Analysis of machine learning mineral prospectivity models at a project-scale using scarce training dataset

Mineral exploration is the necessary first step of any mining project. Mineral prospectivity analysis is a cost and time efficient exercise with for goals to delineate area of high prospectivity or to rank targets. The worldwide tendency in mineral exploration efforts is to focus on brownfield areas where large amount of data is available. Hence, data-driven methods for mineral prospectivity modeling (MPM) is preferred. In this research, two aspects of data-driven MPM are explored to examine the influence of them on the mineral prospectivity map using two different machine learning algorithm (random forest (RF) and support vector machine (SVM)). This research aims at demonstrating that RF algorithm is the best method for MPM in a variety of case scenarios involving different training datasets and input features. This study primary target is epithermal Au deposit in the area of the Iskut property owned by Seabridge Gold Inc. Different training dataset were created using same 18 deposit locations but different set of non-deposit locations: selected non-prospective known mineral occurrences (KMO) for Au deposits and 10 sets of random point pattern. Predictor maps were generated from publicly available and privately-owned geospatial data based on a conceptual exploration model using a mineral system approach and were separated into two input datasets: one exclusively included public data while the other included data from the private and public domain. Data-driven mineral potential models using RF and SVM (using three different kernel function) were compared based on sensitivity to parameter configuration, accuracy and performance. It was found that the accuracy of a model increases when the number of predictor maps increases. This research also showed that non-prospective KMO based uniquely on distance and commodity can introduce a bias. On one hand, almost all the SVM models are overfitting, most likely because of the scarce training dataset. Moreover, they are sensitive to outliers in the training data and require long computational time. On the other hand, the RF is easy to parameter, transparent, less sensitive to outliers and performs well. Hence, RF is the method to opt for in data scarce region over SVM.Applied Science, Faculty ofMining Engineering, Keevil Institute ofGraduat