Integration of auto-encoder network with density-based spatial clustering for geochemical anomaly detection for mineral exploration

摘要

Auto-encoder network can be used for dimensionality reduction of data and for re-construction of sample population with unknown, complex multivariate probability distribution, where small-probability samples have little contribution to the auto-encoder network, leading to high re-construction error. In this paper, the trained auto-encoder networks were used to detect geochemical anomalies. Compared with deep auto-encoder network, the density-based spatial clustering application with noise (DBSCAN) regards noise samples (e.g., geochemically anomalous samples) that differ from core samples (e.g., geochemically background samples) as anomalies. Therefore, the learned representations from the code layer in the auto-encoder network are clustered by DBSCAN to detect noise samples representing geochemical anomalies. As benchmark for evaluating the performance of auto-encoder network and DBSCAN, and in consideration of the compositional nature of geochemical data, the compositional multivariate outlier detection was also applied. We applied these methods to two forms of the geochemical data, namely (1) without any transformation and (2) with isometric log ratio transformation. The similarities of the resulting anomaly maps in terms of data forms indicate that the auto-encoder network is effective for detecting multivariate geochemical anomalies. Differences between the anomaly maps indicate, however, that the compositional nature of geochemical data affects the performance of multivariate geochemical anomaly detection. Nevertheless, the assessment, by receiver operating characteristics analysis, of the geochemical anomalies derived using the different methodologies described implies that the detected geochemical anomalies are related to Au mineralization. Finally, the Youden index, which measures the relationship between binary anomalies and known deposits, was used for optimal threshold selection to create an optimal mineral potential map from the derived continuous geochemical anomaly data. The spatial distribution of geochemical anomalies at/around faults and magmatic rocks provides insights to where further detailed exploration is warranted in the study area.