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

This paper investigates the reliability of deep learning (DL) models for post-disaster damage detection (PDD), with a particular focus on tornado events in the United States. Using the xBD satellite imagery dataset, the authors analyze building damage across four categories: no damage, minor damage, major damage, and destroyed. The study specifically examines three major tornado events—Joplin, Moore, and Tuscaloosa—selected due to their extensive destruction and availability of annotated imagery.

To enhance both model accuracy and interpretability, the research integrates different attention mechanisms—Channel Attention (CA), Spatial Attention (SA), and Multihead Attention (MA)—into eight deep learning architectures. The study evaluates 32 model configurations and applies explainable AI techniques such as Grad-CAM and Saliency Maps to visualize how models make predictions. Results show that models enhanced with Multihead Attention achieve the highest reliability and improved focus on high-severity damage classes.

The findings highlight that incorporating attention mechanisms—particularly Multihead Attention—significantly improves both performance and interpretability in disaster damage detection. The research emphasizes that reliable and transparent AI systems are essential for disaster response and recovery operations, where rapid and trustworthy assessments directly influence resource allocation and decision-making. The study contributes a structured evaluation framework for developing operationally reliable AI tools in disaster risk management.