Performance of two-step technique for gap-filling and reconstruction of basin-scale Terrestrial Water Storage Anomalies (TWSA) 

Abstract

Accurate estimation of terrestrial water storage anomalies (TWSA) is essential for the assessment of hydrological extreme events, managing water resources, and evaluating climate change impacts. In this study, a two-step method is applied for the reconstruction of a gridded TWSA product in two study basins: Godavari (GRB), a tropical river basin in India, and the Murray-Darling river basin (MDRB) in Australia. In the first step, the probabilistic dependence structure of the target TWSA and 15 potential predictor variables is developed through Bayesian Network technique to obtain the optimal features which strongly influence the target. In the second step, the potential of Machine Learning (ML) algorithms is utilized to obtain TWSA values, considering the grid-specific features selected in the first step as input. The input set of potential predictors includes monthly TWSA simulations from Global Land Data Assimilation System (GLDAS) Catchment Land Surface Model (i.e., CTWSA) and the GLDAS Noah Land Surface Model (i.e., NTWSA) as well as meteorological variables such as precipitation and temperature for a lead time of up to 2 months and large scale climate indices such as Dipole Mode Index, North Atlantic Oscillation index, and Oceanic Niño Index (ONI). For both study basins, CTWSA and ONI are prominent features selected by the Bayesian Network that influence TWSA. After obtaining the optimal features, Machine Learning (ML) algorithms such as Convolutional Neural Network (CNN), Support Vector Regression (SVR), Extra Trees Regressor (ETR), and Stacking Ensemble Regression (SER) are employed to derive TWSA values (henceforth named as BNML_TWSA). The performances of BNML_TWSA, as well as CTWSA and NTWSA, are evaluated against GRACE TWSA for both study basins using performance metrics such as the Correlation Coefficient (R), Nash–Sutcliffe Efficiency (NSE), and Root Mean Square Error (RMSE). At GRB, ETR demonstrates superior performance at most of the grids (74.3%), followed by SVR (21.1%). In contrast, at MDRB, all four ML algorithms show similar performance: CNN, SVR, ETR, and SER, each being selected as the best models at 25.9%, 21.4%, 26.1%, and 26.6% of the grids respectively. When evaluated against GRACE TWSA, the median values of R for NTWSA, CTWSA, and BNML_TWSA across all grids are 0.78, 0.90, and 0.93, respectively, at the GRB. Similarly, for the MDRB, these values are 0.79, 0.85, and 0.87, respectively. At the GRB, the best NSE value is obtained for BNML_TWSA (0.84), while the lowest performance is observed for NTWSA (0.475). At the MDRB also, the least performance is shown by NTWSA with an RMSE value of 57.3 mm/month, and the best performance is achieved by BNML_TWSA with an RMSE of 35.0 mm/month. The proposed two-step method offers dependable estimates of TWSA compared to land surface models and hydrological models. Hence, the reconstructed TWSA (1960-2022) proves valuable during the data gap period between GRACE and GRACE-FO and the pre-GRACE period.