CN-122023889-A - Hyperspectral image classification method based on superpixel spectrum similarity measurement

Abstract

The invention discloses a hyperspectral image classification method based on hyperspectral pixel spectral similarity measurement, which is characterized in that a first three principal components are extracted based on principal component analysis to synthesize a false color image, an entropy rate hyperspectral pixel segmentation algorithm is adopted to generate a spatial coherent hyperspectral pixel unit, spatial structural integrity is kept, internal spectral information is kept, a domain transformation recursive filter is adopted to carry out edge-keeping filter processing on an original hyperspectral image, then spectral domain noise separation and data dimension reduction are carried out through maximum noise fraction transformation, minimum-maximum normalization processing is carried out on all spectral features to eliminate dimension difference, a distance matrix is formed by calculating minimum Euclidean distance between a test sample and each class training sample in each hyperspectral pixel unit, a k nearest neighbor average strategy is adopted to filter the most representative test pixel to calculate similarity score, and finally class labels with the highest similarity score are distributed to all pixels in the whole hyperspectral.

Inventors

• ZHANG PANPAN
• LI XINGYU
• WANG YULEI
• ZHAO ENYU
• YU CHUNYAN

Assignees

• 大连海事大学

Dates

Publication Date

20260512

Application Date

20260120

Claims (4)

1. 1. A hyperspectral image classification method based on hyperspectral similarity measurement of super pixels is characterized by comprising the following steps: Carrying out principal component analysis on an original hyperspectral image, and extracting and retaining the first three principal components with the largest variance contribution rate by calculating characteristic values and characteristic vectors; synthesizing the extracted three main components into a false color image, dividing the image into super-pixel units by adopting an entropy rate super-pixel dividing algorithm, and generating a two-dimensional super-pixel dividing map; filtering spatial noise in the standardized hyperspectral image in a transformation domain based on a domain transformation recursive filtering mode, and simultaneously retaining edge information among ground objects; Carrying out maximum noise fraction conversion on the filtered hyperspectral image, and carrying out noise separation and dimension reduction treatment on the hyperspectral image in a spectrum domain to obtain preprocessed hyperspectral data; Fusing the two-dimensional super-pixel segmentation map with the preprocessed hyperspectral data to construct three-dimensional super-pixel cube data with homogeneous region characteristics; carrying out maximum and minimum normalization processing on the three-dimensional superpixel cube data, and scaling the pixel values of each wave band to the [0,1] interval to obtain normalized hyperspectral data; for a test sample to be classified, calculating the spectrum Euclidean distance between the test sample and a training sample set of a known class in a super pixel unit to which the test sample belongs, calculating the distance from the test pixel to all training samples for a certain test pixel, and taking the minimum distance as the similarity measurement of the test pixel and the class; Traversing all pixels to be classified in super pixels in the normalized hyperspectral data, calculating the Euclidean distance between the pixels to be classified and various training sample sets, selecting a minimum distance value from Euclidean distances between the pixels to be classified and various training sample sets as similarity measurement with a certain class, thereby obtaining the similarity between each pixel to be classified and various samples, and obtaining the distance matrix of the super pixels; Selecting the distance values of k most similar test samples from the distance matrix aiming at each type of samples, calculating an average value of the distance values, traversing each type of samples by taking the average value as the final distance value from the whole super pixel to the type of samples, and converting the distance matrix into a distance vector; and distributing the class label with the smallest distance value to all pixels in the super pixel so as to complete the hyperspectral image classification process.
2. 2. The hyperspectral image classification method based on hyperspectral similarity measure as claimed in claim 1, wherein the normalized hyperspectral data is Wherein N is the total pixel number, b is the band number after dimension reduction, and the training sample set of the q-th class is set , Is the j-th pel of the q-th class training sample, Is the number of class q training samples, and for the ith test sample in the p-th superpixel, the Euclidean distance to the class q training sample is calculated as follows: calculating the distance from each training sample in the q-th class to obtain Taking the minimum distance value As a distance measure from the category, the minimum distance between the sample and the samples of other categories is calculated respectively to obtain 。
3. 3. A hyperspectral image classification method based on hyperspectral similarity measure as claimed in claim 1 is characterized by selecting k test samples most similar to each type of sample, averaging them, calculating average value of k minimum values of each row to obtain a column vector ; Matrix is arranged Wherein And For the ith row, find the smallest k values in the ith row, record as a set Vector then Is the ith element of (2) Expressed as: Finally, the distance vector is obtained ; Distance vector And each value of the super pixel represents the distance between the super pixel and each ground object, the smaller the distance is, the higher the similarity is, the category with the largest similarity is selected as the prediction labels of all pixels in the super pixel, and the category label with the smallest distance value is distributed to all pixels in the super pixel, wherein the classification criterion is as follows: 。
4. 4. the hyperspectral image classification method based on the hyperspectral image similarity measure of claim 1, wherein when spatial noise in the original hyperspectral image is filtered in a transform domain based on a domain transform recursive filtering mode, for a given input signal I, the input signal I is converted into a transform domain, the distance between pixels is calculated in the transform domain, and the domain transform process is expressed as: Wherein the method comprises the steps of The domain transformation operator is represented as a function of the domain transformation, And Representing spatial information and distance information of the bilateral filter, respectively, c represents the number of channels of the input signal, Is the derivative of the input signal Defining a recursive edge preserving filter in the transform domain The method comprises the following steps: Wherein the method comprises the steps of Is the result of the output of the filter, As a feedback coefficient, the feedback coefficient is used, Is a neighbor sample in the transform domain And The distance between the two is increased, as d increases, Toward 0, the propagation chain is stopped, leaving edges.

