Hyperspectral Image Classification Based On Feature Learning

Hyperspectral remote sensing has been successfully used in various areas of daily life,including food,agriculture,mining,etc.,which makes the study of hyperspectral images become more and more meaningful,especially hyperspectral image classification.It can be found that the application fields of hyperspectral remote sensing have a certain relationship with the classification task.Therefore,the main purpose of this paper is to study the effective feature learning methods to achieve efficient hyperspectral image classification.The main work is as follows: 1)In traditional feature learning methods,spatial feature information and spectral feature information are separately extracted and then be used for classification.However,these methods ignore the relationship between spectral features and spatial features in hyperspectral image data.As we all know the spatial information is reflected by the spectral band and the neighborhood,so separate extraction will affect the classification accuracy,moreover,these feature learning methods are time consuming and complex.Based on the above reasons,a novel hyperspectral image spectral–spatial classification method based on non-local average filter is proposed.In this method,similarity between pixels is measured by Pearson correlation coefficient,and then is used to construct the filter weights of nonlocal average filter for filtering.In order to eliminate the redundancy,the maximum noise fraction is used to reduce the dimension of the filtered data.Finally,using the SVM for classification.The results show that the method can not only improve the overall classification accuracy,but also improve the accuracy of homogeneity and edge regions.2)In view of the fact that the extended morphology profile method does not have a good protection to the edge in the process of extracting the features of hyperspectral images,this paper proposes a new extended morphology profile method based on guided filer.In this method,the extended morphology profile and the guided filtering are combined,and the edge protection of the guided filter is used to make up the deficiency of the extended morphology profit.In this feature learning methods,firstly,the principal component analysis is used to reduce the dimension of the hyperspectral image data,and the extended morphology profile method is used to extract the morphological features,then the first few principal components as the guided images and the corresponding morphological features as the input images to guide filter for filtering,finally,the filtering result is the feature of hyperspectral image we want to study,which will be classified by SVM.The results show that the proposed method can improve the accuracy of hyperspectral image data.At the same time,it can improve the accuracy of the small area and the edge,and it also improves the traditional EMP method.3)In order to overcome the shortcomings of traditional methods in extracting the spatial information of hyperspectral images,this paper proposes a new method for hyperspectral image classification based on multi-scale feature and deep network.In this method,the spatial features of hyperspectral images are extracted in linear and nonlinear space respectively,which can be better used to replace the original hyperspectral image data.In this feature learning methods,the spectral features and spatial features are extracted by using the deep auto-encoder,the improved Gauss scale space and the directional wavelet transform.The directional wave transform is used to extract the spatial features of the transform domain.Then the extracted features are combined,and will be classified by SVM.The results show that the method has a good classification effect on the detail,homogeneity and edge regions.