Integrated hydrological model (IHM) forecasts provide critical insights into hydrological system states, fluxes, and its evolution of water resources and associated risks, essential for many sectors and stakeholders in agriculture, urban planning, forestry, or ecosystem management. However, the accuracy of these forecasts depends on the data quality of the precipitation forcing data. Previous studies have utilized data-driven methods, such as deep learning (DL) during the preprocessing phase to improve precipitation forcing data obtained from numerical weather prediction simulations. Nonetheless, challenges related to the spatiotemporal variability of hourly precipitation data persist, including issues with ground truth data availability, data imbalance in training DL models, and method evaluation. This study compares three (near) real-time spatiotemporal precipitation datasets to be used in the aforementioned IHM forecast systems: (1) 24 h precipitation forecast data obtained by ECMWF’s 10-day HRES deterministic forecast, (2) H-SAF h61 satellite observations as reference, and (3) DL-based corrected HRES precipitation using a U-Net convolutional neural network (CNN). As high-resolution data, H-SAF is used both as a reference for correcting HRES precipitation data and as a stand-alone candidate for forcing data. These datasets are used as forcing data in high-resolution (~0.6 km) integrated hydrologic simulations using ParFlow/CLM over central Europe from April 2020 to December 2022. Soil moisture (SM) simulations are used as a diagnostic downstream variable for evaluating the impact of forcing data. The DL-based correction reduces the gap between HRES and H-SAF by 49, 33, and 12% in mean error, root mean square error, and Pearson correlation, respectively. However, comparison of SM simulations obtained from the three datasets with ESA CCI SM data reveals better agreement with the uncorrected HRES 24-h forecast data. In conclusion, H-SAF satellite-based precipitation data falls short in representing precipitation used for SM simulations compared to 24 h lead time HRES forecasts. This emphasizes the need for more reliable spatiotemporally continuous high-resolution precipitation observations for using DL correction in improving precipitation forecasts. The study demonstrates the potential of DL methods as a near real-time data pre-processor in quasi-operational water resources forecasting workflows. The quality of the preprocessor is directly proportional to the quality of the applied observation.
