Evaluating different predictive strategies for filling the global GRACE/-FO terrestrial water storage anomalies gap

Abstract

The Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (GRACE-FO) missions provide unprecedented approaches for tracking terrestrial water storage anomalies (TWSA). However, evaluating long-term hydrologic states requires continuous TWSA without the ∼11-month gap between the two GRACE missions. Trend prediction is a challenging problem for TWSA gap-filling. There are three common methods for handling trends in previous efforts, i.e., de-trending the TWSA, adding back the long-term or piecewise trends of GRACE/-FO to the detrended predictions, or not performing such trend replacement. However, a single global application of one of these methods will not produce optimal results. Therefore, we designed a framework to select the optimal trend replacement strategy for each grid in this study. Based on this framework, we better filled the gap (excluding Antarctica) using machine learning techniques adopting the Global Land Data Assimilation System (GLDAS) Noah TWSA, precipitation, and temperature as inputs. The median gridwise Nash-Sutcliffe efficiency of the result generated by our framework improves by 0.08 compared to the result of a single long-term trend replacement strategy. Furthermore, we quantitatively evaluated the impact of three predictive strategies on the results: selection of leader machine learning technique, selection of optimal trend replacement strategy, and selection of most relevant inputs. The results indicate that the selection of trend replacement strategy has the greatest influence, followed by the selection of machine learning technique and then the selection of inputs. In addition, we found that in areas with abundant surface water, utilizing surface water anomalies as an additional predictor benefits the results. Our study is expected to provide suggestions for better TWSA predictive strategies.