Partial view assisted discrimination hyperspectral band selection method and system

Abstract

The application discloses a local view auxiliary discrimination hyperspectral band selection method and system. Firstly, the hyperspectral image is segmented by using the superpixel segmentation technology ERS to form a series of superpixel blocks to capture the spectral feature difference of different ground object distribution; secondly, the application regards the band as the supergraph vertex, constructs the corresponding supergraph to express the multivariate adjacency relationship between the bands, reasonably optimizes the band adjacency structure as far as possible to guide the optimization of the sparse self-representation model, and reduces the redundancy of the band subset; wherein the consensus matrix fuses the coefficient matrix of each superpixel block through iterative updating; finally, the importance of each band to reconstruct the original data is calculated according to the consensus matrix, the band subset is selected, and the purpose of representing the hyperspectral image by using the unified band subset is realized; the spectral feature difference of the heterogeneous region is fully captured, the model pixel space local constraint can be practically enhanced, the quality of the band subset is improved, and the accuracy of subsequent classification is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of hyperspectral image dimension reduction, and particularly relates to a local view assisted discriminant hyperspectral band selection method and system. BACKGROUND

[0002] Hyperspectral remote sensing image has hundreds of adjacent and narrow spectral channels, can provide higher spectral resolution than RGB image, has rich spectral and spatial information, and is more conducive to accurate identification of ground objects. However, the nanoscale spectral resolution of the hyperspectral image also brings some problems to data processing, such as high computational complexity and information redundancy leading to waste of storage space.

[0003] In order to solve these problems, it is necessary to perform dimension reduction preprocessing on the hyperspectral image. Typical dimension reduction methods include feature extraction and band selection. Among them, feature extraction is to project the original high-dimensional data into a low-dimensional space, which changes the physical properties of the original data itself, and some key information is destroyed. Band selection is to select the most discriminative band subset from the original data. Compared with feature extraction, band selection can better preserve the information of the original hyperspectral data without changing the physical properties of the bands, so that the dimension-reduced data has higher interpretability and usability.

[0004] Band selection technology can be roughly divided into two categories: supervised and unsupervised. Supervised band selection needs certain prior information, such as training samples and corresponding labels. The difficulty and high cost of obtaining labels hinder the development of supervised band selection to some extent. In contrast, unsupervised band selection method does not need explicit labels, only uses unlabeled data to develop a learning model, providing a feasible solution for many band selection methods plagued by labels, and is more convenient to apply.

[0005] In recent years, the application of sparse representation theory to the field of hyperspectral has been proven to be reasonable. Sparse-based hyperspectral band selection increases the interpretability of the model, greatly reduces the redundancy of data, improves storage efficiency, and avoids unnecessary waste of resources. However, some existing researches only expand the function of the sparse self-representation model to a certain extent, ignore the influence of pixel spatial information on key feature extraction in the band selection process, and cannot accurately express the real multi-element adjacent structure when describing the band relationship, resulting in a low quality of the final selected band subset. SUMMARY

[0006] In view of the problem that the traditional sparse self-representation wave band selection method is insufficient in local spatial constraints of image elements in the wave band selection process, the application provides a local view assisted discrimination hyperspectral wave band selection method to sufficiently capture the spectral feature difference of heterogeneous regions, enhance the local spatial constraint of the model, improve the quality of the wave band subset, and further improve the accuracy of subsequent classification.

[0007] In order to achieve the above-mentioned purpose, the application adopts the following technical scheme:

[0008] The local view assisted discrimination hyperspectral wave band selection method comprises the following steps:

[0009] Step 1: according to the distribution characteristics of ground objects, the hyperspectral image is segmented by using the superpixel segmentation technology ERS to form pixel-level superpixel blocks for capturing the spectral feature difference of different ground object distributions;

[0010] Step 2: the local spectral-spatial hypergraph of each superpixel block is constructed in combination with the spatial proximity degree and the wave band spectral correlation to express the multivariate adjacency relationship between wave bands and rationalize the wave band adjacency structure;

[0011] Step 3: the local view assisted discrimination hyperspectral wave band selection model RwSSR is constructed in combination with the local spectral-spatial hypergraph of the superpixel block constructed in step 2 and the sparse self-representation model;

[0012] Step 4: the local view assisted discrimination hyperspectral wave band selection model RwSSR is optimized and solved by using the iterative updating method to obtain a consensus matrix and calculate the wave band priority to select a wave band subset.

