Hyperspectral remote sensing image space-spectrum joint feature extraction method

Abstract

The invention relates to the field of remote sensing and computer vision, in particular to a hyperspectral remote sensing image space-spectrum joint feature extraction method, which comprises the following steps of: acquiring a structural similarity index consisting of brightness, contrast and structure in a hyperspectral remote sensing image, representing the similarity between a domain pixel and a central pixel based on the structural similarity index, and extracting the spatial-spectral feature of the hyperspectral remote sensing image; selecting a domain pixel set with the highest similarity, and arranging the center pixels and the domain pixels to generate a one-dimensional extension vector; for an extended vector, selecting a value in a vector range corresponding to a central pixel by using a singular spectrum analysis method, converting an original hypercube into a hyperspectral cube with SISSA pixels, and retaining main spectral information; a two-dimensional principal component analysis framework guided by Gaussian filtering is constructed based on spectral information of a hyperspectral cube, and spectrum-space joint features are further effectively extracted.

Description

Technical Field

[0001] The present invention relates to the fields of remote sensing and computer vision, and in particular to a method for extracting spatial-spectral joint features of hyperspectral remote sensing images. Background Art

[0002] With the rapid development of remote sensing and computer vision, hyperspectral remote sensing imagery is increasingly being applied in diverse research fields, providing strong support for research in Earth observation, biomedicine, military early warning, environmental monitoring, agricultural management, and more. The large number of bands in hyperspectral remote sensing images triggers the Hughes phenomenon (also known as the curse of dimensionality), significantly reducing the performance of classification and other detection techniques. Furthermore, the spectral feature values ​​of hyperspectral remote sensing images contain a large amount of redundant information, which can be partially or fully predicted by other spectral bands, increasing the time and cost of data processing and storage. Feature extraction from hyperspectral remote sensing images primarily mines the implicit patterns, structures, and intrinsic relationships within hyperspectral data, and uses mathematical transformations or complex filtering techniques to convert the raw high-dimensional data into more compact, easier-to-process, low-dimensional data. The key lies in minimizing data complexity while maximizing the preservation of key features and discriminative power.

[0003] Traditional feature extraction methods include principal component analysis, linear discriminant analysis, local information feature extraction algorithms, and nonlinear feature mining methods. Local information feature extraction algorithms break through the traditional global Gaussian distribution assumption and focus on maintaining local neighborhood relationships. Nonlinear feature mining methods enhance the feature extraction capabilities of hyperspectral data in complex scenarios through nonlinear transformations, providing a more stable and reliable classification feature basis. However, traditional hyperspectral image feature extraction methods focus on mining spectral features while neglecting the utilization of the spatial contextual information of the hyperspectral cube. Furthermore, these methods treat individual pixels as independent samples, ignoring the spatial dependencies between adjacent pixels. This makes it difficult for the extracted features to fully reflect the spatial distribution patterns of ground objects. As a three-dimensional data structure, the spatial dimension of the hyperspectral cube not only carries geometric distribution information of the pixel neighborhood but also implicitly implies the spatial contextual associations of ground objects. In fact, spatial information includes geometric features such as texture, edges, and morphology of ground objects, which can provide key spatial structural constraints for ground object classification. Joint spatial-spectral feature extraction is gaining increasing attention. Researchers have used methods such as local binary patterns, Gabor wavelet transform, and two-dimensional singular spectrum analysis to extract more spatial texture features. In general, spectral variation is a key factor restricting the accuracy of hyperspectral classification, which becomes more severe as the spatial resolution increases. The detailed information provided by high-resolution images may lead to increased differences in samples within a category and reduced discrimination between categories, thus affecting the classification effect.

