Definition of Geological Domains with an Ensemble Implementation of Support Vector Classification

Abstract

Building a resource model can be a challenging process because of the complex nature of geology. Employing limited information to build the resource model inevitably gives rise to uncertainties, and therefore uncertainty is an important phenomenon that needs to be taken into account while building the model. In fact, quantifying the uncertainties can be a significant advantage for the evaluation of mineral resources. Uncertainties can be quantified by traditional approaches, such as geostatistical simulation methods. While traditional approaches generally focus directly on domains to build geological models, today huge geochemical datasets obtained during exploration campaigns are seen as an opportunity to better evaluate the geological models. In this context, machine learning algorithms stand as a strong alternative to traditional approaches (explicit or implicit modeling, or geostatistical techniques), because they can easily incorporate large and multivariate datasets to reach consistent and semi-autonomous results. This study proposes an ensemble learning approach to define geological domains and their inherent uncertainty. The study adopts an unsupervised binary clustering method to label the domains of a multivariate geochemical dataset, and the information obtained by this unsupervised method is used to inform a supervised learning algorithm, which is Support Vector Classification. The supervised learning step is applied to assign domaining in locations of a gridded model. The unsupervised and supervised learning steps are repeated with subsets of geochemical variables and with subsets of training samples to realize an Ensemble model. These models are combined to define a strong model, and the uncertainty can be incorporated with the model. Finally, a hierarchical application is applied on the results to build multiple domains. We demonstrate the proposed workflow with an application to a real database from a porphyry copper deposit. The workflow was capable to build a robust model with high accuracy and with well-defined, curvy boundaries.