Improving the spatial resolution of GRACE-based groundwater storage estimates using a machine learning algorithm and hydrological model
The low-resolution characteristic of Gravity Recovery and Climate Experiment (GRACE) satellite data greatly limits their application in many fields at regional or local scales. Aiming to overcome this limitation, the partial least squares regression (PLSR) model is firstly utilized to assess the importance of some independent variables that are commonly employed in GRACE downscaling research. Three kinds of downscaling models are chosen to improve the resolution of GRACE-based water storage estimates from 1 to 0.25°, namely: multivariable linear regression, random forest (RF), and NoahV2.1. Results indicate that terrestrial water storage anomalies are more closely related to four independent variables in the Haihe River Basin, China: these variables are evapotranspiration, land surface temperature, air temperature, and soil moisture. With respect to the spatial distribution, the downscaled results based on the NoahV2.1 and RF models can effectively capture the subgrid heterogeneity while preserving the water storage characteristics at the original scale. By verifying the downscaled results with measured groundwater levels, it can be observed that the correlation coefficient between the RF-based downscaled groundwater storage anomalies (GWSA) and in-situ measurements is increased by 20.55% (Beijing), 9.13% (Tianjin), and 10.48% (Hebei) relative to the downscaled results based on the NoahV2.1 model. The cross wavelet transform illustrates that the meteorological factors have a strong influence on the GWSA series in the Haihe River Basin with an approximately 12-month signal during 2003-2016. This study can provide high-resolution GWSA datasets for water resources management and also provide a reference for the selection of dominant independent variables.