An Unsupervised Classification of Hyperspectral Images Based on Pixels Reduction with Spatial Coherence Property

In order to improve classification and edge accuracy,PRSCP and linear regression analysis are introduced;a new algorithm of unsupervised classification based on PRSCP is proposed.The algorithm procedure starts with the similarity of pixel spectral,and then makes use of minimum related window to combine similar pixels spatially adjacent into a block.Linear expression is applied to model the spectral vector of pixels in the same block,and significance of the linear expression is verified by F-statistic.The basic vector of block is estimated via ODLR,and blocks with similar basic vectors are combined into the same class.AVIRIS data is used to evaluate the performance of the proposed algorithm,which is also compared with K-MEANS and ISODATA.Experimental results show that the proposed algorithm outperforms K-MEANS and ISODATA in terms of classification accuracy,edge and robustness.