[0004] In summary, while traditional methods can effectively mine statistical correlations between bands and achieve dimensionality compression, they generally suffer from insufficient utilization of spatial information. Furthermore, many methods focus solely on independent analysis in the spectral domain, failing to fully consider the geometric constraints between spatially neighboring pixels and the semantic associations of scene context, resulting in feature expression capabilities being limited to the level of local spectral similarity. Therefore, this patent designs a joint and efficient extraction strategy for spectral and spatial features to achieve accurate classification of hyperspectral remote sensing images. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for extracting spatial-spectral features based on hyperspectral remote sensing images to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for extracting spatial-spectral joint features of hyperspectral remote sensing images, comprising:

[0008] S100, obtaining a structural similarity index composed of brightness, contrast, and structure in the hyperspectral remote sensing image, characterizing the similarity between the domain pixels and the center pixel based on the structural similarity index, selecting a domain pixel set with the highest similarity, arranging the center pixel and the domain pixels, and generating a one-dimensional expansion vector;

[0009] S200, for the extended vector, using a singular spectrum analysis method to select a value within the vector range corresponding to the central pixel, converting the original hypercube into a hyperspectral cube with SISSA pixels, and retaining the main spectral information;

[0010] S300, based on the spectral information of the hyperspectral cube, constructs a Gaussian filtering-guided two-dimensional principal component analysis framework to further effectively extract spectral-spatial joint features.

[0011] Preferably, the structural similarity index in S100 is represented by , including three parts: brightness, contrast and structural information;

[0012] The brightness refers to the brightness value of pixels in a hyperspectral remote sensing image; the contrast refers to the degree of change in the brightness value of pixels in a hyperspectral remote sensing image; the structural information refers to the shape and texture characteristics of objects in a hyperspectral remote sensing image. From the perspective of image synthesis, structural information is defined as an attribute independent of brightness and contrast, and reflects the structure of objects displayed in the scene;

[0013] The three components of brightness, contrast and structure information are combined to form the SSIM system, and the SSIM index is determined:

[0014] ;

[0015] in, 、 represents the hyperspectral remote sensing image signal, 、 is the average intensity, indicating the image signal 、 brightness, 、 is the standard deviation, indicating the image signal The contrast ratio, Indicates image signal and The covariance between

[0016] in, , , , , , .

[0017] Preferably, in S100, based on the structural similarity index, the similarity between the domain pixels and the center pixel is characterized, a domain pixel set with the highest similarity is selected, the center pixel and the domain pixels are arranged, and a one-dimensional expansion vector is generated, including:

[0018] S101. For each pixel, select a small neighborhood window , select the area pixel set with the highest similarity to the central pixel to construct the expansion vector;

[0019] S102: Select a subset of pixels with the highest similarity , and sort them according to their SSIM values, and arrange the central pixel and the selected pixels into a one-dimensional extended vector according to the SSIM value. The pixels with higher similarity are closer to the central pixel.

[0020] Preferably, selecting a value within a vector range corresponding to the central pixel using a singular spectrum analysis method in S200 includes:

[0021] S201. For a given one-dimensional signal and an embedding window of size L, the original signal can be mapped into a series of lag vectors, namely the trajectory matrix ;

[0022] S202, will The eigenvalues ​​and their corresponding eigenvectors are denoted as and ; Then the trajectory matrix is ​​reconstructed as the sum of the following basic matrices:

[0023] ;

[0024] ;

[0025] in, and denote the empirical orthogonal function and the trajectory matrix The principal components of the basic matrix;

[0026] After eigenvalue grouping, select A subset of An approximation of , where the subset contains the main trend component and eliminates the noise component.

[0027] Preferably, S300 includes:

[0028] S301, for each spectral channel of the hyperspectral cube , establish a spatial feature mapping mechanism under Gaussian smoothing constraints, and obtain the pixel response after spatial regularization through sliding window convolution operation:

[0029] ;

[0030] ;

[0031] in, represents the weight matrix generated by a two-dimensional separable Gaussian function, Represents the local window coordinates centered on the current pixel, Indicates the radius of the spatial scope;

[0032] Represents the standard deviation parameter, which is used to control the spatial smoothing intensity: When , the filter degenerates into a Dirac impulse function, preserving the original spatial details; As the value increases, the filter exhibits isotropic diffusion characteristics, which suppresses high-frequency noise while enhancing local spatial correlation.

