Recognition of geochemical anomalies using a deep variational autoencoder network

Deep learning (DL) algorithms have received increased attention in various fields. In the field of geoscience, DL has been shown to be a powerful tool for mining complex, high-level, and non-linear geospatial data and for extracting previously unknown patterns related to geological processes. In this study, a deep variational autoencoder (VAE) network was used to extract features related to mineralization; and these features were then integrated as a anomaly map in support of mineral exploration based on geochemical exploration data, which consist of Cu, Pb, Mn, Zn and Fe2O3. Various experiments were conducted to determine the optimal parameters of the VAE. The structure of the VAE, in which the network depth and number of hidden units were 24-12-3-12-24, was built to recognize geochemical anomalies related to Fe polymetallic mineralization in the southwest Fujian Province, China. The geochemical anomalies recognized by the VAE show a close spatial correlation with known Fe polymetallic deposits. Meanwhile, the areas with high probability are located in or around the Yanshanian intrusions and the contact zones of the Carboniferous-Permian formation and Yanshanian intrusions. These results suggest that the anomalous areas identified by the VAE are meaningful for mineral exploration.