Multispectral Image Classification Using Generalized Fully Constrained Least Squares Approach

Fully Constrained Least Squares (FCLS) has been widely used and proven to be a powerful tool for hyperspectral image classification. But for multispectral remote sensing images with only a few bands, the Least-Squares based approaches will all encounter the band number constraint (BNC), which requires the number of bands should be no less than the number of classes. In this paper, we proposed a generalization of the FCLS called generalized FCLS (GFCLS) that relaxes this constraint in such a manner that the FCLS can be extended to multispectral image processing in a supervised fashion. The idea of the GFCLS is to create a new set of additional bands that are generated nonlinearly from original multispectral bands prior to the FCLS classification. The effectiveness of the proposed GFCLS is evaluated by SPOT-5 images. Experimental results show that the generalized FCLS (GFCLS) method outperforms the conventional FCLS approach for multispectral imagery classification.