Class Label Representativeness in Machine Learning-Based Mineral Prospectivity Mapping

Abstract

Mineral prospectivity mapping (MPM) can be deemed a binary classification task, with classifiers trained and validated on labels indicating the presence or absence of the targeted mineralized zones. Using economically viable mineral deposits as positive labels could, in theory, yield prospectivity models with geometallurgical reliability, thereby aiding land management and decision-making. The inherent scarcity of economically viable deposits, however, ultimately affects MPM products. The positive class label, therefore, often requires augmentation with either mineral occurrences (i.e., mineralized sites lacking economic viability) or synthetically generated labels. This paper examines how augmented positive labels and different negative label selection procedures geospatially represent economically viable mineral deposits and affect deep learning-based MPM’s classification performance and its spatial selectivity (i.e., MPM’s capability to efficiently narrow the exploration search space). To achieve this objective, large ensembles of deep learning classifiers were trained and validated with diverse combinations of positive and negative labels. Two positive class label sets were created by augmenting mineral deposits with either synthetic labels, generated using generative adversarial networks, or mineral occurrences, paired with distinct negative label sets selected based on (1) locations distant from known mineral deposits, (2) areas geospatially dissimilar to known mineral deposits, and (3) mineralized areas unrelated to the targeted style of mineralization, resulting in six unique class configurations. This study ultimately provides insights into how different label sets affect MPM's classification performance and spatial selectivity. The results indicate that selecting negative class labels from geospatially different localities enhances classification performance and MPM's spatial selectivity compared to other negative label selection procedures.