Deep Learning For Mineral Prospectivity Mapping Of Lala-type Copper Deposit In The Huili Region,Sichuan

The Huili region of Sichuan Province is located at the western end of the Huili-Dongchuan aulacogen,central section of the Sichuan-Yunnan rift system in the southwestern margin of the Yangtze Platform,which is one of the significant copper supply sources of China.Currently,how to make full use of massive multisource geospatial big data and deep learning algorithms to dig intrinsic and more representative information aiming at improving the effect of prospecting prediction has been the foci in mineral prospectivity mapping.The author conducted research on the application of machine learning algorithms for the prediction of deposits-type sought based on collected multi-source geological data of the Huili region.The emphasis is very much on fabricating samples and introduction of deep convolutional neural network,and five prospecting areas were finally delineated.The research not only exerts great influence on the innovation of mineral prediction,and also has practical application value for the further exploration of Lala-type copper deposits in Huili region(1)On the basis of the geological,geochemical and geophysical thematic database in Huili region,5 variables are constructed for mineral prospectivity mapping including the strata of the Hekou Group,the proximity of mafic intrusion,the Cu content,the second principal component scores of geochemical data based on isometric logarithm ratio transformed,and the RTP aeromagnetic ?T data as independent evident variables.Then comparative analyses of favourability mappings obtained from Weight of Evidence method,Support Vector Machine,Random Forest and Back Propagation Neural Network with a single hidden layer was carried out(2)A set of systematic and standard sample data was constructed,which was the foundation for training neural network models.With planar projection of typical deposits from exploration and taking the gridding units of which as the center,1468 mineralization window samples were produced by sample augmentation and then combining with the same number of non-mineralization window samples obtained randomly to forming training datasets that are subsequently used for deep learning.The research showed that it is feasible to train an artificial neural network model using training samples derived from the orebodies of representative deposits and the trained model is discriminative for deposit type sought,which serves well when it comes to specific deposit prospecting.(3)With the concept of ensemble learning and Combination with Deep Convolutional Neural Network(CNN),the technology of "Random Samples Integrating CNN"(RSI-CNN)was innovatively proposed for mineralization targeting.Complete procedures including the evident variables preprocessing,fabrication and random combination of training samples(mineralized and non-mineralization),CNN modeling,training and prediction can be implemented on the MATLAB platform.Studies have shown that the RSI-CNN increased the diversity of training samples and hence improved the stability of the prediction results in the view of data and modeling perspective.(4)Base learners were assembled by using maximum and mean strategies for mineral potential mapping of Lala-type copper deposits.Combined with the metallogenic geological context,5 prospecting areas including Songzhiba,Luodang-Hongnipo,West Dachangpo,Lihong and Diaojingdong were delineated,which provides insights and evidence for the decision-making for further exploration of copper deposits in the study area.