Hyperspectral remote sensing image recognition method and system based on superpixels and medium

The method addresses mis-segmentation and boundary information challenges in hyperspectral image classification by integrating spectral dimension reduction, superpixel segmentation, and structural feature extraction, resulting in improved classification accuracy and visual effects.

CN119068342AActive Publication Date: 2024-12-03HUNAN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411281811.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-12-03
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

Hyperspectral remote sensing image classification methods face challenges with mis-segmentation at complex edges or structures using superpixel-based approaches, and feature extraction methods fail to consider boundary information, leading to reduced classification accuracy.

Method used

A method involving spectral dimension reduction, superpixel segmentation, spatial optimization of class probabilities, extraction of structural features, and probabilistic fusion of these probabilities using a multi-class support vector machine model to enhance classification accuracy.

Benefits of technology

The method achieves higher ground object classification accuracy and better visual effects by fully utilizing spatial information and edge information, optimizing misclassification issues.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068342A_ABST
    Figure CN119068342A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of hyperspectral remote sensing image processing, in particular to a hyperspectral remote sensing image recognition method and system based on superpixels and a medium, and the hyperspectral remote sensing image recognition method comprises the following steps: 1, carrying out spectrum dimension reduction on a source hyperspectral image to obtain a dimension-reduced hyperspectral image; 2, extracting superpixel information from the source hyperspectral image, obtaining a superpixel-level category probability, and then performing spatial optimization on the superpixel-level category probability to obtain optimization probabilities of different ground features; 3, extracting structural features from the dimensionality-reduced hyperspectral image, and obtaining a feature-level category probability; and 4, carrying out probability fusion on the optimization probabilities and the feature level category probabilities of different ground features to obtain a final classification result. According to the method, the spatial information in the superpixels can be fully and accurately mined, the problem of ground feature misclassification is optimized by fusing the edge information, and meanwhile, higher ground feature classification precision and a better visual effect can be obtained.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing image processing, and in particular to a hyperspectral remote sensing image recognition method, system and medium based on superpixel. Background Art

[0002] Hyperspectral images contain hundreds of narrow and adjacent spectral bands, thanks to the continuous development of spectral imaging equipment and the improvement of spectral resolution. The rich spatial and spectral information in hyperspectral images greatly improves the perception ability of earth observation, making hyperspectral remote sensing technology play a vital role in precision agriculture, space exploration, pollution monitoring and military applications.

[0003] Hyperspectral remote sensing image classification is a technology that uses hyperspectral remote sensing data to classify objects. It aims to assign a unique category label to each pixel in the hyperspectral image through learning methods, thereby completing the semantic analysis of the actual scene. Superpixel-based methods are widely used in hyperspectral image classification because they can reduce data redundancy and enhance local features. However, such methods are more prone to mis-segmentation at complex edges or structures, resulting in reduced classification accuracy. On the other hand, feature extraction-based methods do not take into account the boundary information of objects, making it difficult for the classification results to be consistent with the actual objects. Summary of the invention

[0004] The present invention provides a hyperspectral remote sensing image recognition method, system and medium based on superpixels to solve the technical problems mentioned in the background technology.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] The present invention provides a hyperspectral remote sensing image recognition method based on superpixels, comprising the following steps:

[0007] S1, performing spectral dimension reduction on the source hyperspectral image to obtain a hyperspectral image after dimension reduction;

[0008] S2, extracting superpixel information from the source hyperspectral image and obtaining the class probability at the superpixel level, and then spatially optimizing the class probability at the superpixel level to obtain the optimized probabilities of different ground objects;

[0009] s3. Extract structural features from the hyperspectral image after dimensionality reduction to obtain an optimized structural feature image, and use the trained multi-class support vector machine model to identify the optimized structural feature image to obtain the feature-level category probability of each pixel;

[0010] S4. Probabilistically fuse the optimized probabilities of different objects and the feature-level category probabilities of each pixel to obtain the final classification result.

