Explainable spatial-temporal ensemble modeling for predicting food security with knowledge distillation

Abstract

Accurate prediction of food security is critical for mitigating climate-induced agricultural risks, yet existing methods struggle to reconcile spatiotemporal complexity with interpretability. This study addresses this gap through a comprehensive evaluation of machine learning paradigms, including baseline models, ensemble techniques, and spatialtemporal graph neural networks, for rainfall-driven food security forecasting in Uganda. A unified methodology was implemented, integrating 15 distinct models, ranging from linear regressors (Ridge, Lasso) and tree-based ensembles (XGBoost, Gradient Boost) to a novel Spatial Graph Attention Network (GAT) enhanced with multi-head temporal attention. Implemented Temporal-GNN which integrates graph convolutional networks (GCN) with gated recurrent units (GRU) to model spatiotemporal patterns in rainfall-food security relationships. 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 include: XGBoost achieving state-of-the-art performance (R² = 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 prioritize temperature, whereas ensembles emphasize atmospheric pressure dynamics. Recognizing 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 visualization is introduced to enhance methodological rigor. Policymakers benefit from actionable trade-offs: The Voting Regressor (MAE = 2.2880) balances interpretability and performance, while the Spatial GAT enables localized, climate-resilient planning. This work advances scalable food security analytics by unifying statistical robustness with temporal dependency modeling, setting the stage for hybrid ensemble-GAT architectures in future research.