How do non-deposit sites influence the performance of machine learning-based gold prospectivity mapping? A study case in the Pitangui Greenstone Belt, Brazil

One of the greatest challenges in mineral prospectivity mapping (MPM) research nowadays is to find a solid methodology that ensures the reliability of the prospectivity model during the learning and prediction procedures. Multiple uncertainties such as the location of non-deposit sites or the type of machine learning algorithm (MLA) can bias the MPM. To investigate these effects, we used multiple training datasets with different non-deposits locations, randomly created, and MLAs such as Artificial Neural Network (ANN), Random Forests (RF) and Support Vector Machine (SVM), to model orogenic-Au prospectivity in the Pitangui Greenstone Belt (PGB, Brazil). Regarding the implications in the methodology for MPM, there are great differences between the models' performances in mapping prospective zones when there is a slightly change in the location of negative samples. These changes can be observed by using the Shapley additive explanation metrics (SHAP values), which can help mitigate such effects by choosing an optimal model among all randomly created datasets. The SHAP values of non-deposit sites also showed that ANN and SVM present overfitting problems despite the use of balanced data. RF on the other hand outperformed in all ten datasets and showed great recognition and adjustment to the negative samples. The results presented in this research are also promising to the prospective studies in the PGB, as it shows a map capable to correctly predict 97 % of the known deposits and occurrences in 3 % of the total area and points the new frontiers for gold exploration in the PGB.