The Reconstruction and Extension of Terrestrial Water Storage Based on a Combined Prediction Model

Abstract

Monthly changes in total water storage (△TWS) can be employed for drought and flood monitoring as well as early warning systems and can be obtained from the total water storage anomalies (TWSAs) of the Gravity Recovery and Climate Experiment (GRACE). However, the relatively short GRACE time series limits its application on a wider scale. To this end, this study proposes a combined prediction (CP) model including a support vector machine (SVM) and an artificial neural network (ANN) to reconstruct and extend the monthly TWSAs from 1960 to 2012. Moreover, an innovative input selection strategy is proposed to build a monthly TWSA prediction model. In this strategy, the partial correlation algorithm was used to select the best input variables from the candidate input variables. These candidate input variables included streamflow, precipitation, evaporation, and soil moisture storage (SMS). Yunnan Province, a typical humid area in China, was selected as a case study. The results showed that: (1) the innovative input selection strategy effectively improved the simulation ability of the model, particularly when the candidate input variables influenced each other; (2) the CP model using the innovative input selection strategy yielded the best performance; (3) the monthly △TWS obtained from the extension of TWSAs recorded five of the seven extreme meteorological drought events in Yunnan Province from 1961 to 2001., All of this showed that the reliability of the expanded TWSAs was better than those of the Global Land Data Assimilation System TWSAs. Generally, the findings of this study showed that the CP model using an innovative input selection strategy was a useful and powerful tool to predict monthly TWSA.