A comprehensive high-resolution lake inventory across the highly heterogeneous Tibetan Plateau from 2020 Sentinel-2 imagery using deep learning

Accurate, large-scale, and temporally explicit lake mapping is critical for water resource management, hazard risk assessment, and understanding lake responses to climate change. The Tibetan Plateau (TP) hosts a high density of lakes, many of which are small and difficult to detect due to highly heterogeneous environments, including mountain shadows, glacier and snow cover, clouds, and turbid waters. Existing studies often produce multi-year composite datasets, which enhance the detectability of lakes by emphasizing their long-term occurrence. However, these datasets ultimately remain static, reflecting the long-term aggregated distribution of lakes without capturing their spatial extent in any specific year. In this study, we processed 17,942 Sentinel-2 images acquired from July to October 2020 using an automated, deep learning–based water extraction framework to extract lakes across the TP. The method demonstrated high accuracy under challenging conditions, with Intersection over Union (IoU) values exceeding 92% in cloudy, glacial, and mountainous test areas. Applying this framework, we generated a comprehensive 2020 lake inventory, identifying 57,841 lakes larger than 0.01 km2. Approximately 65% of the total lake coverage was concentrated in the Inner Plateau Basin, with the 4,500–5,000 m elevation exhibiting the highest density (29,200 lakes covering 32,892.51 km2, representing 50.5% of all lakes and 59.3% of total lake area). Compared with previous studies, this dataset improves the detection of small lakes and provides a temporally explicit, year-specific map of lake distributions, distinguishing multiple lake types (large natural lakes, glacial lakes, thermokarst lakes, and reservoirs). This high-resolution, comprehensive dataset constitutes a valuable resource for hydrological, climatic, and ecohydrological research across TP. The lake dataset generated in this study has been archived and made publicly available through Zenodo (https://doi.org/10.5281/zenodo.15639602).