Water depth inversion based on ICESat-2 and Sentinel-2—A case study of Qiagui Co and Ayakekumu Lake on the Tibetan Plateau

Two lakes on the Tibetan Plateau. This study utilizes ICESat-2 data to extract water depth, integrates Sentinel-2 and various machine learning models to invert water depth, and evaluates the accuracy of these models, ensuring robust validation of the results. An improved OPTICS denoising algorithm was applied to extract lake water depths from the ICESat-2 data. The water depths from ICESat-2 and the reflectance data from Sentinel-2 were used to construct the water depth inversion with eight machine learning models, including Quadratic Polynomial, SVR, XGBoost, LightGBM, RF, MLP, Transformer and KAN. The extracted maximum depth of Qiagui Co and Ayakekumu Lake was 14.14 m and 15.96 m, respectively, and the accuracy was high with low RMSE (0.356–0.369 m) by comparing with in-situ bathymetric data. Among the machine learning models, the KAN model exhibited the best inversion accuracy (RMSE: 0.789–0.952 m), followed by the MLP and Transformer models, and the SVR model was the poorest (RMSE: 0.893–1.253 m). Comparison of the water storage changes from the KAN model and SRTM in different periods since 2000, suggested that the accuracy of the KAN model was high with an average error of 12.8% in ∼7 m water depth. This study provides new insights into the lake depth extraction, water storage and change estimation based on the ICESat-2 and Sentinel-2 imagery data. • Water depth extraction from ICESat-2 data showed good agreement to in-situ measurements with a RMSE of 0.356–0.369 m. • Eight methods were used to construct lake water depth inversion model with water depth from ICESat-2 and Sentinel-2 images. • The KAN model outperformed the other methods with an RMSE of 0.789–0.952 m. • The KAN model provided a precise method for water storage changes monitoring with an average error of 12.8 %.