Support vector machine and artificial neural network modelling of orogenic gold prospectivity mapping in the Swayze greenstone belt, Ontario, Canada

Abstract

Exploration for new mineral deposits has become increasingly difficult as new discoveries are being found under progressively deeper cover. To better understand and predict orogenic gold mineralization in the Archean Swayze greenstone belt, the essential ingredients of a mineral system are considered: 1) the source of gold and transport fluid ligands, 2) fluid pathways, 3) traps, and 4) the processes responsible for gold precipitation. The aim of this study is to use a mineral systems approach to help generate exploration targeting models using a spatial statistical method - weights of evidence (WofE) and data-driven machine learning tools, namely radial basis function neural networks (RBFNN) and support vector machine (SVM). The mineral prospectivity maps generated using the RBFNN and SVM machine learning methods were trained using the K-Fold cross-validation approach whereby 10 subsets of the data were used to train and test the model performance. The mean area under receiver operator curve after 10-fold cross-validations were 91% and 94% for the RBFNN models, and the SVM models obtained accuracies of 91% and 87%. Feature importance estimations obtained from both methods indicate that D2 and D3 high-strain zones, lithological contacts and D2 folds (i.e., synclines and anticlines) were found to be important predictor layers for targeting potential prospective zones of gold mineralization. The machine learning algorithms used in this study are novel and pragmatic methods that use the full potential of geoscience datasets in mapping orogenic gold prospectivity.