Multi-geoinformation Integration For Mineral Prospectivity Mapping In The Hezuo-meiwu District,Gansu Province

The study area,Hezuo-Meiwu district in Gansu Province is located in the Western Qinling Orogen and endowed with numerous mineral deposits,such as Zaozigou,Jiagantan etc,forming an important mineral resource base in China.At the age of big data,geo-bigdata and corresponding big data processing methods is one of researching hotspots currently.The thesis focuses on the extraction of mineralized anomaly and mineral prospectivity mapping from the systematically collected geoscientific dataset with emphasis on the interpretability of machine learning algorithms and advantage of deep learning,and finally the joint application of unsupervised convolutional autoencoder network and supervised convolutional neural network was proposed.This contribution not only promotes the application of deep learning algorithms for the synthesis of multi-source geoinformation to obtain mineralized anomaly,but also facilitate the mineral exploration in the field work.The main research contents and conclusions are as follows:(1)The introduction of interpretable machine learning algorithms(Random forest and Maxentropy model)for mineral prospectivity mapping has considered the problem of training small samples and the selection of negative training samples,which indicated the significance of the interpretability of machine learning algorithms for qualitative and quantitative analysis of predictive elements and variables.The research of relevant interpretable machine learning algorithms is of great significance for improving the reliability and reducing the uncertainty of mineral prospectivity mapping based on multisources geoscientific dataset.(2)The deep autoencoder network was applied for outlier detection in geochemical dataset together with traditional statistical analysis.The closure effect in such dataset makes the tradition statistical treatments designed for Euclidean space fail to detect outliers.The compositional outlier detection converts the data from simplex space in to Euclidean by logratio transformations.Here,taking the results of compositional outlier detection as benchmark,deep autoencoder network(DAE)and Density-based spatial clustering of applications with noise(DBSCAN)were applied for detecting outliers.The comparison of results proves the efficacy of DAE for coping with data in simplex space and shows the advantage of deep learning in dealing with geoscientific data.(3)Convolution neural network was introduced for mineral prospectivity mapping and a method for fabrication of training dataset based on unsupervised learning used for training deep learning algorithms has been proposed.The research on the relationship of outliers and known mineral deposits and occurrences indicated the feasibility of the proposed method.