A New Application of Pixel Purity Index to Unsupervised Multispectral Image Classification

Two major challenging issues arise in unsupervised classification. One is how to generate desired knowledge directly from the data in an unsupervised manner. The other is how to find an appropriate follow-up classifier to use the obtained unsupervised knowledge to perform supervised classification. This paper presents a new approach to unsupervised classification for multispectral imagery. To address the first issue the pixel purity index (PPI) which is commonly used in hyperspeftral imaging for endmember extraction is used to find a good set of initial training samples without prior knowledge. To address the second issue the PPI-found samples are then used as training samples for a support vector machine to find a good set of training samples for a follow-up supervised classifier, Fisher’s linear discriminate analysis (FLDA) which performs classification iteratively to produce final results. The experimental results show the proposed approach has great promise in unsupervised classification.