Integrating time-series analysis and deep learning methods to reconstruct the long-term retrogressive thaw slumps dynamics in the Qinghai-Tibet Plateau

Retrogressive thaw slumps (RTSs) represent a severe manifestation of permafrost degradation exacerbated by climate warming. Current monitoring approaches face significant limitations at large spatial and temporal scales since they rely mainly on very-high resolution (VHR) imagery, which is costly to obtain, intermittently available, and often cloud-obscured. Here, we present a data-efficient framework that uses medium-resolution (Landsat) time series to detect abrupt vegetation change and sparse VHR snapshots to confirm RTS occurrence. Compared with conventional approaches, the proposed method achieves a substantial reduction in VHR imagery requirements (about 90% in this study) while additionally providing pixel-level information on RTS development. Implementing this algorithm in the Qinghai-Tibet Plateau (QTP), we identified 3,273 RTSs larger than one pixel (900 m²) as of 2022. Validation results yielded a plateau-wide F1 score of 0.67 and a mean Intersection over Union (IoU) of 0.49, with performance closely tied to the density of vegetation cover. Notably, in regions with favorable vegetation such as the Source Region of the Yangtze River, which accounts for over 70% of identified RTSs, the framework achieved an F1 score exceeding 0.9 and an IoU of 0.51. The temporal segmentation tracking RTS development yielded a root mean square error (RMSE) of 2.1 years. The reconstruction of RTS dynamics shows that the total RTS-affected area expanded from 12 ha in 1995 to approximately 6,400 ha by 2022, with 2002, 2010, and 2016 representing the three peak years of RTS development. Overall, this study provides a low-cost solution for RTS mapping. The proposed datasets will enhance our understanding of the long-term patterns of permafrost degradation and its feedback to climate warming.