Geoscience and Exploration

Clustering Based Anomaly Detection for Exploration Geochemistry

Thursday, November 22, 2018 - 08:40 to 08:59 Theatre 1


F.A. Cabral Pinto (Presenting)
University of Alberta

C.F. Prades Koscina
University of Alberta

H. Falck
NWT Geological Survey

C.V. Deutsch
University of Alberta

Unsupervised learning is the field of machine learning (ML) where there are no response variables to train the algorithm, the purpose is to examine the input data for associations and patterns. This characteristic of unsupervised learning makes it a suitable tool for mineral exploration purposes, where it is commonly not known the location of nearby deposits, and therefore, there is no reliable response variable to train a supervised learning algorithm. One of the most important areas of unsupervised learning is cluster analysis. It is defined as the process of classifying data instances into groups based on their similarities.

In addition to straightforward univariate anomalies, three cluster-based methods for multivariate anomaly detection were developed and applied on stream silt samples from the Mackenzie Mountains: (1) small anomalous clusters, (2) Lack of Uniform Clustering Classification (LUCC) anomalies, and (3) spatial anomalies. The first method uses combinations of clustering and data transformations for finding small anomalous clusters; the second uses different clustering methods on different data transformations for identifying samples that do not clearly belong to any cluster; the third recognizes samples that are different from the surrounding samples in the geographic space.

The analyses were performed on pathfinder elements for tungsten/intrusion deposits. The goal is to identify samples that are anomalous in the multivariate space. The proposed methods are capable of identifying several showings, that is, known mineral deposits in advanced exploration or production stage in the Mackenzie Mountains. Some of these showings are not detected from traditional univariate methods, which supports and motivates the use of multivariate anomaly detection methods for mineral deposit exploration. Although further validation is required, the multivariate methods appear promising to complement conventional analyses.