A Hybrid-Loss and Multi-Supervision-Based Semantic Segmentation Network for Crater Extraction of Airport Runways

Image-based assessment of airport runway damage effects faces two critical challenges in segmenting the main runway and associated damage areas (e.g., craters, bulges): (1) the highly similar textural characteristics between the main runway and connecting taxiways impede their reliable distinction; (2) the significant inter-class discrepancy between the main runway surface and damage regions complicates precise contour extraction, particularly for craters, hindering high-accuracy regional segmentation. To address these challenges, this paper proposes the Swin-HMNet network for extracting airport runway damage areas. The network employs an encoder-decoder architecture based on the Swin Transformer module and integrates an improved multi-level weighted hybrid loss function to enable deep supervision during training. Comparative and ablation experiments demonstrate that the proposed detection strategy significantly outperforms traditional methods in extracting runway damage areas, achieving a mean Intersection over Union (mIoU) of 84.79%. These results underscore the substantial value of the proposed approach for military applications involving damage effect assessment of airport runways using remote sensing imagery.