Wave band selection method based on specific pixel extraction and adaptive neighborhood clustering

Abstract

The invention relates to a band selection method based on specific pixel extraction and adaptive neighborhood clustering, which analyzes common features and specific features among bands by introducing a low-rank self-representation model, mines specific pixel information to enhance the distinguishability of the bands, establishes an adaptive neighborhood band clustering model based on the specific pixels, and improves the band selection efficiency. The establishment and structure optimization of the inter-band similarity Laplacian diagram are realized while the parameter setting is simplified, so that the generalization performance of the algorithm on different spectral data is improved.

Description

Technical Field

[0001] The invention belongs to the field of pattern recognition and relates to a band selection method based on unique pixel extraction and adaptive neighborhood clustering. Background Art

[0002] Hyperspectral images contain rich information about objects and can reflect the characteristics of objects in multiple dimensions such as spectrum and pixels. They are widely used in smart agriculture, environmental protection, disaster detection, military reconnaissance and other fields. However, hyperspectral images have the characteristics of large data volume, high correlation between bands, and much redundant information, which leads to high time complexity of downstream tasks with hyperspectral images as input, and are affected by the "curse of dimensionality", which restricts the technical development and performance improvement of downstream tasks such as classification. As a key technology to deal with the above limitations, hyperspectral image data dimensionality reduction technology is usually divided into two categories: feature extraction and band selection. Among them, band selection selects a certain number of optimal band combinations from the band set to replace the full band data to perform downstream tasks. The clustering-based band selection technology divides all bands into several clusters, selects one or more band combinations from each cluster as the optimal band set, and thus screens out key bands with strong representation ability and low correlation in the original bands.

[0003] Sun Genyun et al. ("Band Selection Algorithm Based on Inverse Neighbor Density Peak Clustering of Hyperspectral Images", Acta Geodaetica et Cartographica Sinica, 2024, 53(1):8-19.) constructed a directed graph based on the relationship between band neighbors and inverse neighbors, and constructed enhanced local density based on the average Euclidean distance and similarity between bands, and then selected the band subset by combining enhanced local density, distance information, and information entropy. However, this method only focuses on band-level similarity evaluation, ignoring the impact of pixel-level similarity evaluation on band correlation, and cannot fully reflect the unique information of single-band images, resulting in weak discriminative analysis capabilities for band images, affecting subsequent band selection and downstream task performance.

[0004] Most of the existing clustering-based hyperspectral band selection methods ignore the pixel-level processing of band images, resulting in inaccurate similarity evaluation results between bands and weak discriminative analysis capabilities of band images. Summary of the invention

[0005] Technical issues to be solved

[0006] In order to avoid the shortcomings of the prior art, the present invention proposes a band selection method based on unique pixel extraction and adaptive neighborhood clustering. To address this problem, a band selection method based on unique pixel extraction and adaptive neighborhood clustering is proposed. By introducing a low-rank self-representation model to analyze the common features and unique features between bands, unique pixel information is mined to enhance the distinguishability of bands, and an adaptive neighborhood band clustering model is established based on unique pixels. While simplifying the parameter setting, the Laplacian graph of similarity between bands is established and the structure is optimized, thereby improving the generalization performance of the algorithm for different spectral data.

[0007] Technical Solution

[0008] A band selection method based on unique pixel extraction and adaptive neighborhood clustering, characterized by the following steps:

[0009] Step 1: Pre-select several bands from the hyperspectral data cube H with dimensions a×b×l, and take the union of pre-selected bands by multiple methods to form a band pre-selected set Where a and b represent the length and width of a single-band hyperspectral image, respectively, and l is the number of spectral bands;

[0010] Step 2: Create a band pre-selection set H p A low-rank self-representation model of each hyperspectral image with a dimension of a×b in the image is used to extract a single-band image from the low-rank self-representation model. Low-rank essential information;

[0011] The low-rank self-representation model is:

[0012]

[0013] HkDJ i =H i Z i +E i

[0014] in, is the low-rank information of a single-band image, It is the unique pixel information of single-band image;

[0015] Then, the pixel information E of the single-band image is i Extract the data matrix Each row of the matrix Represents the unique pixel information of the hyperspectral image in the i-th band;

[0016] Step 3: Use unique pixel information Construct a band selection objective function based on adaptive neighborhood clustering:

[0017]

[0018] Among them, x i ,x j Respectively represent the unique pixels of the i-th and j-th spectral images; is the graph similarity matrix, is the graph Laplacian matrix, is the degree matrix; is the soft label matrix, c is the number of categories, that is, the number of bands finally selected;

[0019] Step 4: Use a two-step alternating optimization algorithm to solve the band selection objective function, select the band closest to the cluster center from each cluster, and combine all bands to obtain the optimal band set, which is the band selection result.

