Building a machine learning classifier for iron ore prospectivity in the Yilgarn Craton

Abstract

High resolution, large-scale geophysical data have recently become readily and freely available for the majority of the Australian continent; yet there have been few efforts to create a synthesis of these datasets for mineral exploration. Considering the rising cost of finding new deposits and the recent economic downturn, there is a focus on using low expenditure, large-scale explorative techniques to assist in finding deposits. Using sophisticated machine learning algorithms coupled with increases in computational power, we present a methodology that tests and trains a classifier using six geophysical datasets in conjunction with 37 iron ore locations in the Pilbara Craton that accurately predicts the locations of iron ore deposits throughout the Yilgarn Craton. Our selected classifier uses principal component analysis and mixture of Gaussian classification with reject option, and it successfully identifies 88% of iron ore locations. We use cross-validation (10 fold, 70% testing 30% training) to ensure the generalisation of our classifier. We apply our classifier to the Yilgarn Craton, an area not used for the training and testing phase, and compare the predictive confidence map to previously published locations of iron ore occurrences. We find that our classifier correctly locates key known Yilgarn iron ore deposits, in addition to highlighting other areas that could potentially be prospective for iron ore.