Remote Sensing-Driven Interpretable Random Forest to Predict Maize End-of-Season WRSI Drought in Kirehe District, Rwanda (2001–2023)

Drought significantly limits rain-fed maize yields and food security in Rwanda’s Kirehe District, where over 30% of the March–June growing cycles experience severe water stress when the end-of-season Water Requirement Satisfaction Index (EOS-WRSI) falls below 70% of the long-term average, with a localised hazard probability of 42%. Current national drought datasets only extend to 2014 and lack mid-season predictive tools. The conventional Extended WRSI diagnostic tends to overestimate water sufficiency in East Africa, and no localised machine learning-based EOS-WRSI forecasting framework exists for Rwandan districts. This study constructs a Random Forest classification model to predict terminal drought severity for maize using hydroclimatic and satellite data collected by the eighth 10-day growth interval, enabling early risk alerts six weeks before harvest. A pixel-scale dataset spanning 20 growing cycles (2001–2023) with 158,920 observations was built by updating historical WRSI outputs via multi-source remote sensing and reanalysis data, paired with FAO-standard crop water balance calculations. Sixteen predictive variables were developed, including raw environmental measurements and season-normalised relative features to mitigate interannual climate bias. Leave-one-year-out cross-validation eliminated temporal data leakage,  and three anomalous growing cycles influenced by extreme hydrological disturbances and human agricultural interventions were removed from the training set to reduce data noise, prevent biased model fitting, and enhance model generalisation performance. Rolling time-window experiments assessed predictive skill at earlier monitoring intervals, and SHAP interpretation quantified dominant drought drivers. The integrated feature model achieves robust classification capacity, achieving cross‑validation overall accuracy of 87% and balanced classification performance, demonstrating robust predictive capability for maize drought forecasting, with the sixth 10-day window delivering balanced performance and an eight-week warning window; marginal skill gains occur between early and late mid-season monitoring, indicating that early growth-stage signals capture most drought-discriminative information. SHAP outputs identify season-normalised topsoil moisture as the primary predictive factor, outperforming rainfall and vegetation proxies. The model errs toward over-warning rather than under-reporting drought events, a favourable bias for humanitarian risk response. This research delivers Rwanda’s first district-level ML mid-season WRSI forecasting workflow, addressing gaps in national drought surveillance and offering an interpretable, adaptable framework for regional agricultural emergency planning.