Hyperspectral Image Classification Based On Spatial Filtering And Sparse Representation

Hyperspectral image(HSI)makes use of the imaging spectrometer to continuously image the target in a spectral wavelength from visible to near-infrared.The spatial geometric information and detailed spectral information of the observation scene are simultaneously collected to realize "integration of map and spectrum".HSI has been widely used in land use/land cover(LULC)mapping,environmental monitoring,resource reconnaissance,precision agriculture,military target detection and other fields.Feature extraction and classification of HSI plays an important role in HSI interpretation,which can extract the intrinsic features of high-dimensional HSI data and comprehensively use spectral information and spatial structure to achieve precise identification of ground objects.However,due to the affect of high dimensionality,strong bandwise correlation,the phenomenon of "the same objects with different spectrum,different objects with the same spectrum" in HSI data,how to achieve precise classification of groundtruth is still a difficult problem to be solved.This paper takes accurate classification of HSI as the main research goal,makes full use of the detailed spectral information and spatial structure of HSI,and conducts research on the feature extraction and spectral-spatial classification of HSI.The research content of this article mainly includes the following three aspects:(1)HSI has high spectral dimensions,non-linear data distribution.The samples belonging to the same class have the characteristics of multi-modal distribution.So the phenomenon of"dimension disaster" is prone to occur in HSI classification with small-size samples.Aiming at this problem,a non-linear feature extraction algorithm called local preserving discriminant analysis(LPDA)is proposed.LPDA separately learns the local manifold structure within the same class and between different classes of HSI,and maps the high-dimensional data to a lowdimensional feature space with compact structure within the same class and separate structure between classes,while still retaining the multi-modal distribution within the same class.This can mine the intrinsic structure of HSI dataset effectively.Because it is difficult to select a reasonable neighborhood structure in local learning,a more robust discriminant sparse preserving embedding(DSPE)algorithm is proposed.DSPE utilizes sparse representation to adaptively construct neighborhood in the same class,and imposes l1 and l2 norm penalty terms on the reconstruction coefficients to guarantee the sparsity and stability of neighborhoods.The neighborhood can describe the similarity of samples in the same class.By mapping this kind of similarity to a low-dimensional feature space,DSPE realizes nonlinear feature extraction of HSI.(2)When HSI is classified pixel by pixel,the classification map is likely to appear"salt and pepper noise".When pixels in a local region belonging to the same class are mistakenly classified,the consistent distribution of the same class is destroyed.Aiming at this problem,a spectral-spatial classification algorithm based on probabilistic classification map with edgepreserving filtering is proposed.The algorithm uses the characteristics of strong correlation between adjacent bands to divide HSI into band subsets,and uses the principal components of each band subset to construct a new feature subset,which can reduce the band correlation while maintaining the global structure of the data.We use the new feature set to train the probabilistic support vector machine to ensure the classification accuracy with small-size training samples,and construct probabilistic classification map with detailed probability attributes of each sample,which can help enhance the intensity of subsequent spatial optimization.In order to preserve the true edges between different classes and the local consistency of the same class,the principal components of the HSI is used to guide the edge-preserving filter to optimize the spatial structure of probabilistic classification map.The final classification map is more in line with the actual situation.(3)The spatial optimization result of HSI is limited by the initial classification accuracy.To avoid the spatial optimization factor amplifying the local misclassification,a superpixel-wise multi-feature joint sparse representation is proposed.The method first extracts multiple pixelwise features of HSI to describe the spectral and spatial characteristics of HSI from different aspects.Then,the entropy rate segmentation algorithm is used to divide the HSI into several superpixels,and multiple pixel-wise features are merged into the superpixel-wise joint sparse representation model.For the same kind of feature,all samples in the superpixel are sparsely represented by common atoms to preserve the similarity between these samples in the same superpixel;for different kinds of features,the reconstructed atoms are adaptively adjusted respectively to maintain the characteristics and advantages of different features.In order to improve the reconstruction accuracy,the objective function of the joint sparse representation model is relaxed into a convex optimization problem,and an alternate iterative solution method for reconstruction coefficient matrix is utilized.Finally,the minimum criterion of weighted sub-dictionary reconstruction error is used to determine the category of each superpixel and high-precision classification of HSI can be achived.