Study On The Classification Of Heperspectral Image Based On Band Subspace Partition

In comparison with traditional remote sensing image,hyperspectral remote sensing image improve the spectral resolution of remote sensing images greatly,hyperspectral image can get continuous spectral curves of objects on the ground,provides abundant spectral information for the identification and classification of objects.Although the spectral resolution of hyperspectral images is higher,the adjacent band image has a large amount of redundant information,so hyperspectral image dimensional reduction before the classification based on hyperspectral image is very necessary.In addtion,the traditional unsupervised classification method is greatly affected by the i:nitial category center,and how to select a better initial classification center is an urgent problem to be solved.Aiming at the problems existing in hyperspectral image dimensional reduction and classification,the main work is as follows:1?Aim at the spectral correlation is large,there are a lot of redundant information,proposed unsupervised band selection method based on subspace partition,mainly includes four steps:mutual information matrix construction,subspace partition,representative band selection,classification and its evaluation.The representative bands selected by our method and three traditional methods were used in classification,experiments show that this method in the vast majority of the number of band selection has higher classification accuracy and Kappa coefficient.For different classifiers,the representative bands selected by our method have better adaptability,and the classification performance can be higher than the other three band selection methods.labor at the time of consumption,this method is slightly more time spent.2?Band selection based on Orthogonal subspace projection method need to build the image vector orthogonal subspace,and have to traverse all the rest bands,will spend a lot of time.To solve these two problems,spectral information divergence and structural similarity are used as similarity measurements between bands and subspace are used to reduce the number of bands traversed.Representative selection method selected by our method and other traditional method are used to classification experiment,and experiments show that under the premise of ensuring the classification accuracy,with the increase of the number of selected bands,the time improved method were less than the OSP method,and the advantages of time first increased and then decreased.3?Aim at the traditional fuzzy clustering method is easily affected by the initial clustering center,therefore the global search performance of PSO is introduced into FCM to improve the clustering performance of FCM method.The representative bands of the above two band selection method are used into the improved FCM clustering method,compared the clustering performance with the traditional and improved FCM clustering.Experimental results show that improves FCM clustering method has higher classification accuracy,and the error rate is less.