Change-centric building damage assessment across multiple disasters using deep learning

Natural catastrophes such as floods, earthquakes, and hurricanes often result in widespread destruction, making post-disaster building damage assessment a challenging task. The xBD dataset offers a vast collection of annotated damaged buildings on a common damage scale, partially overcoming dataset limitations in this area. In this paper, we propose a novel and computationally efficient approach to localize and classify building damage by combining object and change detection methods on the Mask R-CNN architecture. Our change-centric technique focuses on analyzing the differences between pre- and post-disaster images to enhance classification performance. To streamline the process and reduce complexity, we concentrate on a subset of the xBD dataset, covering specific disaster events. This allows us to explore and compare the performance of the model trained on this subset against the one trained on the complete dataset. Additionally, we conduct ablation studies to assess the model’s performance and compare it against a baseline model for localization and classification, highlighting the potential areas of improvement for our approach. Although the change-centric technique was met with mixed results, we made an interesting observation regarding the impact of input data quantity on model performance where an increased number of diverse disaster events correlated with higher F1-scores. This indicates the direct relationship between input data quantity and model performance, suggesting that a more comprehensive dataset yields substantial improvements in damage assessment accuracy.