Convolutional neural networks-driven bias correction of satellite precipitation improves rainfall-runoff-inundation modeling

A convolutional neural network (CNN)-based framework is developed to address systematic biases inherent in satellite precipitation products. The CNN-based model generated a bias-corrected gridded precipitation dataset across Cambodia between 0.05° grid resolution from the Climate Hazards Infrared Precipitation with Station data (CHIRPS) for 24 years (1985–2008). Then, this study coupled these two datasets, the CHIRPS and Corrected-CHIRPS dataset, into the Rainfall-Runoff-Inundation (RRI) model to replicate river discharge and flood inundation in the Tonle Sap Lake Basin (TSLB). With better spatial and temporal correlations, this study observed significant bias reduction with the KGE(RMSE) values from 0.04 (715.97 mm) in 2007 to 0.87 (170.03 mm) decreased by 76 %, and from 0.07 (510.06 mm) in 2008 to 0.87 (152.80 mm) decreased by 70 %. Additionally, the accuracy of the RRI model improved during the simulation period (2000–2008) at all five stations (S1-S5); on average, NSE, RSR, and R 2 were 0.73 (0.85), 51.69 (37.81), and 0.78 (0.88) for CHIRPS (Corrected CHIRPS), respectively. A CNN-based approach for a more accurate and contemporary precipitation dataset is presented to benefit Cambodia's hydrological modeling and flood response strategies.