[0033] S302. Process the hyperspectral image of each band with the help of two-dimensional principal component analysis, solve the optimal projection direction through multivariate linear transformation, and achieve spatial dimension compression of the hyperspectral image.

[0034] Preferably, S302 includes:

[0035] S302-1. Consider each band of the hyperspectral image as a two-dimensional matrix. , where m and n represent the width and height of each band, and L represents the number of bands in the hyperspectral image. Representing an image No. Band;

[0036] Then calculate the average image of each band of the hyperspectral image for: ;

[0037] calculate The covariance matrix of is:

[0038] ;

[0039] calculate The eigenvalues ​​of , select the first d largest eigenvalues and its corresponding eigenvector ,make , It is called the optimal projection matrix;

[0040] S302-2, project each band image onto superior: ;

[0041] in, For band The principal components of each band image The principal components can form a dimensional matrix;

[0042] Because Orthogonal, the reconstructed image is: ;

[0043] in, The size of is usually determined by the cumulative contribution rate of principal component analysis.

[0044] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the above-mentioned method for joint spatial-spectral feature extraction of hyperspectral remote sensing images.

[0045] A computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, the steps in the above-mentioned method for joint spatial-spectral feature extraction of hyperspectral remote sensing images are implemented.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention uses the structural similarity index to analyze the similarity of image pixels and sorts the images, which is beneficial for subsequent classification; selects the value within the vector range corresponding to the central pixel through singular spectrum analysis (SSA), which is beneficial for extracting key features and suppressing noise; and constructs a two-dimensional principal component analysis framework guided by Gaussian filtering to further effectively extract spectral-spatial joint features. Compared with the six representative algorithms in the current field of hyperspectral feature extraction, this method achieves the best in all three comprehensive indicators and has greater reliability and application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0049] Figure 1 A flow chart of the method for implementing the present invention;

[0050] Figure 2 Classification results generated by different methods on the KSC dataset with a training sample ratio of 1%;

[0051] Figure 3 Classification results generated by different methods on the WHU Hi LongKou dataset with a training sample ratio of 0.1%;

[0052] Figure 4 Comparison of experimental results of various methods with different training sample sizes on the KSC dataset: (1) 2DPCA, (2) NGNMF-E2DSSA, (3) OTVCA, (4) KECA, (5) SuperPCA, (6) MSTV, (7) SISSA-GADPCA;

[0053] Figure 5 Comparison of experimental results of various methods with different training sample sizes on the WHU Hi LongKou dataset: (1) 2DPCA, (2) NGNMF-E2DSSA, (3) OTVCA, (4) KECA, (5) SuperPCA, (6) MSTV, (7) SISSA-GADPCA. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] See also Figure 1-Figure 5, the present invention provides a technical solution:

[0056] Example 1: A method for extracting spatial-spectral joint features from a hyperspectral remote sensing image, comprising:

[0057] S100, obtaining a structural similarity index composed of brightness, contrast, and structure in the hyperspectral remote sensing image, characterizing the similarity between the domain pixels and the center pixel based on the structural similarity index, selecting a domain pixel set with the highest similarity, arranging the center pixel and the domain pixels, and generating a one-dimensional expansion vector;

[0058] Preferably, the structural similarity index in S100 is represented by , used to evaluate the similarity between two natural images, the measurement task is divided into three parts: brightness, contrast and structural information;

[0059] The brightness refers to the brightness value of pixels in a hyperspectral remote sensing image; the contrast refers to the degree of change in the brightness value of pixels in a hyperspectral remote sensing image; the structural information refers to the shape and texture characteristics of objects in a hyperspectral remote sensing image. From the perspective of image synthesis, structural information is defined as an attribute independent of brightness and contrast, and reflects the structure of objects displayed in the scene;

[0060] The SSIM system is composed of three components: brightness, contrast, and structure information:

[0061] ;

[0062] and Represents two image signals, Indicates brightness, Indicates contrast, Represents structural information;

[0063] in, ;

[0064] in, is the average intensity, indicating the image signal brightness, is the standard deviation, indicating the image signal The contrast ratio, Indicates image signal and The covariance between , , , , , ;

[0065] in, ; ; ;

[0066] in, 、 represents the grayscale value of the image pixel; therefore, the final form of the SSIM index is obtained:

[0067] .

