Hyperspectral image classification method based on discriminant sub-dictionaries and multi-scale superpixels

Abstract

The invention discloses a hyperspectral image classification method based on discriminant sub-dictionaries and superpixels. According to the method, a KSVD dictionary learning method with consistent labels is improved, and global label information is introduced as a constraint in the sub-dictionary learning process. The method is combined with a multi-scale super-pixel strategy, the dependence ofa dictionary on the number of training samples is broken through, and the problem that the classification effect is poor when the number of certain types of samples is small and the number of the samples is unbalanced is solved on the premise that the total number of the samples is not changed. Experimental results show that global label information is innovatively introduced in the sub-dictionarylearning process, so that the learned dictionary is more discriminative. Meanwhile, multi-scale superpixel strategies in the early dictionary learning process and the later classification process arecombined, so that the advantages of the two algorithms are exerted to the maximum extent, a relatively ideal effect is obtained, high-precision quantitative analysis of hyperspectral image classification is realized, and the method has important practical significance in crop monitoring and the like.

Description

Technical field [0001] The invention relates to the field of hyperspectral image classification methods, in particular to a hyperspectral image classification method based on discriminant sub-dictionary learning and multi-scale superpixels. Background technique [0002] In recent years, the development of hyperspectral image classification is very extensive, and its applications in military, agriculture, environmental monitoring, etc. are becoming more and more urgent. Especially the supervised classification method can make full use of the spectral characteristics and label information of hyperspectral images, such as minimum distance classification, maximum likelihood classification, etc. However, because these methods do not take into account the sparse characteristics and spatial information of the image itself, the classification effect is not good. Inspired by the "sparseness of human visual attention mechanism", sparse representation came into being, and its applications ...

AI Technical Summary

Problems solved by technology

This part of the research mainly focuses on the utilization of image spatial information. Some methods sparsely encode hyperspectral pixels pixel by pixel. Such methods fail to combine the spatial features and spectral features of the image, and the classification effect is poor; other methods consider The test pixel is used as the center pixel to establish a neighborhood, and the sparse coefficients of the pixels in the neighborhood are jointly solved to improve the classification accuracy

[0005] Therefore, in terms of dictionary learning, the quality of dictionary learning is affected by the number of training samples. If the number of training samples of a certain class of samples is much less than that of other classes, or the number of samples between each class is unbalanced, then the classification accuracy of these classes it may be very different

In terms of sparse coding, if we only consider pixel-by-pixel coding or only establish a single-scale neighborhood, we cannot make full use of the differences in different regions (homogeneous regions and edge regions), which will affect the overall classification accuracy of the image.

Images