Rapid assessment of building damage from satellite imagery is crucial for effective disaster response and recovery. This study explores the integration of spatial-domain features with frequency-domain representations, highlighting their complementary capabilities in capturing structural cues such as debris patterns and collapse-induced textures. A controlled comparison of spatial-domain, frequency-domain, and dual-domain deep learning approaches is presented for multi-class building damage classification using post-disaster imagery from the xView2 (xBD) dataset. All models are based on EfficientNet-B0 and trained under identical settings, differing only in input representations and fusion strategies. Performance metrics include accuracy, macro F1-score, per-class metrics, and confusion matrices. Results indicate that dual-domain models outperform single-domain counterparts, with the dual spatial configuration achieving the highest test accuracy (0.4688) and lowest loss, while the spatial-only model yields the best macro F1-score (0.4254), indicating more balanced class performance. In contrast, frequency-only models perform the worst and exhibit overfitting, suggesting limited generalization. Despite improvements in detecting severe damage with dual-domain approaches, challenges remain in identifying subtle damage levels, particularly the Minor class, due to class imbalance and fine-grained visual ambiguity. These findings underscore the benefits and limitations of hybrid representations and motivate future work on data balancing, advanced fusion, and regularization.
Blogger's Review: This research provides valuable insights into disaster damage assessment, especially with the application of dual-domain models showcasing the potential of frequency-domain features. However, the models' generalization capabilities and subtle damage recognition need further enhancement. Future studies could focus on data augmentation and more complex fusion techniques to improve model robustness and accuracy.