[0013] In addition, on the basis of the local view assisted discrimination hyperspectral wave band selection method, the application further provides a local view assisted discrimination hyperspectral wave band selection system suitable therefor, which adopts the following technical scheme:

[0014] The local view assisted discrimination hyperspectral wave band selection system comprises:

[0015] The superpixel block segmentation module is used for segmenting the hyperspectral image by using the superpixel segmentation technology ERS according to the distribution characteristics of ground objects to form pixel-level superpixel blocks for capturing the spectral feature difference of different ground object distributions;

[0016] The local spectral-spatial hypergraph construction module is used for constructing the local spectral-spatial hypergraph of each superpixel block in combination with the spatial proximity degree and the wave band spectral correlation to express the multivariate adjacency relationship between wave bands and rationalize the wave band adjacency structure;

[0017] A hyperspectral band selection model construction module is configured to construct a local view assisted discriminative hyperspectral band selection model RwSSR by combining the constructed local spectral-spatial hypergraph of the superpixel block and the sparse self-representation model;

[0018] A band subset selection module is configured to obtain a consensus matrix and calculate a band priority by optimizing and solving the local view assisted discriminative hyperspectral band selection model RwSSR using an iterative update method, and select a band subset.

[0019] In addition, based on the above-described local view assisted discriminative hyperspectral band selection method, the application further provides a computer device, which comprises a memory and one or more processors.

[0020] The memory stores executable code, and the processor executes the executable code to implement the steps of the above-described local view assisted discriminative hyperspectral band selection method.

[0021] In addition, based on the above-described local view assisted discriminative hyperspectral band selection method, the application further provides a computer readable storage medium, which stores a program. The program is executed by a processor to implement the steps of the above-described local view assisted discriminative hyperspectral band selection method.

[0022] The application has the following advantages:

[0023] As described above, the application provides a local view assisted discriminative hyperspectral band selection method and system. Firstly, the application uses the superpixel segmentation technology ERS to segment a hyperspectral image to form a series of superpixel blocks to capture the spectral feature differences of different ground object distributions. Secondly, the application regards a band as a hypergraph vertex to construct a corresponding hypergraph to express the multivariate adjacency relationship between bands, and reasonably guides the optimization of the sparse self-representation model and reduces the redundancy of the band subset as much as possible. The consensus matrix is fused with the coefficient matrix of each superpixel block through the iterative update method. Finally, the importance of each band in reconstructing the original data is calculated according to the consensus matrix, and a band subset is selected, so as to realize the purpose of representing the hyperspectral image with a unified band subset. The spectral feature differences of heterogeneous regions are fully captured, the local constraints of the model in the pixel space are effectively enhanced, the quality of the band subset is improved, and the accuracy of subsequent classification is improved. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The figure is a flowchart of the local view assisted discriminative hyperspectral band selection method in the embodiments of the application.

[0025] Figure 2 The figure is a process diagram of segmenting a hyperspectral image using the superpixel segmentation technology ERS.

[0026] Figure 3 The average precision comparison chart of each waveband selection method in the embodiment of the application on the Indian Pines data set.

[0027] Figure 4 The overall precision comparison chart of each waveband selection method in the embodiment of the application on the Indian Pines data set. DETAILED DESCRIPTION

[0028] The application will be further described below in conjunction with the drawings and the specific embodiments:

[0029] Embodiment 1

[0030] This embodiment 1 describes a local view assisted discrimination hyperspectral waveband selection method to solve the problem of insufficient local constraints on the pixel space in the waveband selection process of the traditional sparse self-representation waveband selection method.

[0031] As shown in Figure 1 , the local view assisted discrimination hyperspectral waveband selection method comprises the following steps:

[0032] Step 1. According to the distribution characteristics of the ground objects, the hyperspectral image is segmented by using the superpixel segmentation technology ERS to form a pixel-level superpixel block for capturing the spectral feature difference of different ground object distributions.

