Improving the Reliability of the Prediction of Terrestrial Water Storage in Yunnan Using the Artificial Neural Network Selective Joint Prediction Model

Abstract

Although Gravity Recovery and Climate Experiment (GRACE) can provide accurate estimates in water storage, there are about 11 months gaps between GRACE and its successor GRACE-Follow On (GRACE-FO). To improve the accuracy of bridging the gaps, this study combines the partial least squares regression (PLSR) model with the learning-based models for the first time, and construct an artificial neural network selective joint prediction model. To this end, the performance of combination of PLSR and nonlinear autoregressive with exogenous input (NARX) models, the back propagation (BP) models, and multiply linear regression (MLR) models are compared on both grid and region scales. Results indicate the proportion of ''qualified'' and above predicted by PLSR combined with NARX on grid scale increased by 32.71% and 24.65% compared to the other two models. Finally, PLSR joint NARX model performs the best against GRACE observations, then employed to predict and bridge the TWSA gaps from July 2017 to December 2019 in northwest, southwest and central of Yunnan. The NSE values between predicted TWSA (terrestrial water storage anomalies) and GRACE-FO are increased by 55.10%, 93.10% and 133.33% compared to modeled TWSA and GRACE. This study provides a new perspective for filling the GRACE gap, and for selecting optimal variables in other relative hydrologic studies.INDEX TERMS GRACE/GRACE-FO mission, learning-based models, artificial neural network selective joint prediction model, TWSA.

Predicting Changes in Spatiotemporal Groundwater Storage Through the Integration of Multi-Satellite Data and Deep Learning Models

“…There were temporal gaps between GRACE and GRACE-FO with a length of approximately 14 months. Many studies have reported on the gap filling between GRACE and GRACE-FO temporal data using interpolation (linear [48], singular spectrum analysis [48], and cubic spline [49], [50]). Reference [48] applied linear interpolation to fill the gap and confirmed that the trend is positively biased by over 10% compared to the original series.…”

“…Nevertheless, since no data can replace the GRACE and GRACE-FO data, the interpolation method is widely used. In this study, a cubic spline suitable for time series interpolation was applied [49], [50].…”

Normalization Influence on ANN-Based Models Performance: A New Proposal for Features’ Contribution Analysis

“…The impact of the FN method's selection on the algorithm's performance has been experimentally studied by some researchers [30]- [33] to estimate the most appropriate one for a given problem. But, it remains the extended approach of employing the min-max normalisation method before the use of an ANN [34]- [38].…”