Optimal feature selection for improved ML based reconstruction of Global Terrestrial Water Storage Anomalies

Abstract. Understanding long-term Terrestrial water storage (TWS) variations is vital for investigating hydrological extreme events, managing water resources, and assessing climate change impacts. However, the limited data duration from the Gravity Recovery and Climate Experiment (GRACE) and its follow-on missions (GRACE-FO) poses challenges for comprehensive long-term analysis. In this study, we reconstruct TWS anomalies (TWSA) for the period Jan 1960 to Dec 2022 thereby filling data gaps between GRACE and GRACE-FO missions as well as generating a complete dataset for the pre-GRACE era. The workflow involves identifying optimal predictors from land surface model (LSM) outputs, meteorological variables, and climatic indices using a novel Bayesian Network (BN) technique for grid-based TWSA simulations. Climate indices, like the Oceanic Niño Index and Dipole Mode Index, are selected as optimal predictors for a large number of grids globally, along with TWSA from LSM outputs. The most effective machine learning (ML) algorithms among Convolutional Neural Network (CNN), Support Vector Regression (SVR), Extra Trees Regressor (ETR), and Stacking Ensemble Regression (SER) models are evaluated at each grid location to achieve optimal reproducibility. Globally, ETR performs best for most of the grids which is also noticed at the river-basin scale, particularly for the Ganga-Brahmaputra-Meghana, Godavari, Krishna, Limpopo, and Nile river basins. The simulated TWSA (BNMLTWSA) outperformed the TWSA from LSM outputs when evaluated against GRACE datasets. Improvements are particularly noted in the river basins such as Godavari, Krishna, Danube, Amazon, etc. , with median values of the correlation coefficient, Nash-Sutcliffe efficiency, and RMSE for all grids in Godavari, India, being 0. 927, 0. 839, and 63. 7 mm respectively. A comparison with TWSA reconstructed in recent studies indicates that the proposed BNMLTWSA outperforms them globally as well as for all the 11 major river basins examined. The presented dataset is published at https: //doi. org/10. 6084/m9. figshare. 25376695 (Mandal et al. , 2024) and updates will be published when needed.