Study On Theory And Methodology For The Reconstruction/Prediction Of Global Gridded Total Water Storage Change

The Gravity Recovery and Climate Experiment(GRACE)mission opened up a new era for studying the mass transfering of the surface earth,and substantially promoted the research progress in global water cycle,terrestrial hydrology and solid geophysics.Besides,it also made significant contributions to the related research in earth sciences and is the unique method to accurately estimate the spatial and temporal evolution of global Total Water Storage Change(TWSC).Nevertheless,the GRACE mission ended its operation in October 2017 and the GRACE Follow-On mission was launched only in May 2018,leading to approximately one year of data gap.Furthermore,for many eatrh-related applications,it is also desirable to reconstruct TWSC prior to the GRACE period.In this thesis,for bridging the data gap between two GRACE missions and providing a long term and uninterrupted time series of TWSC,the statistical decomposition,time series decomposition,and machine learning techniques were combined in a unified framework to predict the GRACE-like gridded total water storage change.Four groups of global TWSC data sets from January 1992 to June 2019 were predicted based on the combined method proposed in this thesis.In addition,this thesis also studied the hydrological droughts and their spatial patterns(1992-2019)over the Amazon river basin using the predicted TWSC data set.The main research contents are listed as follows:(1)The background and motivation for the prediction of TWSC were introduced.The history of the satellite gravity and the literatures of detecting global TWSC using GRACE satellites were reviewed.The efforts made by previous researchers to fill the data gap between the two generations of GRACE satellites were introduced.The current progress of using data-driven methods to forecast GRACE-like total water storage change was discussed and one of the questions that desired to be solved were proposed.(2)The methods for estimating the TWSC using the GRACE data were introduced.The theories of two statistical decomposition methods Principal Component Analysis(PCA)and Independent Principal Component Analysis(ICA),two time series decomposition methods Least Square(LS)and the Seasonal-Trend decomposition based on Loess(STL),and three machine learning methods Artificial Neural Network(ANN),Auto Regressive model with e Xogenous variables(ARX),and Multiple Linear Regression(MLR)were presented.(3)Two GRACE Mascon data sets were used as the target variables.The global precipitation,land surface temperature,sea surface temperature data,17 kinds of climate indices,and two types of soil moisture data were selected as the drivers for predicting the temporal and spatial evolution of global GRACE TWSC.The GRACE-FO TWSC,Swarm TWSC,12 in-situ ovsevations,and the Palmer Drought Severity Index(PDSI)were selected as the validation data.A time delay parameter was used to improve correlation between the precipitation and GRACE.The results show that the correlation coefficient between precipitation and GRACE data can reach a peak value when the time delay parameter was turning from 1 to 300.(4)The performances of two statistical decomposition methods,two time series decomposition methods,and three machine learning methods in predicting the GRACE-like gridded TWSC over 26 major river basins of the world were compared and analyzed.One most robust combination of these methods was identified for predicting the global GRACE gridded TWSC.The predicting uncertainties and robustness of the identified methods combination were estimated.The characteristic of error propagation in three machine learning models was tested.The spatiotemporal evolutions of GRACE TWSC in the 26 river basins were extrapolated date back to1992 and up to 2019.ANN and ARX perform better in the training period than in the testing period,indicating that there are some overfitting problems in these two methods.While the MLR method performs worse in the training section,it is the most robust one in the testing period,and is more stable than the ANN and ARX methods in coping with random input uncertainties.The performance of PCA and LS methods is better than ICA and STL,respectively.PCA-LS-MLR was identified as the best methods combination.Test computations also suggest that the correlation of predicted TWSC maps with observed ones is more than 0.3 higher than TWSC simulated by hydrological models,at the grid scale of 1° resolution.In addition,the predicted TWSC based on the identified combination-i.e.PCA-LS-MLR-correctly reproduce the El Nino-Southern Oscillation(ENSO)signals.(5)The methods combination PCA-LS-MLR was used to produce 4 global gridded TWSC data sets(0.5°×0.5°).The accuracy of the produced data sets was evaluated using GRACE,GRACE Follow-on,and Swarm data,respectively.The method proposed in this thesis was compared with previous studies.The results show that the produced data set fits well with the GRACE data at each grid point of the world and fits well with the GRACE Follow-on data in most regions of the world.As compared with the previous studies,the method proposed in this thesis can not only assimilate more information to support the TWSC prediction,but also predict a more complete picture of the GRACE TWSC.(6)The produced data set was used to detect the hydrological droughts and their spatial patterns in the Amazon Basin from 1992 to 2019.The water levels observed from 12 in-situ stations and PDSI index were used to verify the detected drought events.The results show that four severe drought events occurred in the Amazon basin in 1996,1998,2011,and 2016,respectively.The spatial distribution of the Amazon drought events in 1998 and 2016 are similar,and they are both related to the strong El Nino events.In addition,the drought patterns interpreted by the produced TWSC data set are reliable as validated by the in-situ water levels and PDSI index.