Identification of multi-element geochemical anomalies using unsupervised machine learning algorithms: A case study from Ag–Pb–Zn deposits in north-western Zhejiang, China

Abstract

Identifying significant geochemical anomalies is a fundamental task in mineral exploration. In terms of stochastic process, a specific ore-forming process can be regarded as a small probability event since the deposit type sought rarely occurs in the crust of the earth. Therefore, machine learning algorithms for anomaly detection are suitable for identification of mineralization-related geochemical anomalies. However, identification of significant geochemical anomaly signatures is often hindered by irrelevant and/or redundant variables in a geochemical dataset. It is then necessary to extract significant features by means of dimensional reduction before applying unsupervised machine learning algorithms for anomaly detection. In this study, six unsupervised machine learning methods including isolation forest (IF) and five hybrid models that combined IF with dimensional reduction techniques were implemented and evaluated in identifying Ag–Pb–Zn polymetallic mineralization-related geochemical anomalies using stream sediment geochemical data from north-western Zhejiang Province, China. The five hybrid models involved combining IF with hierarchical clustering (HC-IF), principal component analysis (PCA-IF), stacked denoising autoencoder (SDAE-IF), both HC and PCA (HC-PCA-IF) and both HC and SDAE (HC-SDAE-IF). Anomalies identified by these six methods were evaluated by the receiver operating characteristic (ROC) curve method. The spatial overlay of known mineral deposits and the identified geochemical anomaly maps showed that a strong spatial association existed between the locations of discovered mineral deposits and high geochemical anomaly areas. Furthermore, the area under the curve (AUC) values suggested that the HC-SDAE-IF hybrid model performed best and the HC-IF and HC-PCA-IF models performed better than the IF, PCA-IF and SDAE-IF models. This study demonstrated that: (i) IF is a useful and efficient unsupervised machine learning algorithm for detection of Ag-Pb-Zn polymetallic mineralization-related geochemical anomalies in the study area; and, (ii) the proper manipulation of dimensional reduction and extraction of mineralization-related geochemical variables can not only significantly improve the performance of unsupervised machine learning algorithms including deep learning methods in geochemical anomaly detection, but also be instrumental in clarifying the geological meaning of the identified geochemical anomaly. In this regard, reducing dimensionality and extracting mineralization-related variables (i.e. elements) from raw geochemical data is strongly recommended before applying unsupervised machine learning methods, particularly the sophisticated deep learning algorithms, in geochemical anomaly detection.