Field-scale digital soil mapping of clay: Combining different proximal sensed data and comparing various statistical models

Knowledge about clay fraction is important for soil management. Laboratory analysis is time consuming, however, because information is required at the topsoil (00.3 m), subsurface (0.30.6 m), and subsoil (0.91.2 m). Herein, we developed digital soil mapping (DSM) of clay by coupling limited laboratory data with proximal sensed data (i.e., elevation, gamma -ray, and electromagnetic). Our first aim was to determine whether individual or combined data were better for modeling using various models (i.e., regression kriging [RK], linear mixed model [LMM], and a Bayesian model named integrated nested Laplace approximation [INLA] with stochastic partial differential equation [SPDE]). We also evaluated DSM of clay and associated errors at different depths using various indices (e.g., Lin's concordance). Digital soil mapping performance was compared by mean square prediction error (MSPE). Considering proximal data, the Akaike information criteria and Log-likelihood showed that RK and LMM performed best by combining proximal data, whereas for INLA-SPDE individual and combined data performed equally well. Considering prediction uncertainty, RK error was larger where predicted clay was larger, whereas LMM and INLA-SPDE had larger errors associated with texture or field boundaries. With increasing depth, model performance decreased, as exemplified by LMM and for topsoil (Lin's = 0.79), subsurface (0.72), and subsoil (0.71). The MSPE showed that the DSM map for topsoil clay was best for LMM (28), followed by INLA-SPDE (29) and RK (39). The LMM DSM was also best in the subsurface (37) and subsoil (84); however, RK was superior to INLA-SPDE.