📦 Datasets: disaster-crossview-datasets — the shared cross-view disaster data backbone for the Rayford-AI org.
Generative AI for Disaster Resilience
Publication-safe public companion repository. This repo shares only high-level concepts, public framing, and lightweight demonstrations while the related manuscript is under development. Detailed experiment results, raw data, prompts, model outputs, and unpublished figures are intentionally withheld until publication.
GenAI4Dresilience explores how generative AI can support disaster resilience across the full disaster-management lifecycle: mitigation, preparedness, response, and recovery.
The central idea is that GenAI should not be treated as a single image-generation or text-generation tool. In disaster resilience, its value comes from combining four capabilities:
| Capability | Public-Safe Role in Disaster Resilience |
|---|---|
| Generation | Scenario construction, data completion, and option exploration |
| Multimodal perception | Linking remote sensing, street-view imagery, sensors, infrastructure records, and text reports |
| Reasoning | Turning heterogeneous evidence into interpretable risk, damage, and recovery rationales |
| Multi-agent collaboration | Representing coordination among agencies, infrastructure operators, planners, and communities |
Together, these capabilities support a shift from isolated perception tasks toward evidence synthesis, scenario reasoning, and decision support.
The public framework maps GenAI capabilities to four resilience phases:
flowchart LR
A[Geospatial data inputs] --> B[GenAI capability layer]
B --> C[Mitigation: risk identification]
B --> D[Preparedness: monitoring and early warning]
B --> E[Response: damage assessment and coordination]
B --> F[Recovery: adaptation and long-term resilience]
The framework emphasizes a phase-specific view:
- Mitigation: use counterfactual and scenario reasoning to identify where future losses may concentrate.
- Preparedness: convert heterogeneous monitoring streams into evolving situation models and warning rationales.
- Response: connect visual evidence with severity assessment, prioritization, and role-specific emergency decisions.
- Recovery: support option generation, trade-off analysis, and participatory planning for long-term resilience.
The ongoing case study focuses on hurricane damage assessment with two complementary evidence settings:
- Cross-view evidence: post-disaster street-view imagery paired with post-disaster remote sensing imagery at matched locations.
- Bi-temporal evidence: pre- and post-disaster street-view imagery from the same local area.
The public workflow is summarized at a high level:
flowchart TD
A[Dataset-specific visual evidence] --> B[Selective image restoration]
B --> C[GenAI-assisted damage recognition]
C --> D[Reasoning and decision support]
D --> E[Human review and publication-ready interpretation]
This repo does not release unpublished damage scores, sample imagery, evaluation tables, full prompts, or model-generated reports. Those materials will be shared only when they are appropriate for publication or reproducible release.
For a longer public-safe summary, see docs/public_overview.md.
GenAI4Dresilience/
├── docs/
│ └── public_overview.md # Publication-safe research summary
├── figure/
│ ├── framework.png # Public conceptual framework
│ └── readme.md
├── code/
│ ├── README
│ └── map.py # Lightweight hurricane location visualization
├── README.md
└── README_zh.md
code/map.py provides a lightweight public visualization of Hurricane Ian and Hurricane Milton locations over Florida. It is a demonstration script only and does not contain the unpublished analysis pipeline.
Dependencies
pip install matplotlib cartopyRun
python code/map.pyTo protect the ongoing manuscript, this repository currently avoids releasing:
- raw or derived image samples used in the case study
- detailed restoration, recognition, and reasoning prompts
- quantitative results and evaluation tables beyond public summaries
- intermediate model outputs or generated disaster reports
- complete experimental code for reproducing unpublished findings
After publication, this repository may be expanded with reproducible materials, citation information, and a clearer release package.
| Project | Description |
|---|---|
| Agent4Disaster | Multi-agent GeoAI pipeline for disaster perception and reasoning |
| Sat2Street-DisasterGen | Satellite-to-street-view synthesis for post-disaster assessment |
| DamageArbiter | CLIP-enhanced multimodal hurricane damage assessment |
| Bi-Temporal-StreetView | Hyperlocal damage assessment via bi-temporal street-view imagery |
| DisasterVLP | Vision-language models for multidimensional disaster damage perception |
Yifan Yang - Texas A&M University
- GitHub: @rayford295
- Email: yyang295@tamu.edu