[0011] Furthermore, the specific implementation of S1 is as follows:

[0012] For the input source hyperspectral image X∈R H×W×O The average fusion method is used to reduce the spectral dimension and obtain the reduced dimension image Y∈R H×W×B , where R represents the real three-dimensional space, H×W is the spatial resolution of the image, O is the number of bands of the source image, and B is the number of bands after dimensionality reduction.

[0013] Furthermore, the S2 specifically includes the following steps:

[0014] S21. Perform principal component analysis and dimensionality reduction on the source hyperspectral image X, and extract the principal component graph X PCA ; Then the principal component graph X PCA Perform superpixel segmentation to obtain a superpixel segmentation map, where the superpixel segmentation map consists of multiple superpixel blocks;

[0015] S22. Use the trained multi-class support vector machine model to identify the hyperspectral image after dimensionality reduction, and obtain the pixel-level category probability of each superpixel block in the superpixel segmentation map. Pixel-level category probability for each superpixel block Perform average calculation to obtain superpixel-level category probabilities

[0016] S23. Use the extended random walk method to classify superpixel-level probabilities Optimize and obtain the optimized probability P of different objects S .

[0017] Furthermore, in S21, an entropy rate superpixel segmentation algorithm is used to perform superpixel segmentation, thereby obtaining a superpixel segmentation map.

[0018] Furthermore, the superpixel level category probability in S22 The formula is as follows:

[0019]

[0020] Among them, N m is the number of pixels in the mth superpixel block, and i represents the i-th pixel in the m-th superpixel.

[0021] Furthermore, the S23 specifically includes the following steps:

[0022] S231, first construct a weight graph G = (V, E) for superpixels, where the vertex V of the graph is a superpixel set, and E is an edge set connecting adjacent superpixels; the weights w of the edges between different vertices in the superpixel weight graph G are mnIt can be defined as:

[0023]

[0024] In the above formula, e represents a natural constant, μ represents a free parameter, and x m and x n Represent the spectral values ​​of different vertices respectively;

[0025] S232, model the spatial relationship between adjacent superpixels, and obtain the optimized probability P of different objects by solving the following energy function: S ;

[0026]

[0027] Among them, E t (P S ) represents the energy function; t represents the categories of different objects; γ represents the parameter for adjusting the weights of the two items; and Respectively represent spatial terms and non-spatial terms; sp represents spatial, i.e. spatial; as represents aspatial, i.e. non-spatial; among which the spatial term With the following equation:

[0028]

[0029] In the above formula, L is the Laplacian matrix of the weight graph, and T represents the transpose of the matrix; L has the following equation:

[0030]

[0031] Non-spatial terms With the following equation:

[0032]

[0033] Among them, Λ q ,Λ t Represent two different diagonal matrices, the diagonal matrix Λ t The values ​​on the diagonal of are determined by the class probabilities at the superpixel level constitute, represents the optimized probability of the qth class; represents the optimized probability of the tth class.

[0034] Furthermore, the S3 specifically includes the following steps:

[0035] S31, the principal component graph X in S21 PCA Perform edge detection to obtain the probability that each pixel belongs to the edge;

[0036] S32, optimizing the weighted least squares algorithm by using an optimized global structure filtering algorithm, and extracting the structural information of the hyperspectral image after dimensionality reduction by using the optimized weighted least squares algorithm to obtain an optimized structural feature image SF; specifically, the optimized global structure filtering algorithm optimizes the similarity function of the weighted least squares algorithm, and the similarity function of the original weighted least squares algorithm can be expressed as:

[0037]

[0038] Among them, d1 ij (g) is the similarity function between pixel i and pixel j; g represents the guide image, i.e., the source hyperspectral image, σ r represents the standard deviation of the Gaussian mixture function; g i , g j Respectively represent the pixel values ​​of pixel i and pixel j in the guide image. The above similarity function is optimized by adding edge information and can be expressed as:

[0039]

[0040] Among them, d2 ij (g) represents the optimized similarity function, E i represents the probability that pixel i belongs to the edge.

