A linear mixed effect (LME) model for soil water content estimation based on geophysical sensing: a comparison of an LME model and kriging with external drift

Abstract

Soil water content (SWC) is a critical attribute in precision irrigation. Direct measurements are costly and relative sparse, so there is interest in methods to predict SWC at unsampled sites from sampled data. The precision of such predictions can be improved, if covariates are incorporated into the predictor. To do that, two efficient ways are: (1) linear mixed effect model (LME), in which the spatial process is obtained by splitting the total variability in a systematic term or mean effect, a spatially correlated component and a random noise, and (2) kriging with external drift (KED), a nonstationary geostatistical technique, assuming covariates to have a linear effect on target variable. Geoelectric sensors provide noninvasive information on soil. Their outputs are quite sensitive to SWC; therefore, they can be used as covariates in SWC predictor. The objective of this work was to compare LME and KED, based on geophysical sensing, to estimate shallow SWC. The surveys were conducted in a south-eastern Italy field, using GPR at two frequencies, 600 and 1,600 MHz, and EMI. Volumetric soil water contents were measured with a Time Domain Reflectometer at 96 locations. Three LMEs, using three correlation functions (spherical, exponential, and Gaussian), were used and compared with KED using a set of cross-validation criteria. The mixed models showed a quite similar behavior even if exponential model outperformed the other two. The covariates used in the mixed models and in KED predictor were different and the results of cross-validation showed a slight out-performance of KED.