Improved wall-to-wall DEM in the subtropical and tropical regions of China by optimized ICESat-2 photon processing and machine learning

Overview This dataset provides a high-precision, wall-to-wall Digital Elevation Model (DEM) for the vast subtropical and tropical regions of southern China. It was developed to address the severe terrain underestimation and systematical bias commonly found in existing global DEM products under dense forest canopies and steep topographic conditions. Methodology The DEM was generated through a novel, multi-stage framework that integrates spaceborne LiDAR and machine learning: Optimized Ground Photon Extraction: We proposed an advanced pipeline integrating ICESat-2 ATL08 and ATL03 data. We successfully eliminated pseudo-ground returns and accurately recovered true ground photons beneath complex, highly occluded forest structures. Machine Learning Error Modeling: The highly accurate ground elevations extracted from ICESat-2 were used as target variables. Using advanced machine learning algorithms, combined with spatial covariates including NASADEM and high-dimensional environmental features from Google Embedding Datasets, we modeled and corrected the systematic errors across the entire study area to produce this spatially continuous bare-earth terrain model. Dataset Characteristics Study Area: Subtropical and tropical regions of southern China (covering provinces such as Fujian, Zhejiang, Anhui, Guangxi, Yunnan, Hainan, etc.). Spatial Resolution: 1 arc-second (~30 meters). Data Format: GeoTIFF (.tif). Coordinate Reference System (CRS): WGS 84 (EPSG:4326).