[0041] The S32 is expressed by a formula, which is as follows:

[0042] SF=SWLS(Y,β,σ r )

[0043] In the above formula, SF represents the structural feature image, SWLS is the structured feature extraction operation, and β is the weight parameter;

[0044] S33, for the structural feature image SF, use the trained multi-class support vector machine model to identify and obtain the feature level category probability P of each pixel F .

[0045] Furthermore, the implementation of S4 is specifically as follows:

[0046] The optimization probability P for different objects S And the feature-level category probability P of each pixel F Perform probability fusion to obtain the final classification result;

[0047] The S4 is expressed as follows:

[0048] P = argmax{(1-α)P S +αP F}

[0049] Among them, P represents the final classification result, represents the optimized probability of the t-th type of feature, represents the feature-level category probability of the t-th type of features, and α represents a free parameter.

[0050] The second aspect of the present invention also provides a hyperspectral image classification system integrating superpixels and structural features, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the hyperspectral remote sensing image recognition method described above.

[0051] The third aspect of the present invention further provides a computer-readable storage medium, in which a computer program is stored. The computer program is used to be programmed or configured by a microprocessor to execute the hyperspectral remote sensing image recognition method described above.

[0052] Beneficial effects of the present invention:

[0053] The hyperspectral remote sensing image recognition method in the present invention can fully and accurately mine the spatial information in superpixels and fuse edge information to optimize the problem of misclassification of ground objects. At the same time, the hyperspectral remote sensing image recognition method in the present invention can obtain higher ground object classification accuracy and better visual effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of the hyperspectral remote sensing image recognition method in the present invention;

[0055] Figure 2 The recognition results of the hyperspectral remote sensing image recognition method of the present invention on the Indian Pines data, where (a) is a pseudo-color image, (b) is a ground truth label, (c)-(g) are the results obtained by several other existing hyperspectral recognition methods, and Figure (h) is the result obtained by the method of this embodiment;

[0056] Figure 3 : The recognition results of the hyperspectral remote sensing image recognition method in the present invention on the Pavia University data, where (a) is a pseudo-color image, (b) is the ground truth label, (c)-(g) are the results obtained by several other existing hyperspectral recognition methods, and Figure (h) is the result obtained by the method of this embodiment. DETAILED DESCRIPTION

[0057] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Preferred embodiments of the present invention are provided in the drawings. However, the present invention can be implemented in many other different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0058] Reference Figure 1 The present invention provides a method for recognizing hyperspectral remote sensing images based on superpixels, comprising the following steps:

[0059] S1, performing spectral dimension reduction on the source hyperspectral image to obtain a hyperspectral image after dimension reduction;

[0060] S2, extracting superpixel information from the source hyperspectral image and obtaining the class probability at the superpixel level, and then spatially optimizing the class probability at the superpixel level to obtain the optimized probabilities of different ground objects;

[0061] S3, extracting structural features from the hyperspectral image after dimensionality reduction to obtain an optimized structural feature image, using the trained multi-class support vector machine model to identify the optimized structural feature image, and obtaining the feature-level category probability of each pixel;

[0062] S4. Probabilistically fuse the optimized probabilities of different objects and the feature-level category probabilities of each pixel to obtain the final classification result.

[0063] In some embodiments, the specific implementation of S1 is:

[0064] For the input source hyperspectral image X∈R H×W×O The average fusion method is used to reduce the spectral dimension and obtain the reduced dimension image Y∈R H×W×B , where R represents the real three-dimensional space, H×W is the spatial resolution of the image, O is the number of bands of the source image, and B is the number of bands after dimensionality reduction. In this embodiment, the number of bands B after dimensionality reduction is set to 20. The average fusion-based method is a simple and effective hyperspectral data dimensionality reduction technology. Its principle is to evenly divide the data into multiple subsets, and then fuse the data in each subset to obtain data with reduced dimensionality. Its advantage is that it can simultaneously retain the physical meaning of the data.