[0068] in, 、 Represents a hyperspectral remote sensing image signal.

[0069] Preferably, in S100, based on the structural similarity index, the similarity between the domain pixels and the center pixel is characterized, a domain pixel set with the highest similarity is selected, the center pixel and the domain pixels are arranged, and a one-dimensional expansion vector is generated, including:

[0070] S101. For each pixel, select a small neighborhood window , select the area pixel set with the highest similarity to the central pixel to construct the expansion vector;

[0071] S102: Select a subset of pixels with the highest similarity , and sort them according to their SSIM values, and arrange the central pixel and the selected pixels into a one-dimensional extended vector according to the SSIM value. The pixels with higher similarity are closer to the central pixel.

[0072] S200, for the extended vector, using a singular spectrum analysis method to select a value within the vector range corresponding to the central pixel, converting the original hypercube into a hyperspectral cube with SISSA pixels, and retaining the main spectral information;

[0073] Singular spectrum analysis is a signal processing technique based on singular value decomposition (SVD). In hyperspectral image classification, SSA extracts key features and suppresses noise by decomposing and reconstructing spectral signals. The main modeling steps are as follows:

[0074] S201. For a given one-dimensional signal and an embedding window of size L, the original signal can be mapped into a series of lag vectors, namely the trajectory matrix ;

[0075] S202, will The eigenvalues ​​and their corresponding eigenvectors are denoted as and ; Then the trajectory matrix is ​​reconstructed as the sum of the following basic matrices:

[0076] ;

[0077] ;

[0078] in, and denote the empirical orthogonal function and the trajectory matrix The principal components of the basic matrix;

[0079] After eigenvalue grouping, select A subset of An approximation of , where the subset contains the main trend component and eliminates the noise component.

[0080] S300, based on the spectral information of the hyperspectral cube, constructs a Gaussian filtering-guided two-dimensional principal component analysis framework to further effectively extract spectral-spatial joint features.

[0081] Preferably, S300 includes:

[0082] S301, for each spectral channel of the hyperspectral cube , establish a spatial feature mapping mechanism under Gaussian smoothing constraints, and obtain the pixel response after spatial regularization through sliding window convolution operation:

[0083] ;

[0084] ;

[0085] in, represents the weight matrix generated by a two-dimensional separable Gaussian function, Represents the local window coordinates centered on the current pixel, Indicates the radius of the spatial scope;

[0086] Represents the standard deviation parameter, which is used to control the spatial smoothing intensity: When , the filter degenerates into a Dirac impulse function, preserving the original spatial details; As the value increases, the filter exhibits isotropic diffusion characteristics, which suppresses high-frequency noise while enhancing local spatial correlation.

[0087] S302. Process the hyperspectral image of each band with the help of two-dimensional principal component analysis, solve the optimal projection direction through multivariate linear transformation, and achieve spatial dimension compression of the hyperspectral image.

[0088] Preferably, S302 includes:

[0089] S302-1. Consider each band of the hyperspectral image as a two-dimensional matrix. , where m and n represent the width and height of each band, and L represents the number of bands in the hyperspectral image. Representing an image No. Band;

[0090] Then calculate the average image of each band of the hyperspectral image for: ;

[0091] calculate The covariance matrix of is:

[0092] ;

[0093] calculate The eigenvalues ​​of , select the first d largest eigenvalues and its corresponding eigenvector ,make , It is called the optimal projection matrix;

[0094] S302-2, project each band image onto superior: ;

[0095] in, For band The principal components of each band image The principal components can form a dimensional matrix;

[0096] Because Orthogonal, the reconstructed image is: ;

[0097] in, The size of is usually determined by the cumulative contribution rate of principal component analysis.

