1Kenex Ltd., Lower Hutt, New Zealand
Random forests represent a machine learning implementation of a decision-tree algorithm that can be applied to data-driven mineral potential mapping. Most published studies using random forests include relatively small numbers of input maps that are typically pre-classified by an expert familiar with the mineral system being targeted. The aim of this study was to investigate how random forests performed using different input parameters in terms of the individual predictive maps and training data. Four different implementations of the random forest algorithm were produced based on a case study using data from the eastern Lachlan Orogen in NSW for the purposes of targeting porphyry Cu-Au mineralisation related to the Macquarie Arc: (1) using a large number of multi-class categorical or non-thresholded predictive maps that have had no favourability criteria applied; (2) using a large number of binary predictive maps that have had statistically valid and geologically meaningful thresholds determined through weights of evidence analysis and expert review; (3) using a subset of the binary predictive maps that were used in a weights of evidence mineral potential mapping study; and (4) using this same subset of binary predictive maps with weighted training data. These results were then compared to the results of an existing weights of evidence mineral potential mapping study.
The results of the random forest analysis demonstrate how both the ranking of the input maps and subsequent mineral potential varies considerably depending on the degree of intervention from an expert in the modelling process. The first approach produced a prospective area that covered 47.7% of the study area, the second approach 6.5%, the third approach 23.4%, and the final approach with the weighted training data 40.4%. In comparison, the weights of evidence study produced a prospective area that covered 15.2% of the study area, however failed to predict one of the training points within this prospective area. Increasing the complexity of the input data improved the predictive capacity of the mineral potential maps for targeting the porphyry Cu-Au mineralisation when expert review was used to determine meaningful thresholds and classifications for the input predictive maps. However, when a large number of multi-class categorical or non-thresholded predictive maps were used as input to the random forest (i.e. no favourability criteria were applied, so the algorithm determined the thresholds rather than an expert), a poor result was obtained. The results also highlight that the main limitation of using random forests (and other machine learning approaches) for mineral potential mapping is the lack of a sufficient number of economically significant deposits which can be used to train a large number of input predictive maps.
The random forest study clearly demonstrates that the use of predictive maps that have statistically valid, geologically meaningful, and practically useful thresholds and reclassifications assigned produce more robust mineral potential maps that can be used for exploration targeting.
Arianne is a spatial data analyst whose focus is on the use of mineral potential mapping and spatial statistics for mineral exploration. She spent more than 10 years as an academic before moving to industry to work on delivering value-added geoscience data to both government organisations and the exploration industry.