Abstract
The sudden and unplanned nature of a natural disaster is a threat to infrastructure and humans. This challenges disaster management functions to find innovative approaches to rapidly understanding the extent of damage as well as the location of damaged buildings once a disaster has occurred. The use of pre-and post-disaster satellite imagery can help disaster management teams understand the impact of the disaster on buildings and deploy resources to aid in disaster recovery. The dissertation investigates the application of deep learning to building damage detection for disaster recovery using very high-resolution pre and post-disaster satellite imagery. This was implemented by investigating the optimal pre-processing approach for satellite imagery and identifying a deep-learning architecture which is suitable for automated building damage detection. The dataset used in the study is the xBD dataset, which contains annotated multi-class damage labels on very high-resolution pre- and post-disaster satellite imagery. A two-step approach to conducting damage detection was implemented by using a U-Net deep learning model that was modified to process pre- and post-disaster images for comparison. The first phase entailed building localisation and the second stage entailed damage classification. The study's findings indicate that applying a cropping strategy to focus on areas in the image that have buildings and augmentations such as colour transformations improves the accuracy of the U-Net model in building damage detection. Additionally, the U-Net model outperformed other deep learning models explored in the study. This makes the U-Net model a suitable deep-learning model for building damage detection. The research contributes to disaster recovery through automated building damage detection using a deep learning architecture called U-Net. The method achieved an overall F1 Score of 70%, a Recall score of 74%, and a Precision score of 68%, which shows an improvement over related studies that relied on an unsupervised approach using only post-disaster images.