[0098] Example 2: To evaluate the performance of this patented method (SISSA-GADPCA) for hyperspectral remote sensing image classification under small sample sizes, experiments were conducted on the KSC dataset and the WHU_Hi_LongKou dataset. Considering the difference in the total number of samples between the two datasets, the training sample size was set as follows: 1% of the KSC dataset was randomly selected as training data, with the remainder used for testing; 0.1% of the WHU_Hi_LongKou dataset was randomly selected as training data (due to the large number of annotated samples, 0.1% was chosen for training), with the remainder used for testing. To comprehensively evaluate the performance of SISSA-GADPCA, this paper employed four performance evaluation metrics: overall accuracy (OA), average accuracy (AA), per-class accuracy (CA), and the KAPPA coefficient. To further enhance the robustness of the experimental results, the experimental results for each method were repeated five times, and the mean of the performance metrics was calculated to effectively reduce the impact of random error. The comparison methods included: 2DPCA, KECA, OTVCA, SuperPCA, NGNMF-E2DSSA, and MSTV.

[0099] Tables 1 and 2 systematically compare the classification performance of the SISSA-G2DPCA method proposed in this patent with six comparison methods on the KSC dataset (1% training ratio) and the WHU_Hi_LongKou dataset (0.1% training ratio), and quantitatively evaluate them using four indicators: category accuracy (CA), global accuracy (OA), average accuracy (AA), and Kappa coefficient.

[0100] Table 1 Comparison of classification performance of different methods on the KSC dataset (training sample ratio is 1%)

[0101]

[0102] As can be seen in Table 1, the SISSA-G2DPCA method significantly enhances the robustness and discriminative power of feature representation through the collaborative optimization mechanism of spectral and spatial features. Specifically, the method demonstrates significant advantages in three comprehensive evaluation indicators: OA (96.85%), AA (95.4%), and Kappa coefficient (96.49%). In particular, it achieves the highest classification accuracy in 9 of the 13 ground feature classification tasks, fully verifying its ability to accurately distinguish in complex spectral distribution scenarios.

[0103] Table 2 Comparison of classification performance of different methods on the WHU Hi LongKou dataset (training sample ratio is 1%)

[0104]

[0105] As shown in Table 2, all methods demonstrated excellent classification results on the WHU_Hi_LongKou dataset, primarily due to the dataset's rich spatial contextual information, which supports feature modeling. The SISSA-G2DPCA method proposed in this patent achieves a qualitative leap in feature representation capabilities through a strategy for collaborative spectral-spatial information extraction, achieving optimal performance in three comprehensive metrics: OA (99.54%), AA (98.55%), and Kappa coefficient (99.40%), representing improvements of 0.98%, 1.7%, and 1.3%, respectively, over the suboptimal method. Furthermore, the best classification accuracy was achieved for five of the nine feature categories, fully validating the patented method's ability to distinguish between categories in complex spectral distribution scenarios.

[0106] In order to further verify the effectiveness of the algorithm intuitively, the complete classification diagrams generated by different methods on the two datasets are drawn ( Figure 2 and Figure 3 ).Depend on Figure 2 It can be seen that through the dual constraint mechanism of spatial continuity and spectral consistency, the SISSA-G2DPCA method not only performs well in classification accuracy, but also has a higher degree of consistency with the actual distribution of objects, effectively avoiding the problem of discriminant information loss that may occur in the process of feature dimensionality reduction in traditional methods. Figure 3 It can be seen that the classification map generated by the SISSA-G2DPCA method has clear boundaries of objects and stronger spatial continuity of similar objects, which shows that the method effectively retains key discriminant information during the dimensionality reduction process and provides reliable technical support for high-precision remote sensing interpretation.

