Hyperspectral image spectral spatial feature extraction method based on improved singular spectrum analysis

Abstract

The invention relates to a hyperspectral image spectral spatial feature extraction method based on improved singular spectrum analysis, and relates to the technical field of hyperspectral image classification. According to the method, firstly, dimension reduction is carried out on a hyperspectral image through a principal component analysis method, a first principal component feature map is obtained, superpixel segmentation is carried out on the first principal component feature map, and region division is carried out on the image; through designing an improved window shape adaptive singular spectrum analysis algorithm, spectral space joint features are extracted from irregular shape superpixels formed after segmentation. The problems of low operation speed and the like caused by fuzzy boundary information of a ground feature region and large calculation amount due to an existing processing method of filling irregular-shaped super pixels into regular shapes are solved. And the extracted spectral space joint feature map is classified by combining an SVM algorithm, so that the calculated amount is reduced while high-precision classification is achieved, and feature extraction and classification can be quickly carried out.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral image classification, and particularly relates to a method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis. Background Art

[0002] Hyperspectral remote sensing is a powerful remote sensing technology for observing surface features. The successful application fields of hyperspectral remote sensing include monitoring natural ecosystems such as agricultural areas, forests, urban areas, ice and snow, the atmosphere, inland waters, and the ocean. Hyperspectral image classification is to divide the pixels of hyperspectral images into surface coverage areas such as forests, cities, bare soil, rivers, etc., which has important research significance in aspects such as earth observation and has been widely applied in military and civilian fields. However, the continuous band reflectances are often highly correlated, resulting in redundant and overlapping spectral information. At the same time, the large number of bands often leads to an excessive amount of spectral space data and feature numbers, increasing the computational complexity. Existing hyperspectral datasets have the Hughes phenomenon with the increase in data dimensions, greatly increasing the processing difficulty.

[0003] Currently, for hyperspectral classification tasks, many methods have been proposed by scholars, especially methods for simultaneously extracting spectral-spatial features for hyperspectral image classification. The singular spectrum analysis algorithm for time series prediction has also been introduced into the field of hyperspectral image feature extraction. Research shows that singular spectrum analysis exhibits good capabilities in denoising and extracting spectral features, and thus has been favored by many researchers. For example, the literature "Jaime Z, Jin chang R, Jiangbin Z, Junwei H, Huimin Z, Shutao L, Stephen M, et al. Novel Two-Dimensional Singular Spectrum Analysis for Effective Feature Extraction and Data Classification in Hyperspectral Imaging[J], IEEE Transactions on Geo science and Remote Sensing, 2015, 53(8): 4418-4433." proposed two-dimensional singular spectrum analysis (2DSSA) based on singular spectrum analysis. It traverses each pixel with a window, constructs a one-dimensional sequence from the pixels within the window, then uses the sequence to construct a lag matrix to extract the spatial structure features of the hyperspectral image, but ignores the spectral features. The literature "A Novel Spectral-Spatial Singular Spectrum Analysis Technique for Near Real-Time In Situ Feature Extraction in Hyperspectral Imaging[J], IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 2214-2225." proposed 1.5DSSA based on singular spectrum analysis, introduced the correlation of pixel vectors within the window to extract the spectral-spatial features of the hyperspectral image, and enhanced the feature extraction effect of singular spectrum analysis to a certain extent, but the effect depends largely on the window size.The literature "Fusion of PCA and Segmented-PCADomain Multiscale 2-D-SS A for Effective Spectral-Spatial Feature Extraction and Data Classification in Hyperspectral Imagery[J], IEEE Transactions on Geoscience and Remote Sensing, 2022, 60:1-14." proposed to extract spatial features through multiscale 2DSSA, and adopted the method of fusing principal component analysis and segmented principal component analysis to extract spectral features, achieving good results. However, this method extracts spectral features and spatial features separately, and the calculation method is relatively complex, increasing the computational complexity. The literature "SpaSSA: Superpixelwise Adaptive SSA for Unsupervised Spatial–Spectral Feature Extraction in Hyperspectral Image[J], IEEE Transactions on Cybernetics, 2022, 52(7):6158-6169." proposed the SpaSSA algorithm, which processes large and small superpixels by setting one-dimensional singular spectrum analysis and two-dimensional singular spectrum analysis for large and small superpixels respectively, greatly improving the feature extraction ability. However, the superpixel shape is irregular. This algorithm expands the irregularly shaped superpixels into regular rectangles, ignoring the boundary information. At the same time, the operation of traversing each pixel in the window has a large computational complexity and takes a long time. Summary of the Invention

