Supervised and unsupervised classification of near-mine soil Geochemistry and Geophysics data

Remotely sensed geoscience data can assist detailed geological field mapping in areas of thick vegetation and poor outcrop. However, the potentially high dimensionality of these data makes it difficult to visually interpret and fully comprehend. Machine learning algorithms provide an efficient semi-automated means of recognising and identifying patterns in data. We use Random Forests for supervised classification of geologic units from airborne geophysical and soil geochemical data in the economically significant Hellyer - Mt Charter region of western Tasmania. A backward-recursive variable selection method is used to select the most relevant and useful data for this problem. This reduces computation cost and enhances interpretation of results without significantly affecting prediction accuracy. Random Forests generates accurate predictions of the spatial distribution of surface geologic units from these data. An example is provided regarding the use of Self-Organising Maps, an unsupervised clustering algorithm, to identify distinct but spatially contiguous clusters within a geologic unit. By visualising cluster spatial distribution and identifying key variable contributions to cluster differences, we interpret the geological significance of intra-class variability.