Thermokarst Lake Extraction and Dynamics With scSE‐U‐Net and Polarimetric Sentinel‐1 Images: A Case Study in the Yellow River Source Region

ABSTRACT Thermokarst lakes are sensitive surface expressions of permafrost degradation and hydrological reorganization in alpine permafrost regions, but their small size and complex surrounding environments make accurate mapping difficult. Accurate extraction of these lake is fundamental for analysing their spatiotemporal dynamics. We develop a semantic segmentation model based on scSE‐U‐Net (U‐Net with Concurrent Spatial and Channel Squeeze and Excitation), utilizing polarimetric Sentinel‐1 images. VGG16 was integrated into the U‐Net architecture, and a scSE attention mechanism was introduced to improve feature representation for small water bodies and complex backgrounds. Additionally, this study investigates the variations in lake water and their associations with environmental factors in the Yellow River Source Region (YRSR), with particular attention to thermokarst lakes. Compared with the baseline U‐Net, scSE‐U‐Net achieved moderate improvements, with mIoU increasing by 0.60% for large lakes and 0.09% for small lakes, and mPA increasing by 0.25% for both lake groups. The fused VV+VH+PCA feature scheme produced better lake extraction results, with mIoU and mPA values of 98.33% and 99.23% for large lakes, and 94.16% and 96.01% for small lakes, respectively. Our analysis reveals that the lake areas in the YRSR from 2014 to 2023 initially diminished, then gradually expanded and later exhibited slight fluctuations. Changes in lake number were mainly driven by thermokarst lakes, whereas the number of large lakes remained relatively stable. Lake‐area variations were positively correlated with precipitation and temperature, suggesting that climate warming and changes in water supply conditions may jointly influence thermokarst lake dynamics through permafrost degradation, seasonal snowmelt and surface‐water redistribution.