This study intends to showcase the application of Extreme Gradient boosting (XGboost), a state-of-the-art ensemble-learning technique, for district-scale mineral potential mapping (MPM) of carbonate-hosted Zn-Pb mineral systems in the Emarat district, W Iran. Notwithstanding the demonstrated effectiveness of XGboost in addressing a range of classification and regression problems, the inadequate number of deposit locations employed as labeled samples, which is intrinsic to district-scale MPM, impedes the practical application of this method. Given this notion, an iterative window-based data augmentation technique was proposed and applied to generate additional labeled samples. The predictive models derived by XGboost were compared to those generated by random forests (RF). Both methods were applied to a set of mineral systems-derived exploration targeting criteria representing critical ore-forming processes. Setting aside the similarities of predictive models derived by both methods, it was recognized that RF is much less sensitive to insufficient labeled samples. It was also recognized that, unlike RF, many parameters require tuning for generating an optimum XGboost-based predictive model. Nevertheless, given the achieved results, XGboost slightly outperformed RF in generating an optimum predictive model. Therefore, XGboost coupled to data augmentation can be deemed a viable alternative for district-scale, data-driven MPM.
