Application of classification trees for improving optical identification of common opaque minerals

Abstract

The recognition of opaque minerals by polarized reflected-light microscopy is a challenging task due to the use of certain qualitative properties that may lead to ambiguities in their identification. The use of a decision tree may simplify and guide the evaluation of these properties for reaching a proper mineral identification. Improvements of such classification trees can contribute to employ fewer properties if the depth of the tree is reduced, and achieve less uncertainty in case of reducing the number of minerals in the terminal nodes. This work describes a proposal for obtaining precise and compact classification trees and its application in the optical identification of minerals. The method builds a decision tree by using machine learning techniques, after grouping the minerals with the same properties. Its classification performance was evaluated in a comparison with different classifiers. Also, the complexity of the resulting tree was compared to a widely used tree diagram. The results show that this method can generate classification diagrams that employ few properties and with a reduced number of minerals in each final group, so decreasing the uncertainty of the identification. Furthermore, the inclusion of an additional property (reflectance) was evaluated, applying it to data of common opaque and rock-forming minerals. The resulting tree presents an improvement in the identification, and without a significant increase in the number of properties needed to identify each mineral group. A web application has also been developed to interactively embody the classification process that follows the decision tree obtained.