Spatial Distribution Pattern Analysis Of Regional Geochemical Elements And Prediction Of Quantitative Metallogenic Prospect

The process of data mining and integration of geological,geochemical and other multi-source geoscience information is an important part of metallogenic prospect prediction,which plays a very important role in mineral exploration and has a direct impact on the effect of metallogenic prediction.Therefore,it is an important research direction in the field of metallogenic prediction that how to mine the deep mineralization information from the massive multi-source information and effectively integrate it.In recent years,metallogenic prediction modeling based on machine learning algorithms have attracted more and more attention.Because of its powerful capability of feature extraction and information integration,it has become the research hotspots and frontiers.Although researchers have proposed a series of machine learning algorithms,however,the difficulty of algorithm design is increasing because of the characteristics of geoscientific data such as multi-source,heterogeneous,high computational complexity and high uncertainty.Therefore,there are still many challenges in the design and theoretical analysis of metallogenic prediction algorithms based on machine learning.In the presents study,the uncertainty of data rasterization,abnormal information extraction,machine learning algorithm model construction and other contents were studied by collecting the multi-source geoscientific data of Xiahe-Hezuo district.The uncertainty evaluation of geochemical element spatial distribution model and metallogenic prospect prediction modeling were carried out by using multi-point geostatistics direct sampling algorithm,local singularity analysis and machine learning algorithm.The main results of the study are as follows:(1)Based on the analysis of regional geological background and typical deposits,the ore-controlling structural characteristics,spatial distribution of ore bodies,wall rock alteration and prospecting criteria were clarified,and the main metallogenic / ore control factors were determined.(2)Based on the geochemical data of stream sediments in the study area,an improved direct sampling algorithm was used to model the uncertainty of the spatial distribution of geochemical elements,and the local singularity analysis was carried out on the modeling results.Then,the singularity–quantile(S-Q)method was used to calculate and determine the singularity index,and ?=1.974 was the optimal segmentation threshold,and then the probability map of gold element anomaly was obtained.The results show that most of the known gold deposits(points)are distributed near the high probability region,which indicates that the probability of anomalies can be used as an important basis for delineating geochemical anomalies.(3)The factor analysis method of composition data was used to determine the symbiotic association between geochemical elements,and the Sb?As?P?Th ?Au element combination closely related to gold mineralization was obtained.The direct sampling algorithm and local singularity analysis were carried out for each element,and ?=2 was taken as the segmentation threshold to obtain the geochemical anomaly probability map of each element.Then,the maximum entropy model was used for multivariate data fusion,and the area under curve(AUC),Kappa coefficient and true skill statistics(TSS)were used to evaluate the performance of the model.All the evaluation indexes show that the model can effectively establish the relationship between the probability of multiple geochemical anomalies and the known gold deposits(points),and can identify the favorable/unfavorable areas of gold mineralization.(4)The one-dimensional convolutional neural network model was applied to mapping mineral prospectivity(MPM).By combining various methods to determine the disadvantageous metallogenic areas,the negative samples were constrained,and SMOTE algorithm was used to expand the over-sampled data of the sample set to construct the training data set.The feasibility and effectiveness of the training sample construction strategy were evaluated by setting different sample groups.The results show that the training data set construction strategy proposed in this paper is feasible.The influence of the important super parameters(network depth,vector convolution kernel size,number of convolution kernels,and dropout rate)involved in the model on the classification performance was studied.The optimal parameter combination of the network model was determined by orthogonal experimental design and 5-fold cross validation method,and the prediction performance and efficiency of the model were evaluated by using ROC curve,fitting curve,AUC value and other indicators.(5)The attention mechanism was introduced into the convolutional neural network model.The attention module calculated the influence weights of different features by the way of feature weighting,increased the weight of important features and decreased the weight of unimportant features,so as to highlight important features and suppress irrelevant information.The stability of the prediction results was improved,and the exploration area of metallogenic prospect area was reduced.Combined with the regional geological background,metallogenic geological conditions and geochemical anomalies in the study area,six metallogenic prospective areas were delineated,which provided the basis for the deployment of mineral exploration work in the study area.