Improving Disaster Response with U-Net and Attention Mechanism in Earthquake Damage Assessment Using Remote Sensing Data

The significance of assessing building damage for post-disaster rescue and reconstruction is acknowledged, and a novel approach is introduced to address the challenges associated with detecting and classifying damage levels. Unlike many existing studies that primarily focus on binary classification, the proposed neural network combines localization and classification of damaged buildings simultaneously. The key component of this network comprises a variety of attention U-Nets with different backbones, which leverage attention mechanisms to prioritize relevant features and channels while minimizing the impact of irrelevant ones. Training is conducted using the extensive xBD dataset, dedicated to advancing building damage assessment, and results are assessed based on balanced F1-scores. Notably, the U-Net with an attention mechanism exhibits the best individual performance, achieving an impressive F1-score of 0.752 and overall score of 0.819.The method’s effectiveness is further validated on the xView2 dataset, a comprehensive building damage assessment dataset, demonstrating precise damage scale classification and building segmentation simultaneously. This breakthrough promises to significantly enhance post-disaster rescue efforts, offering robust and transferable solutions for improved disaster response.