Abstract

Mineral systems are composed of many interacting components that lead to complex, singular and rare properties of geo-data. In mineral prospectivity mapping (MPM), supervised machine learning algorithms, which have advantages in dealing with complex geo-data, usually involve uncertainty resulting from the discretization of continuous evidential maps into arbitrary classes as well as the large data imbalance caused by the rarity of deposit locations. Consequently, the predicted results may be biased. In this paper, a random forest (RF) algorithm based on the bagging technique is used to map the prospectivity of tungsten polymetallic deposits in the Nanling metallogenic belt. Data-driven logistic transformation is employed to obtain continuous evidential maps. Both discretized and continuous evidential maps are used to generate prospectivity models for comparison. To reduce the data imbalance, the under-sampling method and the synthetic minority over-sampling technique (SMOTE) are implemented to generate balanced datasets. The receiver operating characteristic (ROC) curve and improved prediction-area (P-A) plot are applied to evaluate the prospectivity models. The predictive results show that when using the RF algorithm in MPM, the application of continuous evidential maps can improve the performance of prospectivity models and reduce the uncertainty resulting from the discretization of evidential maps. The prospectivity model trained with a balanced SMOTE-generated dataset shows the best overall performance for improving the percentage of deposit locations that are correctly predicted and decreasing the percentage of non-deposit locations that are inaccurately identified as deposit locations to some extent. In addition, the improved P-A plot is superior to the ROC curve because the latter neglects the occupied area, which is critical for mineral exploration and may provide an overly optimistic performance with imbalanced data. However, further testing of the evaluation criteria and the SMOTE approach to reduce data imbalance is warranted to determine fully the universality in MPM from the perspective of data imbalance. Based on prospectivity models, four high-potential areas and five moderate-potential areas are delineated, which indicates good future prospecting for tungsten polymetallic deposits in this region.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call