[0004] The present invention aims to solve the technical problems in the prior art and provides a method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis. The present invention can extract spectral-spatial joint features containing more comprehensive information more quickly and has outstanding performance in hyperspectral image classification tasks.

[0005] To solve the above technical problems, the technical solution of the present invention is specifically as follows:

[0006] The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis of the present invention includes the following steps:

[0007] Step 1, obtain a hyperspectral image dataset;

[0008] Step 2, use superpixel segmentation technology to divide the spatial region of the hyperspectral image according to spatial information;

[0009] Step 3: Use the improved window shape adaptive singular spectrum analysis algorithm to extract the spectral-spatial features of the hyperspectral image according to the superpixel labels;

[0010] Step 4: Use the SVM algorithm with five-fold cross-validation to classify the spectral-spatial features of the extracted hyperspectral image, and verify the effectiveness of the feature extraction method through the classification results.

[0011] In the above technical solution, the hyperspectral image dataset is the Pavia University dataset.

[0012] In the above technical solution, the specific steps of Step 1 are as follows:

[0013] Use the entropy rate segmentation (ERS) algorithm to preprocess the obtained hyperspectral image, specifically, reduce the dimension of the hyperspectral image using the PCA algorithm to obtain the first principal component image.

[0014] In the above technical solution, the specific steps of Step 2 are as follows:

[0015] Use the superpixel segmentation technology to divide and segment the spatial regions of the first principal component image obtained in Step 1 according to the spatial information, and assign different labels to different superpixel regions.

[0016] In the above technical solution, the specific steps of Step 3 are as follows:

[0017] Use the window shape adaptive SSA algorithm to identify the superpixel labels and record their position coordinates.

[0018] In the above technical solution, further, Step 3 includes the following steps:

[0019] First, rearrange the spectral vectors within the superpixel block into a one-dimensional vector according to the recorded position coordinates, construct a lag matrix using the one-dimensional vector, perform singular value decomposition on the constructed lag matrix, group and reconstruct the decomposed matrix to obtain a reconstructed vector with the same dimension as the original one-dimensional vector; rearrange the reconstructed vector according to the recorded position coordinates to reconstruct a superpixel block with the same size and shape as the pre-processed superpixel; the processed hyperspectral image is the feature map containing spectral-spatial information extracted.

[0020] In the above technical solution, furthermore, Step 3 includes the following steps:

[0021] Define a superpixel block as S, which is composed of N pixel vectors, denoted as:

[0022] S = {(x 1 ,y 1 ,bands),(x 2 ,y2 , bands), …, (x N , y N , bands)}

[0023] where (x i , y i ) is the position coordinate of the pixel in the superpixel block. The pixels in the superpixel block are arranged in ascending order of coordinate values to form a one-dimensional spectral vector, denoted as X = [1, 2, …, 200, 201, …, N×B], where B is the HSI spectral dimension; this one-dimensional spectral vector contains the local spatial information and spectral information of HSI; define the one-dimensional signal X = [x 1 , x 2 , …, x N ∈ R N . Select an appropriate window length L (1 < L < N) to perform a lag arrangement on the original one-dimensional signal to obtain a trajectory matrix. The window size is equal to the number of extracted components, and the matrix X is the trajectory matrix of the one-dimensional signal;

