Geologically constrained unsupervised dual-branch deep learning algorithm for geochemical anomalies identification

The identification of geochemical anomalies is crucial in mineral exploration. However, the limited sample size, high-dimensional features, and mixed geochemical information make identifying geochemical anomalies a significant challenge. Machine learning algorithms (MLAs), especially those with spatial and spectrum branches, have been proven to be a high efficiency tools for detecting geochemical anomalies related to mineralization. The spatial branch MLAs take two-dimensional images (pixel-patches) as input and mainly capture the spatial characteristics of geochemical patterns. The spectrum branch MLAs take one-dimensional sequence data (pixels) as input and mainly consider the elemental concentration and assemblies. Simultaneously considering the spatial patterns and geochemical concentrations of the geochemical survey data can mitigate geochemical concentration variations arising from objective factors and amplify subtle mineralization anomalies. This study proposes an unsupervised spatial-spectrum autoencoder (AE), namely dual-AE, which consists of a graph convolutional autoencoder (GCN-AE) and a long short-term memory network autoencoder (LSTM-AE) for geochemical anomalies identification. The spatial branch is constructed using the GCN-AE, which can effectively capture spatial geochemical patterns and extract spatial relationships between neighboring pixels. The spectrum branch consists of an LSTM-AE that can study geochemical elemental assemblies within a single pixel. A key ore-controlling factor was added into the dual-AE as a soft constraint to construct a geologically constrained dual-AE. A case study was conducted to recognize geochemical anomalies associated with iron polymetallic mineralization in Southwest Fujian Province, China. The obtained results demonstrated that (1) the unsupervised spatial-spectrum deep learning algorithm serves as a potent method for detecting geochemical anomalies related to mineralization, (2) the geologically constrained unsupervised spatial-spectrum dual-branch model can improve the accuracy and interpretability of geochemical anomaly identification, and (3) the identified anomalous areas can provide essential clues for further mineral exploration.