[0033] The hyperspectral image is segmented by using the superpixel segmentation technology ERS, and the hyperspectral image is divided into a series of superpixel blocks X s with similar spectral features and non-overlapping, to represent different ground object coverages, and to select the corresponding sensitive waveband set for different ground objects, wherein S represents the number of superpixel blocks, 1≤s≤S.

[0034] ERS maps the first principal component of the hyperspectral image to a graph G(V,E); wherein the vertex set V contains all pixel points, and E is an edge set connecting the pixel points, and the weight w(e ij ) represents the similarity between adjacent pixel points, and the formula is as follows:

[0035]

[0036] wherein v i , v j both represent pixel points, v i , v j ∈V; e ij represents an edge connected by the pixel points v i and v j , e ij ∈E; G sLet it be the s-th superpixel block, 1≤s≤S, that is

[0037] The similarity between pixels within the same superpixel block is expressed as exp(-||v i -v j || / 2δ 2 The similarity between pixels in different superpixel blocks is calculated to be 0, where δ represents the kernel parameter.

[0038] Define a hyperspectral image X, X = [b1, b2, ..., b L ], b l Let b represent the l-th band, l∈[1,L], where L represents the number of bands, and each band b l Contains N pixels, i.e., b l =(x1,x2,...,x N ) T x n This represents the nth pixel, where n∈[1,N].

[0039] The objective function of ERS is defined as shown in equation (2), which seeks a subset A from the edge set E such that the graph It contains all Y connected subgraphs, and the final goal of image segmentation is achieved by removing some edges from the original edge set E.

[0040]

[0041] Among them, Figure This represents the image after ERS segmentation of the first principal component of the hyperspectral image.

[0042] H(A) is the graph The entropy rate of the random walk model ensures that the vertices in each cluster are compact and homogeneous; B(A) is a balance term describing the cluster distribution, ensuring that the size of each cluster is similar.

[0043] μ is a variable weighting factor used to coordinate the proportional relationship between H(A) and B(A). N A Representation diagram The number of connected subgraphs in the graph, where Y represents the preset number of connected subgraphs, and V and A represent the graphs respectively. The set of vertices and a subset of edges.

[0044] The process of segmenting hyperspectral images using the superpixel segmentation technique ERS is as follows: Figure 2 As shown.

[0045] The objective function of ERS can eliminate the connectivity between clusters, thereby segmenting and generating S superpixel blocks. The corresponding hyperspectral image X is then re-represented as X = [X1, X2, ... X].s ,...X S ],1≤s≤S,X s represents a superpixel block.

[0046] Step 2. Combining the spatial proximity and the spectral correlation of bands, a local spectral-spatial hypergraph of each superpixel block is constructed to express the multi-adjacency relationship between bands and rationalize the band adjacency structure.

[0047] Combining the spatial proximity and the spectral correlation of bands, a local spectral-spatial hypergraph of each superpixel block X s is constructed to form a complete hypergraph structure of the image.

[0048] The local spectral-spatial hypergraph of each superpixel block X s is denoted as wherein, represents the vertex set of the local spectral-spatial hypergraph , and represents the hyperedge set of the local spectral-spatial hypergraph .

[0049] The band vector is set as a vertex, and the vertex set is composed of L band vectors, e i represents a hyperedge composed of the K-neighbors of the band vector b i , and e i ∈E s .

[0050] The weight w(e i ) of the hyperedge e i is determined by the connection relationship between all vertices in the hyperedge e i , and the calculation formula is as follows:

[0051]

[0052] wherein, θ represents a balance parameter, and its expression is as follows:

[0053]

[0054] wherein, the band vector b j ∈V s is a vertex, and K represents the number of neighbors of the hyperedge e i . f(b i ) and f(b j ) respectively represent the vertices b i and b j ​​The integral function.

[0055] By b i and b i The superedge e formed around the center i The remaining vertices in the equation are determined, i.e.

[0056] Because the other vertices in the hyperedge are related to b i The affinity varies depending on vertex b. i and vertex b j The affinity between them is weighted as follows:

[0057]

[0058] Wherein, band vector b k ∈V s The vertex is the vertex.

[0059] f ij The value of measures both the spectral correlation of the band and the spectral correlation. and spatial proximity Right now

[0060] and As shown in formulas (4) and (5) respectively:

[0061]

[0062]

[0063] in For balancing parameters.