[0065] In some embodiments, S2 specifically includes the following steps:

[0066] S21. Perform principal component analysis and dimensionality reduction on the source hyperspectral image X, and extract the principal component graph X PCA ; Then the principal component graph X PCA Perform superpixel segmentation to obtain a superpixel segmentation map, where the superpixel segmentation map consists of multiple superpixel blocks;

[0067] In this embodiment, the principal component graph X PCA The dimension is set to 3; Principal Component Analysis (PCA) is a classic data dimensionality reduction and feature extraction method, which aims to convert high-dimensional data into low-dimensional representation through linear transformation while retaining the information in the data as much as possible.

[0068] Superpixel segmentation is a typical image processing technology that divides the original image into a set of closely connected pixels, each of which is usually considered to be a region or object in the image. The main advantage of superpixel segmentation is that it reduces the complexity of the image, improves the efficiency of image processing and analysis, and retains the key information in the image. In this embodiment, the Entropy Rate Superpixel Segmentation (ERS) algorithm is used to generate a compact and regular superpixel segmentation map;

[0069] S22. Use the trained multi-class support vector machine model to identify the hyperspectral image after dimensionality reduction, and obtain the pixel-level category probability of each superpixel block in the superpixel segmentation map. Pixel-level category probability for each superpixel block Perform average calculation to obtain superpixel-level category probabilities

[0070] S23. Use the extended random walk method to classify superpixel-level probabilities Optimize and obtain the optimized probability P of different objects S .

[0071] In some embodiments, the entropy rate superpixel segmentation algorithm is used to perform superpixel segmentation in S21 to obtain a superpixel segmentation map.

[0072] In some embodiments, the superpixel level category probability in S22 The formula is as follows:

[0073]

[0074] Among them, N m is the number of pixels in the mth superpixel block, and i represents the i-th pixel in the mth superpixel. After averaging, the class probabilities of all pixels in the superpixel block are calculated by the superpixel-level class probabilities. Alternative

[0075] Furthermore, the S23 specifically includes the following steps:

[0076] S231, first construct a weight graph G = (V, E) for superpixels, where the vertex V of the graph is a superpixel set, and E is an edge set connecting adjacent superpixels; the weights w of the edges between different vertices in the superpixel weight graph G are mn It can be defined as:

[0077]

[0078] In the above formula, e represents a natural constant, μ represents a free parameter, and x m and x nRepresent the spectral values ​​of different vertices respectively;

[0079] S232, model the spatial relationship between adjacent superpixels, and obtain the optimized probability P of different objects by solving the following energy function: S

[0080]

[0081] Among them, E t (P S ) represents the energy function; t represents the categories of different objects; γ represents the parameter for adjusting the weights of the two items; and Respectively represent spatial terms and non-spatial terms; sp represents spatial, i.e. spatial; as represents aspatial, i.e. non-spatial; among which the spatial term With the following equation:

[0082]

[0083] In the above formula, L is the Laplacian matrix of the weight graph, and T represents the transpose of the matrix; L has the following equation:

[0084]

[0085] Non-spatial terms With the following equation:

[0086]

[0087] Among them, Λ q ,Λ t Represent two different diagonal matrices, the diagonal matrix Λ t The values ​​on the diagonal of are determined by the class probabilities at the superpixel level constitute, represents the optimized probability of the qth class; represents the optimized probability of the tth class.

[0088] In some embodiments, S3 specifically includes the following steps:

[0089] S31, the principal component graph X in S21 PCA Perform edge detection to obtain the probability that each pixel belongs to the edge;

[0090] S32, optimizing the weighted least squares algorithm by using an optimized global structure filtering algorithm, and extracting the structural information of the hyperspectral image after dimensionality reduction by using the optimized weighted least squares algorithm to obtain an optimized structural feature image SF;

[0091] Specifically, the optimized global structure filtering algorithm optimizes the similarity function of the weighted least squares algorithm. The similarity function of the original weighted least squares algorithm can be expressed as:

