Leaf characteristics provide many useful clues for taxonomy. We used a back-propagation artificial neural network (BPANN) and C-support vector machines (C-SVMs) to classify 47 species from 3 sections of genus Camellia (16 from sect. Chrysanthae, 16 from sect. Tuberculata, and 15 from sect. Paracamellia). | Turkish Journal of Botany Turk J Bot (2013) 37: 1093-1103 © TÜBİTAK doi: Research Article Classification of Camellia species from 3 sections using leaf anatomical data with back-propagation neural networks and support vector machines 1,2 3 4 2 1, Wu JIANG , Billur BARSHAN ÖZAKTAŞ , Nitin MANTRI , Zhengming TAO , Hongfei LU * 1 College of Life Science, Zhejiang Sci-Tech University, Hangzhou, . China 2 Economic Crops Laboratory, Institute of Zhejiang Subtropical Crops, Wenzhou, . China 3 Department of Electrical and Electronics Engineering, Bilkent University, Bilkent, Ankara, Turkey 4 School of Applied Sciences, Health Innovations Research Institute, RMIT University, Melbourne, Victoria, Australia Received: Accepted: Published Online: Printed: Abstract: Leaf characteristics provide many useful clues for taxonomy. We used a back-propagation artificial neural network (BPANN) and C-support vector machines (C-SVMs) to classify 47 species from 3 sections of genus Camellia (16 from sect. Chrysanthae, 16 from sect. Tuberculata, and 15 from sect. Paracamellia). The classification model was constructed based on 7 leaf anatomy attributes including, area of adaxial epidermal cell, thickness of adaxial epidermal cell, thickness of palisade parenchyma, thickness of total leaf, thickness of spongy parenchyma, thickness of abaxial epidermal cell, and area of abaxial epidermal cell. Model parameters of C-SVM, comprising regularization parameter (C) and kernel parameter (γ), were optimized by cross-validation. The best classification accuracy of the 3 Camellia sections was achieved by the radial basis function SVM classifier (with parameters C = 32, γ = ), as well as the sigmoid SVM classifier (with parameters C = 32, γ = ), which was up to in the training set and in the prediction set, respectively. Compared with BP-ANN, SVM yields slightly