[0064] and These represent the band vectors b respectively. i and band vector b j Spatial index value.

[0065] Construct each local hypergraph The correlation matrix H s Weight matrix W s degree matrix of vertices and the degree matrix of the hyperedge

[0066]

[0067] W s =diag(w s (e1),w s (e2)…w s (e L)) (7)

[0068]

[0069]

[0070] where w s (e i ) denotes the weight of hyper-edge e i , i∈[1, L].

[0071]

[0072] w s (e j ) denotes the weight of hyper-edge e j , h s (b i ,e j ) denotes the incidence matrix of band vector b i , and diag(·) denotes the diagonal matrix.

[0073] So far, the local spectral-spatial hypergraph H is obtained, and the corresponding Laplacian matrix L s is as follows:

[0074]

[0075] where I is the identity matrix; L s The spectral characteristics of the band itself and the spatial similarity between the band sequence are used to more truly reflect the actual adjacency relationship between the bands in the local spectral-spatial hypergraph H .

[0076] Step 3. A local view assisted discriminative hyperspectral band selection model RwSSR is constructed by combining the local spectral-spatial hypergraph of the superpixel block constructed in step 2 and the sparse self-representation model.

[0077] The local view assisted discriminative hyperspectral band selection model RwSSR is obtained as follows:

[0078] Suppose X = [b1, b2,..., b L ] represents a hyperspectral image, which is composed of L bands, where each band contains N pixels b l = (x1, x2,..., x N ) T , and considering the band redundancy of the hyperspectral image, l 2,1 regularization is added to make the coefficient matrix row sparse, and as few bands as possible, i.e., dictionary columns, are used to approximate the data set X. Then the sparse self-representation model can be expressed as:

[0079]

[0080] Where A represents the sparse coefficient matrix, ||·|| F Let F-norm represent the sparse coefficient matrix. The formula for taking the F-norm is as follows: α represents the regularization parameter, ||·|| 2,1 Indicate l 2,1 The norm, A≥0, is used to ensure the non-negativity of A, and diag(A) is used to prevent it from being represented by itself.

[0081] For any superpixel block X after superpixel segmentation s Its objective function is defined as:

[0082]

[0083] Among them, A s It is every superpixel block X s The local coefficient matrix, A s This indicates the effect of the hyperspectral band on the superpixel block X. s The importance of reconstruction. In order to extract the band reconstruction information of the entire hyperspectral image, this invention uses all local coefficient matrices A s The consensus matrix A, formed by integrating the entire hyperspectral image, constrains the hyperspectral image as a whole, reflecting its local features and thus improving the accuracy of band selection. The objective function is further expressed as:

[0084]

[0085] Where λ1 represents the regularization parameter.

[0086] w s It is an adaptive balancing parameter used to ensure Minimize the overall impact. Specifically, if It is very large. In order to minimize this term, w s It should be reduced; conversely, it should be increased.

[0087] For simplicity, the parameter w is used in this embodiment. s Set to 1 / 2||A s -A||. Pass. The term refers to the coefficient matrix A of each superpixel block. s They are merged into a consensus matrix A, which contains key band information of superpixel blocks.

[0088] Because there is spectral similarity between bands and spatial similarity between adjacent sequence bands, which cannot be reflected in equation (13), equation (13) is optimized according to the definition of the spectral-space hypergraph as follows:

[0089]

[0090] where λ2 is the regularization parameter, L s is the Laplacian matrix of each superpixel block X s , A s is the coefficient matrix of each superpixel block X s , Tr(·) denotes the trace operator, and ||A|| 2,1 denotes the l 2,1 norm of the consensus matrix A.

[0091] So far, the model retains the spatial information at the pixel level and introduces the band local constraint term L s at the spectral level, so that the optimization process can comprehensively consider the information of the spatial and spectral spaces, and improve the solution of the subsequent sparse representation model.

[0092] Step 4. The local view aided discriminative hyperspectral band selection model RwSSR is optimized and solved by using the iterative updating method, the consensus matrix is obtained, and the band priority is calculated, and the band subset is selected.

