Residential Wildfire Structural Damage Detection Using Deep Learning to Analyze Uncrewed Aerial System (UAS) Imagery, Aerial Imagery, and Satellite Imagery

In recent years, wildfires in residential regions have increasingly inflicted significant economic and social losses. Preemptive measures can reduce the damage to public infrastructure and lessen these impacts. Rapid evaluation of residential structures after wildfire is crucial for investigating the overall scope of the damage and establishing an effective disaster mitigation strategy. However, conducting these assessments involves detailed on-site examinations, which require considerable time and workforce. Furthermore, these qualitative assessments can be subjective and prone to error. To overcome these shortcomings, this study suggests a practical methodology for performing damage assessments of housing after a wildfire using deep learning technology. The applications of deep learning to three different image sources for residential areas are analyzed and compared as follows: uncrewed aerial systems imagery, aerial imagery, and satellite imagery. Notably, combinations of these image sources were considered from the training stage, and the impact of changes in training data when applied to each image source was comprehensively investigated. Key results reveal achievable accuracies depending on the various remote sensing data sources used in the training and application phases. This study is expected to provide deep learning researchers working on wildfires with a fundamental resource for the comprehensive use of remote sensing data and to provide valuable insights into the decision-making process for wildfire responders.