[0020] The band pre-selection set in step 1 The pre-selected set is obtained by using enhanced density peak clustering, maximum variance principal component analysis or information entropy sorting method.

[0021] The step 2 is based on the single-band image specific pixel information E i Extract the data matrix Yes: E i Stretched to a 1×n row vector, where n=a×b, thus converting the unique pixel information of p bands into a data matrix Each row of the matrix Represents the unique pixel information of the i-th band hyperspectral image.

[0022] The process of solving the band selection objective function using the two-step alternating optimization algorithm in step 4 is:

[0023] For the band selection objective function The process of solving the band selection objective function using the two-step alternating optimization algorithm is:

[0024] Step (1): Fix the soft label matrix F and update the similarity matrix W:

[0025] When the soft label matrix F is fixed, the optimization problem is equivalent to:

[0026]

[0027] Expanding the matrix trace term in the above problem into the form of 2-norm, the model is further transformed into:

[0028]

[0029] in,

[0030] Thus, the optimization problem becomes:

[0031]

[0032] The Lagrange multiplier method is used to solve the above optimization problem, and its Lagrange function is:

[0033]

[0034] Among them, φ≥0,ξ≥0;

[0035] According to the KKT condition, the closed-form solution is:

[0036]

[0037] in,(·) + =max(·,0);

[0038] D ij Sort in ascending order, according to w i Constraints to obtain:

[0039]

[0040] Among them, k is the number of neighboring points;

[0041] The solution to the similarity graph W is:

[0042]

[0043] Step (2): Fix the similarity matrix W and update the soft label matrix F

[0044] When the similarity matrix W is fixed, the optimization problem is equivalent to the spectral clustering problem:

[0045]

[0046] F T F=I

[0047] Solve the eigenvectors corresponding to the first c smallest eigenvalues ​​of matrix L and combine them to form the optimal solution of matrix F;

[0048] Step (3): By alternately updating the similarity matrix W and the soft label matrix F in steps (1) and (2), adaptive neighborhood clustering of hyperspectral bands is realized, and finally the category division of hyperspectral bands is obtained according to F. Then, the band closest to the cluster center is selected from each cluster, and all bands are combined to obtain the optimal band set, which is the band selection result.

[0049] A verification of a band selection method based on unique pixel extraction and adaptive neighborhood clustering is carried out, characterized in that: a support vector machine (SVM) classifier is selected, the data on the hyperspectral image dataset Indian Pines, the training sample ratio is set to 10%, the classification accuracy ACC is used as the evaluation indicator, and the average result of multiple repeated experiments is taken as the evaluation result.

[0050] The invention discloses an application of a band selection method based on unique pixel extraction and adaptive neighborhood clustering, characterized in that: band selection is performed on a hyperspectral image by using the band selection method based on unique pixel extraction and adaptive neighborhood clustering.

[0051] An electronic device, characterized in that it includes a processor and a memory, wherein the processor is used to implement the steps of data migration in the band selection method and verification when executing a computer program stored in the memory.

[0052] A readable storage medium, characterized in that a computer program is stored on the readable storage medium, and when the computer program is executed by a processor, the steps of data migration in the band selection method and verification are implemented.

[0053] A computer program product, characterized by comprising computer executable instructions, wherein the instructions are used to implement the band selection method and the method of data migration in verification when executed.

[0054] Beneficial Effects

[0055] The present invention proposes a band selection method based on unique pixel extraction and adaptive neighborhood clustering. A low-rank self-representation model is introduced to analyze the common and unique features between bands, unique pixel information is mined to enhance the distinguishability of bands, and an adaptive neighborhood band clustering model is established based on unique pixels. While simplifying the parameter setting, the Laplacian graph of similarity between bands is established and the structure is optimized, thereby improving the generalization performance of the algorithm for different spectral data.

