Bridging Spatial And Frequency Views For Disaster Assessment: Benefits And Limitations

This study evaluates the effectiveness of spatial-domain, frequency-domain, and dual-domain deep learning approaches for building damage classification using satellite imagery from the xView2 dataset. All models, based on an EfficientNet-B0 backbone, were trained under consistent conditions, revealing that dual-domain models achieved the highest accuracy (0.4688) and lowest loss, while frequency-only models underperformed due to overfitting. The findings underscore the potential of hybrid representations for improving disaster assessment but also highlight challenges such as class imbalance and the difficulty in detecting subtle damage levels, indicating areas for future research.