Systematic Literature Review of Graph Neural Networks in Disaster Management: Methods, Applications, and Future Directions

The increasing frequency and complexity of natural disasters underscore the urgent need for intelligent systems that support timely and effective decision-making. Graph Neural Networks (GNNs) have emerged as a powerful deep-learning paradigm for modeling spatial and relational data, offering distinct advantages for disaster management. This study presents a Systematic Literature Review (SLR) of GNN-based approaches for disaster mitigation, emergency response, and post-disaster recovery, covering peer-reviewed publications from 2023 to 2024 and following PRISMA 2020 guidelines. A reproducible search strategy with explicit Boolean strings and database filters was applied across Scopus, IEEE Xplore, SpringerLink, and ACM Digital Library, yielding 50 primary studies after deduplication. Records were screened independently by two reviewers, disagreements were resolved by consensus, and the methodological quality of included studies was assessed using a predefined checklist. The findings show that GCN-based models were most widely applied (≈40%), particularly for flood mapping, landslide susceptibility, and infrastructure assessment. ST-GNNs (≈25%) supported dynamic hazard prediction, especially floods and wildfires, while Graph SAGE (≈10%) and GATs (≈8%) addressed sensor reliability, hazard monitoring, and evacuation planning. Hybrid architectures (≈12%) enabled multi-modal integration of satellite imagery, IoT sensor data, and social media, whereas ≈5% of studies explored transfer-learning or multi-task frameworks and explainable models such as GNN Explainer and GRAPHLIME. Common benchmark datasets included GFED, GHCN, LISFLOOD-FP, Sentinel, and OpenStreetMap, with evaluation metrics spanning RMSE/MAE for regression and Accuracy/F1/AUROC for classification. Key trends indicate a shift toward context-aware, real-time models and greater reliance on heterogeneous data sources. Despite these advances, challenges remain in interpretability, scalability, standardized benchmarking, and validation on real-world disaster datasets.