The Trident project is located in the Domes region of the Central African Copper Belt and hosts a number of mineralised systems including the Sentinel (Ni) and Enterprise (Cu) deposits. The project has received extensive systematic geochemical soil sampling in addition to high resolution airborne geophysical coverage. This data-rich environment enables experimentation with machine learning strategies which aim to produce or refine geological maps from limited direct observations.

In this study we present a series of three case studies that test lithological classification using the supervised Random Forests algorithm. These studies inform the situations encountered in mineral exploration including early stage lithology mapping and more mature stage map refinement. We also present a fourth study, using the unsupervised algorithms k-means and Self-Organising Maps, to identify clusters, potentially associated with lithology in absence of a priori geological information. Our case studies are most relevant to the situation where the geology of a prospect is largely concealed beneath extensive cover rocks, with some rock types being poorly expressed or even absent in outcrop.

We find that sampling from limited outcrop produces a RF lithology prediction that is likely to be incorrect. We demonstrate that balancing sample size through a combination of decimation and bootstrapping can improve results. Additionally, we identify some important indicators in both the predicted geology and uncertainty metrics which could alert an explorer to an inability of their training data to make accurate predictions and to the presence of lithological classes not expressed in outcrop. Sampling from a mature lithology map enables further map refinement and acts as an objective audit of the existing product. Information entropy (H) is calculated as a metric to describe quantitatively the uncertainty associated with classification, provide valuable information on the geological complexity of the mapped region and highlight areas which are potentially misclassified. Clusters obtained using the k-means algorithm produced a result more consistent with lithology in this instance and was faster; however Self Organising Maps remains attractive due to the production of additional metrics to assess algorithm performance. Clustering could be used either in the development of a first pass interpretation, or in the critical appraisal and subsequent refinement of existing interpretations.