Detection of geochemical anomalies related to mineralization using the GANomaly network
Introduction
The complexity of geological systems and the multi-phase nature of ore-forming processes lead to a complex and non-single statistical distribution of geochemical data (Zuo et al., 2019). Traditional methods have limitations in processing the complex distribution of multivariate geochemical data (Xiong and Zuo, 2016; Parsa et al., 2017). Deep learning (DL) algorithms, which have been proven to be powerful tools for mining complex and nonlinear geospatial data and extracting unknown patterns related to geological processes, have been introduced in the field of exploration geochemistry (Luo et al., 2020; Chen et al., 2019; Zhang et al., 2021). Because the mineralization process is singular, mineral deposits are inherently rare (Cheng, 2007). One of the main limitations of using neural networks or other supervised machine learning algorithms is the scarcity of known mineral deposits (Li et al., 2021; Parsa, 2021). In general, the number of ore-bearing units (positive samples) is much smaller than that of ore-free units (negative samples), thus resulting in a data imbalance problem (Xiong and Zuo, 2018). Skewed class distributions will greatly underestimate the prediction performance for minority classes (e.g., ore-bearing units) and provide an inaccurate evaluation of the classification performance (Longadge and Dongre, 2013).
A generative adversarial network (GAN), proposed by Goodfellow et al. (2014), has become a leading method for addressing unsupervised and semi-supervised problems (Akcay et al., 2018). The basic concept of a GAN is derived from the zero-sum game in game theory, which enables discriminator D and generator G to improve the performance of the model during a mutual game (Goodfellow et al., 2014). GANs have been widely applied in the fields of image generation (e.g., Denton et al., 2015; Larsen et al., 2016), image recognition (e.g., Ledig et al., 2017), style transfer (e.g., Taigman et al., 2016; Isola et al., 2017), and anomaly detection (e.g., Schlegl et al., 2017; Zenati et al., 2018). To improve the performance of a GAN, a variety of improved algorithms, such as a conditional GAN (Mirza and Osindero, 2014), deep convolutional GAN (DCGAN) (Radford et al., 2015), and Wasserstein GAN (WGAN) (Arjovsky et al., 2017) have been proposed. In the field of geosciences, the application of a GAN is mainly focused on seismic data interpolation (Oliveira et al., 2018) and remote sensing classification (Merkle et al., 2018).
The deep autoencoder and its variants, which are reconstruction-based approaches, have been widely adopted to investigate geochemical anomalies (Xiong and Zuo, 2016; Luo et al., 2020). Recently, the idea of blending the autoencoder architecture with the adversarial learning have demonstrated promise in anomaly detection problems, thus have received extensive attention. Beggel et al. (2019) adopt adversarial autoencoder architecture for anomaly detection by imposing a prior distribution on the latent representation, and then typically placing anomalies into low likelihood regions. Schlegl et al. (2017) employed the standard GAN model to train normal samples under the assumption that the latent vector of the GAN represents the true data distribution for anomaly detection. Zenati et al. (2018) proposed an Efficient-GAN-Anomaly for anomaly detection by jointly mapping from an original data space to a latent space.
Different from the previous reconstructed-based model for anomaly detection, which calculate the reconstruction error based on the original data and their reconstructions, GANomaly model, proposed by Akcay et al. (2018), calculate the reconstruction error based on latent feature vector z and reconstructed latent feature vector z' by adding an additional encoder structure. During the training phase, the model is aimed to learn the distribution of a normal sample to minimize the difference between the two latent feature vectors of the sample. Thus, during the testing phase, a sample with a large error between the two latent feature vectors is regarded as an abnormality (Akcay et al., 2018). The main aim of this study is to explore the efficient of GANomaly model to deal with multivariate geochemical exploration data and delineate geochemical anomalies associated with mineralization in southern Jiangxi and its adjacent areas in China.
