ST-ResGAT: Explainable Spatio-Temporal Graph Neural Network for Road Condition Prediction and Priority-Driven Maintenance

Abstract Climate-vulnerable road networks demand a transition from reactive, fix-on-failure maintenance toward predictive and decision-ready strategies. Addressing this need, we propose ST-ResGAT, a novel Spatio-Temporal Residual Graph Attention Network that integrates residual graph-attention encoding with GRU-based temporal aggregation to model and forecast pavement deterioration. The framework is explicitly designed for resource-constrained environments and directly maps continuous Pavement Condition Index (PCI) predictions to American Society for Testing and Materials (ASTM)-compliant maintenance priorities. We evaluate the proposed approach on a real-world inspection dataset comprising 750 road segments in Sylhet, Bangladesh, collected between 2021 and 2024. Experimental results demonstrate that ST-ResGAT substantially outperforms conventional non-spatial machine learning baselines, achieving high predictive accuracy ($R^{2} = 0.93$, $RMSE = 2.72$). Ablation analysis further reveals the critical role of spatial topology, confirming that pavement degradation propagates through network connectivity. To enhance interpretability, we incorporate GNNExplainer, showing that the model’s learned decision patterns align with established engineering principles. In addition, we assess classification reliability in practical deployment scenarios, achieving $85.5\%$ exact ASTM class agreement and $100\%$ adjacent-class containment, thereby ensuring bounded and engineer-safe predictions. These results highlight ST-ResGAT as a practical, explainable, and robust solution for intelligent infrastructure management in high-risk, resource-limited settings.