Predicting AquaCrop-Simulated Durum Wheat Yield with Machine Learning: Algorithm Comparison and Agronomic Signal Convergence in the Capitanata Plain

Durum wheat production in the Mediterranean basin faces increasing climate variability and thus the need for computationally efficient tools to support agronomic decision-making at regional scale. Process-based crop models such as AquaCrop provide mechanistically sound yield estimates but require substantial computation time when screening large numbers of soil–climate–management combinations. The present study addresses this constraint by developing and evaluating five machine learning (ML) surrogate models—Linear Regression (LR), Multilayer Perceptron (MLP), Support Vector Machine for regression (SMOreg), RandomTree, and Reduced Error Pruning Tree (REPTree)—trained to emulate the AquaCrop-GIS response surface for durum wheat (Triticum durum Desf.) grain yield across the Capitanata plain (Southern Italy). A dataset of 342 instances was constructed by crossing 25 soil profiles, three sowing dates, and two irrigation regimes across 15 climatic grid cells (2014–2023), evaluated by stratified 10-fold cross-validation. The MLP achieved the highest accuracy (R = 0.983; R2 = 0.966; RMSE = 0.083 t ha−1); the four interpretable models were clustered at R = 0.891–0.907 (RMSE = 0.192–0.203 t ha−1). All models converged on consistent agronomic signals: standard sowing (1 November) yielded +0.53 t ha−1 over late sowing (15 November), supplemental irrigation added +0.17 t ha−1, and fine-textured soils produced superior yields. The convergence of directional signals across linear, kernel-based, and tree-based architectures confirms that ML surrogates trained on process-model outputs can efficiently emulate AquaCrop response surfaces and deliver actionable management guidance for durum wheat producers and agricultural planners in Mediterranean dryland farming systems.