Improve Streamflow Simulations by Combining Machine Learning Pre-Processing and Post-Processing

Addressing uncertainties in hydrological modelling is critical for improving model accuracy. Pre-processing and post-processing techniques offer effective ways to reduce these uncertainties, but strategic selection and combination of these methods with machine learning models require further exploration. In this study, we introduced input uncertainty into a hydrological model using satellite precipitation data as a case study. We investigated the effects of pre-processing using Random Forest (RF), post-processing using RF and Long Short-Term Memory (LSTM) networks, and their combination within a comprehensive modelling framework for daily streamflow simulation across 522 sub-basins. We used the Kling-Gupta Efficiency (KGE) and its three components-correlation, bias, and variability to assess model performance. Our analysis revealed that pre-processing with RF alone substantially improved model performance over raw simulation, especially in terms of KGE and correlation. Pre-processing improved KGE in 96.2% of sub-basins (average increase: 0.38) and enhanced correlation in 100% of sub-basins (average increase: 0.164). Post-processing with RF alone also improved performance but was less effective, enhancing KGE in 83.0% of sub-basins (average increase: 0.24). The combined approach delivered the most significant gains, improving KGE in 92.7% of sub-basins (average increase: 0.42) and achieving 100% improvement in correlation (average increase: 0.161), while also resulting in better bias and variability scores. Comparisons indicated that pre-processing is more effective than post-processing. Further analysis showed that different post-processing models, RF and LSTM, yielded similar results with marginal differences. This study highlights the importance of carefully selecting and combining pre-processing and postprocessing strategies to optimize hydrological model performance.