Research On Data Expansion And Dimensionality Reduction Technology For Small Sample Hyperspectral Image Classification

The classification of ground objects in hyperspectral remote sensing images(Hyperspectral Image,HSI)is one of the important topics in the field of hyperspectral remote sensing image processing.In the problem of hyperspectral image classification,the labeling of training samples is a time-consuming and laborious task,and fewer training samples and higher dimensions of hyperspectral images can easily cause the "Hughes effect".In this regard,in the case of a small sample of hyperspectral images,this paper proposes two algorithms suitable for this situation from two perspectives of data expansion and dimensionality reduction.The main research work and innovative ideas are as follows:(1)A semi-supervised classification algorithm for hyperspectral images under the condition of small samples of hyperspectral images is proposed,that is,Super-pixel-based Secondary Data Augmentation For Hyperspectral Semi-supervised Classification(Super-pixel-based Secondary Data Augmentation For Hyperspectral Semi-supervised Classification).,S-SDA).The algorithm is divided into two parts,namely,the first and second data expansion algorithms based on superpixels.The first data expansion algorithm uses superpixel segmentation technology to divide the other pixels except the boundary in the superpixel block where the training sample is located.Join the training set.In the algorithm,an algorithm for calculating irregular superpixel boundaries and a method for regularization of superpixel blocks are also proposed;the second data expansion algorithm uses the training set after the first sample expansion.All super pixel blocks are classified based on spectral information,and super pixel blocks with higher confidence are added to the training set.Finally,a traditional classifier is used for classification.A small sample of data is used to verify the effectiveness of the algorithm on the IN,IP and SS data sets.The experimental results show that the algorithm has excellent algorithm accuracy and can well solve the Hughes effect.(2)Proposed a dimensionality reduction algorithm suitable for small sample hyperspectral images,namely PCA dimensionality reduction method based on information entropy and superpixel segmentation(IE-SGPCA).The algorithm merges from two perspectives of band selection and feature extraction,and is mainly divided into two parts,based on coarse subdivision group and information entropy band selection,and PCA feature extraction based on global and local information.The first part can retain the original information of the image as much as possible while reducing the redundant information of the hyperspectral image;the second part can use the local and global information to extract the features of the hyperspectral image,so that the reduced image can be fully used Spatial context information takes into account global information at the same time,and it has better distinguishing ability for objects in some small homogeneous areas and some large homogeneous areas with mixed ground objects.A small sample of data is used to verify the effectiveness of the algorithm on the IN,IP,and SS data sets.The experimental results show that the algorithm has excellent algorithm accuracy and can well solve the Hughes effect.