Section snippets
GANomaly network
The GAN network consists of two parts: the generator (G) and discriminator (D) (Goodfellow et al., 2014) (Fig. 1). The input of generator is random noise that follows a certain distribution. The generator learns the distribution of real data from the latent space and generates new examples. The discriminator is a classic classification system structure that attempts to classify examples as either real (from the domain) or fake (generated). The two models are trained together in a zero-sum game;
Study area and data
Jiangxi and its surrounding areas (Fig. 3) are world-class tungsten ore-concentrating areas with high mineralization intensity and high deposit density. Large-scale tungsten mineralization occurred in this area during the middle of the Yanshanian Movement. Among these areas, the southern Jiangxi region is the area with the most abundant tungsten mineral resources and has the longest mining history of wolframite in China. Geochemical concentration distributions show that large amounts of
Results and discussion
The original five geochemical data including W, Sn, Mo, Bi, and Ag were interpolated by the inverse distance weighting with a 1 km × 1 km cell size, resulting in five 336 × 336 raster maps. All five raster maps were combined into a set of input feature vectors at each cell location in the set of grids. 30% of the whole data, randomly selected from the five raster maps, were adopted as the training data. The remaining data in this study area were adopted as the testing data for the recognition
Conclusions
In this study, a method for extracting multivariate geochemical anomalies that require only non-mineralized samples during the training phase can avoid the overfitting problem caused by the imbalance between positive and negative training samples. By comparing with the deep autoencoder, the GANomaly network combining a GAN, an AE, and a CNN has the advantages of learning the spatial characteristics of the data and reducing the influence of noisy data on the extraction of abnormal information.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
We thank Dr. Mohammad Parsa and an anonymous reviewer's comments and suggestions which help us improve this study. This study was supported by the National Natural Science Foundation of China (No. 41772344).
References (46)
- et al.
A Spatially Constrained Multi-Autoencoder approach for multivariate geochemical anomaly recognition
Comput. Geosci.
(2019) Mapping singularities with stream sediment geochemical data for prediction of undiscovered mineral deposits in Gejiu, Yunnan Province, China
Ore Geol. Rev.
(2007)An introduction to ROC analysis
Pattern Recogn. Lett.
(2006)- et al.
Recognition of geochemical anomalies using a deep variational autoencoder network
Appl. Geochem.
(2020) A data augmentation approach to XGboost-based mineral potential mapping: an example of carbonate-hosted ZnPb mineral systems of Western Iran
J. Geochem. Explor.
(2021)- et al.
Recognition of significant multi-element geochemical signatures of porphyry Cu deposits in Noghdouz area, NW Iran
J. Geochem. Explor.
(2016) - et al.
Multifractal analysis of stream sediment geochemical data: implications for hydrothermal nickel prospection in an arid terrain, eastern Iran
J. Geochem. Explor.
(2017) - et al.
Mass accumulation and distribution of metallogenic elements in the South China block
Geochimica
(2013) - et al.
Recognition of geochemical anomalies using a deep autoencoder network
Comput. Geosci.
(2016) - et al.
GIS-based rare events logistic regression for mineral prospectivity mapping
Comput. Geosci.
(2018)
Detection of the multivariate geochemical anomalies associated with mineralization using a deep convolutional neural network and a pixel-pair feature method
Appl. Geochem.
Deep learning and its application in geochemical mapping
Earth Sci. Rev.
Measuring the performance of mineral-potential maps
Nat. Resour. Res.
Ganomaly: semi-supervised anomaly detection via adversarial training
Asian Conference on Computer Vision
Wasserstein gan
Robust anomaly detection in images using adversarial autoencoders
The Study of non-linear analysis method of Geochemical ore-forming anomaly
Prog. Geophys.
Anomaly detection of mechanical systems based on generative adversarial network and auto-encoder
Chin. J. Sci. Instrum.
Deep generative image models using a laplacian pyramid of adversarial networks
Adv. Neural Inf. Process. Syst.
A discussion on the chronology of tungsten mineralization and the time difference of diagenesis and mineralization in southern Jiangxi
Miner. Deposits
Generative adversarial nets
Adv. Neural Inf. Process. Syst.
Delineating anomalies using similarity coefficients based on element assemblage characteristics: an example of the Nanling area
Geology and Exploration
Image-to-image translation with conditional adversarial networks
Proceedings of the IEEE conference on computer vision and pattern recognition
Cited by (45)
Vertical-horizontal latent space with iterative memory review network for multi-class anomaly detection
2024, Knowledge-Based SystemsMineral prediction based on prototype learning
2024, Computers and GeosciencesAuto encoder generative adversarial networks - based mineral prospectivity mapping in Lhasa area, Tibet
2023, Journal of Geochemical ExplorationSpatial extrapolation of downscaled geochemical data using conditional GAN
2023, Computers and GeosciencesAn interpretable attention branch convolutional neural network for identifying geochemical anomalies related to mineralization
2023, Journal of Geochemical Exploration