SearchPublisher and Institutes
You are here: Home :: Area NEM :: Agricultural science
David Nilsson, Ulf Edlund
Pine and spruce roundwood species classification using multivariate image analysis on bark
Keywords: bark, co-occurrence matrix, multivariate image analysis, partial least squares projection to latent structures (PLS), pine, spruce, wavelet, wood species
Wood discs from 67 pine and 79 spruce logs were collected from a forest clearing. Three different 24-bit red-green-blue (RGB) images were acquired from the radial surface of each disc. The first image contained bark, the second image was a mixture of bark and wood surface, and the third image consisted only of wood surface. The image texture was compressed into vectors of Fourier-transformed wavelet coefficients. These were assembled in matrices and analysed by principal component analysis (PCA) and partial least-squares projections to latent structures (PLS). Classification using Fourier-transformed wavelet scales showed that the wood species could be predicted with 90% accuracy. A thorough examination of this classification showed that the predicting power of these models was mostly due to wavelet scales that represented the mean value of each colour channel. The prediction accuracy that could be obtained from coefficients representing image texture was generally low. The use of grey-level co-occurrence matrices prior to the wavelet transformation showed, however, that it is possible to classify the wood species of pine and spruce with an accuracy approaching 100%.
Holzforschung, Walter de Gruyter
Print ISSN: 0018-3830
Volume: 59, 11/2005
Pages: 689 - 695