[0024]

[0025] where K = N - L + 1, and the column vector c i of the trajectory matrix is the lag vector; perform a singular value decomposition on the trajectory matrix X, let S = XX T , λ 1 , λ 2 , …, λ L be the eigenvalues of S, and λ 1 ≥ λ 2 ≥ … ≥ λ L ≥ 0; and U 1 , U 2 , …, U L are the orthonormal vectors of the matrix S corresponding to these eigenvalues; let In this case, the SVD of the trajectory matrix X is written as: X = X 1 + … + X d ; where is called an elementary matrix of rank 1, and and are called the empirical orthogonal function and the component of the trajectory matrix; the matrices constructed by U i and V i are as follows:

[0026] U = (U 1 , U 2 … U L ) ∈ R L×L

[0027] V = (V 1 , V 2 … V L ) ∈ R L×L

[0028] represents the contribution of the i-th eigenvalue to the matrix X;

[0029] To separate the target signal component from other signal components, the subscript set {1, …, d} is partitioned into m non-overlapping subsets I 1 , …, I m , and let I = {i 1 , i 2 , …, i P}, then the synthesis matrix corresponding to I After combination, the trajectory matrix X becomes: Each matrix among them is transformed into a new sequence of length N, that is, the decomposed sequence is obtained; let Y be an L×K matrix with elements yi j , 1 ≤ i ≤ L, 1 ≤ j ≤ K; let L * = min(L, K), K * = max(L, K), N = L + K - 1; if L < K, then Otherwise The average of the diagonal of Y is taken and transformed into the sequence y 1 , y 2 , …, y N , the spectral vectors are separated from this sequence according to the number of pixels and spectral dimensions contained in the superpixel block before reconstruction, and the separated spectral vectors are formed into new superpixel blocks with the same size as the superpixel before feature extraction according to the recorded position coordinates. After the small regions divided by all superpixel segmentation techniques are decomposed and reconstructed by the WSA-SSA algorithm, they are formed into an HSI local spatial-spectral feature image according to the label positions of the superpixels. This feature image contains HSI spectral information and local spatial information.

[0030] In the above technical solution, in step 4, four indicators, namely OA, AA, Kappa coefficient, and running time, are used to evaluate the classification effect.

[0031] The beneficial effects of the present invention are:

[0032] The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis provided by the present invention first reduces the dimension of the hyperspectral image through the principal component analysis method to obtain the first principal component feature map, performs superpixel segmentation on the first principal component feature map, and divides the image into regions. By designing an improved window shape adaptive singular spectrum analysis algorithm to extract spectral-spatial joint features from the irregularly shaped superpixels formed after segmentation, it avoids problems such as the blurring of ground object region boundary information and the slow running speed caused by large computational amounts in the existing processing method of filling irregularly shaped superpixels into regular shapes. Combining with the SVM algorithm, the extracted spectral-spatial joint feature map is classified, achieving high-precision classification while reducing the computational amount, and enabling rapid feature extraction and classification. Description of the Drawings

[0033] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0034] Figure 1 It is the hyperspectral image of the present invention and its first principal component and the image after superpixel segmentation;

[0035] Figure 2 It is the flow chart of the method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis of the present invention;

[0036] Figure 3 It is the data chart of the classification index for hyperspectral images of the present invention;

[0037] Figure 4 It is the classification effect diagram for hyperspectral images of the present invention. Specific Embodiments

[0038] The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis provided by the present invention includes the following steps:

[0039] Step 1: Obtain a hyperspectral image dataset;

[0040] Preprocess the obtained hyperspectral image using the entropy rate segmentation (ERS) algorithm. Specifically, reduce the dimension of the hyperspectral image using the PCA algorithm to obtain the first principal component image.

[0041] In this step, the hyperspectral image dataset is the Pavia University dataset; this dataset is obtained by the ROSIS sensor and is often used for hyperspectral image classification. The sensor has a total of 115 bands, and after processing, there are 103 bands; the size is 610*340. This dataset contains 9 ground object categories.

