UNSUP: An Approach to Unsupervised Classification of Remote Sensing Imagery

Multispectral remote sensing (RS) images are high-dimensional, their dimension varying from 7 (Landsat TM) to 256 or more for hyperspectral data (AVIRIS). A high spatial resolution leads to huge data volumes for RS datasets. Their interpretation, given the usual lack of sufficient ground knowledge, depends on feature detection, requiring efficient unsupervised classification methods. An approach is sketched using fast kd-tree algorithms, with adaptive k-NN density estimators, leading to a 4-step unsupervised classification. In the first step an adaptive, spatially biased learning sample of spectral values is drawn from an RS image to provide an optimal base for class detection. In the second step a density-based cluster analysis detects the class system, for various values of separation and sample coverage. In the third step classes are not just defined as labels but also as linear segments on their singular value decompositions. Finally, in the fourth step the full image is classified by mapping its pixels to the nearest class as spectral mixtures. A prototype was developed with Java JDK 1.1 and tested on a Landsat TM image of the Painted Rock reservoir. Performance was quite satisfactory. The resulting image classifications showed good discrimination and class texture.