Estimating surface soil moisture from satellite observations using a generalized regression neural network trained on sparse ground-based measurements in the continental U.S

Abstract

This study attempted to develop a point-surface collaborative inversion (PSCI) method for the estimation of regional surface soil moisture (SSM) using a generalized regression neural network (GRNN) trained on sparse ground-based measurements. Specifically, GRNN was employed to establish a nonlinear relationship between ground-based measurements from sparse network stations (SNSs) and passive microwave observations from the Soil Moisture Active Passive (SMAP) satellite in the continental U.S. for April 2015 to March 2018. More importantly, the extended triple collocation (ETC) technique was applied to address the scale mismatch issue resulting from the small spatial support of ground-based measurements, whereby individual SNSs’ reliability at the SMAP coarse footprint scale could be determined before fed into GRNN. The 10-fold cross-validation results showed that the GRNN model trained on reliable SNSs obtained a satisfactory performance—the cross-validated R and unbiased RMSE values were 0.88 and 0.050 cm3 cm−3, respectively, which outperformed the back-propagation neural network (BPNN) and the other GRNN model trained on all SNSs. Furthermore, temporal and spatial comparisons between the PSCI-based SSM retrievals and other SSM datasets indicated the former agreed the best with ground measurements both in time and space, suggesting the proposed PSCI method had shown great potential in estimating reliable regional SSM climate records.