Uncovering controlling factors on rock glacier velocities in the Pamir-Karakoram-Kunlun region using explainable machine learning.

Rock glaciers are ice-debris landforms commonly found in high mountain environments. Shaped by long-term creep of ice-rich permafrost, they provide critical information for permafrost studies, mountain hydrology, and hazard assessment. Although the characteristics and controlling factors of rock glacier velocities across various temporal scales have been studied at individual sites, their environmental drivers in the spatial domain over large regions remain poorly understood. In this study, we employ four machine learning methods, i.e. support vector machine, extreme gradient boost, random forest, and backpropagation neural network, to model the relationship between rock glacier velocities and environmental variables for 5,163 rock glaciers in the Pamir-Karakoram-Kunlun region. Subsequently, we use SHapley Additive exPlanations to quantify variable importance. Results show that the upslope connection to a glacier is a critical factor controlling rock glacier velocities. In our study area, glacier-connected rock glaciers exhibit on average faster movement (median velocity = 38 cm/year) than talus-connected ones (median velocity = 28 cm/year). We also find that geomorphological properties exert stronger controls on the spatial variability of rock glacier velocities than regional climate variability. Rock glacier area and slope are identified as the second and third most important variables, with larger areas and steeper slopes associated with higher velocities. The snow cover duration ranks fourth, followed by precipitation, while air temperature shows minimal influence on velocity. Overall, these findings bridge a critical knowledge gap regarding the environmental controls on rock glacier dynamics at the regional scale, extending our understanding of rock glacier kinematics beyond site-specific investigations.