Damaged Building Detection from Post-Earthquake Drone Images Using Deep Learning

Abstract

Especially after earthquakes affecting large areas, rapid damage detection is important for understanding the extent of damage and allocating resources in a fast, effective and more organized manner. In this study, a dataset was prepared for the images taken by drones after the February 6, 2023 Kahramanmaras earthquake and a system was designed to detect and segment damaged buildings using deep learning based methods. For this purpose, a dataset was first created by labeling the buildings according to their damage levels. Then, data diversity was increased by using different augmentation techniques on the images. Different semantic segmentation models were tested for damaged building detection. Highest success for segmenting only damaged buildings was achieved with OneFormer model with 0.77 mIOU, and the highest success for segmenting and classifying buildings according to their damage levels was with SegFormer model with a mIOU of 0.52.