Hyperspectral Image Classification Based On Super-pixel And Dictionary Representation

Compared with color image and multispectral image,hyperspectral image(HSI)has unique advantages in classification and recognition because of its organic combination of spatial information and spectral information.At present,HSI has been widely and successfully used in national defense,urban planning,precision agriculture,environmental detection,and so on.HSI classification for ground material information is a cutting-edge research topic of integrated image processing and remote sensing imaging technology.It is an effective method to improve the classification accuracy of hyperspectral image by using image technology and fully considering the complementarity between spatial information and spectral information of hyperspectral image.In this paper,two methods of HSI classification,which combine the spatial and spectral feature of HSI,are proposed,which can be summarized as follows:(1)The classification method based on super-pixel segmentation and collaborative representation nearest neighbor.In HSI,the neighboring pixels of one pixel have a high correlation with itself.Combining their respective information to classify is more advantageous than single pixel classification.In this method,firstly,the original data of HSI is dimensionally reduced by principal component analysis(PCA).The first three principal components are extracted,and then the super-pixel image is generated by the method of super-pixel segmentation based on entropy rate.Afterwards,in each super-pixel block,the Euclidean distance between each test sample and each class of training samples is calculated by combining collaborative representation and nearest neighbor classification algorithm.The classes of the test samples are predicted.Furthermore,the classification probability matrix of the test sample is placed in the Markov random field,and the final result of classification is calculated by the loopy belief propagation algorithm,which further improves the classification effect.(2)The classification method based on super-pixel fusion and joint sparse representation.In the traditional joint sparse representation(JSR),the neighboring pixels of the test pixels are simply considered to be the same category as them.But in fact,the neighboring pixels may also contain different categories of pixels,for example,at the junction of different categories of pixels,the neighborhood is likely to contain more than two categories of pixels.In this way,the classification performance of the JSR of the neighborhood will be affected.The JSR based on super-pixel can solve this problem,as the boundary of super-pixel can well avoid that there are many kinds of pixels in the joint representation samples.However,in order to ensure the pixels in the same super-pixel being the same class,the super-pixel segmentation algorithm used in HSI may usually over-segment,which makes the super-pixel smaller.In the JSR,more pixels can improve the accuracy of classification,so in this method,based on the possibility fuzzy c-means clustering(PFCM),a super pixel-fusion method is proposed.It merges the adjacent over-segmented super-pixels into larger super-pixels,and then the new super-pixels are used for the JSR method.In order to solve the limitation of JSR based on neighborhood,a new sample set is selected from the super-pixel and the neighborhood,which improves the accuracy of JSR algorithm.Experiments show that this method can improve the accuracy of HSI classification,compared with many other state-of-the-art methods.