DL-Driven Precipitation Correction for Enhanced Hydrological Simulations over Central Europe

Integrated hydrologic models are useful for assessing the impact of climate change on water resources and associated risks. The performance of these models strongly depends on the quality of precipitation forcing data, where errors can significantly affect the simulation accuracy. Therefore, methods such as data assimilation (DA) bias adjustments, and data-driven (e.g., deep learning, DL) methods are in use to improve precipitation simulation data. However, given the high spatiotemporal variability of hourly precipitation, challenges such as availability of “ground truth” measurements, data imbalance, and evaluation of the methods affect the applicability and assessment of these methods. In this study, we correct precipitation data for the first 24h obtained from the  ECMWF HRES 10-day deterministic forecast using EUMETSAT H-SAF h61 satellite observations, by learning the errors using a U-Net convolutional neural network (CNN) as a DL technique. Our findings show good agreement between the corrected precipitation data (HRES-C) and the reference data (H-SAF) with roughly about 49%, 33%, and 12% improvement in mean error, root mean square error, and Pearson correlation, respectively. Additionally, we investigate the impact of original HRES, H-SAF, and HRES-C corrected products used as forcing data in high-resolution (~0.6km) integrated hydrologic simulations using ParFlow/CLM over central Europe in daily and monthly scales from April 2020 to December 2022. We choose soil moisture (SM) as a diagnostic variable for our evaluation. SM simulations produced with uncorrected HRES 24h show a better agreement with ESA CCI SM satellite data compared to SM produced with HRES-C. Further comparison of the three products with in-situ rain gauge measurements over the same period shows superiority of HRES 24h in representing the “ground truth” precipitation.  Our study highlights the need for better precipitation reference data, challenging reliance only on satellite observations (H-SAF) for DL-based correction of precipitation forcing data in hydrological simulations.