Downscaling Of GRACE-derived Groundwater Storage Changes Based On Hierarchical Clustering And Non-linear Regression Model

As an important part of freshwater resources,groundwater resources are largely exploitable and widely distributed.It is an important water source for China's agricultural production,industrial and mining industries.Especially for northern provinces,the lack of surface water resources makes the exploitation and utilization of groundwater resources to increase dramatically.Among all national-level urban agglomerations,Central Plains Urban Agglomeration was approved by the State Council at the end of 2016.The area covers entire Henan province and some parts of neighboring four provinces.It has dense population,developed transportation network,and huge potential for economic development.However,with population growth,industrialization acceleration,and the improvement of agricultural productivity,the amount of groundwater in the Central Plains urban agglomeration area has surged,and it is urgent to monitor groundwater resources in time series to control overexploitation.Traditional monitoring methods typified by in situ well observations and hydrological model assimilations have long been difficult to efficiently obtain groundwater anomalies results.The Gravity Recovery and Climate Experiment Satellite(GRACE)jointly developed and launched by the National Space Agency(NASA)and the German Aerospace Center(DLR)of the two countries,has opened up a new way for terrestrial water storage monitoring.The satellite is characterized by high precision,all-weather capabilities,large monitoring scale,high automaticity,easy data access etc.Its published data has played a huge role in the study of regional water storage changes.However,the low spatial resolution of GRACE data severely limits its application and promotion.In order to obtain the high-resolution groundwater storage distribution and evolution results in the Central Plains urban agglomeration area,this paper first obtained the original low-resolution groundwater storage changes results based on GRACE gravity satellite and GLDAS hydrological model data.Later,the collaboration of hierarchical clustering and machine learning non-linear regression method was used to statistically downscale the original groundwater storage changes results.Finally,the actual measured groundwater level data was used for verification analysis.The main research and conclusion of this paper are as follows:(1)In the existing downscaling studies,the uniform downscaling of all grid values is susceptible to spatial heterogeneity and leakage errors between adjacent grids,resulting in poor model fitting results.In order to overcome these mentioned problems,the grid results of the original groundwater storage changes time series were hierarchically clustered firstly.During the clustering calculation,the correlation coefficient was used as the similarity measurement method,the centroid connection distance was used as the similarity criterion,the Silhouette Coefficient was used as the clustering performance criterion.Finally,all grids were divided into two individual clusters.(2)In order to accomplish the optimal downscaling results,this paper used the 5-fold cross-validation together with grid search method,and took Root Mean Squared Error(RMSE)and Nash-Sutcliffe Efficiency(NSE)as the model's optimal evaluation indicators.NDVI,land surface temperature,evapotranspiration,air temperature and precipitation were the model predictors.The downscaling model performance of the three machine learning methods of Support Vector Regression,Multi-layer Perceptron and Random Forest were compared and analyze,the output results showed that the model performance of the Random Forest method was the best.In addition,the model accuracy of multi-cluster downscaling is compared with global unified downscaling.The results showed that the model performance had been significantly improved after the multi-type cluster downscaling,with Random Forest being the best,indicating that the hierarchical clustering method proposed in this paper can effectively improve the accuracy of the upcoming downscaling regression results.(3)Based on the Random Forest regression method,the two clusters obtained from hierarchical clustering analysis were separately trained.Thus,grid results of groundwater storage changes were successfully downscaled from the original 0.25° × 0.25° to 1 km ×1km(0.008333° × 0.008333°).Later,the correlation analysis between downscaling results and in situ measured data of 5 groundwater level monitoring wells located in the study area was carried out.The correlation coefficient of the 5 sets of verification results were all at the significance level of 0.01,indicating the downscaling method proposed in this paper has high reliability.(4)The result of downscaling groundwater storage change rate was obtained through least squares linear fitting.Combining with the groundwater overexploitation positions published by Henan,Hebei,Shandong and Anhui provinces,the spatial and temporal distribution and evolution characteristics of the downscaled groundwater storage changes were comprehensively analyzed.The results showed that the change rate results was in good agreement with the range of the groundwater overexploitation area.The final downscaled results of GRACE groundwater storage changes have high timeliness and accuracy,indicating that the collaboration of hierarchical clustering and Random Forest downscaling method proposed in this paper can be used as an effective method to evaluate groundwater storage changes in the Central Plains urban agglomeration area.The research algorithm flow in this paper can provide theoretical support and data reference for the survey and monitoring of wide-area groundwater resources,as well as the sustainable use of groundwater and the management of over-extraction and prohibition.