Unsupervised multiscale segmentation of multispectral imagery

Abstract

A method for segmenting high resolution multispectral forestry images acquired from aircraft is described. This method makes use of a hierarchical smoothing network to aggregate pixels. The aggregation process is guided by a nonorthogonal multiscale spatial/spatial frequency texture representation. Texture and spectral similarity measures between and within network levels are used to inhibit smoothing between land cover classes at five different resolutions. Segmentation performance is evaluated in terms of classification accuracy using independent and dependent samples for labeling emergent classes. The hypothesis that the accuracy of the network as it approaches steady state drops when interlayer connections are eliminated or when the texture information is removed is supported. The hypothesis that the segmentation network is more accurate than fuzzy clustering and unsupervised segmentation is verified.<>