Benchmarking attention mechanisms and consistency regularization semi-supervised learning for post-flood building damage assessment

Abstract

Rapid and accurate building damage assessment (BDA) following floods is critical for effective disaster response, yet faces challenges from limited labeled data and subtle damage cues in satellite imagery. Existing deep learning change detection (CD) methods may exhibit low recall or misclassify damage inappropriately when transferred directly to the post-flood BDA (Flood-BDA) task. This study addresses these gaps by establishing the first systematic benchmark evaluating both supervised CD model transfer and semi-supervised learning (SSL) specifically for Flood-BDA. This research investigate image-level consistency regularization SSL to combat data scarcity, finding that strategies using pseudo-label derived reference distributions significantly enhance performance (+1.17% avg. Kappa at 5% labels). Notably, pseudo-label outperform ground-truth label strategies in low-label settings (e.g., +4.84% Kappa at 5% labels). Furthermore, confronting the limitations of transferred CD models (low recall, misclassifying ’destroyed’ as ’no damage’), this paper proposed a simple prior attention disaster assessment Net (SPADANet), a lightweight U-Net incorporating a simple prior attention module designed for Flood-BDA. SPADANet demonstrably improves recall (+9.22% over best CD baseline) and exhibits more favorable error patterns for Flood-BDA, despite a precision trade-off. This work provides crucial benchmarks, validates the need for recall-driven, DA-specific designs distinct from CD, and demonstrates the potential of prior attention and image-level consistency regularization for post-flood building damage assessment. The code will be available at https://github.com/JX-OctoNeko/Flood_BDA_benchmark.git