Research And Application Of Hyperspectral Images Feature Extraction And Classification Under Different Noise Conditions

With the continuous progress of spectral imaging technology,hyperspectral remote sensing images can accurately identify the detected target objects.Each pixel in the image records the spectral information of dozens or even hundreds of continuous bands,which makes it easy to identify each pixel.This characteristic of hyperspectral images makes them of great value in the fields of environmental monitoring,precision agriculture and aerospace.However,whether supervised or unsupervised classification,different types of noise are inevitably introduced in the process of hyperspectral image acquisition and transmission,there are Gaussian noise,salt and pepper noise and other mixed noise in the original image,which reduces the subsequent classification accuracy.On the other hand,for the practical application of false label detection of remote sensing images based on supervised classification,researchers need to label the labels manually.However,it is difficult to realize manual labeling in the real scene.There are noise labels in the training samples identified by human eyes alone.Based on hyperspectral images in the case of two types of noise.This paper focuses on the denoising and classification of images with different types of noise,and puts forward three algorithms of hyperspectral image denoising and feature extraction.The specific research of this paper is as follows:(1)In this paper,a denoising and dimensionality reduction algorithm(LRS-HRFMSuper PCA)based on multi-scale superpixel dimensionality reduction combined spatial domain filtering and transform domain filtering is proposed..The steps are as follows: First using the low rank structure of hyperspectral images and the sparse characteristics of noise,the unobserved part of the original hyperspectral image caused by mixed noise is repaired,block matching 3D algorithm is used to remove mixed noise.Next principal component analysis is used to reduce the dimension of the reconstructed hyperspectral image,and the principal component image is segmented by multi-scale superpixels.All principal component images with superpixels are projected into the reconstructed hyperspectral image in parallel,and the dimension of the image with superpixels is reduced by principal component analysis again.Then hierarchical domain transform recursive filtering algorithm is used to extract the features of each principal component image.Final support vector machine based on decision fusion classifies the image set after feature extraction.Experiments show that the algorithm can effectively remove the original noise in hyperspectral images and improve the classification accuracy.(2)In this paper,a low rank sparse representation combined with normalized spectral angle mapping for hyperspectral image spatial density peak clustering denoising classification algorithm(LRS-Norm SDP)is proposed,which overcomes the problem of noise labels in the training samples of hyperspectral images and decreases the successive classification accuracy under the background of supervised learning.This method took the following steps: First the subspace is estimated by using the low rank characteristics of the image itself,and the estimated signal subspace is reconstructed to obtain the fully observed hyperspectral image.Next the sparse noise in the completely observed image is denoised to obtain the hyperspectral image without sparse noise.Then the normalized spectral angle mapping algorithm is used to get the spectral distance matrix between the two training samples.Considering the sampling of the neighborhood around each pixel,the spatial correlation coefficient between each sample in each class is calculated,and the local density is calculated by density peak clustering algorithm.Finally,the abnormal labels are detected and removed by the decision function based on simple threshold,which improves the quality of training samples.The experimental results on two data sets show that the algorithm improves the detection and classification performance compared with the advanced similar algorithms proposed recently.(3)Aiming at the dimensionality disaster of hyperspectral image and the phenomenon that the image information continues to propagate across the strong edge when the original guided filter processes the image,A new hyperspectral image denoising classification algorithm(BC-IRF)based on joint band clustering and improved domain transform recursive filtering is proposed.The steps are as follows: First K-means clustering based on K-L divergence is used to cluster all spectral bands of hyperspectral images to reduce the dimension of images.The innovation of BC-IRF algorithm is to perform Gaussian filtering on each band in the central band set selected after K-means clustering,and set the guide image as the characteristic image obtained after Gaussian filtering.Then the input image is set as the selected band image after K-means clustering,and the input image remains unchanged.The output image after recursive filtering is updated to the guide image of recursive filtering of the next domain transformation,until the small-scale texture structure is eliminated.Finally,the last output feature image of domain transform recursive filtering is classified by support vector machine.The experimental results show that BC-IRF algorithm can effectively improve the classification accuracy.