Hyperspectral image classification method based on superpixel spectrum similarity measurement

Abstract

The invention discloses a hyperspectral image classification method based on superpixel spectrum similarity measurement, which comprises the following steps of: extracting first three principal components to synthesize a false color image based on principal component analysis, and generating spatial coherent superpixel units by adopting an entropy rate superpixel segmentation algorithm, so that not only is the integrity of a spatial structure maintained, but also the internal spectrum information is reserved; performing edge preserving filtering processing on the original hyperspectral image by adopting domain transformation recursive filtering; spectral domain noise separation and data dimension reduction are carried out through maximum noise fraction transformation, and minimum-maximum normalization processing is carried out on all spectral features to eliminate dimension differences; the minimum spectral Euclidean distance between a test sample and each category of training samples is calculated in each superpixel unit to form a distance matrix, the most representative test pixel is screened by adopting a k-nearest neighbor average strategy to calculate a similarity score, and finally a category label with the highest similarity score is allocated to all pixels in the whole superpixel.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral image classification, and more particularly to a hyperspectral image classification method based on superpixel spectral similarity measurement. Background Technology

[0002] Hyperspectral images contain hundreds of narrow, adjacent spectral bands, capturing subtle spectral variations between surface materials. This capability makes hyperspectral images widely used in land monitoring, agricultural and forestry surveys, ecological environment assessments, and military defense. Classification is one of the main tasks in hyperspectral image processing, aiming to assign a label to each pixel. In the early stages of hyperspectral image research, classification tasks were mainly achieved using classifiers such as support vector machines, decision trees, and random forests. However, due to the high dimensionality of hyperspectral datasets and the limited number of labeled samples, these pixel-by-pixel methods can be affected by the "Hughes phenomenon," causing performance to degrade with increasing dimensionality. To mitigate this problem, dimensionality reduction techniques, such as feature extraction and band selection, are often used before classification. Representative feature extraction methods include principal component analysis, maximum noise fractional transform, and independent component analysis. These linear dimensionality reduction techniques alleviate the challenges posed by high-dimensional data to some extent, but they typically assume that the data lies on or near a linear subspace, making them less efficient in capturing the inherent potential nonlinear structures in the data. In contrast, manifold learning can reveal the inherent nonlinear structures of the data and therefore can serve as a powerful feature extraction method for hyperspectral images. The core idea is that although hyperspectral image data typically resides in a high-dimensional spectral space, the spectral response patterns of different land features often follow underlying low-dimensional nonlinear structures. By discovering and learning this latent manifold, the original high-dimensional data can be mapped to a lower-dimensional embedding space while preserving the basic geometric and topological properties of the data. While the above methods are effective in dimensionality reduction, they often neglect the spatial correlation between pixels, limiting their classification accuracy in complex scenes.

[0003] Besides rich spectral information, hyperspectral images also contain abundant spatial information. Various spatial feature extraction techniques have been applied to hyperspectral classification, such as morphological and texture-based methods. Although spatial structure-based methods have effectively improved classification performance, their design emphasizes the utilization of spatial features and often fails to fully leverage the inherent rich spectral details of hyperspectral images. This limitation can lead to unsatisfactory classification performance in spectrally cluttered or complex scenes. To further improve classification performance, researchers have begun to explore more effective strategies for integrating spatial and spectral information. However, most existing spatial-spectral feature-based methods rely on fixed-size sliding windows for spatial feature extraction, lacking adaptability to local structural changes and easily introducing interfering pixels of different categories within the fixed window. In recent years, deep learning has achieved significant success in hyperspectral image analysis and processing. Various network architectures, represented by Convolutional Neural Networks (CNNs) and Transformers, have the powerful ability to automatically learn multi-level spatial-spectral features, effectively mining deep discriminative information from high-dimensional data through end-to-end training. However, their practical application faces significant challenges: heavy reliance on large-scale labeled datasets, high model complexity, and substantial computational overhead make their deployment difficult and limit their applicability in real-world scenarios with scarce labeled data or limited computing resources. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention discloses a hyperspectral image classification method based on superpixel spectral similarity measurement, specifically including the following steps: Three-dimensional data cube Remodeled into a two-dimensional matrix ,in Then, data standardization is performed:

[0005] in It is the mean of each band. This is the standard deviation. Then, principal component analysis is performed on the standardized hyperspectral image. By calculating eigenvalues ​​and eigenvectors, the top three principal components with the largest variance contribution rates are extracted and retained. The covariance of the standardized matrix is ​​then calculated.