[0056] The beneficial effects of the method of the present invention mainly include:

[0057] (1) By introducing a low-rank self-representation model, the present invention extracts the unique pixel information of the band image, establishes the inter-band similarity graph based on the unique information, and fully exploits the unique features of the band image. Compared with using full pixel information to evaluate the similarity between bands, the unique pixel information is used to enhance the distinguishability between bands and the discriminative analysis ability of the algorithm;

[0058] (2) The present invention adopts an adaptive neighborhood clustering method to perform band clustering, iteratively executes the Laplace graph construction and category matrix learning process, and gradually learns the Laplace similarity with the optimal structure through alternating optimization, thereby enhancing the adaptability and performance of band clustering to different spectral data.

[0059] (3) The unique pixel extraction technology proposed in the present invention can obtain the characteristic information of the band image by performing self-representation reconstruction processing on each band image, thereby enhancing the accuracy of the description of the similarity relationship between bands at the pixel level and enhancing the discriminative analysis capability of the band image. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 : Algorithm flow chart;

[0061] Figure 2 :When the number of optimal bands is 30, the prediction result of the SVM classifier on the optimal band set selected by the method in this paper. Specifically, the black part in the figure is the background pixel, and the other parts are the pixels of various ground objects. Among all kinds of ground objects, the present invention has a good recognition effect;

[0062] Figure 3 : Line graph of prediction accuracy of SVM classifier on full band and optimal band set selected by the method of this paper. Specifically, in the figure, baseline is the experimental result when the original full-band image is used as the classifier input data, and OurMethod is the experimental result when the optimal band subset selected by the band selection method of the present invention is used as the classifier input data. It can be seen from the figure that when the number of selected bands is greater than 12, the effect of using the optimal band set selected by the present invention to classify objects is better than that of using the full-band image. DETAILED DESCRIPTION

[0063] The present invention will now be further described with reference to the embodiments and the accompanying drawings:

[0064] The present invention is implemented by the following technical scheme: a band selection method based on unique pixel extraction and adaptive neighborhood clustering, taking the hyperspectral image dataset Indian Pines as an example to illustrate the specific steps of band selection by the proposed method, but the technical content of the present invention is not limited to the scope described. The hyperspectral image dataset Indian Pines consists of 220 spectral reflectance bands, and each band image pixel is 145×145. The dataset contains pixel-level category label information of 16 ground object instances, which is used for classification and verification of the performance of the band selection method.

[0065] The inter-band discriminative analysis capability is enhanced by extracting band-specific pixels, and the inter-band similarity graph with the optimal structure is constructed using adaptive neighborhood clustering, so as to achieve accurate evaluation of band correlation. The specific steps are as follows:

[0066] Step 1: Enter the Indian Pines data cube The number of band selection targets c, the regularization parameter α of the low-rank self-representation model, the number of neighbors k and the regularization parameter λ of the adaptive neighborhood clustering. According to the enhanced density peak clustering, maximum variance principal component analysis, information entropy sorting and other methods, c bands are pre-selected from H respectively, and the pre-selected bands of multiple methods are combined to form a band pre-selection set.

[0067] The dimension of the hyperspectral data cube H is a×b×l, where a and b represent the length and width of a single-band hyperspectral image, respectively, and l is the number of spectral bands.

[0068] Step 2: Pre-select bands based on the previous step Establishing a low-rank self-representation model to extract single-band images Unique pixel information And stretched to a 1×n row vector, where n = a×b, thereby converting the p band-specific information into a data matrix

[0069] For each hyperspectral image with dimension a×b, a low-rank self-representation model is established to extract a single-band image. The low-rank essential information, its low-rank self-representation model is expressed as follows:

[0070]

[0071] HkDJ i =H i Z i +E i

[0072] in, is the low-rank information of a single-band image,

[0073] is the pixel information unique to a single-band image. i Stretched to a 1×n row vector, where n=a×b, thus converting the unique pixel information of p bands into a data matrix Each row of the matrix Represents the unique pixel information of the i-th band hyperspectral image.

