Multi-task low-rank hyperspectral image classification method

Abstract

The invention discloses a multi-task low-rank hyperspectral image classification method. It mainly solves the problem that the existing methods only use spectral features in hyperspectral image classification, and cannot describe hyperspectral characteristics from multiple perspectives, resulting in low classification accuracy. The steps include: 1) input hyperspectral images; 2) extract spectral gradient features from hyperspectral images; 3) use spectral features and spectral gradient features as input signals and dictionaries of multi-task low-rank models, and obtain two Coefficient matrix; 4) Connect two coefficient matrices by row to get a new coefficient matrix as the new eigenvector matrix of the sample; 5) Select a part of the samples as the training set, and the rest as the test set; 6) Input the training set and the test set Sparse representation classifiers get classification results. Compared with the traditional low-rank model classification method, the invention effectively utilizes the cross feature information, and obtains higher classification accuracy compared with the existing image classification method.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to a target recognition method, in particular to a multi-task low-rank hyperspectral image classification method, which can be applied to remote sensing image ground object classification. Background technique [0002] As a kind of remote sensing image with high spatial and spectral resolution, hyperspectral remote sensing image contains a lot of information in its rich spatial and spectral features, and is widely used in the identification and classification of ground objects in military, agricultural and industrial fields. and feature descriptions. Hyperspectral remote sensing image classification is the process of classifying the pixels in a hyperspectral image into different categories. Hyperspectral remote sensing image classification is based on the classification of remote sensing images. Data is identified and classified at the pixel level. [0003] At present, many...

AI Technical Summary

Problems solved by technology

Many of these methods only use a single spectral feature, but only using spectral features can only describe the characteristics of hyperspectral images from one perspective

And because the hyperspectral image has many bands and the correlation between the bands is high, the same substance will have different spectral features in different regions, but different substances are likely to have similar spectral features, so a single spectral feature is used. It is difficult to avoid the occurrence of the above phenomenon, and it is easy to divide the same substance into different categories, and different substances are considered to be of the same kind, thus reducing the accuracy of image classification

[0004] At present, the most effective method to solve this phenomenon is the space-spectrum joint classifier, which introduces spatial local information, and believes that the samples of the neighbors have a large spatial similarity and belong to the same class of substances with a high probability, but this method is very difficult. Difficult to maintain smooth boundaries, and adjacent different substances are easily misclassified

[0005] Many existing multi-feature hyperspectral image classification methods obtain the correlation matrix between samples separately from each feature, and linearly combine multiple correlation matrices to obtain the final correlation between samples. This processing is insufficient. It ignores the cross-feature information between samples in the process of solving different correlation matrices, so it cannot make good use of the complementary image information contained in multiple features, which affects the effect of introducing multiple feature classification methods, which leads to accurate classification. low rate

Images