[0042] Step 2: Use superpixel segmentation technology to divide the hyperspectral image into spatial regions according to spatial information;

[0043] Using the superpixel segmentation technology, the first principal component image obtained in step 1 is divided into spatial regions according to spatial information, and different labels are assigned to different superpixel regions.

[0044] Step 3: Use the improved window shape adaptive singular spectrum analysis algorithm to extract the spectral-spatial features of the hyperspectral image according to the superpixel labels;

[0045] Specifically, the window shape adaptive SSA algorithm is used to identify the superpixel labels and record their position coordinates.

[0046] The superpixel labels and the recorded label positions are given to the spatial dimension of the hyperspectral image, and the window shape adaptive singular spectrum analysis performs feature extraction on the hyperspectral image with superpixel label and label position information. First, the spectral vectors within the superpixel block are rearranged into a one-dimensional vector according to the recorded position coordinates, a lag matrix is constructed using the one-dimensional vector, the constructed lag matrix is subjected to singular value decomposition, the decomposed matrix is grouped and reconstructed, and a reconstructed vector with the same dimension as the original one-dimensional vector is obtained. The reconstructed vector is rearranged according to the recorded position coordinates to reconstruct a superpixel block with the same size and shape as the superpixel before processing. The processed hyperspectral image is the feature map containing spectral-spatial information extracted.

[0047] More specifically, the implementation steps of step 3 are as follows:

[0048] Define a superpixel block as S, which is composed of N pixel vectors, denoted as:

[0049] S = {(x 1 , y 1 , bands), (x 2 , y 2 , bands), …, (x N , y N , bands)}

[0050] where (x i , y i ) are the position coordinates of the pixel in the superpixel block, the pixels in the superpixel block are arranged into a one-dimensional spectral vector in ascending order of coordinate values, denoted as X = [1, 2, ..., 200, 201, …, N×B], where B is the HSI spectral dimension; the one-dimensional spectral vector contains the local spatial information and spectral information of the HSI; define the one-dimensional signal X = [x 1 , x 2 , …, x N ∈ R N, select an appropriate window length \(L(1 < L < N)\) to perform a lag arrangement on the original one-dimensional signal to obtain a trajectory matrix. The window size is equal to the number of extracted components, and the matrix \(X\) is the trajectory matrix of the one-dimensional signal;

[0051]

[0052] where \(K = N - L + 1\), and the column vector \(c\) of the trajectory matrix i is the lag vector; perform a singular value decomposition on the trajectory matrix \(X\), let \(S = XX\) T , \(\lambda\) 1 , \(\lambda\) 2 , …, \(\lambda\) L be the eigenvalues of \(S\), and \(\lambda\) 1 \(\geq \lambda\) 2 \(\geq \cdots \geq \lambda\) L \(\geq 0\); and \(U\) 1 , \(U\) 2 , …, \(U\) L are the orthonormal vectors of the matrix \(S\) corresponding to these eigenvalues; let In this case, the SVD of the trajectory matrix \(X\) is written as: \(X = X\) 1 + \(\cdots\) + \(X\) d ; where is called an elementary matrix of rank 1, and are called the empirical orthogonal functions and the components of the trajectory matrix; the matrices constructed by \(U\) i and \(V\) i are as follows:

[0053] \(U=(U\) 1 , \(U\) 2 \(\cdots U\) L ) \(\in R\) L×L

[0054] \(V=(V\) 1 , \(V\) 2 \(\cdots V\) L ) \(\in R\) L×L

[0055] represents the contribution of the \(i\)-th eigenvalue to the matrix \(X\);

