evaluated, not only on their apparent accuracy. This thesis investigates whether outputs of the EPIC (Environmental Policy Integrated Climate) process-based crop model provide predictive value when used as machine learning (ML) features for maize yield prediction, using four seasons (2019–2022) and 134 pixels of 30 m at Ca' Bianca, a reclaimed coastal field in the Veneto region, Italy. EPIC used alone, ML models trained on Kriging-interpolated soil properties and remote sensing variables (leaf area index, evapotranspiration, biomass and crop coefficient), and hybrid ML models that additionally used EPIC-simulated seasonal variables were evaluated under spatial cross-validation, in which the four years are pooled, and leave-one-year-out (LOYO) cross-validation. Under spatial cross-validation, the ML models reached a test R² of 0.74–0.81. Under LOYO, the pooled R² of these models fell to between −0.32 and 0.17, largely because pooling the years in spatial cross-validation allowed information from the test year to enter the training data: weather features that are identical for all pixels within a year raised the spatial cross-validation R² from 0.74 to 0.81. Most of the LOYO error of the ML models was an error in the predicted annual mean yield. EPIC, which was calibrated on all four years, reproduced the annual means best (pooled LOYO R² of 0.43) but not the pattern within each year. EPIC-derived features were highly ranked by SHAP and reduced the LOYO error of the hybrid models (mean absolute error from 2.49 to 1.96 t/ha), but their transferability between years remained limited. A Ridge regression model predicting the within-year yield anomaly, with nested cross-validation, reproduced the within-year pattern best (mean per-year R² of 0.16). In this dataset, spatial cross-validation substantially overestimated model performance, and the error in the annual mean yield should be evaluated separately from the error in the within-field pattern. Keywords: Machine Learning; Yield Prediction; EPIC Model; Hybrid Modelling; Leave-One-Year-Out Cross-Validation; Kriging; Remote Sensing; Precision Agriculture