A spatiotemporal graph-based robust optimization framework for data-driven disaster recovery scheduling

Scheduling heterogeneous resources for post-disaster recovery remains a critical challenge, as existing approaches predominantly rely on deterministic models or static heuristics that fail to capture the dynamic, uncertain, and spatially heterogeneous nature of real disaster environments. This study introduces STGR-DRS, a Spatiotemporal Graph-based Robust Disaster Recovery Scheduling framework that utilizes large-scale, multi-source open data to coordinate heterogeneous resources, including repair crews, medical teams, transportation units, and shelter provisions. The framework comprises four stages: (1) a multi-source data fusion pipeline that constructs a dynamic spatiotemporal recovery graph using OpenStreetMap, Humanitarian Data Exchange, Shelter Cluster, and Copernicus datasets; (2) a gradient boosting-based demand urgency estimation module with SHAP interpretability; (3) a multi-objective robust mixed-integer linear programming formulation that jointly minimizes recovery time, transportation cost, and unmet demand while enforcing fairness constraints; and (4) a rolling-horizon re-optimization mechanism that updates schedules dynamically in response to new disruptions. Extensive experiments on the 2023 Türkiye–Syria earthquake case study demonstrate that STGR-DRS reduces total recovery time by 13.4% compared to the best-performing baseline under nominal conditions and achieves a robustness advantage of 15.2% under severe disruption scenarios. Ablation studies illustrate the contribution of each framework component, and scalability analysis confirms computational tractability for networks with more than 5,000 nodes. Key limitations include the reliance on centralized coordination, which may be challenged by communication failures, and the use of synthetic disruption scenarios that may not fully replicate real-time operational complexity. Future work should explore decentralized multi-agent scheduling, integration of predictive aftershock models, and field validation with humanitarian organizations. The framework offers actionable decision support for emergency management agencies and provides a reproducible open-data experimental pipeline for the disaster recovery scheduling research community.