[0056] To separate the target signal components from other signal components, divide the subscript set \(\{1, \cdots, d\}\) into \(m\) non-overlapping subsets \(I\) 1 , \(\cdots\), \(I\) m , let \(I = \{i\) 1 , \(i\) 2 , \(\cdots\), \(i\) P \}, then the synthesis matrix corresponding to \(I\) After combination, the trajectory matrix \(X\) becomes: Each matrix in it Transform it into a new sequence of length N, that is, obtain the decomposed sequence; let Y be an L×K matrix with elements y ij , where 1 ≤ i ≤ L and 1 ≤ j ≤ K; let L * = min(L, K), K * = max(L, K), N = L + K - 1; if L < K, then Otherwise Perform diagonal averaging on Y and convert it into the sequence y 1 , y 2 , …, y N , separate the spectral vectors from this sequence according to the number of pixels and spectral dimensions contained in the superpixel block before reconstruction, and form new superpixel blocks with the same size as the superpixel before feature extraction according to the recorded position coordinates of the separated spectral vectors. After the small regions divided by all superpixel segmentation techniques are decomposed and reconstructed by the WSA-SSA algorithm, they form the HSI local spatial-spectral feature image according to the label positions of the superpixels. This feature image contains HSI spectral information and local spatial information.

[0057] Step 4: Use the SVM algorithm with five-fold cross-validation to classify the spectral-spatial features of the extracted hyperspectral image, and verify the effectiveness of the feature extraction method through the classification results.

[0058] In the said step 4, the classification effect is evaluated by four indicators: OA, AA, Kappa coefficient, and running time.

[0059] The present invention will be described in more detail below with reference to the accompanying drawings and embodiments.

[0060] The following are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

[0061] The terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0062] Combined with Figure 2 the flowchart of, the method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis of the present invention includes the following steps:

[0063] Step 1: Obtain a hyperspectral image dataset

[0064] Preprocess the hyperspectral image using the Entropy Rate Segmentation (ERS) algorithm. As Figure 1 Reduce the dimensionality of the hyperspectral image using the PCA algorithm to obtain the first principal component image;

[0065] In this step, the hyperspectral image dataset is the Pavia University dataset; this dataset was acquired by the ROSIS sensor and is often used for hyperspectral image classification. The sensor has a total of 115 bands, and after processing, there are 103 bands; the size is 610 * 340. This dataset contains 9 land cover classes.

[0066] Step 2: Use the superpixel segmentation technique to perform spatial region division and segmentation on the first principal component image obtained in Step 1 according to spatial information, and assign different labels to different superpixel regions.

[0067] Step 3: Use the window shape adaptive SSA algorithm to identify the superpixel labels and record their position coordinates. Define a superpixel block as S, which is composed of N pixel vectors, denoted as:

[0068] S = {(x 1 , y 1 , bands), (x 2 , y 2 , bands),..., (x N , y N , bands)};

[0069] Where (x i , y i ) are the position coordinates of the pixel in the superpixel block, and bands represents the spectral dimension. Arrange the pixels in the superpixel block in ascending order of coordinate values into a one-dimensional spectral vector, denoted as X = [1, 2, b, ands, ba + nd1s, ×], where B is the HSI spectral dimension. This one-dimensional vector contains the local spatial information and spectral information of the HSI. Define X = [1, 2,..., bands, bands + 1,..., N × B] as the one-dimensional signal X = [x 1 , x 2 , …, x N ∈ R N , and select an appropriate window length L (1 < L < N) to perform a lag arrangement on the original one-dimensional signal to obtain a trajectory matrix, and the window size is equal to the number of extracted components. Matrix X is the trajectory matrix of the one-dimensional signal.

