Modeling Soil Moisture from Multisource Data by Stepwise Multilinear Regression: An Application to the Chinese Loess Plateau

Abstract

This study aims to integrate multisource data to model the relative soil moisture (RSM) over the Chinese Loess Plateau in 2017 by stepwise multilinear regression (SMLR) in order to improve the spatial coverage of our previously published RSM. First, 34 candidate variables (12 quantitative and 22 dummy variables) from the Moderate Resolution Imaging Spectroradiometer (MODIS) and topographic, soil properties, and meteorological data were preprocessed. Then, SMLR was applied to variables without multicollinearity to select statistically significant (p-value < 0.05) variables. After the accuracy assessment, monthly, seasonal, and annual spatial patterns of RSM were mapped at 500 m resolution and evaluated. The results indicate that there was a high potential of SMLR to model RSM with the desired accuracy (best fit of the model with Pearson’s r = 0.969, root mean square error = 0.761%, and mean absolute error = 0.576%) over the Chinese Loess Plateau. The variables of elevation (0–500 m and 2000–2500 m), precipitation, soil texture of loam, and nighttime land surface temperature can continuously be used in the regression models for all seasons. Including dummy variables improved the model fit both in calibration and validation. Moreover, the SMLR-modeled RSM achieved better spatial coverage than that of the reference RSM for almost all periods. This is a significant finding as the SMLR method supports the use of multisource data to complement and/or replace coarse resolution satellite imagery in the estimation of RSM.