HR: 0800h
AN: IN21A-1168 INVITED     [Abstracts]
TI: Tectonic discrimination with classification trees
AU: * Vermeesch, P
EM: pvermees@pangea.stanford.edu
AF: Stanford University, Braun Hall, room 320-305 450 Serra Mall, Stanford, CA 94305 United States
AB: Traditionally, geochemical classification of basaltic rocks of unknown tectonic affinity has been performed by discrimination diagrams. Although easy to use, this method is fairly inaccurate because it only uses bi- or trivariate data. Furthermore, many popular discrimination diagrams are statistically not very rigorous because the decision boundaries are drawn by eye, and they ignore closure, thus violating the rules of compositional data analysis. Classification trees approximate the data space by a stepwise constant function, and are a more rigorous and potentially more effective way to determine tectonic affinity. Trees allow the simultaneous use of an unlimited number of geochemical features, while still permitting visualization by an easy-to-use, two-dimensional graph. Two classification trees are presented for the discrimination of basalts of mid-ocean ridge, ocean island and island arc affinities. The first tree uses 51 major, minor and trace elements and isotopic ratios and should be used for the classification of fresh basalt samples. A second tree only uses high field strength element analyses and isotopic ratios, and can also be used for basalts that have undergone alteration. The probability of successful classification is 89% for the first and 84% for the second tree, as determined by ten-fold cross-validation. Even though these trees use many geochemical features, it is not a problem if some are missing in the unknown sample. Classification trees solve this problem with surrogate variables, which give more or less the same decision as the primary variables.
DE: 1000 GEOCHEMISTRY
DE: 1021 Composition of the oceanic crust
DE: 1065 Major and trace element geochemistry
DE: 1094 Instruments and techniques
DE: 8140 Ophiolites (3042)
SC: Earth and Space Science Informatics [IN]
MN: Fall Meeting 2005