BDA-Net: a novel object detection framework for post-disaster damage assessment using semi-supervised learning and attention mechanisms

Accurate and rapid building damage assessment from satellite imagery is critical for effective disaster response but remains challenging due to the complexities of fine-grained damage classification and severe class imbalance. While deep learning methods offer promise, they often struggle with generalization and require large, labelled datasets. To address these challenges, we propose BDA-Net, a novel object detection framework that integrates Convolutional Block Attention Module (CBAM) and semi-supervised learning (SSL) within a YOLOv8-m backbone. The attention mechanism enhances feature representation, enabling better detection of subtle damage cues, while the SSL pipeline leverages unlabelled data via a Mean Teacher framework to reduce annotation dependency. We evaluate BDA-Net on the xBD-Hurricane dataset, which comprises 2000 image pairs from Hurricanes Michael (1,200) and Harvey (800), covering Florida, Georgia, and Texas (U.S.A.). The ‘minor-damage’ class constitutes 18% of the labelled instances, reflecting significant class imbalance, where it achieves a top-performing detection mAP@0.5:0.95 of 65.8% and anF1-Det score of 77.2%, outperforming baseline models by up to +4.6mAP. Ablation studies confirm the individual contributions of CBAM and SSL, particularly in improving performance on minor-damage instances (+11.6 F1-MiD). Furthermore, our model demonstrates strong cross-domain generalization, achieving 53.5 mAP on the unseen Ida-BD dataset without fine-tuning. With an inference speed of 72 FPS, BDA-Net balances high accuracy and real-time capability, making it suitable for time-sensitive disaster scenarios. This work advances practical AI-driven damage assessment under realistic data constraints