Integrating time-series analysis and deep learning method to reconstruct the RTS dynamics in the Tibetan Plateau over the past nearly four decades

Retrogressive thaw slumps (RTSs) represent a severe form of permafrost degradation exacerbated by climate warming. Current automated monitoring methods primarily depend on object-based deep learning (DL) approaches using very high-resolution (VHR) remote sensing imagery and spectral trajectory time-series analysis applied to medium-resolution imagery. DL methods are constrained by publicly available VHR data and computational limitations, their application at large areas in decadal scale remains challenging, whereas time-series analysis often suffers from low detection accuracy due to noise caused by climate fluctuation or human activities. Here, we propose a novel framework that integrates time-series analysis with DL method. The algorithm first conducts a pixel-based detection of surface disturbance through analysis of abrupt changes in vegetation indices derived from Landsat images. Then, the features of these disturbed clusters in VHR imagery are analyzed to confirm the RTS occurrence by using an object-based DL method