Downscaling and bias-correcting satellite precipitation using a hybrid machine learning framework for flood modelling in ungauged basins.

Hari Prakash^a,  Pramod Soni ^b K .K Pandey^c

^aResearch Scholar, Department of Civil Engineering IIT (BHU), Varanasi (U.P),221005,India,Email:hariprakash.rs.civ23@iitbhu.ac.in

^bAssistant Professor,Department of Civil Engineering IIT (BHU), Varanasi(U.P),221005, India. Email: pramod.civ@iitbhu.ac.in

^cAssociate Professor,Department of Civil Engineering,IIT(BHU),Varanasi(U.P),221005,India

Email: kkp.civ@iitbhu.ac.in

* Corresponding author: hariprakash.rs.civ23@iitbhu.ac.in

Accurate estimation of flood peaks in ungauged and data-scarce basins critically depends on the accuracy of rainfall inputs, still remains challenging due to the limited availability of ground observations and inherent uncertainties in satellite precipitation datas. Although datasets such as CHIRPS and GPM IMERG provide high-resolution rainfall information, their direct application in hydrological modelling is often constrained by regional bias, spatial scale mismatch, and temporal inconsistencies. Moreover, physically consistent representation of large-scale atmospheric variables is rarely incorporated in conventional bias-correction approaches.To address these limitations, this study proposes an integrated and scalable framework that combines satellite precipitation, ERA5 reanalysis variables, machine learning, and process-based hydrological modelling for flood peak estimation in ungauged basins. The framework is demonstrated over the Varuna River Basin (Varanasi, India). To resolve spatial scale mismatch, ERA5 atmospheric variables are spatially aggregated within an approximately 30 km buffer around each CHIRPS grid point prior to their use as predictors. A time-aware artificial neural network (ANN) is then developed to integrate multi-pixel GPM IMERG rainfall and aggregated ERA5 predictors, using CHIRPS as a reference dataset to generate physically informed, bias-corrected daily rainfall fields. Model robustness is ensured by systematically testing different network architectures with varying numbers of hidden neurons. The framework is implemented over more than one thousand grid cells, ensuring spatial consistency while maintaining computational efficiency.The corrected rainfall products are subsequently used to drive the SWAT hydrological model, and streamflow simulations are calibrated and validated using SWAT-CUP, with particular emphasis on reproducing peak discharge and high-flow extremes. At the daily scale, the proposed framework achieves coefficient of determination (R²) values of up to 0.76 for rainfall estimation, and leads to substantial improvements in streamflow simulation compared to uncorrected satellite rainfall, including reduced bias, improved temporal variability, and markedly enhanced simulation of flood peaks.