Towards reliable deep learning for post-disaster damage Assessment: An XAI-based evaluation

Abstract

The increasing frequency and severity of natural hazard-induced disasters necessitate rapid and reliable post-disaster damage detection (PDD) to inform disaster response and recovery. Deep learning (DL) models, when paired with remote sensing (RS) data, have shown potential in this domain, but challenges persist due to limited interpretability and inconsistent reliability, particularly for high-severity damage classes. This study investigates the use of attention mechanisms—Channel Attention (CA), Spatial Attention (SA), and Multihead Attention (MA)—to enhance the accuracy and interpretability of state-of-the-art DL models. Utilizing the xBD dataset, we evaluated eight DL architectures and their attention-augmented configurations, in total 32 model, using explainable AI (XAI) models, i.e., Grad-CAM and Saliency Maps to visualize decision-making processes. Results indicate that models enhanced with MA achieve the highest reliability, with MA_ShallowNetV2 and MA_InceptionV3 achieving accuracies of 81.9 % and 80.0 %, respectively. Grad-CAM analysis demonstrated precise localization of damaged areas, while Saliency Maps revealed well-concentrated pixel-level focus. Specifically, MA generally improved interpretability abd reliability in our evaluation, particularly for identifying high-severity damage levels in post-disaster scenarios. In contrast, models with CA or certain SA configurations struggled with misplaced or diffused attention. These findings underscore the importance of incorporating explainable and interpretable AI approaches in disaster risk management.