Mineral Prospectivity Mapping Of Tungsten Deposits In Southern Jiangxi Province Based On Weights Of Evidence And Machine Learning

Quartz vein-type wolframite is a dominant resource with regional characteristics in Southern of Jiangxi Province.However,tungsten resources in Southern of Jiangxi Province have been gradually depleted in recent years after long-term high-intensity mining.A new round of mineral exploration and increase of tungsten resources have become a major demand to guarantee regional resource security and protect regional economic development.Mineral prospectivity mapping is an important precursor of resources exploration.Space exploration information technology in the development and exploration data acquisition methods increasingly diverse background and under the new normal,through the analysis of the spatial data driven and prediction algorithm,the spatial distribution of targets can be more objectively and accurately reflected,thus provide a more accurate prediction results.It is an important developing trend of mineral prospectivity mapping.The quantitative prediction work in this paper can provide portable method system and quantitative model support for the metallogenic prediction work in the future.In this paper,through a series of spatial techniques such as fractal analysis,Fry analysis and spatial distribution analysis,the spatial distribution characteristics and the controlling factors of tungsten deposits in the study area were analyzed and expressed quantitatively.Based on the result of spatial distribution analysis and mineral system analysis,eight tungsten minerality-related elements were selected,which integrated geological data,geophysical and geochemical information.Using Arc GIS for data processing and integration,then a multi-source metallogenic information system was established.The buffer analysis method was used to obtain the optimal buffer distance of the weights of evidence analysis,and the modelling of mineral prospectivity based on the weights of evidence was established.Through the grid search method and 10 fold cross-validation method,the modelling of mineral prospectivity in southern Jiangxi Province Based on machine learning is trained.Optimizing the best model for each type of machine learning methods,using confusion matrix,the receiver operating characteristic curve(ROC curve)to verify this model,with a success rate curve to evaluate model predicted results.Finally,the metallogenic potential area of tungsten deposits in Southern of Jiangxi Province was divided by the success rate curve of the results of modelling of mineral prospectivity based on weights of evidence and machine learning method,and the prospectivity map of the study area was obtained,and the geological interpretation of the prediction results was carried out.Based on the above work,the main achievements of this paper are as follows:(1)The fractal analysis results show that the spatial distribution of tungsten deposits in the study area has two self-similar zones: 0 < r ≤ 3723 m and r > 3723 m,revealing the scope of ore-forming mechanism at different scales of region and local(mining)area.Fry analysis further resolve the dominant direction of ore point distribution at different scales.At the regional scale(r > 3723 m),ore points are restricted by the EW direction of ore control,while at the mining area scale(0 < r ≤3723 m),the NE direction of ore control plays a dominant role.The results of distance distribution analysis indicate the correlation degree between different orientation faults and the spatial distribution of tungsten deposits in the study area.EW-and NE-trending faults are positively correlated with the location of ore points,which has a significant restriction effect on the distribution of ore points,while NW-trending faults are almost unrelated to the locations of ore points,which should be post-mineralization faults.The results of quantitative analyses of three spatial points confirm each other,revealing the multi-scale self-similarity of tungsten deposits and the differential ore-controlling effect of regional multi-stage faults.(2)According to the results of spatial analysis and the analysis of metallogenic system in the study area,selected the reflect the tungsten mineralization source area,migration,precipitation funnel,key process of eight information layer: Yanshanian granites,regional fracture,fracture intersection,abnormal magnetic anomaly,gravity anomaly,tungsten,manganese and iron anomaly metallogenic information system in the study area is established.(3)The weight analysis of evidence quantitatively reveals the correlation degree between oreforming factors and ore points in the study area.Tungsten anomaly has the highest correlation value C,followed by manganese anomaly and iron anomaly,indicating the important significance of geochemical anomaly in the study area.Yanshanian granites,regional faults and intersections of regional faults have large C values,showing close correlation with ore points,while gravity anomalies and magnetic anomalies have small C values and weak correlation.Using the buffer analysis to determine the optimal buffer distance of each ore-forming factors,and combined with weights of evidence,the mineral prospectivity map was drawn,and seven high potential metallogenic target areas were delineated.(4)The training process of machine learning model shows that the performance of random forest and convolutional neural network model is more stable than that of support vector machine and artificial neural network model in dealing with parameter changes.The results of model verification show that convolutional neural network has higher sensitivity(87.62%),specificity(97.14%),positive sample prediction rate(96.94%),negative sample prediction rate(88.72%),classification accuracy(92.38%)and Kappa coefficient(0.856)than other machine learning methods.Therefore,it has the best classification accuracy,followed by random forest,support vector machine and artificial neural network.The AUC curve indicated that random forest had the highest prediction accuracy(AUC = 0.95509),followed by support vector machine(AUC = 0.93922),convolutional neural network(AUC = 0.93337)and artificial neural network(AUC = 0.92232).The high potential area prediction efficiency of four machine learning models was evaluated by success rate curve.The prediction efficiency of random forest(prediction curve slope is 6.7797),followed by support vector machine model(prediction curve slope is 6.4241),convolutional neural network(prediction curve slope is 5.3232)and artificial neural network(prediction curve slope is 5.1835)are relatively low,indicating that random forest model has the highest prediction efficiency.Combining the results of model training,validation and evaluation,convolutional neural network has the best classification accuracy,while random forest has the higher overall prediction accuracy and the prediction efficiency of high potential area.At last,random forest is selected as the most suitable model for mineral prospectivity mapping in Southern of Jiangxi Province among the four machine learning models.(5)The evaluation results of random forest and weights of evidence model by success rate curve show that the random forest model and weights of evidence model successfully predict 66.95% and64.41% of known tungsten deposits in the research area with 9% and 8% target area respectively.The slope of success rate curve of high potential area of random forest model is 6.7797.The slope of the success rate curve of the high-potential area of the metallogenic prediction model based on the weights of evidence method is 6.8411.The classification accuracy,prediction accuracy and prediction efficiency of the model were integrated.Finally,the prediction results of random forest model and weights of evidence model are integrated to obtain the mineral prospectivity map of the study area.(6)The results of geological interpretation show that wolframe-related geochemical anomalies have the largest contribution weight to the prediction results,and are the most effective prospecting indicators in the area;Yanshanian granites and NE-and EW-trending faults have important influence on the prediction results,which reflects the significant ore-controlling effect of magmatic and tectonic elements.The significance of geophysical anomaly in this study is relatively weak.It is particularly noteworthy that manganese anomaly contributes to the prediction results only second to tungsten anomaly in all prediction models.This important prediction contribution of such a factor that has been neglected for a long time in previous exploration may reveal that ore-forming materials in surrounding rocks play an important role in the formation of wolframite.This conclusion can provide important reference for studying the genesis mechanism of tungsten veins and future tungsten resources exploration.