A Decision-Oriented Multi-Modal Framework for Disaster Risk Assessment and Resource Optimization

Abstract

Natural disasters cause significant human, economic, and infrastructural losses worldwide, yet traditional disaster response systems often rely on static risk assessments and isolated data sources that fail to capture evolving severity, spatial dependencies, and real-time operational constraints such as infrastructure capacity and transportation logistics. To address these challenges, this work proposes an integrated multi-modal disaster response framework that combines satellite imagery analysis, social media–derived situational intelligence, graph neural network (GNN)–based spatial risk prediction, and reinforcement learning–driven resource allocation. Disaster severity is first estimated through multimodal fusion of remote sensing damage indicators and social media urgency signals. A GNN then models spatial dependencies across disaster zones to generate refined risk scores. Based on these risk estimates, a structured disaster resource database is constructed using population grounding derived from WorldPop 1 km spatial grid density alongside a simulated infrastructure and transport resource database. Resource allocation is formulated as a sequential decision-making problem and optimized using a Proximal Policy Optimization (PPO) reinforcement learning agent, with Deep Q-Network (DQN), greedy heuristic, Mixed Integer Linear Programming (MILP), A* search, and Priority Queue dispatching approaches serving as comparative baselines. Experimental evaluation demonstrates that the PPO-based allocation strategy achieves strong humanitarian impact, with an overall population rescue rate of 92%, including 92.0% of injured individuals and 91.9% of displaced populations successfully assisted across simulated disaster zones, outperforming the DQN baseline by 2.9 percentage points and the greedy heuristic by 5.5 percentage points. These results indicate improved resource utilization efficiency and adaptive allocation compared to baseline methods while suggesting potential operational feasibility under simulated infrastructure and transport constraints. Visualization-driven analysis further enhances interpretability and supports transparent decision-making. Future work will emphasize the integration of dynamic infrastructure information, including sources from OpenStreetMap (OSM), and extend the framework toward operational decision-support deployment in large-scale disaster response scenarios.