Kriging and Simulation in Gaussian Random Fields Applied to Soil Property Interpolation

Abstract

Spatial modeling is increasingly prominent in many fields of science as statisticians attempt to characterize variability of the processes that are spatially indexed. This paper shows that the Gaussian random field framework is useful for characterizing spatial statistics for soil properties. A sample of soil properties in 94 spatial locations are taken from a field (186.35m×211.44m) wide in northern Ethiopia, Karsa-Malima. We use observations of organic carbon (OC) from the site in our study. Box-Cox transformation is used because of OC follows non-Gaussian distributions. We develop ordinary kriging which is universal kriging with unknown trend models which enables us to predict any point within the field even outside the field up to the "Range" of the model. In this thesis work we predict 100×100 grids (10000 points) using kriging interpolation models. More over in each of these 10000 locations 1000 conditional simulations are made. Interestingly prediction using universal kriging and mean of conditional simulations agree in expectation and kriging variance. For covariance and/or variogram modeling and for parameter estimation we used least square principle and maximum likelihood estimation method. The classical geostatistical approach known as kriging is used as a spatial model for spatial prediction with associated spatial variances. Moreover, conditional simulation is performed. From ordinary kriging model results, predictions are accurate when predictions are close to observation locations. Prediction variance in the observed locations is close to the nugget effect.

Technical note: Determining Arctic Ocean cold halocline and cold halostad layer depths based on vertical stability

“…The latter is approximated with a so called semivariogram model. In this analysis only the spherical semi-variogram model is used (Mert and Dag, 2017;Reshid, 2019). The weights take into account the clustering of the observations as it uses the inverse of the correlation-distance matrix between input points and the correlation-distance vector between input points and the point of interest.…”