Abstract
The nature of targets extracted through data-driven mineral prospectivity mapping (DD-MPM) is dependent on geodata annotation. However, geodata annotation is noisy because mineral systems are complex systems and the definition of positivity changes with annotator, jurisdiction and time. Consequently, historically annotated data (HAD), which are the exhaustive collection of labeled samples (sites) that have accumulated over the history of exploration, is a noisy and subjective source of ground truth. However, the spatial sensitivity and selectivity of DD-MPM products to annotation is quantitatively unknown. Although HAD are always used at large scales of MPM, the cost-to-benefit analysis of improving and augmenting them is unclear because of inherent limitations of DD-MPM, including: (1) scientific confounds that ironically bias targeting toward the known as opposed to truly greenfield; (2) research design challenges that create circular logic during validation; and (3) geoscience and implementation uncertainty that make DD-MPM products heuristic. We conducted two experiments to: (1) verify the distribution of noisy positive samples in HAD and test whether a minority of samples carry a majority of information; and (2) assess the effectiveness of mineral system constraints by annotating maximally separable mineral systems of a commodity and relating their target overlap with class separability. Using a HAD database containing all copper samples in Canada, we observed that: (1) about 14% of the cleanest samples in HAD generated 71% of all spatial targets; and (2) fully separable mineral systems generated spatially overlapping targets, which is a hypothesis-generating finding. Together, these results (a) provide scientific evidence for annotation noise, its characteristics and impact; and (b) highlight the value of HAD and the need for more rigorous frameworks of mineral system-based DD-MPM.