Reliability of rice yield monitoring under climate non-stationarity in Uganda: A hybrid modelling approach

Climate change increasingly threatens the reliability of agricultural monitoring systems by altering the environmental conditions under which historical climate–yield relationships were established. In climate-sensitive smallholder production systems, monitoring models calibrated under past conditions may become unreliable as climatic patterns shift. This study assessed the reliability of statistical, machine-learning, and hybrid modelling approaches for monitoring rice yields in Uganda under climate non-stationarity. A 30-year district-level dataset (1995–2024) from Bugiri, Butaleja, Lira, and Nwoya was analysed using Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX), Random Forest (RF), Support Vector Machine (SVM), and a hybrid statistical–machine-learning framework. Reliability was evaluated using out-of-sample predictive performance, statistical comparison, residual bias diagnostics, spatial transferability, climatic stress robustness, and spatial vulnerability under heterogeneous perturbation conditions. The hybrid framework demonstrated the strongest overall monitoring reliability, with the lowest systematic bias (mean error = 0.18), stable residual behaviour, and strong spatial transferability. Although Random Forest showed marginally lower perturbation-related error escalation, the hybrid framework maintained superior multidimensional monitoring reliability overall. In contrast, ARIMAX exhibited substantial positive bias (mean error = 1.24) and marked deterioration under intensified climatic perturbations (61.7% increase in RMSE). Diebold–Mariano tests confirmed that ARIMAX had significantly poorer predictive performance than adaptive alternatives (P < 0.05). SPI, SPEI, and temperature-related variables emerged as influential predictors. These findings show that predictive accuracy alone is insufficient for evaluating agricultural monitoring under changing climatic conditions. Adaptive hybrid monitoring frameworks provide a stronger basis for climate-risk monitoring, early warning, and food security planning.