Classification of rocks radionuclide data using machine learning techniques

The aim of this study is to assess the performance of linear discriminate analysis, support vector machines (SVMs) with linear and radial basis, classification and regression trees and random forest (RF) in the classification of radionuclide data obtained from three different types of rocks. Radionuclide data were obtained for metamorphic, sedimentary and igneous rocks using gamma spectroscopic method. A P-type high-purity germanium detector was used for the radiometric study. For analysis purpose, we have determined activity concentrations of ²³²Th, ²²⁶Ra and ⁴⁰K radionuclides, published elsewhere (Rafique et al. in Russ Geol Geophys 55:1073–1082, 2014), in different rock samples and built the classification model after pre-processing the data using three times tenfold cross-validation. Using this model, we have classified the new samples into known categories of sedimentary, igneous and metamorphic. The statistics depicts that RF and SVM with radial kernel outperform as compared to other classification methods in terms of error rate, area under the curve and with respect to other performance measures.