Effects of Random Negative Training Samples on Mineral Prospectivity Mapping

Supervised data-driven mineral prospectivity mapping (MPM) usually employs both positive and negative training datasets. Positive training datasets are typically created using the locations of known mineral deposits, whereas negative training datasets can be generated using the locations of random points. However, not all the negative points can be treated as true negative samples, which means that the selection of random negative training points creates uncertainty. This study provides a framework for addressing the effects of random negative training points on MPM. Fifty negative training datasets were generated using random point locations, and 50 mineral potential maps were created using logistic regression model. The area under the receiver operating characteristic curves (AUC) was used to evaluate the MPM performance. The mean of the AUC was employed to represent the average spatial correlation between mineralization and the selected spatial patterns, and the standard deviation of the AUC was used to indicate uncertainty relating to the use of random negative training samples. Additionally, a risk and return analysis was conducted to explore the uncertainty of the mineral potential map due to the use of random negative samples, and an odds ratio of probability was employed to describe the chances of both the occurrence and non-occurrence of a mineral deposit. The risk and return maps were obtained using the average and standard deviation of the log odds ratio per location. The mean of the log odds ratio was transformed into a mean probability, which can be regarded as the final mineral potential value that considers uncertainty due to random negative samples. Initial mineral exploration can thus be prioritized in areas that have low risk and high return compared to other areas. The proposed framework was demonstrated by mapping potential for skarn Fe mineralization in southwestern Fujian, China.