Evaluating machine learning models for enhanced permafrost distribution mapping using rock glaciers: A case study in Shaluli Mountain, Southeast Tibetan Plateau

Rock glaciers are widely used as indirect indicators for modeling permafrost distribution, particularly in remote mountain regions with limited in-situ observations. However, previous studies have often relied on empirically selected models and predictor variables, leaving their impacts on mapping accuracy unclear. In this study, we focus on the Shaluli Mountain region in the southeastern Tibetan Plateau to conduct permafrost distribution mapping driven by rock glaciers, with an emphasis on model evaluation. Using an interferometric synthetic aperture radar (InSAR)-assisted method, we compiled an inventory of 236 active and 229 relict rock glaciers in the study area. We then evaluated multiple machine learning models and environmental predictors, identifying logistic regression (LR) with mean annual air temperature (MAAT) and potential incoming solar radiation (PISR) as the most effective combination. The optimal model achieved 82 % accuracy (Kappa = 0.64), producing a 90 m resolution permafrost favorability index (RG-PFI) map. Our results estimate permafrost coverage at 1554 km² (20.2 % of the study area), primarily between 4750 and 5200 m elevation. Compared to four existing permafrost maps, RG-PFI demonstrated a 3 %–13 % improvement in classification accuracy. This study underscores the importance of integrating robust statistical modeling with high-quality rock glacier inventories to enhance permafrost mapping in data-scarce regions. Additionally, our findings highlight the urgent need to address permafrost degradation risks posed by climate warming, which threaten critical infrastructure such as the under-construction railway.