[0107] In order to verify the adaptability of the SISSA-G2DPCA method under different training samples, a small sample training scenario is constructed by randomly selecting 1%, 2%, and 5% of labeled samples from each category in the KSC dataset. The test results are shown in Figure 2. Figure 4 As shown; in the WHU HiLongKou dataset, a very small sample condition is set, using only 0.1%~0.5% samples for each class, and the test results are as follows Figure 5 As shown. Figure 4 It can be seen that the classification accuracy of all methods increases with the increase of training samples, but the patented method always maintains a significant lead, and each evaluation index is better than the comparison method, indicating that it has stronger data utilization efficiency and feature discrimination ability. Figure 5As can be seen, when only 0.1% of the training samples are used, the proposed method still maintains an absolute advantage; as the number of training samples increases, the performance advantage continues to increase. This experiment further confirms the effectiveness of the SISSA-G2DPCA method, demonstrating that the spectral-spatial coupling feature extraction strategy proposed in this patent can significantly improve the discrimination ability in small sample scenarios, and the setting of the Gaussian filtering mechanism can effectively suppress noise interference while significantly enhancing the robustness of the model.

[0108] In summary, after multi-dimensional experimental verification, SISSA-G2DPCA has significant advantages in hyperspectral classification tasks, showing efficient feature extraction capabilities and stable generalization performance under small sample conditions, providing a more practical solution for remote sensing image classification applications and has great application potential.

[0109] Example 3: The computer-readable storage medium of this embodiment stores a computer program, which, when executed by a processor, implements the steps of the method for joint spatial-spectral feature extraction of hyperspectral remote sensing images in Example 1.

[0110] The computer-readable storage medium of this embodiment may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal; the computer-readable storage medium of this embodiment may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, etc. equipped on the terminal; further, the computer-readable storage medium may also include both an internal storage unit of the terminal and an external storage device.

[0111] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0112] Example 4: The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for joint spatial-spectral feature extraction of hyperspectral remote sensing images in Example 1 are implemented.

[0113] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The memory can include read-only memory and random access memory, and provide instructions and data to the processor. A part of the memory can also include non-volatile random access memory. For example, the memory can also store information about the device type.

[0114] Those skilled in the art will appreciate that the disclosed embodiments may be provided as methods, systems, or computer program products. Therefore, the present solution may take the form of a hardware embodiment, a software embodiment, or a combination of software and hardware embodiments. Furthermore, the present solution may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage) containing computer-usable program code.

[0115] The present solution is described with reference to the flowcharts and / or block diagrams of the methods and computer program products according to the embodiments of the present solution. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as the combination of the processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or methods Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0116] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or methods Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or methods Figure 1 A step that specifies a function in one or more boxes.

[0118] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0119] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for joint spatial-spectral feature extraction of hyperspectral remote sensing images, characterized by: The method comprises: S100, obtaining a structural similarity index composed of brightness, contrast, and structure in the hyperspectral remote sensing image, characterizing the similarity between the domain pixels and the center pixel based on the structural similarity index, selecting a domain pixel set with the highest similarity, arranging the center pixel and the domain pixels, and generating a one-dimensional expansion vector; S200, for the extended vector, using a singular spectrum analysis method to select a value within the vector range corresponding to the central pixel, converting the original hypercube into a hyperspectral cube with SISSA pixels, and retaining the main spectral information; S300, based on the spectral information of the hyperspectral cube, constructs a Gaussian filtering-guided two-dimensional principal component analysis framework to further effectively extract spectral-spatial joint features.

