A comprehensive classification of wood from thermogravimetric curves

Wood is one of the most complicated materials to be classified in different classes or species. In this paper, the thermogravimetric (TG) curves of 49 wood samples are used to classify them in 7 predetermined species. Different functional and multivariate statistical supervised classification methods are used for this task: a nonparametric Nadaraya-Watson kernel functional estimator (K-NFDA), using the complete TG curves, and multivariate supervised classification approaches, such as linear discriminant analysis (LDA), k Nearest Neighbors (k-NN), Naive Bayes (NBC), Neural Networks (NN), and Support Vector Machines (SVM). Before applying the multivariate techniques, the TG curves are discretized using principal component analysis (PCA) or fitting a four-component generalized logistic model. The results show that the classical method of LDA using the logistic parameters had the best performance, although high correct classification percentages were also obtained with the rest of the approaches. The work is completed with a comprehensive simulation study, comparing the classification techniques in different scenarios. Artificial TG curves are generated using the logistic model and additional conclusions on wood classification are established. Due to the heterogeneity of wood, this simulation study is very useful to describe worst-case scenarios and to assess more accurately the proposed classification methodologies. (C) 2012 Elsevier B.V. All rights reserved.