مدل‌ سازی پتانسیل کانی سازی مس و طلای پورفیری با به‌ کارگیری روش یادگیری نیمه‌ نظارتی در پهنه اکتشافی دهسلم، شرق ایران

شناسایی نواحی امیدبخش معدنی در اکتشافات ناحیه ای برای برنامه ریزی عملیات اکتشاف تفصیلی با به‌کارگیری و تحلیل داده های اکتشافی موجود در قالب مدل‌سازی پتانسیل معدنی توسعه‌یافته است. در این پژوهش، برای مدل‌سازی پتانسیل مس و طلای پورفیری در پهنه اکتشافی دهسلم واقع‌ در جنوب بلوک لوت، شرق ایران، از روش یادگیری ماشین بردار پشتیبان نیمه‌نظارتی استفاده‌شده است. روش های یادگیری نیمه‌نظارتی در مرحله یادگیری، از داده های برچسب دار و بدون برچسب اکتشافی در الگوریتم محاسباتی خود بهره می برند. در این مقاله، با به‌کارگیری الگوریتم ماشین بردار پشتیبان نیمه‌‌نظارتی بر روی داده های اکتشافی منطقه دهسلم شامل داده های زمین شناسی (سنگ‌شناسی و ساختاری)، ژئوشیمی رسوبات آبراهه ای، تصاویر ماهواره ای و مغناطیس هوابرد، مناطق هدف اکتشافی مس و طلای پورفیری شناسایی شد. در ادامه، نتیجه به‌کارگیری این مدل با خروجی روش ماشین بردار پشتیبان در حالت نظارت‌شده مقایسه و ارزیابی عملکرد مدل های تولیدشده با استفاده از نمودارهای منحنی مشخصه عملکرد سیستم و میزان تغییرات پیش بینی مساحت بهبودیافته، بررسی شد. بر این اساس، مدل پتانسیل نیمه نظارتی عملکرد بهتری را در شناسایی اهداف اکتشافی مس و طلای پورفیری داشته است. نواحی اهداف پتانسیل شناسایی‌شده در مدل نیمه نظارتی، تمامی اندیس های معدنی شناخته‌شده در منطقه مورد بررسی را در 9.2 درصد از مساحت ناحیه مورد بررسی، به درستی پیش بینی کرده است. اهداف اکتشافی معرفی‌شده، اغلب هم‌راستا با روند گسل های اصلی منطقه، در راستای شمال‌غربی جنوب‌شرقی و مرتبط با واحدهای ولکانیک نظیر ریولیت، آندزیت، داسیت و ریوداسیت هستند. نتیجه حاصل از این پژوهش نشان دهنده برتری روش یادگیری نیمه نظارتی در شناسایی نواحی هدف معدنی برای برنامه ریزی عملیات تفصیلی اکتشافی است.

Porphyry Cu-Au prospectivity modelling using semi-supervised learning algorithm in Dehsalm district, eastern Iran

