Spatiotemporal prediction of soil moisture content using multiple-linear regression in a small catchment of the Loess Plateau, China

Abstract

Three basic multiple-linear regression models with different sets of more readily observed environmental variables (land use, topographic and meteorological factors) were developed for the prediction of soil moisture content in space and time. The model performances were evaluated in the Danangou catchment (3.5 km 2) on the Loess Plateau, China, with soil moisture content measurements. The soil moisture content measurements were performed biweekly at five depths in the soil profile (0–5, 10–15, 20–25, 40–45 and 70–75 cm) from May to October 1998 and from May to September 1999, using a Delta-T theta probe. It was indicated that the regression models could describe the relationships of soil moisture content with environmental attributes. It was found that the spatiotemporal-SM model showed the best goodness of fit since it explained the greatest fraction of soil moisture variation in both space and time, and the predicted mean, standard deviation, minimum and maximum soil moisture were closest to the observed values. This model was also either the most precise or the most economical in prediction of soil moisture content in space and time since it gives the lowest values in mean absolute error of prediction (MAE) and Akaike information criterion (AIC). There is little difference in performance and cost-benefit between the spatial-GM model and landuse-GM model. The superior robustness of the spatiotemporal-SM model over the other two models is most significant in the prediction of soil moisture content at 0–5 cm, and decreases with increasing soil depth.