RS-BDNet: A Parameter Efficient Hybrid CNN Transformer Network for Multi-Hazard Building Damage Assessment

Timely and accurate building damage assessment is essential for effective disaster response and recovery planning. Recent advances in deep learning have improved the identification and classification of damaged buildings using pre- and post-disaster satellite imagery. However, many high-performing approaches rely on ensembles of independently trained models, resulting in substantial model complexity and increased computational demands. To address this challenge, we propose RS-BDNet, a parameter-efficient hybrid Convolutional Neural Network (CNN)–Transformer architecture for joint building localization and damage severity assessment. The proposed framework integrates Siamese ResNet-34 U-Net and Swin-Tiny branches to capture complementary local structural features and broader spatial contextual information from bi-temporal satellite imagery. Bidirectional cross-attention is employed to model changes between pre- and post-disaster observations, while CNN–Transformer feature fusion, building-guided damage prediction, and auxiliary ordinal supervision support the classification of buildings into four damage severity categories. The proposed architecture is independently trained and evaluated on six datasets, including the multi-hazard xBD benchmark and five event-specific datasets covering earthquakes, volcanic eruptions, tornadoes, hurricanes, and floods. Experimental results show that RS-BDNet achieves a damage Macro-F1 score of 0.7756 on xBD using a single model with 55.3 million parameters, representing reductions of 96.0%, 94.2%, and 89.5% compared with the first, second, and third-place xView2 ensemble solutions, respectively. These findings highlight the potential of combining convolutional and Transformer-based representations within a single network to support multi-hazard building damage assessment with substantially reduced model complexity.