A GRDC observation-constrained terrestrial water budget dataset under a discrete global grid framework (2000-2020)

By introducing observed basin outlet runoff data from the Global Runoff Data Centre (GRDC), this study used spatial overlay and area-weighted aggregation between basin boundaries and ISEA3H10 grid cells to convert GRDC basin outlet runoff observations into supervised information usable by the gridded model. A physics-constrained deep learning model was then applied to jointly optimize precipitation, evapotranspiration, runoff, and terrestrial water storage change.

The resulting data product improves the physical consistency among different water-cycle variables while preserving major hydrological signals and spatiotemporal variability.Four data products were generated under different loss-weighting schemes: Hex0.05, Hex0.15, Hex0.25, and Hex0.5. Among them, Hex0.15 shows the most balanced performance, effectively reducing water balance residuals while maintaining consistency with the original data characteristics.

Compared with direct combinations of existing multi-source water-cycle products, this dataset substantially improves water balance consistency. Among 672 original combinations of precipitation, evapotranspiration, runoff, and terrestrial water storage change products, the best-performing combination in terms of closure achieved a global monthly water balance residual RMSE of 43.37 mm/month. In contrast, the optimized product reduced the residual RMSE to 11.77 mm/month, with a global mean residual of −0.44 mm/month.

The optimized gridded runoff shows a correlation coefficient close to 0.90 with GRDC observed runoff, and its RMSE is reduced by approximately 31%–49% compared with the original runoff proxy products. Independent validation based on FLUXNET sites further indicates that the optimized precipitation and evapotranspiration maintain accuracy comparable to mainstream input products or show more stable error characteristics.The dataset is released in both Zarr and shapefile formats and can be opened and processed using common scientific computing and geospatial analysis software such as Python and QGIS.

Each scheme has a data volume of approximately 5 GB, and the total volume of the four schemes is approximately 20 GB. Users may select data products generated under different loss-weighting schemes according to their research needs, with Hex0.15 recommended as the scheme that best balances water balance consistency and preservation of original data characteristics.