Vision Transformer based Damage Assessment from Post-Disaster Satellite Imagery: An Applied Study on Hurricane Harvey
Preprint
- 12 February 2026
- preprint
- Published by Springer Nature in Research Square
Abstract
The timely and accurate evaluation of building damage is vital for post-disaster response. This study evaluates the efficacy of Vision Transformers (ViT-B32) compared to a baseline Convolutional Neural Network (EfficientNet-B0) for the binary classification of damaged versus undamaged buildings, utilizing 128×128 RGB satellite imagery from Hurricane Harvey. As opposed to CNNs that encode local features, ViT structures rely on self-attention mechanisms for capturing global spatial relationships, a property that is crucial in spotting intricate damage patterns in visually noisy disaster sites. Experiments show that the ViT model outperforms the CNN baseline in terms of classification accuracy (97.85% vs 96.90%) and prove to be more robust on unbalanced data as shown with F1-score and AUC than respective values of close competitors. Furthermore, the study highlights the model's interpretability: we generate attention heatmaps that visualize the specific image regions driving the classification decisions. These visualizations provide actionable insights by precisely localizing structural damage, thereby offering a valuable tool for prioritizing recovery efforts in disaster management workflows.Keywords
All Related Versions
- Version 1, 2026-02-12, Research Square
- Published version: Earth Science Informatics, 19 (7), 96.
This publication has 36 references indexed in Scilit:
- Machine edge-aware IoT framework for real-time health monitoring: Sensor fusion and AI-driven emergency response in decentralized networksAlexandria Engineering Journal, 2025
- Hyperlocal disaster damage assessment using bi-temporal street-view imagery and pre-trained vision modelsComputers, Environment and Urban Systems, 2025
- The transformative roles of generative artificial intelligence in vision techniques for structural health monitoring: A state-of-the-art reviewAdvanced Engineering Informatics, 2025
- Rapid damage assessment using satellite images after hurricane harvey with gradient boosting modelsSpatial Information Research, 2025
- An enhanced discovery of multiple natural disasters using machine learning modelEarth Science Informatics, 2025
- LandslideSegNet: an effective deep learning network for landslide segmentation using remote sensing imageryEarth Science Informatics, 2024
- Computer vision tools for early post-disaster assessment: Enhancing generalizabilityEngineering Applications of Artificial Intelligence, 2024
- Explainable artificial intelligence in disaster risk management: Achievements and prospective futuresInternational Journal of Disaster Risk Reduction, 2023
- Building damage detection from satellite images after natural disasters on extremely imbalanced datasetsAutomation in Construction, 2022
- Deep learning for remote sensing image classification: A surveyWIREs Data Mining and Knowledge Discovery, 2018