Mapping Mineral Prospectivity For Tin-Copper Polymetallic Mineralization By Machine Learning Methods In Gejiu,Yunnan Province,China

After decades of exploration and development,most of the mines in the Gejiu region of Yunnan,a large Tin-Copper Polymetallic mining area,are facing the problem of insufficient reserves,which has led to serious constraints on the economic development of Gejiu.Therefore,the urgent problem is how to discover new valuable Tin-Copper Polymetallic deposits in the Gejiu area.Therefore,this paper establishes a Tin-Copper Polymetallic prospecting model for the Gejiu region by studying the geological background of the region,the genesis characteristics of typical deposits and the mineralisation pattern,and carries out a quantitative study of the geochemical,geophysical and geological data and other multi-source information,Supported the ArcGIS and GeoDAS platform,through the spatial analysis technology,Nonlinear analysis models,such as Concentration Area fractal model(C-A),Energy Spectrum Density Area fractal model(S-A)and Supervised and Semi-Supervised learning for the mining of multi-source mineralization information and quantitative research on the prediction of mineralization.The main research components and result are as follows:(1)In the aspect of mining multi-source ore finding information,the shadow map and geomorphic rendering map generated by DEM data from Gejiu area are superimposed with the existing fault structure data,and the supplementary interpretation of fault structure is carried out.The geochemical data of 1:200000 in Gejiu Mining Area are different from traditional expert knowledge.Ag,As,Bi,Cu,Sn,Zn,Au and Pb are selected as mineralization indicator elements through Receiver Operating Characteristic Curve,which make the result more objective,and extracts the combination information of mineralized elements by using robust principal component instead of classical principal component,The mineralization favorable degree fitting curve and scatter plot of metallogenic probability show that the effect of robust principal component is better than that of classical principal component.On this basis,the background and anomaly are separated by S-A analysis.The gravity data of 1:200000 in the study area are extracted by spatial extension and matching filtering,and the results are combined with geochemical principal elements to predict granite bodies in the mining area by Random Forest method.The results show that the granite body exposed in the mining area has a strong spatial consistency with the inferred rock mass,and there is also abnormal indication in the area without outcropping,It indicates that there may be concealed rock mass in these areas,which provides more accurate prediction elements for the future metallogenic prediction.(2)In the aspect of mineral quantitative prediction,in order to solve the problem that the number of negative samples is small in the process of supervised learning or whether the negative sample labels are credible in the way of negative sample selection,the Positive and Unlabeled learning(PU)method is introduced.The results show that the Pu method not only ensures the accuracy,but also reduces the prediction area and unnecessary exploration cost,In addition,based on the quantitative analysis of ROC curve and mineralization favorable degree fitting curve,Pu method is obviously superior to supervised learning Support Vector Machine and Random Forest.On this basis,four A-level prospective areas,two B-level prospective areas and four C-level prospective areas are delineated.