Abstract
Mineral exploration in the Arctic is challenging, resulting in decreased survey and mineral discovery rates. Data-driven mineral prospectivity mapping is intended to reduce exploration risk by delineating promising areas. However, data sparsity, especially of deposit locations, is a critical bottleneck. In this study, we developed and tested a method to conduct data-driven mineral prospectivity mapping in regions where annotated data are too sparse. Our method is predicated on the existence of large-scale mineral prospectivity frameworks (data, models and methods), which are fine-tuned using regional ground truth to produce locally refined predictions. This is possible because the spatial transferability of trained machine learning models is incomplete. We explicitly tested the hypothesis of limited spatial transferability. To within limitations of our experimental design, our results are supportive of our hypothesis that the incorporation of new positive samples into an existing training dataset preferentially influences the local outcome. Using a suite of critical minerals combined with published data and models in Canada, we demonstrated that by treating large-scale mineral prospectivity frameworks as foundational, it is productive to fine-tune trained models using regional data. Although national-scale maps exist prior to this study, they did not incorporate a substantial amount of ground truth from Nunavut. Therefore, through regional refinement, we produced the inaugural version of a set of pan-Canadian mineral prospectivity maps that are specifically refined for Nunavut for: (1) flake graphite; (2) magmatic Ni (±Cu ±Co ±PGE); and (3) Mississippi Valley-type Pb–Zn. The regional refinement method is a perspectivist and federated approach to MPM that extracts additional value from large-scale mineral prospectivity frameworks and provides an interoperable method for national, regional and inter-regional mineral exploration.