Detection of collapsed buildings from post-earthquake imagery using mask region-based convolutional neural network

Abstract

After large-scale natural disasters such as earthquakes, tsunamis, and floods, the rapid identification of collapsed buildings from high-resolution imagery plays a crucial role in post-disaster damage assessment, reconstruction, and emergency rescue operations. Deep learning (DL) architectures, widely applied across various scientific domains, have also been used for extracting damaged buildings from aerial and satellite images. This study is focused on identifying collapsed buildings using a DL algorithm applied to remotely sensed data collected after the February 6, 2023, Kahramanmaraş earthquake in Türkiye. To achieve this, post-earthquake WorldView-3 image with a spatial resolution of 0.3 m were obtained to establish a building dataset, from which the boundaries of collapsed and intact buildings were manually outlined. The Mask R-CNN model was then trained and validated using various hyperparameter combinations to optimize its performance. Experimental results revealed that the Mask R-CNN model with a ResNet-50 backbone yielded the most accurate results, successfully distinguishing between intact and collapsed buildings with an Average Precision (AP) of approximately 81% and 69%, respectively. The findings of the study illustrate the promising potential of using Mask R-CNN with high-resolution imagery for the detection and mapping of collapsed buildings following earthquake events. This application is particularly significant for post-disaster operations and mitigation studies.