[0006] The eigenvalues ​​and eigenvectors of the covariance matrix are calculated as follows:

[0007] in It is a diagonal matrix, and the elements on the diagonal are eigenvalues. , representing the variance of each principal component. Each column of V is the corresponding eigenvector, defining the transformation direction from the original band space to the principal component space. Finally, the eigenvalues ​​are sorted from largest to smallest, and their corresponding eigenvectors are the most important principal component directions. This process transforms the standardized original data... Projected onto the front l On the new subspace spanned by the eigenvectors, we obtain the new data after dimensionality reduction:

[0008] The three extracted principal components are synthesized into a false-color image. An entropy rate superpixel segmentation algorithm is then used to segment the image into superpixel units, generating a two-dimensional superpixel segmentation map. First, a graph is constructed. Its apex V It is the set of edges for each pixel in the image. E It consists of similarity weights for each pair of pixels. ERS selects an edge subset. Divide the graph into several connected subsets such that the resulting graph... It consists of small connected components. The objective function of ERS incorporates the entropy rate. and balance function The optimal subset is found using the following objective function:

[0009] in These are the weights of the balance term. Entropy rate tends to form compact and uniform clusters, while the balance function aims to ensure that the generated clusters have similar sizes. Since ERS is primarily designed for RGB images with three bands, while hyperspectral images have tens to hundreds of bands, we adopt the processing method commonly used in superpixel-based hyperspectral image classification. Specifically, we select PCA to extract the first three principal components of the hyperspectral image, synthesize them into a false-color image, and then use the segmentation method described above to divide the hyperspectral image into... P Superpixels ,in It is the first p The first superpixel in the nth superpixel i One pixel, It is the first p The number of pixels in a superpixel.

[0010] Recursive filtering using domain transform is employed to remove spatial noise from standardized hyperspectral images in the transform domain while preserving edge information between ground features. For a given input signal... I The data is then transformed to the transform domain, and the distance between pixels is calculated in the transform domain. The domain transformation process is represented as follows:

[0011] in Represents the domain transformation operator, and These represent the spatial and distance information of the bilateral filter, respectively. c This indicates the number of input signal channels. It is the input signal. The derivative of . For discrete signals. In the transform domain, the recursive edge-preserving filter is defined as follows:

[0012] in It is the filter output. It is a feedback coefficient. Neighbor samples in the transform domain and The distance between them. With d The increase, When the value approaches 0, the propagation chain stops, thus preserving the edges. Therefore, the edges are preserved, and pixels on the same side of the same edge usually have similar filtering effects.

[0013] The filtered hyperspectral image undergoes maximum noise fraction transform, followed by noise separation and dimensionality reduction in the spectral domain to obtain preprocessed hyperspectral data. Given a dataset... , d It is the feature dimension. n This represents the number of samples. First, the data needs to be decentralized. ,in It is the mean vector of each feature. It is the first in the dataset One sample. Assume the observed signal. From signal and noise Composition, that is ,therefore The covariance matrix is ​​expressed as ,in , and These are the covariance matrices of Z, X, and N, respectively. It is estimated using the maximum / minimum autocorrelation factor method. The objective of the maximum noise fractional transform is to find a projection matrix W that maximizes the signal-to-noise ratio of the data in the transform domain. Its objective function is:

[0014] Solving using the Lagrange multiplier method, we know that W is derived from... of b A matrix consisting of the eigenvectors corresponding to the largest eigenvalues. bThe dimension of the projected data is Y. After dimensionality reduction using MNF, the data Y is represented as:

[0015] The two-dimensional superpixel segmentation map generated by the entropy rate superpixel segmentation algorithm is fused with the hyperspectral data preprocessed by joint spatial-spectral filtering to construct a three-dimensional superpixel cube data with homogeneous region characteristics. Max-min normalization is applied to the fused low-dimensional hyperspectral data, scaling the pixel values ​​of each band to the [0,1] interval to eliminate the influence of dimensions, thus obtaining normalized hyperspectral data. For any original pixel in each band of the dimensionality-reduced data... x The following transformation formula is applied for calculation:

[0016] For a test sample to be classified, calculate its spectral Euclidean distance to each training sample set within its superpixel unit. For a given test pixel, calculate its distance to all training samples, and take the minimum distance as its similarity measure with that class. The normalized hyperspectral data is as follows: ,in N It is the total number of pixels. b This is the number of bands after dimensionality reduction. Given the th... q Training sample set of the class , It is the first q The first training sample j One pixel, It is the first q The number of training samples for each class. p The first superpixel in the nth superpixel i The test sample, compared with the first q The Euclidean distance between the training samples is calculated as follows:

[0017] Calculation and the first q The distance between each training sample in the class is obtained Take the minimum distance value This serves as a distance metric to the class. The shortest distances between this sample and samples from other classes are calculated separately. ; By iterating through all pixels to be classified within a superpixel in the normalized hyperspectral data, and using the Euclidean distance between each pixel and samples of different classes as a similarity measure, the distance matrix of that superpixel is obtained. Each row represents the distance between each test pixel in the superpixel and a specific category.

[0018] Considering that test pixels within the same superpixel may exhibit spectral variations, not every pixel is representative. Therefore, the pixel most similar to each type of sample is selected. k The distance values ​​of each test sample are averaged, i.e., the minimum distance value in each row is calculated. k The mean of the values ​​is used to obtain a column vector. Let the matrix be... ,in and For the first i Okay, first find the first i The smallest in the row k Let the set contain a set of values. Then the vector The i element It can be represented as:

[0019] Finally obtained .

[0020] The classification is completed by assigning the category label with the smallest distance value (i.e., the highest similarity) to all pixels within that superpixel. (Vector) Each value represents the distance between the superpixel and each category of land cover; the smaller the distance, the higher the similarity. The category with the highest similarity is selected as the predicted label for all pixels in that superpixel. The classification criteria of the proposed method are as follows:

[0021] By employing the above technical solutions, this invention provides a hyperspectral image classification method based on superpixel spectral similarity measurement. This method effectively solves the problems of introducing heterogeneous pixels into the fixed rectangular window and the difficulty in adapting to complex scene spatial structure transformations in hyperspectral classification. First, the hyperspectral image is processed sequentially using domain transformation recursive filtering and maximum noise score to reduce spatial and spectral domain noise and dimensionality reduction, respectively. At the same time, superpixel segmentation is performed to synthesize a three-dimensional superpixel cube. Subsequently, the spectral Euclidean distance between each superpixel and all class training samples is calculated, and all pixels within each superpixel are classified into the class with the highest similarity. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the method disclosed in this invention; Figure 2 a- Figure 2 l represents the false-color images, ground truth, and classification results of the proposed classification method (SLSSM) and the contrasting classification methods from the Indian Pines dataset; Figure 3 a- Figure 3 l represents the false-color images, ground truth, and classification results of the proposed classification method (SLSSM) and the contrasting classification methods from the Pavia University dataset; Figure 4 a- Figure 4 l represents the false-color images, ground truth, and classification results of the proposed classification method (SLSSM) and the contrasting classification methods from the Salinas dataset; Figure 5 a- Figure 5 f is a schematic diagram of the overall accuracy (OA) and Kappa coefficient curves of the proposed classification method (SLSSM) and the contrastive classification method under different numbers of training samples. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] like Figure 1 The hyperspectral image classification method based on superpixel spectral similarity measurement shown includes the following steps: S1: Perform principal component analysis on the original hyperspectral image, and extract and retain the top three principal components with the largest variance contribution rates by calculating eigenvalues ​​and eigenvectors; The following methods were specifically adopted in S1: Three-dimensional data cube Remodeled into a two-dimensional matrix ,in Then, data standardization is performed:

