GRDL: A new Global Reservoir area-storage-depth dataset derived through Deep Learning-based bathymetry reconstruction

Abstract: Reservoirs play a critical role in the global water cycle by regulating the flow of water from the environment into human systems. Accurate estimation of the area-storage-depth relationships for global reservoirs is essential for effective hydrological modeling and reservoir storage monitoring. Bathymetry reconstruction presents a promising approach to derive this information. Current bathymetry methods either rely on simple approximations or are constrained by dependence on altimetry data or field survey data. To overcome these limitations, this study presents a pioneering approach involving training a deep learning model to reconstruct bathymetry and establish precise area-storage-depth relationships. We trained the deep learning model with approximately 160,000 simulated labels derived from SRTM and fine-tuned the model based on the DEM in hundreds of reservoirs that were constructed after 2000. By employing the trained model and SRTM, the bathymetry of 7,250 reservoirs in the Global Reservoir and Dam Database (GRanD) were subsequently reconstructed. The method is validated against comprehensive reference datasets, including 54 test reservoirs with full bathymetry profiles, 118 satellite altimetry-based reservoirs, and 68 LiDAR-based reservoirs. The reconstructed bathymetry achieves a mean absolute error (MAE) of 7.87 m and a mean error (ME) of +2.05 m for the test reservoirs. The validation against satellite altimetry and LiDAR-based references significantly outperforms previous geometric approximation techniques, with median normalized root mean square error (NRMSE) values of 20.6% for area-storage and 22.1% for area-level curves. Additionally, the storage variations of reservoirs are estimated with precision, outperforming previous methods. The proposed deep learning-based approach presents a robust solution for accurate reservoir bathymetry estimation and establishes more reliable area-storage-depth relationships for reservoirs worldwide. Dataset description: The dataset includes files (GRAND_ID.csv) containing area (km2)-storage (million cubic meters)-depth (m) profiles for each reservoir, with maximum area derived from JRC GSW maximum extent data and depth starting from the lowest elevation of the reconstructed bathymetry. A separate file (grand_area_ratio.csv) provides the area coverage ratio for each reservoir, with values closer to 0 indicating higher quality as described in the paper