The applicability of using NARX neural network to forecast GRACE terrestrial water storage anomalies

The Gravity Recovery and Climate Experiment (GRACE) satellite has proven adept at monitoring, characterizing, and predicting hydrological variables. This paper attempted to investigate the applicability of using nonlinear autoregressive with exogenous input (NARX) neural network to forecast GRACE terrestrial water storage anomalies over 11 basins. By using six hydrological indicators as external inputs and forming all possible combinations of these variables, we have found the appropriate external inputs of the optimal NARX model for each basin. In addition, the number of acceptable time delays and the number of hidden neurons are adjusted to train out the optimal NARX model. The results reveal that the NARX models with one time delay perform better than the models with higher than one delay, and the structures of the optimal NARX models vary with basins. The performance of the optimal NARX models of the 11 basins falls into “very good” category whether during training, validating, or testing period. In comparison with GRACE Follow On (GRACE-FO) results, the predictions from the optimal NARX models are satisfactory, with the highest correlation coefficient of 0.97 in the Amazon basin and the lowest correlation coefficient of 0.62 over the Yangtze basin. The results shown here not only bridge the data gap between GRACE and GRACE-FO but also facilitate the applications of using the NARX neural network to predict climate extremes like droughts and flooding.