Reconstruction and inversion of lake water depth based on ICESat-2 photon and Sentinel-2 — A case study of Caiduochaka Lake on the Tibetan Plateau

Caiduochaka Lake (CK) on the Tibetan Plateau (TP). This study reconstructs the bathymetry of CK by integrating Sentinel-2's large-area, spatially continuous spectral data with precise but spatially discontinuous depth references from ICESat-2. Three machine learning models were developed to invert water depth from Sentinel-2 spectral reflectance, trained on ICESat-2-derived and in situ bathymetry. The aim is to evaluate whether ICESat-2-derived bathymetry can substitute field measurements, and to demonstrate the combined value of both datasets for robust, large-scale bathymetric mapping. Reconstruction of water depth in CK revealed a maximum depth of 14.73 m and an average depth of 3.90 m, with ICESat-2-derived bathymetry showing strong agreement with in situ bathymetry (R 2 =0.985, RMSE=0.534 m). When using ICESat-2-derived bathymetry as training data, the KAN model yielded the best performance (R 2 =0.911, RMSE=1.064 m), whereas the RF model trained on in situ bathymetry achieved the highest overall accuracy. The bathymetry and water storage estimates obtained from both approaches were highly consistent, indicating that ICESat-2-derived bathymetry can reliably substitute for traditional field measurements in shallow water areas. From 2000–2023, lake water storage increased by 0.529–0.555 km 3 , reflecting significant long-term hydrological changes in the region. These findings provide a robust and scalable approach for reconstructing lake water depth, monitoring lake water resources, and evaluating hydrological responses to climate variability on the TP. • Extracted the water depth from ICESat-2 photons with RMSE of 0.811 m. • Established water depth inversion models with ICESat-2 and Sentinel-2 data. • The KAN model achieved an RMSE of 1.064 m in water depth inversion. • ICESat-2 enables shallow water depth reconstruction without in situ bathymetry.