Comparison features importance for temporal and spatial-temporal approaches in GRACE and GRACE-FO signal reconstruction based on GLDAS data

Machine learning algorithms can effectively learn the complex relationships between various input variables from the global land data assimilation system (GLDAS) and the total water storage (TWS) observed by gravity recovery and climate experiment (GRACE) and GRACE-FO (follow-on) missions. As the TWS depends on various features, a serious question arises about the importance of used variables for reconstruction. Furthermore, will the variables used for the reconstruction be equally significant for grid-based and basin-based analyses? This work examined the importance of individual predictors for the temporal and spatial-temporal approach over 254 river basins using GRACE and GRACE-FO data as target and GLDAS data as predictors. The extreme gradient boosting (XGBoost) algorithm was used to reconstruct TWS. Results were evaluated with root-mean-square error, normalised root-mean-square error, Pearson correlation coefficient, Nash-Sutcliffe efficiency, and Kolmogorov-Smirnow-test metrics. Model output influence was checked by the model-agnostic version of the feature importance and by Shapley additive explanations (SHAP).