Dual-backbone DeepLabV3+ integrating ResNet50 and vision transformer for freeze-thaw hazard extraction in the Sanjiangyuan region

Under global warming, permafrost degradation in the Sanjiangyuan region has intensified freeze-thaw hazard risks. Deep-learning-based pixel-level extraction from high-resolution remote sensing imagery supports hazard mapping and ecological conservation. However, freeze-thaw hazards in complex alpine landscapes show irregular boundaries and severe background confusion, and single-network models cannot balance local details and global semantics. We propose a dual-backbone DeepLabV3+ integrating ResNet50 and Vision Transformer (ViT) in parallel. Shallow Convolutional Neural Network (CNN) features provide boundary-texture information, whereas deep Transformer features capture long-range context; the decoder fuses multi-scale semantics. Compared with DeepLabV3+ (MobileNetV2), PSPNet, TransUNet, SegNet, and HRNet, our model achieves a mean Intersection over Union (mIoU) of 97.85%, an Intersection over Union (IoU) of 96.44%, and an Overall Accuracy (OA) of 99.38%. Cross-region testing obtains an mIoU of 88.39%. Ablation verifies 4.52%–12.95% IoU gains over single-backbone schemes, demonstrating local-global feature fusion improves boundary continuity and object integrity for alpine freeze-thaw hazard extraction.