Assessment of Thaw-Hazard susceptibility in permafrost regions using multisource data and machine learning: A case study in Qinghai Province, China

Abstract

Thaw hazards in high-latitude and glaciated regions are becoming increasingly frequent because of global climate warming and human activities, posing significant threats to infrastructure stability and environmental sustainability. However, despite these risks, comprehensive investigations of thaw-hazard susceptibility in permafrost regions remain limited. Here, this gap is addressed by a systematic and long-term investigation of thaw hazards in China's Qinghai Province as a representative permafrost area. A detailed inventory of 534 thaw-hazard sites was developed based on remote sensing, field verification, and surveys by a UAV, providing critical data for susceptibility analysis. Eleven environmental factors influencing thaw hazards were identified and analyzed using information gain and Shapley additive explanation. By using the random forest model, a susceptibility map was generated, categorizing the study area into five susceptibility classes: very low, low, moderate, high, and very high. The key influencing factors include precipitation, permafrost type, temperature change rate, and human activity. The results reveal that 17.5 % of the permafrost region within the study area is classified as high to very high susceptibility, concentrated primarily near critical infrastructure such as the Qinghai–Tibet Railway, potentially posing significant risks to its structural stability. The random forest model shows robust predictive capability, achieving an accuracy of 0.906 and an area under the receiver operating characteristic curve of 0.965. These findings underscore the critical role of advanced modeling in understanding the spatial distribution and drivers of thaw hazards, offering actionable insights for hazard mitigation and infrastructure protection in permafrost regions under a changing climate.