Tectonic discrimination and application based on convolution neural network and incomplete big data

摘要

Many classical discriminant diagrams have been employed by geologists on certain represented samples of different tectonic areas to distinguish the tectonic settings of igneous rocks in the last few decades. However, a high misjudgment rate was noted under the big data test and the classification results among the various diagrams are inconsistent. In order to avoid using a few indicators for judgment or typical sample data to discriminate, a new discriminant tool has emerged as a powerful platform to improve the accuracy and reliability of the classification of tectonic settings of igneous rocks using multi-element indicators and comprehensive big data. This paper explores the feasibility of using convolution neural network (CNN) in deep learning technology to synthesize multi-element indicators containing incomplete data. Thirty kinds of geochemical major and trace elements were extracted from GEOROC database, and the gray scale two-dimensional code of more than 1.09 million rock samples was constructed by using the incomplete data. The CNN model obtained in this study can effectively classify igneous rock samples in 11 tectonic settings and has good generalization ability. The model has a Top1 classification accuracy of 94.3% for the test dataset, which is substantially better than other algorithms such as decision trees, discriminant analysis, naive Bayes classifiers and support vector machines. An application based on this model is also provided, thus serving as a potentially important daily tool to assist the manual identification in tectonic discrimination.