Post-Earthquake Building Damage Detection using Earth Observation and Machine Learning

Abstract

This study introduces a methodology for automating the estimation of building damage levels following the February 2023 earthquakes in southeastern Turkey. The carefully curated dataset consists of very high resolution (VHR) satellite imagery combining pre- and post-event optical and Synthetic Aperture Radar (SAR) data with the corresponding building footprints. The approach utilizes state-of-the-art deep learning for binary and multiclass classification, resulting in a reproducible pipeline for damage assessment. The main goals of this study is to investigate the most suitable fusion technique for the different inputs, the advantage of different backbones and pretrained weights as well as the impact of using pre-event images in comparison to only using post-event data. Furthermore, this study also investigates the ability of the model to distinguish three damage classes by comparing it to binary classification with similar model architecture.