Hyperlocal disaster damage assessment using bi-temporal street-view imagery and pre-trained vision models
This repository presents a bi-temporal street-view image analysis framework for hyperlocal disaster damage assessment. By fusing pre- and post-disaster imagery through a dual-channel architecture with Swin Transformer and ConvNeXt backbones, the framework improves both classification accuracy and spatial interpretability of damage detection.
Key results:
- 2,249 labeled street-view image pairs with severity annotations
- Accuracy: 66.14% (post-only) → 77.11% (bi-temporal fusion)
- Grad-CAM visualization confirms improved attention focus with pre-disaster context
.png)
Pipeline: paired image normalization → feature extraction (Swin Transformer / ConvNeXt) → dual-channel fusion → severity classification (mild / moderate / severe) → Grad-CAM interpretation.
Focused on Horseshoe Beach, Florida, severely impacted by Hurricane Milton (2024).
| Study Area | Damage Heatmap |
|---|---|
 |
 |
Pre- and post-disaster street-view image pairs with georeferenced annotations and severity labels.
- Figshare: 10.6084/m9.figshare.28801208.v2
- Hugging Face: Rayford295/BiTemporal-StreetView-Damage
Presented at AAG Annual Meeting 2025 — GISS Specialty Group Paper Competition, Honorable Mention Session: 360, Level 3, Huntington Place — March 25, 2025, 10:10–11:30 AM
This work forms the foundational framework later extended into DamageArbiter (AAG 2026).
@article{YANG2025102335,
title = {Hyperlocal disaster damage assessment using bi-temporal street-view imagery
and pre-trained vision models},
journal = {Computers, Environment and Urban Systems},
volume = {121},
pages = {102335},
year = {2025},
doi = {10.1016/j.compenvurbsys.2025.102335},
author = {Yifan Yang and Lei Zou and Bing Zhou and Daoyang Li and
Binbin Lin and Joynal Abedin and Mingzheng Yang}
}Yifan Yang — Department of Geography, Texas A&M University yyang295@tamu.edu · rayford295.github.io