| Mineral resources are an indispensable and important resource for national development.With the progress of the investigation of mineral resources,a large amount of mineral resources information data has been accumulated.It is increasingly important to use the geophysics,geochemistry,and remote sensing information to comprehensively forecast minerals.Machine learning is more and more widely used in the field of comprehensive mineral prediction.In this study,the point pattern analysis was used to determine the location of 32 non-mineral sites,and a total of 64 sample data sets were merged with the discovered ore and mineralization sites.Using the evidence weight model,the metallurgical prediction grid was evaluated with two states.The importance of the 20 ore-controlling factors in the study area was assessed by the random forest algorithm.Ten important factors were selected as model training characteristics according to the evaluation results.Two ensemble strategies,weighted majority voting and Subensemble ensemble,were adopted to integrate the random forest model,support vector machine model,and logistic regression model,and they were applied to the Eerguna region in the northern part of the Daxinganling.Through the ROC curve area(AUC),AUC standard deviation(SAUC),ROC curve area and random area curve area 0.5 statistical difference(ZAUC),the most approximate index(MYI),the most approximate index threshold value of polymetallic minerals forecast Percentage of Existence Area(PSAO)and Polymetallic mineral predicted area including percentage of polymetallic deposits and mineralization points(PDDC)found in the study area 7 evaluation indicators Two ensemble metallogenic prediction models and the other three models were evaluated.Subensemble ensemble metallogenic model superior performance the weighted majority vote integrates metallogenic prediction models.Finally,the two-level metallogenic prediction prospects are delineated based on the most youden index,and the similarities and differences between two ensemble metallogenic prediction models are compared and analyzed.The metallurgical prediction python tool set is developed based on arcpy and scikit learn.The main results and innovations are as follows:(1)Identifying and training an equal number of non-mineralized points based on point pattern analysis and constructing small sample data sets for mineral forecasting are more reasonable and effective in solving the problem of data imbalance.(2)Firstly,based on the weight of evidence method,two-state assignment of grid prediction grids is carried out,and then the importance of ore-controlling factors is evaluated based on the random forest algorithm.The two combined to play their respective advantages to better achieve the optimization of ore-controlling factors.(2)For the first time,the weighted majority voting integration model and Subensemble integration model were applied to mineralization prediction.Both of them showed excellent performance,indicating that the two ensemble learning models are suitable for small sample training sets,and Subensemble ense is superior to weighted majority voting integration.The performance of the ensemble learning model is better than that of a single classifier.The Subensemble ensemble model fully mines the training data set information and passes it to the meta-classifier,so the model has strong generalization ability.(3)Describes the general process of model prediction for the Python machine learning library scikit-learn,introduces the scikit-learn API and integrates the ArcGIS Python toolset with Arcpy.Scikit-learn highly abstracts and refactors the machine learning process,providing a unified API for machine learning model development,so understanding the scikit-learn architecture is important for developing new machine learning evaluators and learning scikit-learn. |