Introduction The identification of potentially mineralized areas has progressed with the use and interpretation of all available exploratory data in the form of mineral potential modeling (MPM) (Yousefi and Nykänen, 2017). Recently, machine learning methods have been a popular research topic in the field of MPM ((Chen and Wu, 2016). Machine learning algorithms that have been used in MPM generally fall into the categories of being supervised or unsupervised. Supervised models, use the location of the known mineral occurrences as training sites (or labeled data). Therefore, these models suffer stochastic bias and error (Zuo and Carranza, 2011). Unsupervised models classify mineral prospectivity of every location based solely on feature statistics of individual evidential data layers ((Abedi et al., 2012). The semisupervised learning models are a hybrid of supervised and unsupervised learning models that use both labeled and unlabeled data to extract the hidden structure of the data, as well as the relation between the input exploration layers and the output labeled data (Fatehi and Asadi, 2017). The Dehsalm study area forms a part of the Lut metallogenic block of eastern Iran, which is characterized by the subduction zone setting and extensive magmatism (Beydokhti et al., 2015). The objective of this research is to present a prospectivity model to delineate exploration target areas for porphyry CuAu mineralization in the study area. For generating a prospectivity model, we used TSVM algorithm, a semisupervised learning integration technique, to identify the anomalous areas related to the porphyry CuAu mineralization. The input layers are selected based on a conceptual model for porphyry CuAu mineral system. The performance of the mineral prospectivity maps (MPMs) is evaluated using the various techniques, including the receiver operating characteristic (ROC) curve, an area under curve (AUC) metric.   Materials and methods To apply a processbased understanding of porphyry coppergold deposit system on the mapping of prospectivity, a conceptual model must be first developed (Fatehi and Asadi, 2017). Such a model should depict critical scaledependent processes involved in the mineral deposit formation, and a mineral system approach can be followed to aid understanding where, when and why mineral deposits form (Parsa et al., 2016). The spatial data sets used to model porphyry CuAu prospectivity of the study area include geological, remote sensing, geophysical, and geochemical data. In this study, semisupervised support vector machine (TSVM) prospectivity technique is utilized to model porphyry CuAu target areas. The TSVM is an extension of SVM that uses the unlabeled data to improve the performances of the classifier. The aim of TSVM algorithm is to find the decision hyperplane subject to maximize the margin distance in labeled and unlabeled data.   Result In the present study, TSVM and SVM models were applied to CuAu prospectivity modeling in the study area. The models were trained based on the location of known CuAu mineralization occurrences and nondeposit location using e1071 package in R opensource statistical software (70% of the labeled data were used in training and 30% in the testing phase of learning in both algorithms). The RBF kernel function were used and the optimal values of the kernel parameters were assessed using a 10fold crossvalidation procedure and the best learning performance was selected by correct classification rate. The output of the models highlighted the target areas for porphyry CuAu mineralization in the study area. The receiver operating characteristics (ROC) analysis shows that both models perform well, however, the TSVM model yields the best performance.   Discussion To identify exploratory target areas on a regional scale, various supervised and unsupervised approaches have been developed in mineral potential modeling. Supervised methods such as SVM use labeled data to classify exploratory datasets. In this research, a new semi supervised learning method, TSVM, was applied to model the mineral potential for porphyry CuAu mineralization in the Dehsalm exploration zone. The introduced target areas by TSVM method, within the known mineral indices, covered smaller areas than targets identified by SVM model, so planning the detailed exploration phase will be optimal. The result of this research demonstrates the superiority of the semisupervised learning method in identifying the target areas for planning the exploratory operations.   Acknowledgements We would like to thank the Geological Survey of Iran for providing the exploration data used in this research. The financial support of the South Khorasan Industry, Mine Trade Organization is gratefully thanked.   References Abedi, M., Norouzi, G.H. and Bahroudi, A., 2012. Support vector machine for multiclassification of mineral prospectivity areas. Computers Geosciences, 46: 272–283.  https://doi.org/10.1016/j.cageo.2011.12.014 Beydokhti, R.M., Karimpour, M.H., Mazaheri, S.A., Santos, J.F. and Klötzli, U., 2015. UPb zircon geochronology, SrNd geochemistry, petrogenesis and tectonic setting of Mahoor granitoid rocks (Lut Block, Eastern Iran). Journal of Asian Earth Sciences, 111: 192–205. https://doi.org/10.1016/j.jseaes.2015.07.028 Chen, Y. and Wu, W., 2016. A prospecting costbenefit strategy for mineral potential mapping based on ROC curve analysis. Ore Geology Reviews, 74: 26–38. https://doi.org/10.1016/j.oregeorev.2015.11.011 Fatehi, M. and Asadi, H.H., 2017. Data integration modeling applied to drill hole planning through semisupervised learning: A case study from the Dalli CuAu porphyry deposit in the central Iran. Journal of African Earth Sciences, 128: 147–160. https://doi.org/10.1016/j.jafrearsci.2016.09.007 Parsa, M., Maghsoudi, A., Yousefi, M. and Sadeghi, M., 2016. Prospectivity modeling of porphyryCu deposits by identification and integration of efficient monoelemental geochemical signatures. Journal of African Earth Sciences, 114: 228–241. https://doi.org/10.1016/j.jafrearsci.2015.12.007 Yousefi, M. and Nykänen, V., 2017. Introduction to the special issue: GISbased mineral potential targeting. Journal of African Earth Sciences, 128: 1–4. https://doi.org/10.1016/j.jafrearsci.2017.02.023 Zuo, R. and Carranza, E.J.M., 2011. Support vector machine: a tool for mapping mineral prospectivity. Computers Geosciences, 37(12): 1967–1975. https://doi.org/10.1016/j.cageo.2010.09.014