Improved estimation of terrestrial water loading deformation in GNSS vertical time series based on a hybrid machine learning method: A case study in Amazon River basin

Accurate estimation of terrestrial water load deformation (TWLD) from GNSS observations is of great significance to terrestrial water storage (TWS) inversion and tectonic deformation detection. However, it is difficult to separate the signals effectively from the complex and noisy GNSS observation series due to a variety of events and noises mixed together. We here present a hybrid estimation and prediction method for TWLD combining multi-scales decomposition and deep learning algorithm. Non-tidal atmospheric load and non-tidal ocean load are firstly corrected from the original GNSS time series. The residuals are then decomposed into main intrinsic mode components by variational mode decomposition algorithm, where grey wolf optimization and sample entropy is adopt to search for the optimal model parameters. Finally, every component is used as input to extract the data features suitable for BiLSTM model. We apply the method to estimate TWLD of the Amazon River Basin based on 28 GNSS sites. The results show that the hybrid model greatly enhance feature distinguishability and prediction accuracy of GNSS time series, with an average improvement of 73.20%, 26.87%, 73.40% and 51.14%, 7.83%, 59.55% in RMSE, R² and MAPE, compared to the single model of LSTM and BiLSTM. The TWS derived from GNSS agrees well with that of Gravity Recovery and Climate Experiment (GRACE) and Global Land Data Assimilation System (GLDAS), with the correlations of over 0.71 and 0.81, respectively. The joint TWS integrating GNSS, GRACE and GLDAS has a better consistency than single-source data, with the correlations of above 0.91. We find that the larger changes in the middle and lower reaches of the northeastern Amazon Basin, gradually tapering off towards the upstream areas near the Andes Mountains in the southwest. The significant variations area of water volume corresponds to the estuary and the downstream section of the Amazon River Basin. Moving southward from the aforementioned region, the magnitude of water storage changes gradually becomes more moderate. The research provides an important reference for the precise monitoring and efficient supervision of regional water resource.