[0027] in It is the mean of each band. That's the standard deviation. Next, we calculate the covariance of the standardized matrix:

[0028] The eigenvalues ​​and eigenvectors of the covariance matrix are calculated as follows:

[0029] in It is a diagonal matrix, and the elements on the diagonal are eigenvalues. , representing the variance of each principal component. Each column of V is the corresponding eigenvector, defining the transformation direction from the original band space to the principal component space. Finally, the eigenvalues ​​are sorted from largest to smallest, and their corresponding eigenvectors are the most important principal component directions. This process transforms the standardized original data... Projected onto the front l On the new subspace spanned by the eigenvectors, we obtain the new data after dimensionality reduction:

[0030] S2: Combine the three extracted principal components into a false-color image, and use the entropy rate superpixel segmentation algorithm to segment the image into superpixel units to generate a two-dimensional superpixel image; S2 specifically adopts the following method: First, construct a graph. Its apex V It is the set of edges for each pixel in the image. E It consists of similarity weights for each pair of pixels. ERS selects an edge subset. Divide the graph into several connected subsets such that the resulting graph... It consists of small connected components. The objective function of ERS incorporates the entropy rate. and balance function The optimal subset is found using the following objective function:

[0031] in These are the weights of the balance term. Entropy rate tends to form compact and uniform clusters, while the balance function aims to ensure that the generated clusters have similar sizes. Since ERS is primarily designed for RGB images with three bands, while hyperspectral images have tens to hundreds of bands, we adopt the processing method commonly used in superpixel-based hyperspectral image classification. Specifically, we select PCA to extract the first three principal components of the hyperspectral image, synthesize them into a false-color image, and then use the segmentation method described above to divide the hyperspectral image into... P Superpixels ,in It is the first p The first superpixel in the nth superpixel i One pixel, It is the first p The number of pixels in a superpixel.

[0032] S3: Use domain transform recursive filtering to remove spatial noise from the original hyperspectral image in the transform domain, while preserving edge information between ground features; The specific method used in S3 is as follows: For a given input signal I The data is then transformed to the transform domain, and the distance between pixels is calculated in the transform domain. The domain transformation process is represented as follows:

[0033] in Represents the domain transformation operator, and These represent the spatial and distance information of the bilateral filter, respectively. c This indicates the number of input signal channels. It is the input signal. The derivative of . For discrete signals. In the transform domain, the recursive edge-preserving filter is defined as follows:

[0034] in It is the filter output. It is a feedback coefficient. Neighbor samples in the transform domain and The distance between them. With d The increase, When the value approaches 0, the propagation chain stops, thus preserving the edges. Therefore, the edges are preserved, and pixels on the same side of the same edge usually have similar filtering effects.

[0035] S4: Perform maximum noise fraction transformation on the filtered hyperspectral image, and perform noise separation and dimensionality reduction in the spectral domain.

[0036] S4 specifically adopts the following method: Given a data , d It is the feature dimension. n This represents the number of samples. First, the data needs to be decentralized. ,in It is the mean vector of each feature. It is the first in the dataset One sample. Assume the observed signal. From signal and noise Composition, that is ,therefore The covariance matrix is ​​expressed as ,in , and These are the covariance matrices of Z, X, and N, respectively. It is estimated using the maximum / minimum autocorrelation factor method. The objective of the maximum noise fractional transform is to find a projection matrix W that maximizes the signal-to-noise ratio of the data in the transform domain. Its objective function is:

[0037] Solving using the Lagrange multiplier method, we know that W is derived from... of b A matrix consisting of the eigenvectors corresponding to the largest eigenvalues. b The dimension of the projected data is Y. After dimensionality reduction using MNF, the data Y is represented as:

[0038] S5: The two-dimensional superpixel segmentation map generated by the entropy rate superpixel segmentation algorithm is fused with the hyperspectral data processed by joint spatial-spectral filtering to construct a three-dimensional superpixel cube with homogeneous region characteristics. S6: Perform max-min normalization on the fused low-dimensional hyperspectral data, scaling the pixel values ​​of each band to the [0,1] interval to eliminate the influence of dimensions. For any original pixel in each band of the dimensionality-reduced data... x The following transformation formula is applied for calculation:

[0039] S7: For the test sample to be classified, calculate the spectral Euclidean distance between it and the training sample set of known categories within its superpixel unit. The specific method used in S7 is as follows: For a given test pixel, calculate its distance to all training samples, and take the minimum distance as its similarity measure with that class. The normalized hyperspectral data is... ,in N It is the total number of pixels. b This is the number of bands after dimensionality reduction. Given the th... q Training sample set of the class , It is the first q The first training sample j One pixel, It is the first q The number of training samples for each class. p The first superpixel in the nth superpixel i The test sample, compared with the first q The Euclidean distance between the training samples is calculated as follows:

[0040] Calculation and the first q The distance between each training sample in the class is obtained Take the nearest neighbor distance This serves as a distance metric to the class. The shortest distances between this sample and samples from other classes are calculated separately. ; S8: Calculate the distances between other test samples and each category in the superpixel sequentially to obtain the distance matrix of the superpixel. Each row represents the distance between each test pixel in the superpixel and a specific category.

[0041] S9: To effectively overcome potential spectral variations and outlier interference within superpixels, for each category, select from the distance matrix... k Find the smallest distance value and calculate its average, then use this average as the final distance value from the entire superpixel to that category.

[0042] In S9, the following method is used: Considering that test pixels within the same superpixel may exhibit spectral variations, not every pixel is representative. Therefore, the pixel closest to each type of sample is selected. k For each test sample, calculate its mean, that is, calculate the minimum mean for each row. k The mean of the values ​​is used to obtain a column vector. Let the matrix be... ,in and For the first i Okay, first find the first i The smallest in the row k Let the set contain a set of values. Then the vector The i element It can be represented as:

[0043] Finally obtained .

[0044] S10: Assign the category label with the smallest distance value (i.e., the highest similarity) to all pixels within that superpixel to complete the classification. Vector Each value represents the distance between the superpixel and each category of land cover; the smaller the distance, the higher the similarity. The category with the highest similarity is selected as the predicted label for all pixels in that superpixel. The classification criteria of the proposed method are as follows:

[0045] The hyperspectral image classification method based on superpixel spectral similarity measurement provided by this invention will be evaluated and assessed on three publicly available real hyperspectral image datasets.

[0046] The first set of experimental data is the Indian Pines real hyperspectral image dataset, which consists of images acquired by the AVIRIS sensor over the Indian Pines test site in northwestern India. The dataset has a spatial size of 145×145 pixels and a spatial resolution of 20 meters. It contains 200 spectral bands covering a spectral range of 0.4–2.5 μm (20 bands were discarded due to noise and water absorption), and includes 16 different land cover categories. The second set of experimental data is the Pavia University real hyperspectral image dataset, which consists of images collected by the ROSIS-03 sensor in the urban area surrounding the University of Pavia, Italy. This dataset has a spatial size of 610×340, a spatial resolution of 1.3 meters, and contains 103 spectral bands covering a spectral range of 0.43–0.86 μm. The dataset includes 9 land cover categories.

[0047] The third set of experimental data is the Salinas real hyperspectral image dataset, which consists of images collected by the AVRIS sensor in the Salinas Valley region of California, USA. The data has a spatial size of 512×217 and a spatial resolution of 3.7 meters, containing 204 spectral bands (20 water absorption bands were discarded), and includes 16 land cover classes.