[0092]

[0093] Among them, d1 i,j (g) is the similarity function between pixel i and pixel j; g represents the guide image, i.e., the source hyperspectral image, σ r represents the standard deviation of the Gaussian mixture function; g i , g j Respectively represent the pixel values ​​of pixel i and pixel j in the guide image. The above similarity function is optimized by adding edge information and can be expressed as:

[0094]

[0095] Among them, d2 i,j (g) represents the optimized similarity function, E i Represents the probability that pixel i belongs to the edge;

[0096] The S32 is expressed by a formula, which is as follows:

[0097] SF=SWLS(Y,β,σ r )

[0098] In the above formula, SF represents the structural feature image, SWLS is the structured feature extraction operation, and β is the weight parameter.

[0099] S33, for the structural feature image SF, use the trained multi-class support vector machine model to identify and obtain the feature level category probability P of each pixel F .

[0100] In some embodiments, the implementation of S4 is specifically as follows:

[0101] The optimization probability P for different objects S And the feature-level category probability P of each pixel F Perform probability fusion to obtain the final classification result;

[0102] Due to the fusion of complementary features and information at different levels, the probability of each category after fusion will be higher than the category probability of a single method, thereby achieving more accurate recognition of various types of objects. The final category probability uses a weighted decision fusion rule to take the maximum value of the two category probabilities, which can be expressed as:

[0103]

[0104] Among them, P represents the final classification result, represents the optimized probability of the t-th type of feature, represents the feature-level category probability of the t-th type of features, and α represents a free parameter.

[0105] In order to verify the hyperspectral image classification method and system of the present embodiment that combines superpixels and structural features, the real data sets Indian Pines and Pavia University are used to test the method of the present invention in this embodiment. The public data set Indian Pines is obtained by shooting over the Indian Pines test site in Indiana, USA, using the AVIRIS hyperspectral sensor system. The data space size is 145 pixels × 145 pixels, and the spatial resolution is 20 meters / pixel. After removing the noise-interfered bands and water absorption bands, the remaining 200 spectral bands are used for subsequent recognition tasks. In addition, the data set contains a total of 16 types of ground object categories. The public data set Pavia University is obtained by shooting over the University of Pavia, Italy, using the ROSIS-3 imager. The data space size is 610 pixels × 340 pixels, and the spatial resolution is 1-3 meters / pixel. After removing 12 noise-interfered bands, the remaining 103 spectral bands are used for subsequent classification experiments. In addition, the data set contains a total of 9 types of ground object categories. Tests were conducted on two real data sets, and the performance of different algorithms was evaluated using three typical objective evaluation indicators: OA (overall accuracy), AA (average accuracy), and Kappa coefficient. The compared algorithms included Image Fusion and Recursive Filter (IFRF), superpixel-based PCA (SuperPCA), PCA-based Edge Preserving Features (PCA-EPFs), multi-scale total variation (MSTV) and superpixel-based Classification via Multiple Kernels (SC-MK).

[0106] In the Indian Pines dataset, the number of training samples for each type of land object is 5. The objective evaluation indicators of five typical hyperspectral image classification methods (IFRF, SuperPCA, PCA-EPFs, MSTV, SC-MK) and the method proposed in this embodiment on the Indian Pines dataset are shown in Table 1. As can be seen from Table 1, the method in this embodiment has obtained the highest values ​​in the three objective evaluation indicators of OA, AA and Kappa coefficient, that is, the method in this embodiment can achieve the best recognition accuracy, which is about 5 to 20 percentage points higher than other comparison methods. In addition, Figure 3 The classification results of five typical classification methods (IFRF, SuperPCA, PCA-EPFs, MSTV, SC-MK) and the method of this embodiment are shown, from which it can be seen that the method of this embodiment can effectively utilize a small number of training samples, and by fusing superpixel-level and feature-level information, it can eliminate misclassified samples to a large extent and maintain the boundaries of objects, thereby further improving the classification accuracy and visual effect.

