Combining unsupervised and knowledge-based methods in large-scale forest classification

Data analysis and physical reasoning, frame-to-frame variability, and the need to minimise operator interaction because of the large number of frames, led the authors to develop a fully automatic classification scheme based on ISODATA concepts within the SIBERIA project. This used a multi-variate Gaussian model for the data, and was adapted to accept different initialisation procedures and to be able to form both maximum likelihood and maximum a posteriori classifications. A further improvement was to use its output to drive an iterated contextual classifier, hence exploiting spatial information