Weighted Neighborhood Composite Kernel For Hyperspectral Image Classification

With the rapid development of hyperspectral remote sensing system technology,hyperspectral image classification has become one of the key research directions of hyperspectral remote sensing.Hyperspectral image contains a wealth of spectral information,which provides powerful data support for the classification of ground objects.However,hyperspectral image classification technology is also deeply affected by many factors such as image noise,limited labeled samples,and high redundant information,which makes it difficult to obtain ideal analysis results.Based on the classic composite kernel classification method,this paper mines more accurate spatial information to reduce noise interference.The specific research contents are as follows:(1)A composite kernel method based on anisotropic filtering is proposed to overcome the weakness that composite kernel method extracts inaccurate spatial information at the edge through isotropic neighborhood.The structure tensor information is used to construct an anisotropic weight function,which assigns different weights to the neighborhood pixels of the target pixel.The higher the similarity between the pixels in the neighborhood and the current pixel,the greater the weight,which reduces the negative impact of heterogeneous pixels in the neighborhood,enhances the ability to maintain edge details,and improves classification accuracy.(2)A composite kernel method based on weighted adjacent superpixels is proposed to overcome the weakness that composite kernel method is difficult to fully extract spatial features through the same scale neighborhood.The shape and size of superpixels can be adaptively changed according to the image texture structure,so as to improve the robustness of the model under different scale targets.The spatial information and location information of adjacent superpixels are used to enhance the extracted spatial features.Furthermore,a composite kernel method based on multiscale weighted adjacent superpixels is proposed.This method fuses the spatial features extracted at different scales through a multiscale superpixel segmentation strategy,which makes full use of the spatial information at different segmentation scales,and further improves the classification accuracy.The classification results on real hyperspectral image datasets verify the effectiveness of the proposed method.