The Inversion Of Terrestrial Water Storage Anomaly Based On Deep-learning Algorithm And Fusion Of GNSS And GRACE

With the continuous maturity and improvement of global navigation satellite system(GNSS),it can accurately monitor the crustal vertical deformation,which provides a data basis for analyzing the crustal motion and inversion the surface load.However,it is difficult to extract GNSS vertical deformation sequence feature terms due to the high frequency noise in GNSS sequences.Meanwhile,the spatial distribution of GNSS stations is extremely uneven,which greatly limits the application of GNSS deformation sequences to invert the terrestrial water storage anomaly(TWSA).To solve these problems,this study proposes a new correlation feature extraction correction method(CFECM),a new machine learning loading inverted method(MLLIM),and a deep-learning weight loading inverted model(DWLIM).These methods have important theoretical significance for effective extraction of feature terms and accurate inversion of TWSA based on sparse GNSS arrays,which are summarized as follows.(1)This study constructs the CFECM.The correlation coefficients between the modal components and the original series are introduced based on the principle of traditional variational mode decomposition(VMD).The original sequence is decomposed twice by using the VMD decomposition method.Then the residual terms in the sequence are extracted based on the energy spectral index and the spectral analysis of Lomb-Scargle(L-S).The sequences of GNSS and Gravity Recovery and Climate Experiment(GRACE)are extracted by the CFECM and Empirical Mode Decomposition(EMD),respectively.The results show that the CFECM is more reliable and stable than the EMD.The trend feature extraction of GNSS sequences based on the CFECM contains an accuracy of 97.87%.In the extraction of sequence seasonal terms,the CFECM also significantly outperforms the EMD feature extraction method.The mean NCC value of CFECM is 0.83,which is 58.10%higher than that of the EMD;and the mean value of signal to noise ratio(SNR)of CFECM is 11.76 d B,which is higher than the EMD decomposition method at 3.38 d B.(2)The new machine learning load inversion method is constructed.Firstly,the sequences of surface temperature and pressure are used as the input variables and GNSS vertical sequence is utilized as the output data.Secondly,the crustal vertical deformation sequences of unobserved grid are simulated based on the random forest(RF)algorithm.Finally,all the corrected sequences are input into the crustal loading model to derive the TWSA over southwest China.The TWSA results of MLLIM are compared with the traditional GNSS inversion results,GRACE mascon dataset,and the global land data assimilation system(GLDAS).The results show that the MLLIM can accurately detect the raised regions of TWSA,meanwhile,the raised regions are consistent with the GRACE and GLDAS.The Pearson correlation coefficient(PCC)between the MLLIM and GRACE,GRACE-FO are equal to 0.91 and 0.88,respectively.The R-squared(R~2)are equal to0.71 and 0.58,respectively,which are higher than those of the traditional GNSS inversion method.The results show that the MLLIM more accurate than the traditional GNSS inverted method in PCC and R~2 with 9.72%and 6.67%,respectively.(3)The new deep learning load inversion method is constructed.Firstly,the study region is divided into 1°×1°grids.The grids are divided into observed and unobserved grids,where the observed grids include GNSS stations and the unobserved grids do not contain GNSS stations.Then,the input data(surface temperature and atmospheric pressure)are decomposed to 20components.The long short-term memory(LSTM)algorithm and inverse distance weights are combined to simulate the crustal vertical deformation in the unknown grid.Then the correction of atmospheric and non-oceanic tidal are applied to the crustal sequences.Finally,the corrected sequences are utilized the input data for the crustal load model to derive the TWSA over mainland China.The DWLIM results are compared with the traditional GNSS inversion results,GRACE and GLDAS datasets.The results show that the DWLIM can effectively invert the raised regions of the anniversary amplitude in the mainland China.And the DWLIM can greatly weaken the speckle effect caused by the insufficient radius of disc expansion.Meanwhile,the DWLIM results of annual amplitude regions are consistent with the GRACE and GLDAS results.Based on the comparison of DWLIM inversion results with GRACE and GLDAS results,the results show that the maximum PCC,Nash Sutcliffe Efficiency(NSE),and root mean squared error(RMSE)reach0.81,0.62 and 2.18 cm,which is an average improvement of 67.11%,128.15%and 22.75%,respectively.Overall,the CFECM method proposed can effectively separate the feature sequences in GNSS signals and has significant advantages over the traditional EMD.Secondly,the MLLIM and DWLIM proposed combine the machine learning and deep learning to invert the TWSA,respectively.Meanwhile,the speckle effect problem caused by the disc expansion effect is also substantially suppressed based on the MLLIM and DWLIM.There are 41 figures,8 tables and 209 references.