Dimension Reduction And Classification Of Hyperspectral Remote Sensing Images

The classification technology of hyperspectral remote sensing image realizes the category information of surface material,more accurate than traditional techniques,which has important applications especially in precision agriculture,geological survey,military reconnaissance and identification of camouflage and so on.However,as the bands increase,the new challenge results in the storage,transmission and processing with the increased of inter-band correlation,data redundancy and data volume created of hyperspectral remote sensing images.For example,the feature corresponding is not obvious enough to some categories in the processing of feature classification,which will result in the reduction of the precision of subsequent classifiers.What's more,the overall classification accuracy of the classifier will decrease with the increase of the feature dimension,during the training samples are limited.Therefore,it has become a hot topic how to reduce the dimension more effectively,and how to establish the classifier which is unaffected or less affected by dimension of hyperspectral remote sensing image at present.In this paper,the band selection algorithm are proposed based on information content and the classification of spectral-based hyperspectral remote sensing image.The real objects hyperspectral data are used to compare the proposed algorithm and the common algorithm.The application value of the proposed algorithm is verified.The main innovations of this paper include the following aspects:1.A band selection algorithm with simultaneous selection and elimination is proposed for hyperspectral remote sensing image.Firstly,the hyperspectral data set is transformed by PCA,and the first N principal components are used as the reference band source,and then the mutual information is used to select the best as the similarity measure of bands selection.Secondly,the R-KL coefficient is introduced to remove the bands with high similarity to the current optimal band and less loss after the elimination of the data.Finally,the reference band is updated to select the current optimal band,so the iterative selection of the optimal band set contains lowcorrelation with a large amount of information in the low-dimensional data.2.At the same time,the spectral angle matching classification algorithm is also proposed based on wavelet transform.First,the multi-class reference spectrum and each pixel spectrum are decomposed by wavelet transform,and the approximate coefficient,and the detail coefficient are extracted.Then,the coefficients of wavelet decomposition of each pixel spectrum and the corresponding coefficients of each kind of reference spectrum are calculated for the angle.Finally,weighted the summation of the angle,and the pixel is placed in the category with the smallest angle.