Explainable Spatiotemporal Ensemble Modeling Using Knowledge Distillation to Forecast Food Security

Abstract

Accurate prediction of food security is crucial for managing climate-induced agricultural hazards, but current approaches struggle to include spatiotemporal complexity while producing interpretable and understandable results. This study fills this gap by conducting a thorough examination of machine learning paradigms for rainfall-driven food security forecasting in Uganda, including the use of baseline models, ensemble approaches, and spatiotemporal graph neural networks. A unifying framework was established, incorporating 15 various models, including linear regressors (Ridge and Lasso), tree-based ensembles (XGBoost and Gradient Boost), and a unique Spatial Graph Attention Network (GAT) with multi-head temporal attention. Temporal-GNN was implemented to model spatiotemporal trends in rainfall-food security interactions by combining graph convolutional networks (GCNs) and gated recurrent units (GRUs). To bridge the transparency gap in AI-driven predictions, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are systematically applied across all models, quantifying the influence of climatic drivers such as immediate rainfall intensity (r1h), seasonal fluctuations, and pressure gradients (rfq). Key advancements in this study include: XGBoost achieving state-of-the-art performance ( $\mathrm{R}^{2}=0.9949$ , MAE = 1.4067), surpassing baseline models by 14.3% in predictive accuracy; The Spatial GAT model demonstrating robust temporal dependency capture (MSE = 0.2092, MAE = 0.337), offering granular spatiotemporal insights despite a lower R2 (0.7801) and model-agnostic explainability revealing divergent feature importance. Linear models prioritise temperature, whereas ensembles emphasise atmospheric pressure dynamics. Recognising the practical challenges of deploying complex models, this work successfully implemented Knowledge Distillation (KD) to create a lightweight, efficient model from the best-performing XGBoost. The distilled student model achieved a compression ratio of 37.5% (reducing from 80 to 50 estimators) while maintaining 98.7% of the teacher model’s R2 score (0.9866 vs 0.9949). Despite a moderate increase in error metrics (MAE increased from 1.4067 to 2.3001, MSE from 3.8404 to 10.0075), the student model preserved the same feature importance hierarchy, with r1h and r1h_avg remaining the most influential predictors post-distillation. A reproducible pipeline incorporating spatial cross-validation, CUDA-accelerated training, and interactive XAI visualisation is introduced to enhance methodological rigour. Policymakers benefit from actionable trade-offs: the Voting Regressor (MAE = 2.2880) balances interpretability and performance, while the Spatial GAT enables localised, climate-resilient planning. This work advances scalable food security analytics by unifying statistical robustness with temporal dependency modelling, setting the stage for hybrid ensemble-GAT architectures in future research.