An ECDF-based Method for Generating DEM Training Data Simulating Sub-waterbody Elevation
Open Access
- 31 December 2025
- journal article
- Published by GeoAI Data Society in GEO DATA
- Vol. 7 (4) , 811-823
Abstract
Satellite-derived digital elevation models (DEMs) are indispensable for hydrologic analysis, yet riverbed topography beneath water bodies is systematically missing because of water reflectance/dispersion and water-level variability between satellite revisits. To address this structural limitation, we propose an automatic dataset construction method for DEM inpainting. For each DEM patch, the empirical cumulative distribution function of elevations is compared with a linear baseline; the threshold is chosen at the knee point where their vertical deviation is maximized. Elevations at or below this threshold are masked as likely water-occupied terrain, while the original DEM serves as ground truth. Using the Copernicus DEM GLO-30, we sample patches from 30°S to 30°N, enforce a water-area ratio of 5-15%, and apply morphological post-processing. A patch size of 64×64 is adopted, yielding 10,000 input-ground-truth pairs in GeoTIFF format (EPSG:4326). Dataset validity is assessed with a lightweight multi-scale U-Net, evaluation is restricted to the masked (missing) regions. On the test set, the model achieves PSNR 13.44 dB, SSIM 0.78, and MAE 2.80 m. Qualitative analyses show stable reconstruction of topographic continuity near shorelines, with most vertical errors confined to within ±5 m around boundaries. The proposed approach enables standardized, automated generation of large training datasets that reflect local terrain characteristics without field-grade ground truth, providing a practical basis for hydrologically complete DEMs for riverbed restoration, water-level estimation, and large-scale hydrologic simulation. Future work will extend to mid- and high-latitude regions and incorporate hydrological-consistency constraints to jointly improve structural similarity and absolute accuracy.Keywords
Funding Information
- Ministry of Environment
- Institute for Information and Communications Technology Promotion (RS-2024-00397964)
- Ministry of Science and ICT
This publication has 12 references indexed in Scilit:
- Deep learning-based bathymetry retrieval without in-situ depths using remote sensing imagery and SfM-MVS DSMs with data gapsISPRS Journal of Photogrammetry and Remote Sensing, 2025
- Comparison of different DEMs for hydrological studies in the mountainous areasThe Egyptian Journal of Remote Sensing and Space Sciences, 2021
- Sensitivity of Bistatic TanDEM-X Data to Stand Structural Parameters in Temperate ForestsRemote Sensing, 2019
- Morphodynamics of rivers strongly affected by monsoon precipitation: Review of depositional style and forcing factorsSedimentary Geology, 2015
- Approach to Digital Elevation Model Correction by Improving Channel ConveyanceJournal of Hydrologic Engineering, 2015
- Impact of DEM accuracy and resolution on topographic indicesEnvironmental Modelling & Software, 2010
- TanDEM-X: A Satellite Formation for High-Resolution SAR InterferometryIEEE Transactions on Geoscience and Remote Sensing, 2007
- An evaluation of void‐filling interpolation methods for SRTM dataInternational Journal of Geographical Information Science, 2007
- Effects of Digital Elevation Model Accuracy on Hydrologic PredictionsRemote Sensing of Environment, 2000
- A Generalization of the Glivenko-Cantelli TheoremThe Annals of Mathematical Statistics, 1959