unsupervised anomaly detection

Recent advancements in autoencoders and their variants have notably enhanced the detection of multi-element geochemical signatures linked to ore occurences. This research employed a convolutional autoencoder algorithm (CAE) to identify geochemical anomalies, leveraging the algorithm’s ability into account the spatial correlation within the geochemical dataset. In this framework, two stream sediment datasets were generated in the Feizabad district using a conceptual modelling approach alongside a big data analysis strategy. These datasets were individually fed into the CAE model to identify multi-element geochemical anomalies based on the reconstruction error in an unsupervised manner. A comparative analysis of two geochemical prospectivity models and the simplified geological map of Feizabad demonstrates a strong spatial correlation between the identified anomaly regions and known mineral occurrences, which are distributed across andesite, tuff, and Eocene-Oligocene intrusive rocks. However, a quantitative assessment using prediction-area plots indicates that the multi-element geochemical map derived from the conceptual model exhibits a higher prediction rate (72%) compared to the geochemical prospectivity map generated through the big data approach (63%).