Description

Hyperspectral image classification method based on superpixel spectrum similarity measurement Technical Field The invention relates to the field of hyperspectral image classification, in particular to a hyperspectral image classification method based on hyperspectral similarity measurement. Background Hyperspectral images contain hundreds of narrow adjacent spectral bands, capturing subtle spectral changes between surface materials. The capability enables the hyperspectral image to be widely applied to the fields of land monitoring, agriculture and forestry investigation, ecological environment evaluation, military affairs prevention and the like. Classification is one of the main tasks of hyperspectral image processing, aiming at assigning a label to each pixel. In the early stage of hyperspectral image research, classification tasks are mainly realized through classifiers such as support vector machines, decision trees, random forests and the like. However, due to the high dimensionality of the hyperspectral dataset and the limited number of labeled samples, these pixel-by-pixel methods may suffer from the "housin phenomenon" resulting in performance degradation with increasing dimensionality. To alleviate this problem, dimension reduction techniques such as feature extraction and band selection are typically used prior to classification. Representative feature extraction methods are principal component analysis, maximum noise fraction transformation, and independent component analysis. These linear dimension reduction techniques alleviate some of the challenges presented by high-dimensional data, but they generally assume that the data is located on or near a linear subspace, which makes them less efficient in capturing potentially nonlinear structures inherent in the data. Manifold learning, in contrast, reveals the inherent nonlinear structure of the data and thus can be a powerful feature extraction method for hyperspectral images. The core idea is that while hyperspectral image data typically resides in a high-dimensional spectral space, the spectral response patterns of different features tend to follow the underlying low-dimensional nonlinear structure, and by finding and learning such potential manifolds, the original high-dimensional data can be mapped into a lower-dimensional embedded space while preserving the fundamental geometric and topological properties of the data. Although the above approach is effective in dimension reduction, spatial correlation between pixels is often ignored, limiting their classification accuracy in complex scenes. In addition to the rich spectral information, the hyperspectral image also contains rich spatial information. Various spatial feature extraction techniques have been employed in hyperspectral classification, such as morphological and texture-based approaches. While spatial structure-based approaches effectively improve classification performance, their design emphasizes the use of spatial features, often without fully exploiting the rich spectral details inherent to hyperspectral, which limitation may lead to poor classification performance in spectrally chaotic or complex scenarios. To further improve classification performance, researchers have begun to explore more efficient strategies for integrating spatial and spectral information. However, most of the existing methods based on space-spectrum features rely on a sliding window with a fixed size for space feature extraction, lack of adaptability to local structural changes, and are easy to introduce different types of interference pixels in the fixed window. In recent years, deep learning has been remarkably successful in hyperspectral image analysis and processing, and various network architectures represented by Convolutional Neural Networks (CNNs) and transformers have strong capability of automatically learning multi-level space-spectrum characteristics, and deep discrimination information in high-dimensional data is effectively mined in an end-to-end training mode. However, its practical application faces significant challenges that severely rely on large scale marker data sets, high model complexity, and large computational overhead, which makes their deployment difficult and limits their applicability in real-world scenarios with scarce marker data or limited computational resources. Disclosure of Invention According to the problems existing in the prior art, the invention discloses a hyperspectral image classification method based on hyperspectral similarity measurement of super pixels, which specifically comprises the following steps: Cube three-dimensional data Remodelling into a two-dimensional matrixWhereinData normalization is then performed: Wherein the method comprises the steps of Is the mean value of each band of wavelengths,Is the standard deviation. And then carrying out principal component analysis on the standardized hyperspectral image, and extracting and retaining the first thre