Abstract: Orthogonal subspace projection (OSP) has beensuccessfully applied to hyperspectral image processing.In order for OSP to be effective, the number of bandsmust be no less than that of signatures to beclassified so that there are sufficient dimensions toaccommodate individual signatures to discriminate oneanother via orthogonal projection. This intrinsicconstraint is not an issue for hyperspectral imagessince they generally have hundreds of bands which aremore than the number of signatures resident withinimages. It, however, may not be true for multispectralimages where the number of signatures to be classifiedis greater than the number of bands such as 3-band SPOTimages. This paper presents a generalization of OSP,called generalized OSP (GOSP) to relax this constraintin such a fashion that OSP can be extended tomultispectral image processing in an unsupervisedfashion. The idea of GOSP is to create new additionalband images nonlinearly from original multispectralimages so as to achieve sufficient dimensionality priorto OSP classification. It is then followed by anunsupervised OSP classifier, called automatic targetdetection and classification algorithm (ATDCA) forclassification. The effectiveness of the proposed GOSPis evaluated by a 3-band SPOT and a 4-band Landsat MSSimages. The experimental results has shown that GOSPsignificantly improves the classification performanceof OSP. !13
展开▼