2. The method for joint spatial-spectral feature extraction of hyperspectral remote sensing images according to claim 1, wherein: The structural similarity index in S100 is referred to as , including three parts: brightness, contrast and structural information; The brightness refers to the brightness value of pixels in a hyperspectral remote sensing image; the contrast refers to the degree of change in the brightness value of pixels in a hyperspectral remote sensing image; the structural information refers to the shape and texture characteristics of objects in a hyperspectral remote sensing image. From the perspective of image synthesis, structural information is defined as an attribute independent of brightness and contrast, and reflects the structure of objects displayed in the scene; The three components of brightness, contrast and structure information are combined to form the SSIM system, and the SSIM index is determined: ； in, 、 represents the hyperspectral remote sensing image signal, 、 is the average intensity, indicating the image signal 、 brightness, 、 is the standard deviation, indicating the image signal The contrast ratio, Indicates image signal and The covariance between in, , , , , , .

3. The method for joint spatial-spectral feature extraction of hyperspectral remote sensing images according to claim 1, wherein: In S100, the similarity between the domain pixels and the center pixel is characterized based on the structural similarity index, a domain pixel set with the highest similarity is selected, the center pixel and the domain pixels are arranged, and a one-dimensional expansion vector is generated, including: S101. For each pixel, select a small neighborhood window , select the area pixel set with the highest similarity to the central pixel to construct the expansion vector; S102: Select a subset of pixels with the highest similarity , and sort them according to their SSIM values, and arrange the central pixel and the selected pixels into a one-dimensional extended vector according to the SSIM value. The pixels with higher similarity are closer to the central pixel.

4. The method for extracting spatial-spectral joint features of hyperspectral remote sensing images according to claim 1, wherein: The singular spectrum analysis method is used in S200 to select a value within the vector range corresponding to the central pixel, including: S201. For a given one-dimensional signal and an embedding window of size L, the original signal can be mapped into a series of lag vectors, namely the trajectory matrix ; S202, will The eigenvalues ​​and their corresponding eigenvectors are denoted as and ; Then the trajectory matrix is ​​reconstructed as the sum of the following basic matrices: ； ； in, and denote the empirical orthogonal function and the trajectory matrix The principal components of the basic matrix; After eigenvalue grouping, select A subset of An approximation of , where the subset contains the main trend component and eliminates the noise component.

5. The method for joint spatial-spectral feature extraction of hyperspectral remote sensing images according to claim 1, wherein: The S300 includes: S301, for each spectral channel of the hyperspectral cube , establish a spatial feature mapping mechanism under Gaussian smoothing constraints, and obtain the pixel response after spatial regularization through sliding window convolution operation: ； ； in, represents the weight matrix generated by a two-dimensional separable Gaussian function, Represents the local window coordinates centered on the current pixel, Indicates the radius of the spatial scope; Represents the standard deviation parameter, which is used to control the spatial smoothing intensity: When , the filter degenerates into a Dirac impulse function, preserving the original spatial details; As the value increases, the filter exhibits isotropic diffusion characteristics, which suppresses high-frequency noise while enhancing local spatial correlation. S302. Process the hyperspectral image of each band with the help of two-dimensional principal component analysis, solve the optimal projection direction through multivariate linear transformation, and achieve spatial dimension compression of the hyperspectral image.

6. The method for extracting spatial-spectral joint features of hyperspectral remote sensing images according to claim 5, wherein: The S302 includes: S302-1. Consider each band of the hyperspectral image as a two-dimensional matrix. , where m and n represent the width and height of each band, and L represents the number of bands in the hyperspectral image. Representing an image No. Band; Then calculate the average image of each band of the hyperspectral image for: ; calculate The covariance matrix of is: ； calculate The eigenvalues ​​of , select the first d largest eigenvalues and its corresponding eigenvector ,make , It is called the optimal projection matrix; S302-2, project each band image onto superior: ; in, For band The principal components of each band image The principal components can form a dimensional matrix; Because Orthogonal, the reconstructed image is: ; in, The size of is usually determined by the cumulative contribution rate of principal component analysis.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for joint spatial-spectral feature extraction of hyperspectral remote sensing images according to any one of claims 1 to 6 are implemented.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the steps of the method for joint spatial-spectral feature extraction of hyperspectral remote sensing images according to any one of claims 1 to 6 are implemented.

Images