Block-based maximum likelihood classification for hyperspectral remote sensing data

In this paper, a block-based maximum likelihood classification is developed. The proposed method is based on individual class statistics to break the complete set of bands into several highly correlated subgroups. By ignoring low correlations between subgroups, the maximum likelihood methods is then employed for each subgroup independently. To accommodate the flexibility of using different segmentations for different classes, a progressive two-class decision procedure is used. This method also overcomes the problem caused by inadequate training samples for small classes. Experiments using a hyperspectral remote sensing data set were carried out. The results show that the block-based maximum likelihood method is an effective and practical alternative to conventional maximum likelihood classification for small class identification. The classification accuracies given by the proposed method are significantly higher than using a minimum distance classifier, which is one of the only viable techniques for hyperspectral data sets.