ASTER GDEM Correction Based on Stacked Ensemble Learning and ICEsat-2/ATL08: A Case Study from the Qilian Mountains

ASTER GDEM provides the fundamental data for remote sensing identification of snow cover in mountainous areas. Due to its elevation accuracy being easily affected by optical stereo images and local terrain, many studies have utilized machine learning (ML) models for correction. However, most correction methods rely on a single ML model, which limits the improvement of DEM accuracy. Stacked ensemble learning (SEL) is a newly developed method of improving model performance by combining multiple ML models. This study proposes a DEM correction method based on SEL and ICESatand affiliations. -2/ATL08 products. Taking the Babao River Basin in Qilian Mountains as the study area, five ML models with good DEM correction effects (XGBoost, AdaBoost, LightGBM, BPNN, and CatBoost) were selected and trained using land cover and various terrain factors to obtain DEM errors, respectively. Then, the SEL algorithm was used to integrate the DEM errors of the five ML models and correct GDEM. Using 740 CORS measurements and 48,000 ATL08 points for accuracy validation, the results showed that the SEL achieved higher DEM accuracy than any single ML model. The root mean square error (RMSE) of the corrected GDEM decreased from 7.15 m to 4.13 m, while the mean absolute error (MAE) and mean bias error (MBE) values both decreased about by 38%. Furthermore, unmanned aerial vehicle (UAV) DEM data from five sample areas were selected for profile analysis, and it was found that the corrected GDEM was closer to the real surface. Further analysis revealed that the influence of slope, aspect, and land cover types on corrected DEM was weakened, with the most significant improvement in DEM accuracy observed in areas with slope ≥5°, north orientation, and bare land. This study can provide high-precision DEM scientific data for quantitative remote sensing, flood prediction, and other research.