| Investigating the ocean environment where seafloor massive sulfide(SMS)deposits are located is not only technically intricate but also costly.Therefore,improving prospecting efficiency by reducing the exploration search space through mineral prospectivity mapping is desirable.Although some traditional predictive models for mineral prospectivity are suitable for a small number of prospect locations,they cannot fully utilize the mineralization information contained in predictor variables.In this paper,we propose a machine learning algorithm called the fuzzy forest model based on the advantages of the random forest and isolation forest models.The proposed model is not only suitable for a small number of prospect locations,but it can also simultaneously utilize binary,multiclass categorical,and continuous data as predictor variables to reflect the mineralization information effectively.The fuzzy forest model is obtained by combining the predicted results of the random forest model with those of the isolation forest model using fuzzy operators in a fuzzy logic method.We implemented the fuzzy forest model with the predictor variables of topography,geology,and hydrothermal plumes to map SMS prospectivity of the 48.7°–50.5°E segments on the Southwest Indian Ridge.Eight prospecting targets were delineated,the areas of which were smaller than the former predictions,meaning that the proposed method can further narrow the scope of exploration.Therefore,the fuzzy forest model is a potentially useful mineral prospectivity mapping method for small training locations.In order to further narrow the exploration scope,this study selects the most important area in the Dragon Horn area with the highest metallogenic probability and relatively abundant data in the prospective area delineated by the 48.7°-50.5°E ridge in the southwest Indian ridge to continue mineral prospectivity mapping.Compared with the data of the 48.7°E-50.5°E ridge in the Southwest Indian Ocean,the Dragon Horn area has fewer known hydrothermal fields and there are missing values in several feature variable layers.It is worth exploring how to apply layers with missing data to metallogenic prediction of SMS.Currently,there are a variety of mineral prospectivity mapping methods that can deal with missing data,such as the weight of evidence method,the comprehensive weight of evidence method.The above methods have achieved good results in the prediction of terrestrial mineral resources,but the lack of seabed data is more complicated than that of the land.The specific manifestations are that the area of missing data is large,and the location of missing data in different layers is located.Inconsistent,there is a problem of missing data in multiple layers at the same time,and due to the low degree of oceanic investigation,there are few known mining areas,which further increases the difficulty of mineral prediction.Therefore,according to the characteristics of ocean data,we need to explore a set of solutions that can not only deal with complex missing data,but also make metallogenic prediction in the case of few known mining areas.According to the data characteristics of Dragon Horn area,this study proposes a forecasting scheme that combines MICE multiple imputation method and isolation forest model.The prediction results of the isolation forest or fuzzy logic model are compared.The insights gained from this study are as follows:(1)According to the genesis of SMS deposits on the seafloor of ultra-slow spreading mid-ocean ridge,collect data on hydrothermal fields,water depths,geology and hydrothermal plume anomalies.Based on the information directly or indirectly related to SMS mineralization,the favorable conditions for prospecting are summarized and a quantitative prediction model for mineral prospectivity mapping of SMS mineralization of the 48.7°E-50.5°E on the Southwest Indian ridge is established.(2)The fuzzy forest mineral prospectivity mapping method proposed in this study is a machine learning model that integrates the prediction results of random forest and isolation forest by using fuzzy operators in the principle of fuzzy logic.By comparing the quantitative prediction results of SMS mineralization by the fuzzy forest model in the 48.7°E-50.5°E ridge in the southwest Indian ridge with the results based on the weight of evidence method,it is found that the fuzzy forest model is not only suitable for small samples,but also allows,according to the different indicators of metallogenic indication,binary classification,multi-classification and continuous data are used as characteristic variables to reflect the metallogenic information to the greatest extent.(3)The importance ranking of feature variables combined with the weight of evidence and fuzzy forest models shows that the top 10 important feature variables based on the fuzzy forest model and the weight of evidence method are consistent,so large faults(Etype fractures,detachment faults),oceanic crust thickness and ridge axis are important prospecting indicators for SMS in the study area.(4)We will superimpose the prospective area delineated by the fuzzy forest and the prospective area delineated based on the right of evidence.Except for the marginal parts of the prospective areas C and E based on the fuzzy forest,other prospective areas are in the prospective area based on the right of evidence.and the overall area is significantly smaller than that of the prospective area based on the right of evidence,indicating that the fuzzy forest model is more accurate in delineating the prospective area than the method of the weight of evidence.In this study,eight prospective areas delineated by the fuzzy forest model were regarded as the final prospective areas for quantitative prediction of polymetallic sulfide mineralization in the 48.7°E-50.5°E on the Southwest Indian ridge.(5)In order to find an imputation method suitable for this study,we carried out missing value simulations on the complete observation dataset according to the missing mechanism,missing rate and missing pattern of the real dataset,using the currently most recognized k-nearest neighbors,miss Forest and The MICE multiple imputation method was used to impute the simulated missing data set,and the imputation performance of the above three methods on the missing data set of this study was compared.The results show that the MICE multiple imputation method performs significantly better than the other two methods,so the MICE multiple imputation method will be used to impute the missing values in the real data set.(6)Aiming at the characteristics of missing values in many characteristic variables in Dragon Horn area and the missing rate is close to 50%,this study proposes a prediction scheme combining MICE multiple imputation method and isolation forest model.On the real data set,comparing the prediction scheme using MICE multiple imputation and the isolation forest model(scheme 1)with the prediction directly using the fuzzy logic model(scheme 2),it is found that scheme 1 has achieved better results.Comparing the prediction performance of the isolation forest model on the MICE multiple imputation data set and only the real complete data set,it is found that the evaluation indicators of the isolated forest model on the real complete data set are lower than the data set after MICE multiple imputation.The above evaluation results show that adding layers with missing values to the dataset can help to improve the accuracy of quantitative prediction of mineralization.(7)The Mineral prospectivity map drawn by the MICE multiple method and the isolation forest model is used as the final result in the Dragon Horn area.The five known hydrothermal fields in the Dragon Horn area are all included in the high-value area of metallogenic probability,and there is no known hydrothermal area in the highvalue area in the southwest Indian ridge,which can be used as follow-up key exploration targets. |
PDF Full Text Request