A Multi-Model Ensemble Approach for Gold Mineral Prospectivity Mapping: A Case Study on the Beishan Region, Western China

Kaijian Wang,Dongya Liu,Gongwen Wang,Ning Cui,Xinqi Zheng

doi:10.3390/min10121126

Abstract

Mineral prospectivity mapping (MPM) needs robust predictive techniques so that the target zones of mineral deposits can be accurately delineated at a specific location. Although an individual machine learning algorithm has been successfully applied, it remains a challenge because of the complicated non-linear relations between prospecting factors and deposits. Ensemble learning methods were efficiently applied for their excellent generalization, but their potential has not been fully explored in MPM. In this study, three well-known machine learning models, namely random forest (RF), support vector machine (SVM), and the maximum entropy model (MaxEnt), were fused into ensembles (i.e., RF–SVM, RF–MaxEnt, SVM–MaxEnt, RF–SVM–MaxEnt) to produce a final prediction. The paper aims to investigate the potential application of stacking ensemble learning methods (SELM) for MPM. In this study, 69 hydrothermal gold deposits were split into two parts: 70% for the training model and 30% for testing the model. Then, 11 mineral prospecting factors were selected as a spatial dataset constructed for MPM. Finally, the models’ performance was assessed using the receiver operating characteristic (ROC) curves and five statistical metrics. Compared with other single methods, the SELM framework showed an improved predictive performance in the model evaluation. Therefore, this finding suggests that the SELM framework is promising and should be selected as an alternative technique for MPM.

Highlights

The demand for mineral resources has grown significantly in recent years [1], mainly due to rapid industrial development in developing countries, such as China, India, and Brazil [2]
There are three main contributions of this work: First, we aimed to investigate the potential application of ensembles of random forest (RF), support vector machine (SVM), and maximum entropy (MaxEnt) algorithms (i.e., RF–SVM, RF–MaxEnt, SVM–MaxEnt, and RF–SVM–MaxEnt) for predictions of mineral prospectivity in the Beishan region, west China
These results show that mineral prospectivity mapping (MPM) has a much larger area under the curve (AUC) value in the ensembles, indicating that heterogeneous stacking ensemble learning methods (SELM) will help to reduce the uncertainty of multiple variables to some extent in mineral potential modelling

Summary

Introduction

The demand for mineral resources has grown significantly in recent years [1], mainly due to rapid industrial development in developing countries, such as China, India, and Brazil [2]. The prediction of mineral prospectivity is a multivariable decision-making tool used to draw and rank zones that have the highest potential for mineral exploration in unexplored regions [3]. Mineral predictive modelling is a vital but challenging step for the mapping of undiscovered prospective deposits in mineral prospectivity mapping (MPM). Effective modelling techniques of mineral resource exploration are increasingly critical for contributing to sustainable economic growth on the national level. As such, mapping new mineral prospectivity has become imperative, and predictive modelling provides a scientific means for delineating the intricate spatial patterns of features that are closely related to mineralisation [4].

Objectives

Discussion

Conclusion