Classification of forestry species using singular value decomposition

A method is defined and tested for the classification of forest species from multi-spectral data, based on singular value decomposition (SVD) and key vector analysis. The SVD technique, which bears a close resemblance to multivariate statistic techniques has previously been successfully applied to the problem of signal extraction from marine data. In this study the SVD technique is used as a classifier for forest regions, using SPOT and landsat thematic mapper data. The specific region chosen is in the County Wicklow area of Ireland. This area has a large number of species, within a very small region and hence is not amenable to existing techniques. Preliminary results indicate that SVD is a fast and efficient classifier with the ability to differentiate between species such as Scots pine, Japanese larch and Sitka spruce. Classification accuracy's using this technique yielded excellent results of > 99% for forest, against four background classes. The accuracy's of the individual species classification are slightly lower, but they are still high at 97 - 100% for the SPOT wavebands. When the Landsat TM bands 3, 4, and 5 were used on their own, accuracies of 95 - 100% were achieved.