Deep Learning Models for Hazard-Damaged Building Detection Using Remote Sensing Datasets: A Comprehensive Review

Building collapse is a leading cause of casualties and economic losses during disasters. Accurate and timely assessment of building damage is critical for effective emergency response and recovery efforts. This article provides a comprehensive review of 242 papers, summarizing the progress and challenges in using deep learning models with remote sensing images for building damage detection. The study specifically examines datasets, deep learning models, and applications. The article highlights the scarcity of publicly available training datasets for hazard-damaged building detection, with only 13 datasets identified. This scarcity significantly limits the data accessible for deep learning training. Building damage detection models can be classified into single-temporal and bitemporal methods. The majority of studies focus on single-temporal methods, which utilize postevent images and employ convolutional neural network algorithms. These methods have shown promising results in comprehensive performance evaluation. Building damage detection systems typically involve two levels (damaged or undamaged) or four levels (degree of damage) classification. The performance accuracy of these systems varies widely, ranging from 52.2% to 95.82% for two-level classification and 14% to 86.5% for four-level classification. The differences in accuracy can be attributed to both the detection methods and the classification systems used, illustrating the complexity of building damage detection. Based on the current research, challenges in building damage detection include the insufficiency of datasets and limited accuracy in classifying the four damage categories. Future studies could focus on improving detection accuracy by considering different damage categories and exploring the fusion of multisource data.