Estimating Terrain Elevations at 10 M Resolution by Integrating Random Forest Machine Learning Model and ICESat-2, Sentinel-1, and Sentinel-2 Satellite Remotely Sensed Data

Accurate mapping of terrain elevations at a large scale and fine resolution can characterize the detailed surface height and geomorphic changes and is very critical for the studies of the internal motions and external forces of the earth. The emergence of the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) offers unprecedented possibilities for global elevation mapping with high vertical accuracy using three-dimensional photon points. However, the ICESat-2 photon points are still sparse in terms of spatial/horizontal resolution, making it unable to satisfy the high-resolution demand of terrain elevation mapping and digital elevation model production. A few previous studies have attempted to estimate elevations/topography in regions with single landscape and landcover (e.g., forest, shallow water, and polar regions) by combining ICESat-2 data with other passive satellite remotely sensed data. However, the potential and capability of ICESat-2 for mapping elevations for spatially continuous large regions with multiple complicated land cover types remains unknown. In this study, a spatially continuous large-scale terrain elevation estimation method is developed under multiple land covers based on the random forest model and the freely accessed satellite data of ICESat-2, Sentinel-1, and Sentinel-2. The core principle is to construct a random forest model that can characterize the complicated relationships of the ICESat-2 ATL03 terrain elevations and their corresponding land cover related polarization characteristics and spectral variables from Sentinel-1 and Sentinel-2, respectively. Integrating the superiorities of the data of these three different satellites enables the proposed method to extrapolate the terrain elevations with decimeter-level vertical accuracy and 10 m spatial/horizontal resolution simultaneously without any prior in situ data or manually set parameters. The proposed method is tested using the elevations from 2021 to 2022 at the third largest island (Chongming Island, Shanghai) in China. The estimated terrain elevations are locally validated with the airborne LiDAR-derived elevations. Moreover, they are compared with the ICESat-2 ATL08 height_terrain_bestfit data and Global Ecosystem Dynamics Investigation L2A elev_lowestmode data from the global perspectives. The predicted elevations exhibit a high correlation with the measured elevations from the two airborne LiDAR validation regions with root mean square errors (RMSE) of 0.34 and 0.59 m. The averaged RMSEs of the predicted elevations at different land covers are 1.26 and 1.18 m when compared with those derived from ATL08 and GEDI L2A, respectively. No remarkable abnormal predicted elevations are observed. This finding suggests the satisfactory robustness performance of the proposed method under different land covers and a relatively good consistency between the predicted elevations and the actual terrain of the entire island. As far as we know, the present work is the first to map elevations at 10 m resolution based only on the newly available satellite active and passive remotely sensed data without any ground truth surveys, manual intervention, and prior knowledge. Different with existing studies for terrain elevation mapping only at single landcovers, the proposed method demonstrates the capability and effectiveness of ICESat-2 for any landforms and landcovers and shows great potential for high-accuracy and high-resolution time-series terrain elevation estimation and updating at regional/national/global scales.