Permafrost distribution mapping using rock glaciers data and machine learning models in the Gaizi River Basin, China

Permafrost is one of the main components of the cryosphere. Due to global warming, water frozen reservoir shrinking and a large number of natural hazards (debris flows, water quality deterioration, etc.) related to permafrost degradation are increasing. The permafrost mapping of the Gaizi River Basin located in the East Pamir where the significant Karakoram Highway traverses was important for understanding the situation of its water resource and related hazards. Three machine learning models (Artificial Neural Network, ANN; Random Forest, RF; Extreme Gradient Boosting, XGBoost), were trained to generate permafrost probability distribution based on rock glaciers from Chinese Gaofen-1. Rock glaciers are commonly used as direct indicators of mountain permafrost distribution, particularly in alpine regions with limited in-situ data. Sixteen independent factors were used for permafrost distribution mapping that includes elevation, aspect, slope, solar radiation, topographic roughness index, topographic wetness index, profile curvature, distance from rivers, distance to glaciers, distance to water bodies, geology, fault density, LULC, NDVI, precipitation, temperature difference. The performances of the models have been evaluated by the area under the receiver operating characteristic curve (AUROC) and by known rock glaciers. Random Forest outperformed other two models, with Receiver Operating Characteristic curve values of 0.97. The permafrost map covers almost all the rock glaciers (98%), which also shows the permafrost map is reasonable. The permafrost distribution covers 1853 km², about 16.81% of the total river basin, mainly between 3233 and 5026 m elevation. The results could be used as baseline information for studying the impact of permafrost degradation and its related hazards because of climate warming, which threaten the Karakoram Highway.