Mineral prospectivity mapping based on Support vector machine and Random Forest algorithm – A case study from Ashele copper–zinc deposit, Xinjiang, NW China

Abstract

The utilization of machine learning (ML) techniques in conjunction with multi-source geoscience datasets for comprehensive metallogenic prognosis (MP) has emerged as a novel means of geological prospecting. Nevertheless, the representativeness of features and composition of datasets employed in constructing the ML model for MP may substantially impact the prediction model's overall performance and augment the metallogenic prediction's uncertainty. In this study, the Ashele copper-zinc deposit was chosen to conduct the case study to resolve these challenges. In order to improve the representativeness of features employed in ML modeling, the ore-controlling geological and rock geochemical exploration criteria were digitized and subjected to spatial analysis to extract information closely linked to mineralization. To investigate the impact of differences in dataset composition on the performance of predictive models, multiple labeled datasets were generated by randomly sampling from known boreholes. The predictive SVM and RF models were trained using the labeled datasets, validated using 10-fold cross-validation, and assessed using multiple metrics. The evaluation findings indicate that all ML models, subsequent to the enhancement of feature representativeness, exhibited satisfactory performance, as evidenced by a consistent predictive accuracy of approximately 0.88. Additionally, the study revealed that variations in the labeled datasets could adversely affect the predictive models' performance. However, this effect could be alleviated by selecting merit-based predictive models constructed from diverse dataset compositions. A comparative analysis of the ML predictive models revealed that the RBF kernel SVM and RF models demonstrated exceptional performance, with an AUC of 0.972 and 0.968, respectively. As a result, an integrated prediction model was developed by overlaying the RBF kernel SVM and RF models, successfully identifying 98% of the mineralized boreholes in the top 15% of the study area. The spatial distribution of gravity and magnetic anomalies in the study area corresponded well with the prospecting areas identified by the machine learning predictive model, further validating the efficacy of ML methods in MP.