Building damage assessment in storm-induced disasters using multi-source remote sensing data and NBDANet

—Rapid identification of building damage distribution during the critical rescue phase is essential for efficient emergency response. However, most existing studies focus on earthquake-like structural hazards, limiting their generalization across different disaster scenarios. In addition, current methods often struggle to effectively integrate cross-temporal features from pre- and post-disaster imagery, while effective links between pixel-level classification and regional decision-making information remain limited. To address these issues, this study proposes an improved deep learning framework, NBDANet, developed based on BDANet for building damage assessment. NBDANet adopts a two-stage architecture with Spatial Pyramid Pooling (SPP) modules embedded in both stages to enhance multi-scale contextual representation. The first stage performs building localization, while the second stage introduces a Dual-input Triplet Attention (DTA) module to improve feature interaction between pre- and post-disaster images and enhance damage classification accuracy. The model is trained and evaluated on the xBD dataset and further applied to the Derna case using high-resolution GeoEye-1 imagery. Validation is supplemented by Sentinel-derived Normalized Difference Built-up Index (NDBI) and Damage Proxy (DP) maps, confirming spatial consistency. In addition, Global Human Settlement Layer (GHSL) data are integrated to estimate affected populations and delineate rescue-priority zones. Results show that 22% of buildings were damaged, including approximately 20% of all buildings classified as Major damage or Destroyed . Estimated affected populations in Major damage and Destroyed zones were approximately 5,200 and 2,650, respectively. The proposed approach demonstrates strong performance in building localization and damage classification, providing a scalable solution for multi-source disaster assessment and recovery planning.