On data quality in mineral potential modelling: A case study using random forest and fractal techniques

Abstract

Summary Here we report the application of a transformed predictor map-based random forest (RF) technique to mineral potential modelling (MPM) of orogenic gold mineral systems in the well-endowed Granites-Tanami Orogen (GTO), Northern Territory, Australia. In this study, we compared a previously developed random forest (RF) model based on 19 predictor maps ( Roshanravan et al., 2020 ) against a new RF model employing an expanded set of 23 predictor maps with the four new predictors representing more detailed geophysical data not previously available. We found the newly constructed RF model performed better than the previous version of Roshanravan et al. (2020 ), illustrating the value of high-quality input data to better reflect the mappable ingredients of the targeted mineral deposit type. To further constrain the new RF modelling results, the gold prospective domains were delimited using the concentration-area (C-A) fractal technique. The resulting first order targets occupy only c. 2% of the study area while capturing 76% of the known gold occurrences. The area captured by the first order targets represents a >50 times (i.e., an order of magnitude) reduction of the search space, the hallmark of a well-performing, practically useful targeting technique.