Rock glaciers manifest the creep of mountain permafrost occurring in the past or at present. Their presence and dynamics are indicators of permafrost distribution and changes in response to climate forcing. There is a complete lack of knowledge about rock glaciers in the Western Kunlun Mountains, one of the driest mountain ranges in Asia, where extensive permafrost is rapidly warming. In this study, we first mapped and quantified the kinematics of active rock glaciers based on satellite Interferometric Synthetic Aperture Radar (InSAR) and Google Earth images. Then we trained DeepLabv3+, a deep learning network for semantic image segmentation, to automate the mapping task. The well-trained model was applied for a region-wide, extensive delineation of rock glaciers from Sentinel-2 images to map the landforms that were previously missed due to the limitations of the InSAR-based identification. Finally, we mapped 413 rock glaciers across the Western Kunlun Mountains: 290 of them were active rock glaciers mapped manually based on InSAR and 123 of them were newly identified and outlined by deep learning. The rock glaciers are categorized by their spatial connection to the upslope geomorphic units. All the rock glaciers are located at altitudes between 3,390 m and 5,540 m with an average size of 0.26 km2 and a mean slope angle of 17°. The median and maximum surface downslope velocities of the active ones are 17±1 cm yr-1 and 127±6 cm yr-1, respectively. Characteristics of the inventoried rock glaciers provided insights into permafrost distribution in the Western Kunlun Mountains.

Rock glaciers manifest the creep of mountain permafrost occurring in the past or at present. Their presence and dynamics are indicators of permafrost distribution and changes in response to climate forcing. Knowledge of rock glaciers is completely lacking in the West Kunlun, one of the driest mountain ranges in Asia, where widespread permafrost is rapidly warming. In this study, we first mapped and quantified the kinematics of active rock glaciers based on satellite Interferometric Synthetic Aperture Radar (InSAR) and Google Earth images. Then we trained DeepLabv3+, a deep learning network for semantic image segmentation, to automate the mapping task. The well-trained model was applied for a region-wide, extensive delineation of rock glaciers from Sentinel-2 images to map the landforms that were previously missed due to the limitations of the InSAR-based identification. Finally, we mapped 413 rock glaciers across the West Kunlun: 290 of them were active rock glaciers mapped manually based on InSAR and 123 of them were newly identified and outlined by deep learning. The rock glaciers are categorized by their spatial connection to the upslope geomorphic units. All the rock glaciers are located at altitudes between 3,389 m and 5,541 m with an average size of 0.26 km2 and a mean slope angle of 17°. The mean and maximum surface downslope velocities of the active ones are 24 cm yr-1 and 127 cm yr-1, respectively. Characteristics of the rock glaciers of different categories hold implications on the interactions between glacial and periglacial processes in the West Kunlun.

The comprehensive understanding of the occurred changes of permafrost, including the changes of mean annual ground temperature (MAGT) and active layer thickness (ALT), on the Qinghai-Tibet Plateau (QTP) is critical to project permafrost changes due to climate change. Here, we use statistical and machine learning (ML) modeling approaches to simulate the present and future changes of MAGT and ALT in the permafrost regions of the QTP. The results show that the combination of statistical and ML method is reliable to simulate the MAGT and ALT, with the root-mean-square error of 0.53°C and 0.69 m for the MAGT and ALT, respectively. The results show that the present (20002015) permafrost area on the QTP is 1.04 × 106 km2 (0.801.28 × 106 km2), and the average MAGT and ALT are -1.35 ± 0.42°C and 2.3 ± 0.60 m, respectively. According to the classification system of permafrost stability, 37.3% of the QTP permafrost is suffering from the risk of disappearance. In the future (20612080), the near-surface permafrost area will shrink significantly under different Representative Concentration Pathway scenarios (RCPs). It is predicted that the permafrost area will be reduced to 42% of the present area under RCP8.5. Overall, the future changes of MAGT and ALT are pronounced and region-specific. As a result, the combined statistical method with ML requires less parameters and input variables for simulation permafrost thermal regimes and could present an efficient way to figure out the response of permafrost to climatic changes on the QTP.