In hydrologic modeling, forecasting peak floods is a complex and high-priority task, particularly in regions with limited in situ observations. This research investigates the potential of utilizing satellite-based precipitation products with near real-time updates (NRT-SbPP), together with deep learning, to improve peak flow estimation and flood forecasting. Using Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG), the Final version as a reference, four NRT-SbPPs were evaluated: Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN)-Cloud Classification System (CCS), PERSIANN-Dynamic- Infrared– RainRate (PDIR), the IMERG-Early, and IMERG-Late. This study evaluated the performance of these datasets for streamflow simulation using the Hydrologic Modeling System (HEC-HMS) in a sub-basin of the Russian River watershed. Bias correction was performed using Long Short-Term Memory (LSTM) networks, with IMERG-Final serving as the reference precipitation dataset. This correction improved the quality of the input rainfall data and led to more accurate streamflow simulations. For example, the Root Mean Square Error of IMERG-Late decreased from 1.98 to 1.48 mm with LSTM, resulting in performance metrics similar to those of observed discharge (Nash-Sutcliffe Efficiency: 0.82). Most importantly, PDIR exhibited significant enhancement, with a 36% correlation coefficient increase (from 0.52 to 0.81), as well as high rates for extreme event detection. The further findings establish the potential of employing LSTM methods and using IMERG-Final as a reference to incorporate NRT-SbPPs within real-time flood forecasting and early warning frameworks. This method leverages the new technologies of satellite-based meteorological data and offers an efficient, cost-effective option to enhance flood prediction and disaster risk management, especially in data-scarce regions.

Keywords