Post-Disaster Affected Area Segmentation with a Vision Transformer (ViT)-based EVAP Model using Sentinel-2 and Formosat-5 Imagery
Abstract
We propose a vision transformer (ViT)-based deep learning framework to improve disaster-affected area segmentation from satellite images, supporting the Emergent Value Added Product (EVAP) system developed by the Taiwan Space Agency (TASA). The process begins with a small number of manually labeled regions. We then use principal component analysis (PCA) to expand these labels with a confidence interval, creating a weakly supervised training set. Our model, which takes multi-band input from Sentinel-2 and Formosat-5 satellites, is trained to distinguish disaster-affected areas using these expanded labels. We adopt several strategies to increase accuracy when only limited supervision is available. To evaluate performance, our predictions are compared to higher-resolution EVAP results to measure spatial accuracy and consistency. Experiments on the 2022 Poyang Lake drought and the 2023 Rhodes wildfire demonstrate smoother and more reliable delineations. Quantitative evaluation is conducted against manually refined ground truth provided by the Taiwan Space Agency (TASA), with EVAP baseline reported as an operational baseline for comparison.