[0074] Step 3: Based on the band-specific pixel information matrix X in the previous step, the band selection objective function based on adaptive neighborhood clustering is constructed as follows:

[0075]

[0076] Among them, x i ,x j Respectively represent the unique pixels of the i-th and j-th spectral images; is the graph similarity matrix, is the graph Laplacian matrix, is the degree matrix; is the soft label matrix, c is the number of categories, and the present invention sets it to the number of bands finally selected. In the above minimization problem, the first two terms together constitute the neighborhood similarity graph construction model, and the third term is used to solve the sample category according to the constructed Laplacian graph, and then used to update the similarity graph, thereby realizing adaptive neighborhood clustering. We use a two-step alternating optimization algorithm to solve the above problem.

[0077] (4) Fix the soft label matrix F and update the similarity matrix W

[0078] When the soft label matrix F is fixed, the optimization problem is equivalent to:

[0079]

[0080]

[0081] Expanding the matrix trace term in the above problem into the form of 2 norm, the model can be further transformed into:

[0082]

[0083] in, Thus, the optimization problem becomes:

[0084]

[0085] The Lagrange multiplier method is used to solve the above optimization problem, and its Lagrange function is:

[0086]

[0087] Among them, φ≥0,ξ≥0. According to its KKT condition, the closed-form solution can be obtained as:

[0088]

[0089] in,(·) + =max(·,0). Further, d ij In ascending order, you can sort by w i Constraints to obtain:

[0090]

[0091] Among them, k is the number of neighboring points. Therefore, the solution of the similarity graph W can be obtained as:

[0092]

[0093] (5) Fix the similarity matrix W and update the soft label matrix F

[0094] When the similarity matrix W is fixed, the optimization problem is equivalent to the spectral clustering problem:

[0095]

[0096] F T F=I

[0097] At this time, the eigenvectors corresponding to the first c smallest eigenvalues ​​of matrix L are solved and combined to form the optimal solution of matrix F.

[0098] (6) By alternately updating the similarity matrix W and the soft label matrix F in steps (4) and (5), adaptive neighborhood clustering of hyperspectral bands is realized, and finally the category division of hyperspectral bands is obtained according to F. Then, the band closest to the cluster center is selected from each cluster, and all bands are combined to obtain the optimal band set, which is the band selection result.

[0099] In a specific embodiment, the band clustering soft label matrix is ​​initialized Combined with the data matrix X constructed in the previous step, initialize the inter-band adaptive distance evaluation matrix D 0 ,in

[0100] Update the Laplace similarity matrix W between bands, define t as the number of iterations, then:

[0101]

[0102] in,

[0103] Step 4: Based on the similarity matrix W in the previous step t , calculate the degree matrix With the Laplacian matrix For L t Perform eigenvalue decomposition and combine the eigenvectors corresponding to the smallest k eigenvalues ​​to obtain the soft label matrix F t .

[0104] Step 5: Repeat steps 3 and 4 until the algorithm results are stable and output the final band label matrix Perform k-means clustering on F row by row to obtain band categories. According to the band category classification results, select a band closest to the cluster center from each cluster to form an optimal band set. So far, the method proposed in this invention has completed the band selection task on the Indian Pines hyperspectral image dataset.

[0105] In order to verify the effectiveness of the band selection method of the present invention on the hyperspectral image dataset Indian Pines, the support vector machine SVM classifier is selected to perform the ground object classification task, the training sample ratio is set to 10%, the classification accuracy ACC is used as the evaluation index, and the average result of ten repeated experiments is taken as the experimental result. Baseline is the experimental result when the original full-band image is used as the classifier input data, and Our Method is the experimental result when the optimal band subset selected by the band selection method of the present invention is used as the classifier input data. Figure 2 This is the prediction result of the SVM classifier on the optimal band set selected by the method in this paper. The number of selected bands is 30. Figure 3 It is a comparison chart of the prediction accuracy of the SVM classifier on the full band and the optimal band set selected by the method of this paper. The number of bands ranges from 3 to 30, and the step size is 3. It can be seen from the comparison chart that the band selection method of the present invention combined with the classification method can obtain a higher classification accuracy. When the number of extracted bands is greater than or equal to 12, the corresponding SVM classification accuracy is higher than the classification accuracy using the full-band image.

