Multivariate Geochemical Anomaly Recognition Using Spatial Constrained Autoencoders

Mineral resources are of great significance towards China's economic development as important pillars.At present,the contradiction between supply and demand of mineral resources is becoming increasingly acute.It is imperative to develop new technologies to achieve major breakthroughs in geological prospecting.Exploration geochemical methods play an important role in practical mineral exploration.As a result,geochemical data processing and geochemical anomaly identification related to mineralization have aroused much more public attention.In recent years,significant progress has been made in the studies of geochemical anomaly identification.The machine learning method is applied to simulate complex and unknown multivariate geochemical distribution models and extract meaningful features related to mineralization,because of its unique merits that independence on data distribution,no requirements on linear correlation between geological variables and predicted values,and excellent characterization ability for nonlinear relationships.However,there are still some problems remained to be solve in the application of machine learning in geochemical anomaly identification currently,which do hinder the development of machine learning in the field of geochemical exploration and limit its anomaly identification ability.The lack of consideration that spatial characteristics of geochemical data is a key issue in the current application of machine learning in the field of exploration geochemistry.If not solved,the problem may limit the data processing advantages of machine learning,prevent application of the method from deepening,and even affects the secondary exploitation and utilization of prospecting information.In the application of geochemical anomaly identification,the machine learning method is mainly classified as supervised anomaly identification using known mine points as label data and unsupervised anomaly identification with unlabeled data.Since the supervised learning methods consider less the internal structure of the data and misse the potential mineralization information,this study mainly discusses how to integrate the spatial information of geochemical data,in the unsupervised machine learning methods,to improve the ability of geochemical exploration of abnormal recognition of machine learning in the application of geochemical anomaly identification.This study starts from two aspects to study the unsupervised anomaly recognition machine learning methods that takes the spatial characteristics into account.One is to use the relationships between multiple elements as the reconstructed component to complete the background reconstruction,and the other is to use the spatial information of the element to do so.Both of them use the large probability sample as the component to reconstruct the background,thus separating the background and abnormal.The specific research works are as follows:(1)The main problem of the multivariate relationship reconstruction method is that the spatial heterogeneity of the geochemical background is neglected,resulting in the invalidated identification of valuable geochemical anomalies.This paper proposes a Spatial Constrained Multi-Autoencoder(SCMA)method for multivariate geochemical anomaly identification.This method conquer the puzzle of anomaly identification caused by spatial heterogeneity according to local conditions.Taking the chemical similarity and spatial continuity of geochemical samples into account,it divides the space of different backgrounds by means of multivariate clustering,spatial filtering and spatial fusion.In order to reduce the effects of random initialization of weights in the encoder neural network,multi-autoencoders is used to learn and reconstruct the geochemical background of each sub-domain.The anomaly score is then calculated as the difference between the observed geochemical features and the reconstructed features.(2)The spatial structural characteristics of geochemical background are often neglected.How to effectively extract the spatial structure information of elements is a key issue.This paper proposes a Multi-Convolutional-Autoencoders(MCAE)method to deal with multi-geochemical anomaly identification.The method includes three unique steps as follows:(1)Eliminating the correlation between geochemical elements and avoiding the effective background information being diluted by redundant data;(2)Using the global Moran's I to definite identification domain of the element background spatial structure to ensure that the model effectively extracts the element background spatial structure information;(3)Using Multi-Convolutional-Autoencoder to learn and reconstruct non-interactively the spatial structure of the background structure,avoiding the interference of multiple elements in the learning process.Finally,the anomaly are calculated based on the difference between the observed geochemical data and the reconstructed data.(3)The methods proposed in this paper are all applied to the iron-polymetallic ore-forming area in southwestern Fujian,China.The multivariate anomaly identification of the research area is based on Cu,Mn,Pb,Zn and Fe2O3 elements in the 1:200,000 sediment samples.The results show that both SCMA and MCAE models are superior to the existing methods in all aspects,by which those identified anomalies are highly correlated with the known Fe deposits in the area,and the accuracy of anomaly identification is effectively improved.Generally speaking,the contribution of this paper is to propose a flexible machine learning framework that takes spatial information into account,as well as two machine learning methods that considering spatial characteristics.Besides,it discusses the factors that contributes to spatial domain partitioning suitable for auto-encoder neural network to learn,and definite the the key element of identification domain of the geochemical element background.This research and its gain not only enrich the understanding of the application of machine learning methods in geochemical anomaly identification,but also provide new ideas and ways for evaluating anomaly identification application of other spatial data.