Ho Yu-Feng, Grohmann Carlos H, Lindsay John, Reuter Hannes I, Parente Leandro, Witjes Martijn, Hengl Tomislav
OpenGeoHub, Doorwerth, Gelderland, Netherlands.
Institute of Astronomy, Geophysics and Atmospheric Sciences, Universidade de São Paulo, São Paulo, Brazil.
PeerJ. 2025 Jul 24;13:e19673. doi: 10.7717/peerj.19673. eCollection 2025.
Production and validation of an open global ensemble digital terrain model (GEDTM30) and derived terrain variables at 1 arc-s spacing grid ( 30 m spatial resolution) is described. Copernicus DEM, ALOS World3D, and object height models were combined in a data fusion approach to generate a globally consistent digital terrain model (DTM). This DTM was then used to compute 15 standard terrain variables across six scales (30, 60, 120, 240, 480 and 960 m). A global-to-local transfer learning model framework with 5° × 5° tiling leveraged globally distributed lidar datasets: ICESat-2 ATL08 (best-fit terrain height) and GEDI02 (lowest mode elevation), totaling over 30 billion training points. A global model was initially fitted using ICESat-2 and GEDI, followed by locally optimized models per tile, ensuring both global consistency and local accuracy. Independent validation shows that GEDTM30 reduces Copernicus DEM RMSE by about 25.4% in built-up areas, 10.0% in regions with 10-50% tree cover, and 27.3% in areas with over 50% tree cover. Compared to state-of-the-art DTMs (MERIT DEM, FABDEM and FathomDEM), GEDTM30 achieves the lowest vertical errors when assessed with GNSS station records, yielding a standard deviation of 7.77 m, an RMSE of 10.69 m, and a mean error of 7.34 m. FathomDEM exhibited the lowest vertical RMSE when validated against independent reference DTMs. GEDTM30 was further used to generate multiscale variables of topography and hydrology through an optimized tiling workflow ( 800 tiles of 600 × 600 km with 16% overlap) based on the Equi7 grid system. The entire workflow was implemented in Python using GDAL and Whitebox Workflows. Visual inspection confirmed the absence of boundary artifacts and the preservation of hydrologic connectivity. The tiling-based implementation significantly reduces computational costs of generating large-scale DTMs and derived terrain variables. The GEDTM30 dataset and code are publicly available as Cloud-Optimized GeoTIFFs Zenodo and the OpenLandMap STAC. Further fusion with local lidar-based DTMs and national DTMs is recommended to enhance local accuracy and level of detail.
描述了一种开放的全球集合数字地形模型（GEDTM30）及其在1弧秒间距网格（30米空间分辨率）下派生地形变量的生成与验证。哥白尼数字高程模型（Copernicus DEM）、先进陆地观测卫星世界3D（ALOS World3D）和物体高度模型通过数据融合方法相结合，以生成全球一致的数字地形模型（DTM）。然后使用该DTM在六个尺度（30、60、120、240、480和960米）上计算15个标准地形变量。一个采用5°×5°分块的全局到局部迁移学习模型框架利用了全球分布的激光雷达数据集：ICESat-2 ATL08（最佳拟合地形高度）和GEDI02（最低模式海拔），总计超过300亿个训练点。最初使用ICESat-2和GEDI拟合一个全局模型，随后对每个分块进行局部优化模型，确保全局一致性和局部准确性。独立验证表明，GEDTM30在建成区将哥白尼数字高程模型的均方根误差（RMSE）降低了约25.4%，在树木覆盖率为10%-50%的区域降低了10.0%，在树木覆盖率超过50%的区域降低了27.3%。与现有最先进的DTM（MERIT DEM、FABDEM和FathomDEM）相比，在使用全球导航卫星系统（GNSS）站记录进行评估时，GEDTM30实现了最低的垂直误差，标准差为7.77米，RMSE为10.69米，平均误差为7.34米。在与独立参考DTM进行验证时，FathomDEM表现出最低的垂直RMSE。GEDTM30还通过基于Equi7网格系统的优化分块工作流程（800个600×600千米的分块，重叠率为16%）用于生成地形和水文的多尺度变量。整个工作流程使用GDAL和Whitebox Workflows在Python中实现。目视检查确认没有边界伪影且水文连通性得以保留。基于分块的实现显著降低了生成大规模DTM和派生地形变量的计算成本。GEDTM30数据集和代码以云优化地理TIFF格式在Zenodo和OpenLandMap STAC上公开可用。建议进一步与基于局部激光雷达的DTM和国家DTM进行融合，以提高局部准确性和细节水平。