[0070]

[0071] Where K = N - L + 1, and the column vector c i of the trajectory matrix is the lag vector. Perform singular value decomposition on the trajectory matrix X. Let S = XXT , λ 1 , λ 2 ,..., λ L are the S eigenvalues, and λ 1 ≥ λ 2 ≥ … ≥ λ L ≥ 0; and U 1 , U 2 ,..., U L are the orthonormal vectors of matrix S corresponding to these eigenvalues. Let (note that in the actual sequence, usually d = L * , L * = min{L, K}), In this case, the SVD of the trajectory matrix X can be written as: X = X 1 +... + X d . Where is called an elementary matrix of rank 1, and are called the empirical orthogonal function and the components of the trajectory matrix. The matrices constructed by U i and V i are as follows:

[0072] U = (U 1 , U 2 ... U L ) ∈ R L×L

[0073] V = (V 1 , V 2 … V L ) ∈ R L×L

[0074] represents the contribution of the i-th eigenvalue to matrix X.

[0075] To separate the target signal components from other signal components, the subscript set {1, …, d} is partitioned into m non-overlapping subsets I 1 , …, I m , let I = {i 1 , i 2 , …, i P}, then after combining the synthesis matrix corresponding to I, the trajectory matrix X becomes: After transforming each matrix in it into a new sequence of length N, the decomposed sequence is obtained. Let Y be an L×K matrix with elements y ij , 1 ≤ i ≤ L, 1 ≤ j ≤ K. Let L * = min(L, K), K * = max(L, K), N = L + K - 1. If L < K, then Otherwise Perform diagonal averaging on Y and convert it into a sequence y 1 , y 2 , …, y N , separate the spectral vectors from the sequence according to the number of pixels and spectral dimensions contained in the superpixel block before reconstruction, and form new superpixel blocks with the same size as the superpixel before feature extraction according to the recorded position coordinates of the separated spectral vectors. After the small regions divided by all superpixel segmentation techniques are decomposed and reconstructed by the WSA-SSA algorithm, they form an HSI local spatial-spectral feature image according to the label positions of the superpixels. This feature image contains HSI spectral information and local spatial information.

[0076] Step 4: Use the SVM algorithm with five-fold cross-validation to verify the effectiveness of the feature extraction method by the classification effect, and evaluate the classification effect through four indicators: OA, AA, Kappa coefficient, and running time. Through Figure 3 As can be seen from the table of Figure 4 , it can be seen from the three indicators of OA, AA, and Kappa coefficient that the present invention is superior to that before improvement. The present invention has a small computational amount and a fast classification speed, far superior to the method before improvement. From

[0077] Obviously, the above embodiments are only examples for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis, characterized in that, it includes the following steps: Step 1, obtain a hyperspectral image dataset; Step 2, use the superpixel segmentation technology to divide the hyperspectral image into spatial regions according to spatial information; Step 3, use the improved window shape adaptive singular spectrum analysis algorithm to extract the spectral-spatial features of the hyperspectral image according to the superpixel labels; Step 4, use the SVM algorithm with five-fold cross-validation to classify the spectral-spatial features of the extracted hyperspectral image, and verify the effectiveness of the feature extraction method through the classification results.

2. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 1, characterized in that, the hyperspectral image dataset in Step 1 is the Pavia University dataset.

3. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 1, characterized in that, Step 1 includes the following steps: Use the entropy rate segmentation (ERS) algorithm to preprocess the obtained hyperspectral image, specifically, reduce the dimension of the hyperspectral image using the PCA algorithm to obtain the first principal component image.

4. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 3, characterized in that, Step 2 includes the following steps: Use the superpixel segmentation technology to divide the first principal component image obtained in Step 1 into spatial regions according to spatial information, and assign different labels to different superpixel regions.

5. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 4, characterized in that, Step 3 includes the following steps: Use the window shape adaptive SSA algorithm to identify the superpixel labels and record their position coordinates.

6. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 5, characterized in that, Step 3 includes the following steps: First, rearrange the spectral vectors within the superpixel block into a one-dimensional vector according to the recorded position coordinates, use the one-dimensional vector to construct a lag matrix, perform singular value decomposition on the constructed lag matrix, group and reconstruct the decomposed matrix to obtain a reconstructed vector with the same dimension as the original one-dimensional vector; rearrange the reconstructed vector according to the recorded position coordinates to reconstruct a superpixel block with the same size and shape as the pre-processed superpixel; the processed hyperspectral image is the feature map containing spectral-spatial information extracted.

7. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 6, characterized in that, Step 3 specifically includes the following steps: Define a superpixel block as S, which is composed of N pixel vectors, denoted as: S = {(x 1 , y 1 , bands), (x 2 , y 2 , bands),..., (x N , y N , bands)} where (x i , y i ) is the position coordinate of the pixel in the superpixel block. The pixels in the superpixel block are arranged in ascending order of coordinate values into a one-dimensional spectral vector, denoted as X = [1, 2, ..., 200, 201, ..., N×B], where B is the HSI spectral dimension; this one-dimensional spectral vector contains the local spatial information and spectral information of HSI; define the one-dimensional signal X = [x 1 , x 2 , …, x N ∈ R N . Select the window length L to perform a lag arrangement on the original one-dimensional signal to obtain a trajectory matrix, where 1 < L < N, the window size is equal to the number of extracted components, and the matrix X is the trajectory matrix of the one-dimensional signal; where \(K = N - L+1\), and the column vector \(\mathbf{c}\) of the trajectory matrix i is the lag vector; perform singular value decomposition on the trajectory matrix \(\mathbf{X}\), let \(\mathbf{S}=\mathbf{X}\mathbf{X}^T\) T , \(\lambda\) 1 , \(\lambda\) 2 , \(\cdots\), \(\lambda\) L are the eigenvalues of \(\mathbf{S}\), and \(\lambda\) 1 \(\geq\lambda\) 2 \(\geq\cdots\geq\lambda\) L \(\geq0\); and \(\mathbf{U}\) 1 , \(\mathbf{U}\) 2 , \(\cdots\), \(\mathbf{U}\) L are the orthonormal vectors of matrix \(\mathbf{S}\) corresponding to these eigenvalues; let In this case, the SVD of the trajectory matrix \(\mathbf{X}\) is written as: \(\mathbf{X}=\mathbf{X}_1+\cdots+\mathbf{X}_r\) 1 +\cdots+\mathbf{X}_r d ; where is called the elementary matrix of rank 1, and are called the empirical orthogonal function and the component of the trajectory matrix; the matrices constructed by \(\mathbf{U}\) i and \(\mathbf{V}\) i are as follows: U = (U 1 , U 2 … U L ) ∈ R L×L V = (V 1 , V 2 … V L ) ∈ R L×L represents the contribution of the i-th eigenvalue to the matrix X; To separate the target signal component from other signal components, the subscript set {1, …, d} is partitioned into m non - overlapping subsets I 1 , …, I m , let I = {i 1 , i 2 , …, i P}, then the synthesis matrix corresponding to I After combination, the trajectory matrix X becomes: Each of the matrices is transformed into a new sequence of length N, that is, the decomposed sequence is obtained; let Y be an L×K matrix with elements y ij , 1 ≤ i ≤ L, 1 ≤ j ≤ K; let L * = min(L, K), K * = max(L, K), N = L + K - 1; If L < K, then Otherwise Perform diagonal averaging on Y and convert it into a sequence y 1 , y 2 , …, y N , separate the spectral vectors from the sequence according to the number of pixels and spectral dimensions contained in the superpixel block before reconstruction, and form new superpixel blocks with the same size as the superpixel before feature extraction according to the recorded position coordinates of the separated spectral vectors. After decomposition and reconstruction of the small regions divided by all superpixel segmentation techniques using the WSA-SSA algorithm, an HSI local spatial-spectral feature image is formed according to the label positions of the superpixels. This feature image contains HSI spectral information and local spatial information.

8. The method for extracting spectral-spatial features of hyperspectral images based on improved singular spectrum analysis according to claim 1, characterized in that, in Step 4, the classification effect is evaluated by four indicators: OA, AA, Kappa coefficient, and running time.

Images