<p>Based on GEDTM120-meter terrain data, terrain data of six regions: Xinjiang, Kazakhstan, Tajikistan, Kyrgyzstan, Uzbekistan and Turkmenistan are obtained after mask processing, with a data resolution of 120 meters. </p>
This dataset was released by He Yufeng's team from OpenGeoHub in the Netherlands on the Zenodo website in February 2025. It has been updated in four versions since February this year. This dataset is simply referred to as GEDTM30. Since the 30-meter dataset has not yet been officially published, the data papers are being reviewed on the Research Square platform. The data author has only opened up the 120-meter DEM terrain data, and the 30-meter terrain data will not be available until the review is completed.
The author of the paper proposed a global-local migration learning framework to generate a gap-free global DEM (GEDTM30) with a resolution of 1 arcsec (about 30m), and extracted 15 types of surface morphology and hydrological parameters such as slope, curvature, and catchment area. Research integrates multi-source heterogeneous data (optical imaging, radar, lidar), generates initial DEM through median filtering and outlier removal, and combines ICESat-2 (2 billion points) and GEDI (1 billion points) lidar data to build training samples, and uses GNSS station data to verify model accuracy. For areas with sparse data (such as Pacific islands), innovatively adopt a global-local migration learning strategy: the global model is based on 10% random samples to train random forests (RF) to capture common characteristics, and the local model is segmented by 5°×5°, integrating localized samples such as elevation anomalies and vegetation coverage fine-tune parameters to significantly improve regional adaptability. In the surface parameterization stage, Whitebox Tools combined with the Equi7 projection system is used to realize efficient calculation of multi-scale parameters (30 - 960m), optimize hydrological connectivity and calculation efficiency, and finally generate a global terrain benchmark dataset with both vertical accuracy (forest area RMSE 23.2 m) and regional adaptability (50% reduction in coastal error).
The RMSE in forest area is 23.2 m, and the coastal error is reduced by 50%.
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
10.12389/ariddc.01156
