Extraction of typhoon-induced building damage area from satellite images

Abstract

We attempted to extract building damage areas from satellite images using semantic segmentation. As a result of experiments using satellite images after the 2019 Typhoon No. 15 (Faxai), the recall of the blue sheet area was 0.48 and that of the building area with damage was 0.35 in the area-based evaluation. When the building shape was obtained from the ground truth data and the number of buildings was evaluated, the recall of buildings with damage was 0.62. The challenge is to improve the accuracy by training data from multiple regions.