Towards reliable deep learning for post-disaster damage Assessment
An XAI-based evaluation
Bibliographic Data
| ID | 22028172 |
|---|---|
| Authors | Umut Lagap (0000-0003-4498-1501, National Science and Technology Center for Disaster Reduction, corresponding author), Saman Ghaffarian (0000-0001-9882-4603, National Science and Technology Center for Disaster Reduction, corresponding author), Sophie Gelinas-Gagne (National Science and Technology Center for Disaster Reduction), Jasmin Jilma (0009-0007-8550-461X, National Science and Technology Center for Disaster Reduction), Zhiyu Liu (0000-0002-6024-7836, National Science and Technology Center for Disaster Reduction), Zhiyuan Luo (National Science and Technology Center for Disaster Reduction) |
| Year | 2025 |
| Volume | 130 |
| Pages | 105839 |
| Publication date | 2025-11-01 |
| Open Access | Yes |
| Type | ARTICLE |
| Venue | International Journal of Disaster Risk Reduction (JOURNAL) |
| Journal identifiers | ISSN: 2212-4209 |
| Publisher | Elsevier BV (PUBLISHER) |
| DOI | 10.1016/j.ijdrr.2025.105839 |
| OpenAlex | W4414559937 |
| Language | EN |
| Citations received | 2 |
| References cited | 46 |
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. 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. Specifically, MA generally improved interpretability and reliability in our evaluation, particularly for identifying high-severity damage levels in post-disaster scenarios
Deep learning · Interpretability · Natural disaster · Anomaly Detection Techniques and Applications · COVID-19 diagnosis using AI · Seismology and Earthquake Studies
Densely Connected Convolutional Networks
ImageNet
ImageNet classification with deep convolutional neural networks
A survey on Image Data Augmentation for Deep Learning
Explainable artificial intelligence in disaster risk management
Discrete-Outcome Analysis of Tornado Damage Following the 2011 Tuscaloosa, Alabama, Tornado
| Citations per year | 2 |
|---|---|
| Citation span | 2026 - 2026 (1) |
| Highly cited | No |