[0106] The above description is only a specific implementation mode 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 various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A band selection method based on unique pixel extraction and adaptive neighborhood clustering, characterized in that Here are the steps: Step 1: Pre-select several bands from the hyperspectral data cube H with dimensions a×b×l, and take the union of pre-selected bands by multiple methods to form a band pre-selected set Where a and b represent the length and width of a single-band hyperspectral image, respectively, and l is the number of spectral bands; Step 2: Create a band pre-selection set H p A low-rank self-representation model of each hyperspectral image with a dimension of a×b in the image is used to extract a single-band image from the low-rank self-representation model. Low-rank essential information; The low-rank self-representation model is: s.d.H i =H i Z i +E i in, is the low-rank information of a single-band image, It is the unique pixel information of single-band image; Then, the pixel information E of the single-band image is i Extract the data matrix Each row of the matrix Represents the unique pixel information of the hyperspectral image in the i-th band; Step 3: Use unique pixel information Construct a band selection objective function based on adaptive neighborhood clustering: Among them, x i ,x j Respectively represent the unique pixels of the i-th and j-th spectral images; is the graph similarity matrix, is the graph Laplacian matrix, is the degree matrix; is the soft label matrix, c is the number of categories, that is, the number of bands finally selected; Step 4: Use a two-step alternating optimization algorithm to solve the band selection objective function, select the band closest to the cluster center from each cluster, and combine all bands to obtain the optimal band set, which is the band selection result.

2. The band selection method based on unique pixel extraction and adaptive neighborhood clustering according to claim 1, characterized in that: The band pre-selection set in step 1 The pre-selected set is obtained by using enhanced density peak clustering, maximum variance principal component analysis or information entropy sorting method.

3. The band selection method based on unique pixel extraction and adaptive neighborhood clustering according to claim 1, characterized in that: The step 2 is based on the single-band image specific pixel information E i Extract the data matrix Yes: Yes: E i Stretched to a 1×n row vector, where n=a×b, thus converting the unique pixel information of p bands into a data matrix Each row of the matrix Represents the unique pixel information of the i-th band hyperspectral image.

4. The band selection method based on unique pixel extraction and adaptive neighborhood clustering according to claim 1, characterized in that: The process of solving the band selection objective function using the two-step alternating optimization algorithm in step 4 is: For the band selection objective function The process of solving the band selection objective function using the two-step alternating optimization algorithm is: Step (1): Fix the soft label matrix F and update the similarity matrix W: When the soft label matrix F is fixed, the optimization problem is equivalent to: Expanding the matrix trace term in the above problem into the form of 2-norm, the model is further transformed into: in, Thus, the optimization problem becomes: The Lagrange multiplier method is used to solve the above optimization problem, and its Lagrange function is: Among them, φ≥0,ξ≥0; According to the KKT condition, the closed-form solution is: in,(·) + =max(·,0); D ij Sort in ascending order, according to w i Constraints to obtain: Among them, k is the number of neighboring points; The solution to the similarity graph W is: Step (2): Fix the similarity matrix W and update the soft label matrix F When the similarity matrix W is fixed, the optimization problem is equivalent to the spectral clustering problem: m F inTr(F T LF) s.t.F T F=I Solve the eigenvectors corresponding to the first c smallest eigenvalues ​​of matrix L and combine them to form the optimal solution of matrix F; Step (3): By alternately updating the similarity matrix W and the soft label matrix F in steps (1) and (2), adaptive neighborhood clustering of hyperspectral bands is realized, and finally the category division of hyperspectral bands is obtained according to F. Then, the band closest to the cluster center is selected from each cluster, and all bands are combined to obtain the optimal band set, which is the band selection result.

5. Verification of a band selection method based on unique pixel extraction and adaptive neighborhood clustering, characterized by: The support vector machine (SVM) classifier was selected for the data of the Indian Pines hyperspectral image dataset. The training sample ratio was set to 10%. The classification accuracy (ACC) was used as the evaluation indicator, and the average result of multiple repeated experiments was taken as the evaluation result.

6. An application of a band selection method based on unique pixel extraction and adaptive neighborhood clustering, characterized in that: Band selection is performed on hyperspectral images through a band selection method based on unique pixel extraction and adaptive neighborhood clustering.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor is used to implement the step of data migration in the band selection method and verification as claimed in any one of claims 1 to 6 when executing the computer program stored in the memory.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the step of data migration in the band selection method and verification as described in any one of claims 1 to 6 is implemented.

9. A computer program product, characterized in that The method comprises computer executable instructions, which are used to implement the band selection method and the method of data migration in verification as described in any one of claims 1 to 6 when executed.

Images