[0048] This invention provides a hyperspectral image classification method based on superpixel spectral similarity measurement. The classification results are shown in the figure below, which represents the classification results of three sets of real hyperspectral image data. Figure 2 , Figure 3 and Figure 4 As shown, the classification results are quantitatively analyzed using overall classification accuracy. For the three datasets, the method provided in this invention achieved classification accuracies of 99.52%, 98.81%, and 98.77%, respectively, all of which are superior to other comparative methods. Figure 5(a)-(f) show the OA and Kappa values ​​of the proposed method and the comparison method under different numbers of training samples in three datasets.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hyperspectral image classification method based on superpixel spectral similarity measurement, characterized in that... include: Principal component analysis was performed on the original hyperspectral image. By calculating eigenvalues ​​and eigenvectors, the top three principal components with the largest variance contribution rates were extracted and retained. The three extracted principal components are combined to form a false-color image. The entropy rate superpixel segmentation algorithm is used to segment the image into superpixel units and generate a two-dimensional superpixel segmentation map. Based on the domain transform recursive filtering method, spatial noise in the standardized hyperspectral image is filtered out in the transform domain while preserving the edge information between ground features; The filtered hyperspectral image is subjected to maximum noise fraction transformation, and noise separation and dimensionality reduction are performed in the spectral domain to obtain preprocessed hyperspectral data. The two-dimensional superpixel segmentation map is fused with the preprocessed hyperspectral data to construct a three-dimensional superpixel cube data with homogeneous region characteristics; The three-dimensional superpixel cube data is subjected to max-min normalization processing, and the pixel values ​​of each band are scaled to the [0,1] interval to obtain normalized hyperspectral data; For a test sample to be classified, within its superpixel unit, calculate the spectral Euclidean distance between it and the training sample set of known categories. For a test pixel, calculate the distance from the test pixel to all training samples, and take the minimum distance as the similarity measure between the test pixel and the category. Traverse all pixels to be classified within a superpixel in the normalized hyperspectral data, calculate the spectral Euclidean distance between each pixel and the training sample sets of each class, select the minimum distance value from the Euclidean distances with the training sample sets of each class as the similarity measure with a certain class, thereby obtaining the similarity between each pixel to be classified and the samples of each class, and obtaining the distance matrix of the superpixel. Select from the distance matrix for each class of samples k The distance values ​​of the most similar test samples are calculated and their average is used as the final distance value from the entire superpixel to the category. The samples of each category are traversed and the distance matrix is ​​transformed into a distance vector. The category label with the smallest distance value is assigned to all pixels within that superpixel, thus completing the hyperspectral image classification process.

2. The hyperspectral image classification method based on superpixel spectral similarity measurement according to claim 1, characterized in that: [The following is a partial translation of the original text, which is not possible without further context.] Normalized hyperspectral data are ,in N It is the total number of pixels. b It is the number of bands after dimensionality reduction, set as the first... q Training sample set of the class , It is the first q The first training sample j One pixel, It is the first q The number of training samples for class 1 p The first superpixel in the nth superpixel i The test sample, compared with the first q The Euclidean distance between the training samples is calculated as follows: Calculation and the first q The distance between each training sample in the class is obtained Take the minimum distance value As a distance metric to this category, the minimum distance between this sample and samples from other categories is calculated. .

3. The hyperspectral image classification method based on superpixel spectral similarity measurement according to claim 1, characterized in that: Select the sample most similar to each type of sample. k For each test sample, calculate its mean and then calculate the minimum value in each row. k The mean of the values ​​is used to obtain a column vector. ; Let matrix ,in and For the first i Okay, find the first one. i The smallest in the row k Let the set contain a set of values. Then the vector The i element Represented as: Finally, the distance vector is obtained. ; Distance vector Each value represents the distance between the superpixel and each category of land cover. The smaller the distance, the higher the similarity. The category with the highest similarity is selected as the predicted label for all pixels in the superpixel. The category label with the smallest distance value is assigned to all pixels within the superpixel. The classification criterion is as follows: 。 4. The hyperspectral image classification method based on superpixel spectral similarity measurement according to claim 1, characterized in that: When using domain transform-based recursive filtering to remove spatial noise from the original hyperspectral image in the transform domain: for a given input signal I The data is then transformed to the transform domain, and the distance between pixels is calculated in the transform domain. The domain transformation process is represented as follows: in Represents the domain transformation operator, and These represent the spatial and distance information of the bilateral filter, respectively. c Indicates the number of channels for the input signal. It is the derivative of the input signal... for discrete signals. Define a recursive edge-preserving filter in the transform domain. for: in It is the filter output. For feedback coefficients, Neighbor samples in the transform domain and The distance between them, as d The increase, When the value approaches 0, the propagation chain stops, thus preserving the edge.

Images