Research On Hyperspectral Image Classification Algorithm Based On Multi-scale Local Binary Pattern

With the development of the detector technology,the spatial resolution and spectral resolution of hyperspectral image have been improved greatly.In the spectral domain it shows that characteristic spectrum is continuous spectrum which can be researched quantitatively.The higher spatial resolution causes the clustering phenomenon,which means that the ground truth of same class are more likely in continuous region,has been verified in hyperspectral remote sensing image processing.Hyperspectral images contain abundant spatial information,radiation information and spectral information,which has brought great opportunities for the development and application of classification algorithm model in the remote sensing image classification.From the spatial feature of hyperspectral image,this thesis uses multi-scale LBP operator to extract the spatial texture feature of hyperspectral image,and extracts the multi-scale LBP characteristics for specially selected wavelengths or characteristic wavelengths,and then classifies them by using the multiple kernel framework multinomial logistic regression.The AVIRIS hyperspectral remote sensing and the ROSIS hyperspectral image are selected as the experimental data,and the LBP operator and the multi-scale LBP operator are respectively used to extract the spatial texture feature of hyperspectral image,then their classification accuracy are compared.The experiment shows that the spatial texture feature extracted by multi-scale LBP operator can obtain the image spatial texture feature at different scales which improves the classification accuracy obviously.In this dissertation,the spectral curve of each spatial pixel is considered one-dimensional vector based on the spatial location from the viewpoint of Spectral vector.The spatial texture feature extraction of multi-scale LBP operator can be extended to vector operation,and the extraction method of spatial texture feature based on the spectral vector multi-scale local binary pattern is proposed.In this algorithm,all spectral features are applied to spatial texture feature extraction,which does not require the band selection of hyperspectral image or the band extraction of the feature.The theoretical and experimental analysis show that the classification accuracy of hyperspectral image has been improved significantly.