Broadacre cropping operations are gathering more spatial data on the soil resource using high resolution proximal sensors such as electromagnetic induction (EMI) and gammaradiometric (GR) instruments. These instruments quantify the magnitude and spatial pattern of variation in properties of the soil (apparent soil conductivity (ECa), magnetic susceptibility and natural gamma emission) that often show spatial relationships with variation in crop production. However the properties being measured are influenced by numerous agronomically significant soil properties, and so a ground-truthing regime is required to convert the sensor data to useful soil property maps. Such a sampling regime needs to cover the distribution of the data in all the covariate layers to be most effective. To achieve this in a both a rigorous and parsimonious way using numerous data layers, this research applies a multivariate latin hypercube physical sample site identification methodology. Latin hypercube sampling schemes have been shown to better represent the overall distribution of data sets when compared to random sampling. The multivariate latin hypercube sample site identification process was instigated on 5 farm-wide multi-covariate data sets (1600 to 2000 hectare farms), running 25 iterations of sample schemes over each farm for a defined range of sample site numbers (5 to 20). Each covariate used in the process is scaled and the distribution quartiles of each sample site iteration is calculated and subtracted from the quartiles of the full data set to provide a difference measurement. The sum of these differences for each sample site number are averaged over all iterations. Comparisons of these averages over the sample size range is used to determine the lowest sample site number that minimises the difference. The process is demonstrated for each farm and it is proposed that if the mean difference of multiple hypercube iterations is calculated for a farm (or field), then the optimum sample size in terms of financial impost, deployment effort and statistical rigour could be retrieved in any instance. With these site locations identified, soil profile sampling is then undertaken to be used in conjunction with the original spatial data to map relevant soil properties. Regression techniques are a useful tool for the prediction of soil properties on the farm scale, but with a limited sample size (as per the hypercube procedure) achieving an optimal result is often tenuous. A solution to this issue is proposed here that involves the prediction of soil properties in a continuous fashion in the vertical direction from soil cores which have been subsampled sampled by horizon. The method fits spline functions to the horizon data which can then be divided into any user defined virtual samples over the depth range, which creates a larger data pool with which to make predictions for mapping. Cubist modelling of soil properties with the soil sensor data is then performed on the virtual data pool. The process is demonstrated for 1 farm where texture, CEC, pH and porosity were calculated over the across the farm at 5cm depth increments from 0 to 60cm. This mapped data can then be used for calculating available water capacity, drainage and estimated soil strength and fertility. These combined processes will be a useful tool in the effort to increase the efficiency of farming operations of any scale.