Future and present susceptibility to thermokarst hazards in the Northern Hemisphere using an interpretable CNN method

Abstract

Degradation of ice-rich permafrost throughout the Northern Hemisphere has triggered the formation of thermokarst hazards, such as thermokarst lakes and thaw slumps, which are rapidly developing under the influence of climate change. However, assessing the potential spatial distribution of thermokarst lakes and thaw slumps on a global scale remains a challenge. Therefore, a deeper understanding of thermokarst hazards is necessary to predict their dynamics and global implications under climate warming. In this study, we constructed thermokarst hazard inventories and multiple conditioning factors and predicted the thermokarst hazard susceptibility map in the Northern Hemisphere based on interpretable convolutional neural network (CNN) models. The results indicated that CNN models performed well with an AUC value (0.974) for thaw slumps and an AUC value (0.857) for thermokarst lakes. The multi-hazard susceptibility map showed that high and very high susceptibility regions were identified covering 11.45% of the permafrost area in the Northern Hemisphere, mainly distributed in the central Qinghai-Tibet Plateau (QTP) and the circum-Arctic lowlands. In the future (2041–2060), high and very high susceptibility regions were projected to decrease to 93.80% of the present areas for thermokarst lakes and 78.09% of the present areas for thaw slumps under the SSP585 scenario. Furthermore, interpretable results indicated that thawing degree days (TDD) and topographic wetness index (TWI) were key factors for thermokarst lakes, while freezing degree days (FDD) and TDD were key factors for thaw slumps. The results not only provide insights into understanding thermokarst dynamic processes in response to climate change, but also provide reference for hazard risk management.