ExplAIn-City: Explainable Spatio-Temporal Deep Learning for Urban Resilience and Risk Governance

Urban resilience modeling often falls short in delivering interpretable, transferable, and cascade-aware predictions under constrained data regimes. Existing frameworks typically neglect the alignment between prediction fidelity and decision-support interpretability, particularly across jurisdictionally diverse regions. This study introduces ExplAIn-City, a deep learning framework that integrates an Explainability-Integrated Spatio-Temporal Encoder with a Resilience Cascade Unit (RCU) and a Policy-Aware Meta-Updater (PAMU). The encoder processes temporal and spatial dependencies through dual-attention mechanisms and localized saliency refinement, using attention-weighted sequence modeling to prioritize informative time steps and spatial neighborhoods. The RCU models dynamic interdependencies using temporal graph convolution and resilience-informed risk propagation, capturing time-varying edges, multi-lag cascade memory, and gated state transitions. PAMU enables few-shot transfer via governance-conditioned meta-adaptation by fusing region-specific policy embeddings with gradient updates to steer parameter modulation under data scarcity. The system further incorporates an interpretability fidelity pipeline that generates attention maps aligned with disruption triggers via gradient saliency and structural attribution mechanisms, providing localized relevance diagnostics for policy-sensitive evaluation. Empirical results report a mean absolute error of 0.141, an R² of 0.921, and explainability gains with GTAA and L-IOU reaching 0.813 and 0.793 across six disruption scenarios and nine baselines under cross-city adaptation. Ablation analyses attribute performance gains to the dual-attention encoder, the cascade memory within the RCU, and policy-conditioned updates within PAMU.