Abstract: Using the ICESat−2 as a reference data source, the study analyzed the overall elevation accuracy of NASA's Digital Elevation Model(DEM) in the complex terrain of the central and western parts of the Hengduan Mountains. It also investigated the relationship between DEM errors and surface variables such as slope, aspect, roughness, vegetation cover, and land use types. Subsequently, to correct elevation errors in NASA DEM, 4 machine learning methods were introduced: Random Forest, Back Propagation Neural Network, Extreme Gradient Boosting, and Light Gradient Boosting Model. These were combined with surface variables to construct an elevation error correction model, and the corrected NASA DEM was analyzed for accuracy. The results showed that the average elevation error of NASA DEM was –3.5 m, the mean absolute error was 8.2 m, and the root mean square error was 11.7 m, with elevation errors following a normal distribution. The elevation accuracy of NASA DEM decreased with increasing slope, roughness, and vegetation cover, and varied significantly across different aspects and land use types. All 4 machine learning models were able to correct the elevation accuracy of NASA DEM, improving the average error, mean absolute error, and root mean square error; the average errors after correction by RF, XGBoost, and LightGBM were all 0 m; the lowest mean absolute error after BPNN correction was 6.9 m; the lowest root mean square error after BPNN and XGBoost correction was 10.6 m; BPNN achieved the highest R² value post-correction; and BPNN showed the best correction accuracy across different land use types. These results provide precision references for the selection and application of NASA DEM in geoscience research, and through the introduction of various machine learning models, offer effective methodologies for the elevation error correction research of NASA DEM and other global DEM products.