A Spatio-Temporal Graph Transformer Framework for Real-Time Multimodal Fusion of Social Media, Satellite Imagery, and Sensor Feeds in Disaster Response

• Ada John

  ladoke Akintola university of technology

  Author

Rapid and accurate situational awareness during disasters is critical for minimizing casualties and optimizing resource allocation, yet existing systems remain limited by unimodal data streams and fragmented processing pipelines. This study addresses the gap in unified spatio-temporal modeling that can dynamically fuse heterogeneous disaster signals—social media text and imagery, satellite observations, and ground sensor feeds—within a single real-time inference framework. We propose the Spatio-Temporal Graph Transformer (STGT) framework, which represents disaster-affected regions as dynamic graph structures where nodes encode multimodal embeddings and edges capture evolving spatial dependencies. The methodology integrates graph convolutional networks for spatial reasoning, transformer architectures for temporal attention across modalities, and a reliability-conditioned fusion mechanism that weights evidence based on cross-modal agreement. Evaluated on a benchmark constructed from historical disaster events including floods, wildfires, and earthquakes, STGT achieves 89.4% classification accuracy for damage severity assessment, outperforming unimodal baselines by 14.2 percentage points and reducing alert latency by 37% compared to sequential processing pipelines. The framework demonstrates that alignment-aware spatio-temporal fusion enables more reliable and timely disaster intelligence than modality-specific approaches. These findings have practical implications for emergency operations centers seeking automated decision support and establish a replicable architecture for multimodal disaster response systems.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.