A Spatially Constrained Multi-autoencoder Approach for Multivariate Geochemical Anomaly Recognition

Separating and recognizing geochemical anomalies from the geochemical background is one of the key tasks in geochemical exploration. Many methods have been developed, such as calculating the mean ±2 standard deviation, and fractal/multifractal models. In recent years, deep autoencoder, a deep learning approach, have been used for multivariate geochemical anomaly recognition. While being able to deal with the non-normal distributions of geochemical concentrations and the non-linear relationships among them, this self-supervised learning method does not take into account the spatial heterogeneity of geochemical background and the uncertainty induced by the randomly initialized weights of neurons, leading to ineffective recognition of weak anomalies. In this paper, we introduce a spatially constrained multi-autoencoder (SCMA) approach for multivariate geochemical anomaly recognition, which includes two steps: spatial partitioning and anomaly score computation. The first step divides the study area into multiple sub-regions to segregate the geochemical background, by grouping the geochemical samples through K-means clustering, spatial filtering, and spatial constraining rules. In the second step, for each sub-region, a group of autoencoder neural networks are constructed with an identical structure but different initial weights on neurons. Each autoencoder is trained using the geochemical samples within the corresponding sub-region to learn the sub-regional geochemical background. The best autoencoder of a group is chosen as the final model for the corresponding sub-region. The anomaly score at each location can then be calculated as the euclidean distance between the observed concentrations and reconstructed concentrations of geochemical elements.The experiments using the geochemical data and Fe deposits in the southwestern Fujian province of China showed that our SCMA approach greatly improved the recognition of weak anomalies, achieving the AUC of 0.89, compared with the AUC of 0.77 using a single deep autoencoder approach.