An Open-Source Machine Learning Framework for GRACE based River Discharge Estimation and Applicability Analysis

This study explores the application of advanced machine learning techniques for hydrologic modeling, focusing on predicting global river discharges using satellite-derived Total Water Storage Anomalies (TWSA) data. We address three key challenges: (1) regionalizing conceptual hydrologic model parameters, (2) predicting spatial applicability of TWSA-discharge relationships, and (3) identifying temporal windows where these relationships can be used.
For parameter regionalization, we compare Gaussian Process Regression (GPR), Gradient Boosting (GB), Support Vector Regression (SVR), and Artificial Neural Networks (ANN). GPR outperforms other methods, achieving R² values of 0.96 and 0.89 for the two required model parameters, respectively, on the test data. For predicting spatial applicability, XGBoost demonstrates superior performance with a macro-averaged F1 score of 0.73 on the test data. For temporal applicability, Random Forests and Extra Trees Classifiers show comparable performance, both achieving F1 scores of approximately 0.75.
To support these tasks, we developed an open-source tool that integrates processed TWSA data with preloaded regionalized model outputs, enabling users to generate discharge estimates at gauge locations interactively. This work contributes to hydrologic modeling by demonstrating the effectiveness of machine learning in handling complex, non-linear relationships in large-scale hydrologic systems and providing interpretable results that can inform water resource management strategies.