A Comparative Analysis of Deep Learning Models for Automated Mapping of Permafrost Thaw Slumps on the Qinghai‐Tibet Plateau

Global permafrost is undergoing accelerated degradation due to climate warming, particularly on the Qinghai‐Tibet Plateau, where retrogressive thaw slumps (RTSs) are expanding rapidly. These geomorphological disturbances threaten critical infrastructure and amplify climate change by releasing stored carbon. While deep learning offers a promising avenue for automated RTS mapping, existing studies typically focus on single models, lacking the systematic comparison needed to guide best practices. This study addresses this critical gap by establishing a comprehensive framework to evaluate five state‐of‐the‐art deep learning architectures combined with 25 distinct backbones. We identify two superior models: a Segment Anything Model fine‐tuned with Low‐Rank Adaptation (SAMLORA) using a Vision Transformer (ViT‐H) backbone, and a Pyramid Scene Parsing Network (PSPNet) with a DenseNet161 backbone. Following rigorous optimization and evaluation on a held‐out test set, the SAMLORA (ViT‐H) model achieved the highest overall performance with a precision of 0.725, recall of 0.808, F1‐score of 0.765, and IoU of 0.681. The PSPNet (DenseNet161) model demonstrated strong performance, achieving a precision of 0.728, recall of 0.715, F1‐score of 0.722, and IoU of 0.560, while exhibiting advantages in boundary delineation. Furthermore, cross‐site validation in an independent RTS region demonstrated promising transferability of the optimized SAMLORA model, which maintained a precision of 0.681, recall of 0.754, F1‐score of 0.717, and IoU of 0.597 without additional fine‐tuning. Our experiments reveal a key architectural trade‐off that the transformer‐based SAMLORA excels at capturing the global context of large RTS features, making it ideal for comprehensive inventory mapping, while the CNN‐based PSPNet provides a robust, high‐precision alternative for applications requiring exact boundary definition. This systematic comparison establishes a critical performance baseline and provides a validated, scalable technical solution for the automated monitoring of permafrost geohazards.