HSGANet: a deep learning network for post-earthquake damaged road extraction

Accurate extraction of damaged roads from post-earthquake remote sensing imagery is essential for emergency response and transportation recovery. However, post-earthquake roads are often affected by occlusion, structural fragmentation, irregular damage patterns, and complex background interference, which makes reliable extraction difficult for existing deep learning methods. To address these challenges, we propose HSGANet, a deep learning network for post-earthquake damaged-road extraction. HSGANet incorporates a Hybrid Strip Context Aggregation (HSCA) block to strengthen long-range dependency modeling and preserve road continuity, and a Multi-Scale Large-Strip Gated Attention (MS-LSGA) module to enhance the representation of irregular damaged roads while suppressing background interference during feature fusion. To evaluate the model under both synthetic and real post-disaster conditions, we construct and use two datasets: a DeepGlobe-based Hybrid Noise Synthetic Post-Earthquake Damaged Road Dataset and a Türkiye Post-Earthquake Damaged Road Dataset derived from post-event UAV imagery. Experimental results show that HSGANet achieves overall macro-averaged F1-scores of 90.03% and 96.46% and IoU scores of 82.71% and 93.31% on the two datasets, respectively, demonstrating its potential for more reliable and structurally consistent damaged-road extraction in post-earthquake accessibility assessment and rescue-route identification.