Assessment of hyperspectral data analysis methods to classify tree species

One of the most challenging issues in forest inventory based on remote sensing data is identification of tree species. Hyperspectral remote sensing data provides information which considerably facilitates tree species recognition. The objective of the research is to evaluate different hyperspectral data analysis methods to classify tree species in Latvian forest conditions. The study site is a forest in the central part of Latvia, Jelgava district (56º39’ N, 23º47’ E). The area consists of a mixed coniferous and deciduous forest. During research 598 trees were analyzed in 70 sample plots. Remote sensing data are 64 hyperspectral bands in the 400 - 970 nm spectral range. Two different classification techniques: linear discriminant analysis (LDA) and artificial neural networks (ANNs) were used. In LDA species classification was done by stepwise and using principal components of hyperspectral bands. In stepwise LDA 18 hyperspectral bands were used. LDA using principal components and ANNs used all 64 hyperspectral bands. The best results show stepwise LDA where 82.4% of the data were correctly classified. Scots pine was classified 94.8%, Norway spruce 83.5%, Silver birch 77%, European aspen 71.4% and Black alder 56.3%. Classification with ANN’s best results showed for Scots pine, Norway spruce and Silver birch – respectively 81%, 84%, 86%. With LDA using principal components Scots pine’s classification showed best results with 85.1% correctly classified trees.