Mineral Prospectivity Mapping based on Isolation Forest and Random Forest: Implication for the Existence of Spatial Signature of Mineralization in Outliers

Known mineralized locations and randomly chosen non-mineralized locations are used traditionally as training samples in data-driven mineral prospectivity mapping (MPM). In this paper, we took advantage of (a) the variable importance and partial dependence plot, which enable interpretation of random forest (RF) modeling, and (b) the synthetic minority over-sampling technique, and investigated the efficacy of outlier-based training samples used for data-driven MPM in contrast to traditional practice of using known mineralized locations as positive training samples. The prediction maps obtained by RF modeling based on different sets of training samples suggest bias toward known mineralized locations in data-driven MPM. The proposed outlier-based training samples for data-driven MPM involve both unsupervised learning and supervised learning. The former aims at outlier detection, while the latter uses the resulting outliers as positive training samples to investigate the following: firstly, the delineation of prospective area or spatial signature of existing mineral system by avoiding the bias arising from the known mineralized locations in data-driven MPM and secondly, the coherence of spatial signature of outliers, which justifies their feasibility as positive training samples for data-driven MPM by RF modeling. Analyses of receiver operating curves and correlations of the resulting prediction maps imply that outliers derived by isolation forest show consistent spatial signature as the known mineralized location and, thus, were effective in narrowing down the prospective target areas just like in traditional data-driven MPM.