Lake topography, which serves as a crucial basis for water resource monitoring, has been extensively applied in hydrological and geomorphological research. However, monitoring lake dynamics in data-scarce regions remains challenging due to the limited revisit frequency of altimetry satellites and uncertainties in estimating submerged depths based on surrounding terrain. Focusing on the Shapotou region of Ningxia, this study collected bathymetric data using an unmanned surface vessel and employed interpolation methods and machine learning (XGBoost) to determine the most effective approach for constructing the lake digital elevation model (DEM). The relationship curve between water area, water level, and water reserve, derived from DEM analysis and calculation, is integrated with remote sensing imagery, thus facilitating the efficient monitoring of lake water dynamics. This approach provides valuable scientific and practical support for water resource management in remote or data-scarce regions lacking conventional hydrological infrastructure. The results indicate that: (1) Despite the theoretical potential of machine learning for underwater terrain prediction, this study demonstrates that traditional spatial interpolation methods offer greater advantages in areas characterized by data scarcity and significant anthropogenic terrain modification, thereby providing empirical evidence to guide method selection under similar conditions. (2) The average annual water storage of lakes in the study area was estimated at 336.249 x 104 m3 by integrating relationship curves with remote sensing imagery. Total storage reached a minimum of 307.246 x 104 m3 in 2016 and a maximum of 411.802 x 104 m3 in 2024. Water level fluctuations were generally less than 1 m, revealing the relative stability of lakes in arid regions.