Temporal and Spatial Generalization of Machine Learning Models for Rice Production Forecasting Across Indonesian Provinces

Accurate provincial rice production forecasting across Indonesia is vital for maintaining national food security, mitigating market volatility, and coordinating strategic grain reserves. While supervised machine learning (ML) architectures are widely celebrated for high in-distribution predictive accuracy, their capability to generalize under geographic and environmental distribution shifts remains a critical blind spot in agricultural policy. In this empirical study, we rigorously investigate the extrapolation fidelity of five representative ML paradigms: Ordinary Least Squares (OLS), Ridge Regression, Random Forest (RF), Gradient Boosting Regressor (GBR), and Support Vector Regression (SVR), using official 2024 agricultural census data across all 38 Indonesian provinces. Evaluated under an out-of-distribution (OOD) spatial transfer benchmark—training exclusively on 32 non-Java provinces and predicting the high-density Java granary—our findings reveal a catastrophic out-of-distribution performance collapse for non-linear models: RF collapses from an in-distribution R² of 0.9385 down to R² = 0.1342 (MAE surging tenfold to 2,752,732 tons), GBR plummets to R² = 0.3187, and SVR drops to R² = -0.9135 due to rigid leaf-clamping and bounded kernel saturation. Conversely, regularized linear models exhibit exceptional structural resilience, with Ridge regression retaining R² = 0.9657 and MAE = 551,875.5 tons by preserving global unbounded gradient trends. These quantitative disparities expose the perils of uncritically deploying standard tree ensembles for national food logistics and establish the necessity of linear-residual hybrid frameworks for robust agricultural policymaking.