Research On Feature Extraction And Classification Of Hyperspectral Images Based On Super-pixel

In recent decades,the technology related to hyperspectral remote sensing has developed rapidly.Hyperspectral images contain both spectral and spatial information,and have more application value than other images,and have been widely used in such fields as crop detection,geological exploration,biomedical image analysis,environmental monitoring and protection.The early classification methods only use spectral features for classification,which can extract the structural information of hyperspectral images well,but the information extracted is not sufficient due to noise,and the extraction of spatial information is neglected,which affects the classification performance.The method based on super-pixel segmentation divides the image into homogeneous regions,which is more conducive to the information retention of feature edges.Therefore,this paper proposes two hyperspectral image classification methods based on the super-pixel segmentation method,the main contents of which are:(1)The traditional classification method only uses spectral information,the feature edge information retention is poor and the classification performance is easily disturbed by mixed image elements.To address this problem,a joint spatial-spectral classification method incorporating hyperpixel and extended multi-attribute profile features(EMAP)is proposed.Named SP-EMAP method,it firstly combines principal component analysis with superpixel segmentation method to extract potential low-dimensional local features to better preserve feature edge information,followed by extended morphological contour method to characterize spatial information from multiple attribute structures,then linearly fuses the two information features and uses recursive filtering method to remove the fused redundant information,and finally uses support vector machine The classification is carried out.In the experimental stage,the SP-EMAP method obtained 95.71% Kappa coefficients on the Indian Pines dataset,95.37% on the University of Pavia dataset and 96.84% on the KSC dataset,all of which have a significant advantage over the comparison methods.(2)To address the problem that the super-pixel segmentation method cannot fully exploit the HSI spatial information at a single scale,a post-processing classification method for hyperspectral images based on multi-scale segmentation is proposed based on the research in the previous chapter.Named MSP-IC method,on the basic framework of classification post-processing,the multi-scale spatial neighbourhood information is obtained by the super-pixel segmentation method,each scale will be input to the support vector machine to obtain the initial classification label map,the initial probability is optimally adjusted by the guided filtering method,which makes up for the shortage of feature extraction of small-scale objects under a single scale,and finally,the decision fusion method is used to obtain the optimal After the experiments,the kappa coefficients of the MSP-IC method reached 97.04% on the Indian Pines dataset,98.45% on the University of Pavia dataset and 96.31% on the KSC dataset,respectively.It is verified that the MSP-IC method can fully extract the multi-scale information features of HSI and effectively improve the classification accuracy.