<p>Data engineering is a challenge in mineral prospectivity mapping using supervised data-driven methods because of a general paucity of true samples, both positive and negative. Naïve random sampling of unlabeled samples as presumptive negative samples is a common method, but it creates label crossover, which is the condition that otherwise true positives would be reversed in class designation. For mineral prospectivity mapping, label crossover has two main effects: (1) it artificially decreases the area of positive prospectivity because label crossover is asymmetrical; and (2) it increases model variance by increasing model complexity and, therefore, weakens the spatial continuity of predictions. These effects reduce the realism and trustworthiness of spatial analytic products (e.g., maps), which hinders their adoption. Here, we propose a data-driven and recursive annotation method based on positive and unlabeled learning to annotate negative samples. Our method progressively biases the negative labeling of unlabeled samples using prospectivity scores (as a class prior). The optimization of the number of recursions occurs through a joint maximization of the variable (aspatial) and spatial domain objectives of prospectivity mapping. Reducing label crossover improves the performance of a deep ensemble (of unique workflows) and increases spatial continuity of the resulting prospectivity map. Only two iterations of annotation were necessary for the refinement of a graphite prospectivity map in Canada. By its design,&#xa0;recursive annotation is broadly compatible with a range of workflows.</p>