Key points are not available for this paper at this time.
Accurate mapping of thermokarst lakes in the Qinghai–Tibet Engineering Corridor is important for understanding permafrost degradation and assessing environmental risks to major infrastructure. However, thermokarst lake extraction from medium-resolution satellite imagery remains challenging because these lakes are often small, fragmented, seasonally variable, and spectrally confused with wetlands, shadows, and other surface water bodies. In this study, Sentinel-2 imagery from the 2020 thaw season was used to map thermokarst lakes in the Qinghai–Tibet Engineering Corridor. A 16-feature dataset was constructed by integrating spectral bands, water indices, texture features, and topographic variables, and a convolutional neural network (CNN) was compared with five conventional machine learning classifiers. In the pixel-based validation, the CNN slightly outperformed the other evaluated models, achieving the overall accuracy of 98.04% and an F1-score of 97.18%. Independent polygon-based validation using the Jilin-1 visual interpretation reference showed that the final CNN-derived inventory achieved an IoU of 0.79, with omission and commission ratios of 0.15 and 0.09, respectively. The CNN more effectively suppressed salt-and-pepper noise, reduced fragmented and serrated lake boundaries, and improved the spatial continuity of mapped water bodies compared with traditional machine learning classifiers. SHAP-based attribution suggested that CNN predictions were more strongly associated with water indices and texture features, whereas Random Forest predictions were mainly associated with near-infrared and shortwave-infrared bands. Thermokarst lakes showed higher lake area proportions and densities in areas with relatively high ground ice content, gentle slopes, thicker active layers, warmer permafrost, and unstable permafrost conditions. These results demonstrate the potential of Sentinel-2 imagery and convolutional models for regional thermokarst lake mapping and permafrost degradation monitoring.
Ma et al. (Mon,) studied this question.