Retrogressive thaw slumps recognition and occurrence analysis using deep learning with satellite remote sensing in the central Qinghai-Tibet Plateau

Abstract

Due to permafrost degradation, retrogressive thaw slumps (RTSs) have increasingly occurred in the past decade, altering landscape patterns and ecological environments in permafrost regions while accelerating soil carbon emissions. This study introduces a framework for RTS recognition and occurrence analysis based on deep learning and change analysis techniques. We propose a hybrid convolution and transformer neural network (HCT-Net) to recognize RTSs automatically. The network employs Synthetic Aperture Radar (SAR) and optical imagery for RTS detection. To effectively integrate the feature extraction capabilities of Convolutional Neural Network (CNN) and Vision Transformer (ViT), a Cross Feature Fusion Module (CFFM) has been developed. With the generated RTS map, Landsat series satellite images spanning from 2001 to 2023 were acquired for long-term observation and analysis of RTSs evolution. The occurrence years of detected RTSs were estimated employing the LandTrendr model, enabling a year-to-year dynamics analysis of RTS development. Focusing on the central Qinghai-Tibet Plateau (QTP), including the Beiluhe basin, Hoh Xil Hill, and Mt. Fenghuo area, covering approximately 49,000 km², experiments have been conducted for RTS recognition and multi-temporal analysis. Results demonstrate that the proposed method achieves an Intersection over Union (IOU) and F1 score of 72.95 % and 88.30 % for RTS recognition, outperforming other semantic segmentation methods such as Deeplabv3+, Segformer, UNet, and Swin-UNet. Based on the RTSs detected within the study area during 2001–2023, the occurrence year of the RTSs was estimated. The findings reveal a continuous expansion of RTSs since 2001 and confirm concentrated outbreaks in the years 2010 and 2016. The proposed framework offers a good approach for RTSs recognition and analysis, which can support further research into permafrost degradation.