A Multi-Agent Framework for Multimodal Disaster Damage Assessment

GeoAgent4Disaster is an autonomous multi-agent GeoAI framework designed for hyperlocal, interpretable, and near–real-time disaster assessment.
The system integrates multimodal inputs—satellite imagery, street-view imagery, textual cues, and temporal change information—and performs a full pipeline of:

• Disaster perception
• Image restoration
• Damage recognition
• Disaster reasoning & recovery recommendation

By leveraging vision–language foundation models and agent-based orchestration, the framework enables cross-view understanding, zero/few-shot disaster analysis, and automated report generation without task-specific retraining.

This repository hosts the project materials associated with our research paper.

Note: The full source code is currently not publicly available because the paper has not yet been released on a public platform.

• Multi-agent collaboration for perception, enhancement, recognition, and reasoning
• Multimodal disaster interpretation (RSI + SVI + text)
• Zero-shot cross-view disaster severity assessment
• Structured JSON outputs for downstream analytics
• Automated disaster situation reporting for the “golden 36 hours”
• Evaluation across cross-view, bi-temporal, and multi-hazard datasets

The GeoAgent4Disaster pipeline consists of four core agents:

Detects hazard type, identifies image mode, and plans downstream processing.

Enhances degraded SVI/RSI for better structural clarity and reliable analysis.

Performs object-level detection, severity classification, and change-based reasoning.

Synthesizes structured outputs to generate high-level causal interpretation and actionable recovery recommendations.

Figure 2. Example of object-level detection using vision–language models.

Figure 3. Final agent-produced output including severity, object detection, and reasoning.

The framework supports and evaluates multiple multimodal disaster datasets, including:

• Cross-view hurricane imagery (paired SVI + RSI)
• Bi-temporal street-view imagery (pre- vs. post-disaster)
• Multi-hazard SVI datasets (wildfire, flooding, earthquake, etc.)

Dataset details and preprocessing scripts will be released alongside the paper.