Geochemical anomaly identification and uncertainty quantification using a Bayesian convolutional neural network model

Abstract

Geochemical prospecting plays an important role in mineral exploration. In recent years, deep learning algorithms (DLAs) have been applied in mapping geochemical anomalies associated with mineralization. However, few of them evaluated the effects of data and model uncertainty on geochemical anomaly identification, which can introduce bias and risk in geochemical prospecting. In this study, a Bayesian convolutional neural network model (BCNN), which applies Bayes by Backprop to a convolutional neural network, was employed to extract geochemical anomalies associated with mineralization and quantify the two types of uncertainties, i.e., aleatoric (data-related) and epistemic (model-related) uncertainty. A case study on recognizing geochemical anomalies related to gold polymetallic mineralization in northern Sichuan Province of China was conducted. The results show that the geochemical anomalies extracted by BCNN are highly spatially correlated with known gold deposits. Meanwhile, both aleatoric and epistemic uncertainty linked to the identified geochemical anomalies were quantified. Areas with high uncertainty are mainly distributed at the boundaries of the geochemical high anomalous zones. Therefore, future mineral exploration should focus on regions with high anomalies but low uncertainty. The obtained results provide an important decision basis for the prospecting of gold polymetallic deposits in the study area.