Uncertainty Reduction in Near Real-time Satellite Precipitation Estimates by Integrating Soil Moisture and Potential Evapotranspiration Using a Machine Learning Approach

Abstract

Near-real-time (NRT) satellite precipitation data inherits complex and random errors due to various reasons. The primary objective of this research is to utilize satellite-based precipitation data for hydrological modelling in ungauged areas. The novelty of this study lies in the development of a hybrid stacking-based machine learning framework that integrates hydrologically meaningful predictors: root-zone soil moisture, potential evapotranspiration (PET), and their time-lagged representations to reduce uncertainty in near-real-time satellite precipitation (GSMaP-NRT). Unlike conventional bias-correction approaches that rely primarily on statistical adjustment between satellite and gauge rainfall, this study incorporates physically relevant catchment-state variables to improve predictive skill, with a focus on the Ovens River basin in Australia. A calibrated GR4H hydrological model was used to simulate the runoff of the catchment. Six objective functions were used to evaluate the performance of the approach. The results demonstrate that stacking machine learning algorithms reduces the Mean Absolute Error of GSMaP-NRT satellite precipitation data by 36% and the corresponding modelled streamflow error by 44% for lower precipitation events (< 2 mm/hour). All six objective functions achieved optimal performances within the low precipitation events. However, RMSE remained high for intermediate and heavy precipitation events. The model-estimated major streamflow peaks for the years 2010 and 2016, based on gauged precipitation and ML-corrected satellite precipitation, are 41% and 48% lower than the observed streamflow peaks, respectively. The reasons were the inability of the GR4H model to capture the perfect initial conditions and the x4 time parameter during the calibration process.