🛰️ BiTemporal-StreetView-Damage

Hyperlocal disaster damage assessment using bi-temporal street-view imagery and pre-trained vision models

---

📘 Overview

This repository presents a bi-temporal street-view image analysis framework for hyperlocal disaster damage assessment.
By integrating pre- and post-disaster imagery through pre-trained vision and vision–language models, this approach improves both classification accuracy and interpretability of damage detection.

🔍 Key Contributions

• ✅ Dual-channel architecture for pre– and post-disaster fusion.
• 📸 2,249 labeled street-view image pairs, annotated with detailed impact levels.
• 📈 Performance Gain: Accuracy increased from 66.14% (post-only) → 77.11% (bi-temporal).
• 🔥 Grad-CAM visualization demonstrates improved attention focus using pre-disaster inputs.
• 🏙️ Supports fine-grained and rapid damage mapping for climate-resilient urban planning.

---

🧩 Methodology

Figure 1: Dual-channel architecture for bi-temporal disaster damage assessment.

Model Pipeline

1. Pre-processing: Normalize paired street-view images (pre-/post-disaster).
2. Feature Extraction: Use pre-trained Swin Transformer & ConvNeXt backbones.
3. Dual-Channel Fusion: Fuse embeddings via a feature-fusion head for comparative reasoning.
4. Classification: Predict severity levels (mild, moderate, severe).
5. Visualization: Apply Grad-CAM to interpret key spatial attention areas.

---

🌍 Study Area

The study focuses on Horseshoe Beach, Florida, which was severely impacted by Hurricane Milton (2024).
Bi-temporal street-view imagery was collected to model the extent and types of disaster damage across different locations.

Figure 2: Heatmap visualization of disaster severity distribution across Horseshoe Beach, Florida.

---

📂 Dataset

You can access the bi-temporal street-view disaster dataset via the DOI below:

📁 Yang, Yifan (2025).
Perceiving Multidimensional Disaster Damages from Street–View Images Using Visual–Language Models.
figshare. Dataset. https://doi.org/10.6084/m9.figshare.28801208.v2

or

The primary hosting platform is Hugging Face Datasets, which provides a version-controlled repository for convenient access, inspection, and integration with machine learning workflows:

🔗 https://huggingface.co/datasets/Rayford295/BiTemporal-StreetView-Damage

Dataset Contents:

• Paired pre-/post-disaster street-view images
• Location and damage-type annotations
• Severity labels: Mild, Moderate, Severe
• Sample imagery from Horseshoe Beach, FL (Hurricane Milton, 2024)

---

🏛️ Conference Presentation (AAG 2025)

This work was accepted for presentation at the 2025 Annual Meeting of the American Association of Geographers (AAG 2025).

• Session: GISS Specialty Group — Paper Competition II
• Official AAG Link: https://aag.secure-platform.com/aag2025/organizations/main/gallery/rounds/131/details/83465

Presentation Schedule:

• Date: 3/25/2025
• Time: 10:10 AM - 11:30 AM
• Location: 360, Level 3, Huntington Place

Paper Title:
Hyperlocal Disaster Damage Assessment Using Bi-Temporal Street-View Imagery and Pre-Trained Image Processing Models

This presentation represents an early-stage foundation of the bi-temporal disaster perception framework, which later evolved into the DamageArbiter disagreement-driven arbitration framework presented at AAG 2026.

---

🧠 Paper Reference

If you use this repository, please cite both the Computers, Environment and Urban Systems article and the arXiv preprint.

📖 APA Citation (click to expand)

Yang, Y., Zou, L., Zhou, B., Li, D., Lin, B., Abedin, J., & Yang, M. (2025). Hyperlocal disaster damage assessment using bi-temporal street-view imagery and pre-trained vision models. Computers, Environment and Urban Systems, 121, 102335. https://doi.org/10.1016/j.compenvurbsys.2025.102335

🧾 BibTeX (click to expand)

@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},
issn = {0198-9715},
doi = {https://doi.org/10.1016/j.compenvurbsys.2025.102335},
url = {https://www.sciencedirect.com/science/article/pii/S0198971525000882},
author = {Yifan Yang and Lei Zou and Bing Zhou and Daoyang Li and Binbin Lin and Joynal Abedin and Mingzheng Yang},
keywords = {Disaster resilience, Street-view imagery, Dual-channel neural network, Pre-trained vision model, Damage estimation}
}