Targeting greenfield mineral prospectivity using scarce data-constrained AI framework and dynamic confusion matrix: A case study from Yulong district, China

Mineral prospectivity mapping (MPM), spurred by data-driven paradigm and artificial intelligence (AI) algorithms in the recent two decades, plays a crucial role in mineral exploration. MPM has demonstrated its great performance in data-rich brownfield regions, however, it faces tremendous challenges when applied in greenfield areas where scarce data are available. To address the challenges, we proposed a MPM framework based on few-show AI modeling and novel dynamic confusion matrix, applied to Yulong district, which has overall low regional exploration extent but proved to have significant mineral potential due to the discovery of the giant Yulong deposit. Hydrothermal alteration and lineament were extracted from OLI and ASTER data to enrich the data source for MPM, which were then integrated with geological and geochemical evidence information to build the input feature dataset for AI algorithms. The synthetic minority oversampling technique (SMOTE) was employed to augment the scarce data of mineral deposit samples, and the t-SNE algorithm was subsequently used to visualize and verify the feature distribution of the synthetic samples, confirming that the augmented samples conform to the distribution pattern of real mineralization features. Both classic machine learning models (support vector machine, SVM and random forest, RF) and cutting-edge deep learning models (convolutional neural network, CNN and deep forest, DF) were employed to train predictive models. A novel evaluation metric called dynamic confusion matrix, which iteratively adjusts classification thresholds to capture continuous model performance variations, was proposed to assess the model efficacy and delineate the exploration targets under the scarce data-constrained condition. The results indicate that CNN model achieves the best classification performance, indicated by its highest values of the area under the curve (AUC) for dynamic evaluation curves of accuracy (0.9648), recall (0.9974), F1-score (0.9662), and Kappa coefficient (0.9308). DF model captures 100% known mineral deposits within 2.28% delimited area, exhibiting the highest predictive efficiency. Spatial cross-validation (SCV) results further validate that the CNN is the model with a positive improvement in classification accuracy, with its a-AUC rising from 0.9648 to 0.9683. Meanwhile, its r-AUC, F1-AUC and k-AUC decrease by 6.11%, 0.58% and 0.58% respectively, with all evaluation indicators ranking first in SCV, demonstrating the strongest spatial generalization capacity for unexplored greenfield areas. Shapley Additive Explanations (SHAP) analysis revealed RF’s clear decision path distinguishing deposits from non-deposits. The geochemical features exert the most significant influence on model prediction, and the remote sensing-derived features contribute considerably to model output, which is especially highlighted under the proposed dynamic evaluation metric. Four target areas were delineated based on the thresholds derived from dynamic confusion matrix, taking consideration of uncertainty analysis. The proposed scenario of exploration targeting was compared with conventional methods of success-rate curves and prediction-area (P-A) plots. It is inferred that the proposed method exhibits better performance and greater practicality in scarce data-constrained MPM tasks. The framework proposed in this study provides an effective solution for AI-driven predictive modeling in greenfield areas, facilitating the decision-making for exploration targeting under the condition of extremely scarce labeled data.