Improving the resolution of GRACE-based water storage estimates based on machine learning downscaling schemes

The applications of the Gravity Recovery and Climate Experiment (GRACE) on local scales are obstructed owing to the coarser spatial resolution of GRACE observations. Much attempts recently have been taken to improve the resolution of GRACE-based water storage estimates based on machine learning algorithms, focusing on new algorithm development. However, there are still two deficiencies in previous GRACE downscaling research, namely the selection of input variables and the scale of model construction. In this study, the partial least squares regression (PLSR) model firstly is employed to identify the representative predictors associated with GRACE observations. Then, the performance of two different downscaling schemes (namely pixel-scale and regional-scale models) are comprehensively investigated, based on a machine learning algorithm known as random forest, to enhance the resolution of GRACE-based water storage estimates to the grid resolution as small as 5 km. The downscaled results are validated against hydrological model simulations and a number of in-situ groundwater level measurements within one of most rapidly urbanized basin in China, Haihe River Basin. The PLSR model recognizes four variables (namely evapotranspiration, temperature, land surface temperature, and soil moisture) as the predominant factors, acting as the predictors of downscaling models. Starting with the GRACE observations, two kinds of pixel and regional downscaling schemes are developed. The downscaled results were consistent each other and with the original GRACE data at a broad scale with the correlation up to 0.98. It was found that there was 3.20 times deviation of the results from the model simulation in computation of groundwater depletion rates within plain areas. In-situ water level measurements highlight that the downscaling models are improved by 36.95 % and 23.25 % in correlation relative to the original GRACE data and the simulated groundwater storage anomalies, respectively. Generally, the pixel-scale model is slightly better than the regional-scale model with 69 % (171 out of 249 observation wells) of higher correlation and 53 % (131 out of 249 observation wells) of smaller root mean squared error. The high-resolution results presented in this study can lead to better understanding on regional water resources and provide quantitative information to water management considering irrigation water use and groundwater consumption.