Machine learning-driven mineral prospectivity mapping: A predictive performance analysis in Janja, Iran

Geochemical analysis is an effective technique for detecting mineral deposits by examining element concentrations. Various statistical techniques have been developed to differentiate abnormal values from background values. A more accurate analysis can be obtained by employing multivariate statistical methods. The use of these methods enables the simultaneous analysis of changes in multiple variables. This research utilized correlation coefficients, cluster analysis, and factor analysis to demonstrate the genetic connections among various elements. The factor analysis method was additionally applied to generate multivariable maps and comprehensive multivariable results. Moreover, the stepwise factor analysis (SFA) method, an enhanced version of traditional factor analysis, was utilized to produce geochemical distribution maps. This technique entails initially recognizing and removing non-representative elements, followed by identifying the most important and impactful representative factors. This study demonstrates the efficacy of the SFA method when applied to geochemical data. This approach removes superfluous elements and increases the variance attributed to the predictive mineralization factor, thereby improving the geochemical halos. Additionally, this research evaluated multivariate analysis approaches alongside machine learning techniques. To achieve this, a multilayer perceptron neural network (MLP) was used to evaluate the levels of gold, silver, copper, lead, and zinc in the study area. The output variable represented the grade of a particular element individually, whereas the input variables encompassed the grades of the remaining four elements. To optimize the model, different quantities of hidden layers and a range of activation functions were applied. Ultimately, an ideal model was developed for each element. The model achieved accuracies of 95%, 88%, 73%, 80%, and 72% for the gold, silver, copper, lead, and zinc, respectively. The results show the significant computational efficiency of this method in assessing element grades. Finally, the element distribution maps generated by both methods indicate that the MLP approach identified the anomalous areas with higher accuracy.