[0107] Table 1: Objective performance indicators of the method of this embodiment and five typical hyperspectral imaging methods on the Indian Pines dataset.

[0108]

[0109] In the Pavia University dataset, the number of training samples for each type of land object is 10. The objective evaluation indicators of five typical hyperspectral image classification methods (IFRF, SuperPCA, PCA-EPFs, MSTV, SC-MK) and the method proposed in this embodiment on the Pavia University dataset are shown in Table 2. As can be seen from Table 2, the objective evaluation indicators of the method in this embodiment are better than other methods, and are improved by about 3 to 30 percentage points compared with other comparison methods. The method in this embodiment can effectively fuse superpixel-level and feature-level information. By combining semi-supervised methods, noise labels are eliminated to a large extent and the boundaries of land objects are maintained, thereby further improving the classification accuracy and visual effects.

[0110] Table 2: Objective performance indicators of the method of this embodiment and five typical hyperspectral imaging methods on the Pavia University dataset.

[0111]

[0112] The hyperspectral remote sensing image recognition method in the present invention can fully and accurately mine the spatial information in superpixels and integrate edge information to optimize the problem of misclassification of ground objects. Experimental results show that in the case of small samples, the hyperspectral remote sensing image recognition method in the present invention can obtain higher ground object classification accuracy and better visual effects.

[0113] In addition, the second aspect of the present invention also provides a hyperspectral image classification system integrating superpixels and structural features, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the hyperspectral remote sensing image recognition method described above.

[0114] In addition, the third aspect of the present invention further provides a computer-readable storage medium, in which a computer program is stored, and the computer program is used to be programmed or configured by a microprocessor to execute the hyperspectral remote sensing image recognition method described above.

[0115] It should be understood by those skilled in the art that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, compact disc read-only memory, optical storage, etc.) containing computer-usable program codes. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0116] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention. In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A hyperspectral remote sensing image recognition method based on superpixels, characterized in that: The steps include: S1, performing spectral dimension reduction on the source hyperspectral image to obtain a hyperspectral image after dimension reduction; S2, extracting superpixel information from the source hyperspectral image and obtaining the class probability at the superpixel level, and then spatially optimizing the class probability at the superpixel level to obtain the optimized probabilities of different ground objects; S3, extracting structural features from the hyperspectral image after dimensionality reduction to obtain an optimized structural feature image, using the trained multi-class support vector machine model to identify the optimized structural feature image, and obtaining the feature-level category probability of each pixel; S4. Probabilistically fuse the optimized probabilities of different objects and the feature-level category probabilities of each pixel to obtain the final classification result.

2. The hyperspectral remote sensing image recognition method according to claim 1, characterized in that: The specific implementation of S1 is: For the input source hyperspectral image X∈R H×W×O The average fusion method is used to reduce the spectral dimension and obtain the reduced dimension image Y∈R H ×W× B , where R represents the real three-dimensional space, H×W is the spatial resolution of the image, O is the number of bands of the source image, and B is the number of bands after dimensionality reduction.

3. The hyperspectral remote sensing image recognition method according to claim 1, characterized in that: The S2 specifically includes the following steps: S21. Perform principal component analysis and dimensionality reduction on the source hyperspectral image X, and extract the principal component graph X PCA ; Then the principal component graph X PCA Perform superpixel segmentation to obtain a superpixel segmentation map, where the superpixel segmentation map consists of multiple superpixel blocks; S22. Use the trained multi-class support vector machine model to identify the hyperspectral image after dimensionality reduction, and obtain the pixel-level category probability of each superpixel block in the superpixel segmentation map. Pixel-level category probability for each superpixel block Perform average calculation to obtain superpixel-level category probabilities S23. Use the extended random walk method to classify superpixel-level probabilities Optimize and obtain the optimized probability P of different objects S .

4. The hyperspectral remote sensing image recognition method according to claim 3, characterized in that: In S21, an entropy rate superpixel segmentation algorithm is used to perform superpixel segmentation, thereby obtaining a superpixel segmentation map.

