Modeling lake bathymetry and water storage from DEM data constrained by limited underwater surveys

Lake bathymetry, which provides crucial information for water resource management, has been widely used in hydrological, ecological, and geomorphological studies. Restricted by the high cost of conventional full-covered lake field surveys and the large uncertainty of spatial prediction method that models lake underwater depths from the surrounding exposed terrains, the knowledge on lake bathymetry worldwide is scarce and inconsistent. This study aims to solve the problem by investigating how the optimized spatial prediction methods are developed by combining limited field survey data. Two methods, namely, the skeleton-based interpolation method that extends the exposed topography toward the lake underwater area with a constraint of field surveys, and the machine learning method (XGBoost) that establishes the decision rule between measured water depths and multiple geospatial variables to predict depths of unknown underwater areas, were developed and tested in twelve representative lakes in the Tibetan Plateau. Our results suggest that both methods can provide acceptable estimations of underwater topography by comparing with measured data, with the mean R2 of 0.70 approximately. The overall performance of the machine learning method (XGBoost) is more reliable, with the biases less than 20% in water volume estimates for all lake cases. In comparison, the skeleton-based interpolation method outperforms lakes in long and narrow shapes. This study is expected to provide an efficient approach for modeling lake bathymetry and to improve the monitoring capacity of freshwater mass changes, especially for ungauged lakes in harsh environments.