Spatial filtering of electrical resistivity and slope intensity: Enhancement of spatial estimates of a soil property

Abstract

To best utilize the electrical resistivity data and slope intensity derived from a Digital Elevation Model, the kriging spatial components technique was applied to separate the nuggets and small- and large-scale structures for both resistivity and slope intensity data. The spatial structures in the resistivity and slope intensity data, which are poorly correlated with soil thickness (ST), are then filtered out prior to integrating the resistivity data and slope intensity into soil thickness estimation over a 12ha area located in the south-western Parisian Basin (France). ST was measured at 650 locations over the study area by manual augering. Twenty percent of the observations (131 points) were randomly selected to constitute the validation dataset. The remaining 80% of the dataset (519 points) was used as the prediction dataset.The resistivity data represent a set of 7394 measurement points for each of the three investigated depths over the study area. The methodology involves successively (1) a principal component analysis (PCA) on the electrical measurements and (2) a geostatistical filtering of the small-scale component and noise in the first component (PC1) of the PCA. The results show that the correlation between ST and PC1 is greatly improved when the small-scale component and noise are filtered out, and similarly, the correlation between ST and slope intensity is greatly improved once the geostatistical filtering is carried out on the slope data. Thus, the large scales of both slope intensity and the electrical resistivity's PC1 were used as external drifts to predict ST over the entire study area. This prediction was compared with ordinary kriging and kriging either with a large scale of slope intensity or with a large scale of the electrical resistivity's PC1 taken as an external drift. The first prediction of ST by ordinary kriging, which was considered as our reference, was also compared to those achieved by kriging using the raw secondary variables: PC1 and slope intensity as external drifts; slope intensity as an external drift; and PC1 as an external drift. The results indicate a reasonably low bias of prediction for all of the methods, in particular in the case of kriging using the large scales of both slope intensity and PC1 as external drifts. The root mean square error shows that kriging accounting for the large scales of two secondary exhaustive variables is the most accurate prediction method. The relative improvement of the accuracy is at least equal to 29% between the approach accounting for both large scale components of secondary attributes in the spatial estimates of ST and the other approaches of estimates considered in this study.