5. The hyperspectral remote sensing image recognition method according to claim 3, characterized in that: The superpixel level category probability in S22 The formula is as follows: Among them, N m is the number of pixels in the mth superpixel block, and i represents the i-th pixel in the m-th superpixel.

6. The hyperspectral remote sensing image recognition method according to claim 3, characterized in that: The S23 specifically includes the following steps: S231, first construct a weight graph G = (V, E) for superpixels, where the vertex V of the graph is a superpixel set, and E is an edge set connecting adjacent superpixels; the weights w of the edges between different vertices in the superpixel weight graph G are mn It can be defined as: In the above formula, e represents a natural constant, μ represents a free parameter, and x m and x n Represent the spectral values ​​of different vertices respectively; S232, model the spatial relationship between adjacent superpixels, and obtain the optimized probability P of different objects by solving the following energy function: S ; Among them, E t (P S ) represents the energy function; t represents the categories of different objects; γ represents the parameter for adjusting the weights of the two items; and Respectively represent spatial terms and non-spatial terms; sp represents spatial, i.e. spatial; as represents aspatial, i.e. non-spatial; among which the spatial term With the following equation: In the above formula, L is the Laplacian matrix of the weight graph, and T represents the transpose of the matrix; L has the following equation: Non-spatial terms With the following equation: Among them, Λ q ,Λ t Represent two different diagonal matrices, the diagonal matrix Λ t The values ​​on the diagonal of are determined by the class probabilities at the superpixel level constitute, represents the optimized probability of the qth class; represents the optimized probability of the tth class.

7. The hyperspectral remote sensing image recognition method according to claim 6, characterized in that: The s3 specifically includes the following steps: S31, the principal component graph X in S21 PCA Perform edge detection to obtain the probability that each pixel belongs to the edge; S32, optimizing the weighted least squares algorithm by using an optimized global structure filtering algorithm, and extracting the structural information of the hyperspectral image after dimensionality reduction by using the optimized weighted least squares algorithm to obtain an optimized structural feature image SF; specifically, the optimized global structure filtering algorithm optimizes the similarity function of the weighted least squares algorithm, and the similarity function of the original weighted least squares algorithm can be expressed as: Among them, d1 i,j (g) is the similarity function between pixel i and pixel j; g represents the guide image, i.e., the source hyperspectral image, σ r represents the standard deviation of the Gaussian mixture function; g i , g j Respectively represent the pixel values ​​of pixel i and pixel j in the guide image; the above similarity function is optimized by adding edge information and can be expressed as: Among them, d2 i,j (g) represents the optimized similarity function, E i Represents the probability that pixel i belongs to the edge; The S32 is expressed by a formula, which is as follows: SF=SWLS(Y, β, σ r ) In the above formula, SF represents the structural feature image, SWLS is the structured feature extraction operation, and β is the weight parameter; S33, for the structural feature image SF, use the trained multi-class support vector machine model to identify and obtain the feature level category probability P of each pixel F .

8. The hyperspectral remote sensing image recognition method according to claim 7, characterized in that: The implementation of S4 is specifically as follows: The optimization probability P for different objects S And the feature-level category probability P of each pixel F Perform probability fusion to obtain the final classification result; The S4 is expressed as follows: Among them, P represents the final classification result, represents the optimized probability of the t-th type of feature, represents the feature-level category probability of the t-th type of features, and α represents a free parameter.

9. A hyperspectral image classification system integrating superpixels and structural features, characterized in that: The invention comprises a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the hyperspectral remote sensing image recognition method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored in the computer-readable storage medium, and the computer program is used to be programmed or configured by a microprocessor to execute the hyperspectral remote sensing image recognition method according to any one of claims 1 to 8.

Citation Information

Patent Citations

  • Hyper-spectral image classification device and hyper-spectral image classification method

    CN106228186A

  • Remote sensing image sub-pixel positioning method based on extended random walk algorithm

    CN110298883A