Ultrahigh dimensional variable selection for interpolation of geostatistical data: Case studies in soil carbon modelling

This thesis explores statistical methodologies for predicting maps of soil carbon levels from small numbers of soil core observations. Each of these methods improves the accuracy of the mapping by discovering and exploiting empirical relationships between soil carbon observations and data on large numbers of potentially related environmental characteristics. In tandem, data visualisation techniques are applied in novel ways to represent the roles of the many environmental characteristics used in these models of soil carbon distributions. This thesis also holds relevance beyond soil carbon mapping to the widespread task of leveraging maps of potentially related, ancillary data when predicting maps from point referenced observations