Unsupervised detection of multivariate geochemical anomalies using a high-performance deep autoencoder Gaussian mixture model

Abstract

It is of great significance to construct an efficient geochemical anomaly detection model for the successful accomplishment of a mineral exploration process in a complex geological environment. However, the complex geological environment of the prospecting area often results in the high-dimensional unknown complex population distribution of geochemical exploration data. This complex distribution is difficult to fit with a theoretical probability distribution model. As a result, it becomes a challenge to carry out an effective detection of geochemical anomalies. Therefore, to develop an anomaly detection model that can effectively fit the complex population distribution of geochemical exploration data is the key for accurately detecting geochemical anomalies. For this reason, the deep autoencoder Gaussian mixture model (DAGMM) was adopted to model the geochemical exploration data obtained in the 1:200,000 geological survey conducted in the Baishan area (Jilin, China) to check its superiority in identifying multivariate geochemical anomalies. As an innovative deep learning framework for unsupervised anomaly detection, DAGMM ingeniously combines the data dimensionality reduction and compression capabilities of a deep autoencoder (DAE) with the probability density estimation advantage of the Gaussian mixture model (GMM). The DAGMM model can deeply explore the deep-level features of geochemical exploration data and effectively model the complex unknown data distribution through the synergistically work and joint optimization strategy in training the DAE and GMM model, so it can accurately identify geochemical anomalies. To show the superiority of the DAGMM model in detecting polymetallic geochemical anomalies, the DAGMM model was compared with the GMM and DAE models. The receiver operating characteristic (ROC) curves of the three models were plotted, and the areas under the ROC curves (AUCs) and lift indices were calculated. The ROC curve of the DAGMM model dominates that of the DAE model and GMM model. The DAGMM model has an AUC of 0.904 and a lift index of 10.44, respectively, which are much larger than those of the GMM model (AUC = 0.858, lift index = 3.63) and DAE model (AUC = 0.83, lift index = 5.31). Therefore, the DAGMM model significantly outperforms the other two models in detecting multivariate geochemical anomalies and the polymetallic geochemical anomalies detected by the DAGMM model contain all the known polymetallic deposits. Compared with DAE and GMM, DAGMM is more efficient and more powerful in detecting multivariate geochemical anomalies in complex geological environments.