Abstract

Retrogressive thaw slumps (RTSs) serve as key indicators of global climate change and also pose significant risks to critical infrastructure along the Qinghai–Tibet Engineering Corridor (QTEC). Accurate automatic segmentation of RTSs using Sentinel-2 imagery is of great value for climate change research and risk assessment, owing to the dataset’s ready availability and extensive spatiotemporal coverage. However, this segmentation task remains challenging due to the complex morphology and variable sizes of RTSs, as well as their low contrast and fuzzy boundaries against the surrounding landscape in medium-resolution satellite imagery. To deal with these challenges, this study proposes the Multi-Scale Object-aware Context Attention Network (MOCA-Net), which enhances the Swin Transformer backbone through two critical components: the Feature Enhancement Network and Enhanced Decoder. Evaluation metrics show that MOCA-Net outperforms seven mainstream baseline models, achieving a Mean Intersection over Union (mIoU) of 0.8609 and an RTS-class IoU of 0.7473. The qualitative visual evaluation further confirms MOCA-Net’s improved performance in delineating RTSs through more accurate morphologies and boundaries. Ablation studies confirm that each designed component contributes to the MOCA-Net’s segmentation performance, and their combination yields more balanced results. This model unlocks the capability of Sentinel-2 imagery for accurate RTS segmentation, making it promising for applications over large spatiotemporal extents.

IPC Classification

Keywords