Research On Hyperspectral Image Classification Based On Feature Fusion

Hyperspectral images contain a large amount of spectral information and spatial information,which makes hyperspectral images more valuable than ordinary images.With this advantage,hyperspectral images have been successfully applied in many fields,such as crop growth monitoring,geological survey and investigation,biomedicine,environmental monitoring and so on.Hyperspectral classification is a hot topic in the field of remote sensing technology.If only spectral features are used for classification,it will seriously affect the classification performance.Although the features obtained during feature extraction can well represent the spatial structure of the image,it is often necessary to adjust the parameters in practical application,and it is difficult to determine the optimal parameters.Therefore,small-scale objects are often ignored in this process,resulting in poor classification effect,In addition,a single feature will lead to the omission of image internal structure information.In order to reduce the impact of above problems on the classification effect,this thesis proposes a hyperspectral classification method based on feature fusion.The main contents are as follows:（1）Aiming at the problem that the neglect of small-scale objects in spatial feature extraction will affect the classification effect,this thesis proposes a hyperspectral image classification method based on edge preserving filtering and extended random walk.The method includes two stages: feature extraction and spatial optimization.In the feature extraction stage,in order to overcome the disadvantage that global PCA can not extract potential low-dimensional features well,we use Super PCA for dimension reduction preprocessing,which can not only remove the noise in the image,but also directly use the spatial information of the image.Then the edge preserving filter is used to extract the features of the reduced dimension image.Although the edge preserving filter can remove the noise in the image and weaken the edge details while maintaining the overall structure of the image,due to the existence of objects with different scales in the image,and the feature extraction often needs to adjust the parameters,it is inevitable to lose small-scale objects in this process,resulting in the inability to fully characterize the spatial information of the image and affect the classification effect,Therefore,spatial optimization is used to capture small-scale objects to make up for the deficiency of feature extraction stage.Finally,the final result is obtained by decision fusion.This method can show good classification effect on three authoritative data sets.（2）Aiming at the problem that a single feature will lead to the omission of image internal structure information and affect the classification effect,a hyperspectral image classification method based on extended random walk and multi feature fusion is proposed in this thesis.This method includes three different modules.Gabor filtering,structural profile and Super PCA are used to extract the spatial features of hyperspectral images,and then the Gabor features and structural profile features are superimposed with spectral features to form fusion features.This process not only effectively represents the spatial structures of different scales and directions,but also makes full use of spectral spatial information,Then the obtained features are input into the classifier to obtain the initial probability value.In order to maximize the utilization of spatial information,Extended random walk algorithm is used to spatially integrate the initial probability,and the spatial information of the image is encoded to obtain a weighted graph to optimize the initial probability at the pixel level.Finally,the final result is obtained by probability fusion.Good classification results can be obtained on the three data sets.