[0093] The objective function is solved by using the iterative updating algorithm, and the process of obtaining A s , A, and w s is as follows:

[0094] Step 4.1. First, fix A and w s , update A s , and the fixed objective function is converted into:

[0095]

[0096] Take the derivative of A s , and the update equation of A s is obtained as:

[0097]

[0098] where n represents the current iteration number; A s (n+1) represents the value of A s in the n+1th iteration, A (n) represents the value of A in the nth iteration, and w s represents the value of w in the nth iteration.

[0099] Step 4.2. Secondly, fix A s and w s , update A, and the fixed objective function is converted into: ​

[0100]

[0101] Set U = diag(u1, u2, …, un) is an L x L diagonal matrix, L

[0102] where a i represents the i-th row of consensus matrix A, ||a i ||2represents the l2-norm of a i , and ζ is a number in u i to avoid zero denominator; accordingly, formula (17) is rewritten as:

[0103]

[0104] Take the derivative of A and let The update equation of A is obtained as:

[0105]

[0106] where A (n+1) represents the value of A at the n+1th iteration.

[0107] Step 4.3. Fix A s and A, update w s , and the update equation of w s is obtained as:

[0108]

[0109] where, represents the value of w s at the n+1th iteration.

[0110] When the iteration is updated to a specified number of iterations, or ||A (n+1) -A (n) || is less than a set threshold, stop, at this time, each row a i in the obtained consensus matrix A represents the contribution of the i-th wave band to the reconstruction of the original data X.

[0111] where r i = ||a i2 ||, the greater the value, the more important the wave band.

[0112] Therefore, r i is sorted in descending order, and the first n BS wave bands are selected as the final wave band subset to serve subsequent classification.

[0113] In addition, in order to verify the effectiveness of the method proposed in the present application, the following experiments are also carried out: ​

[0114] 1. Sample data: Indian pines hyperspectral data is from the Indian Pine Test Site in the United States, which is taken by an airborne visible infrared imaging spectrometer (AVIRIS). The data set has 220 bands, a spectral range from 0.4 μm to 2.5 μm, a size of 145x145x220, and a total of 16 target ground objects, including no-till corn, tillage soybean, forest, etc. The data set contains 21025 pixel points, of which the total number of target pixels is 10249, and the background contains 10776 pixel points.

[0115] 2. Experimental setup:

[0116] Comparison method: Support vector machine (SVM) is used as a classifier for classification, and seven different comparison methods are used to verify the reliability and accuracy of the method. The comparison methods are maximum variance principal component analysis (MVPCA), enhanced fast density peak clustering (E-FDPC), optimal class clustering framework (OCF), adaptive subspace partitioning strategy (ASPS), scalable one-pass self-learning method (SOP-SRL), graph regularized space-spectral subspace clustering (GRSC), and full bands.

[0117] Evaluation index: Two indexes are used to evaluate the quality of the band subset, namely average accuracy (AA) and overall accuracy (OA). The larger the value of the evaluation index, the better the classification effect and the more accurate the selected band subset. In the experiment, 10% of the samples are randomly selected as the training set, and the rest are used for testing. All experiments are repeated five times, and the average value is calculated.

[0118] 3. Parameter analysis:

[0119] There are five parameters in the experiment that need to be adjusted, which are S, K, λ1, λ2 and α. Among them, K is the number of neighbors of each hyperedge, and the parameter range is set to {3, 5, 7, 9}. The parameter range of λ1, λ2 and α is {1e -3 ,1e -2 ,1e -1 ,1,1e 1 ,1e 2 ,1e 3}. The range of superpixel segmentation block number S is {10, 50, 100}. The voting method is used to determine the optimal parameters. First, the optimal parameter set under different bands is selected, then the parameter with the most occurrences is selected by voting, and the result obtained under this parameter is recorded as the final result. For the Indian pines data set, the final values of the five parameters are shown in Table 1.

[0120] Table 1

[0121] Parameter S K ​ ​ α Optimum value 100 3 1e -1 ]] 1e -3 ]] 1e -3 ]]

[0122] 4. Experimental results:

[0123] Figure 3 and Figure 4 show the overall accuracy and average accuracy of the comparison results on the Indian pines data set under different numbers of bands, respectively. It can be seen from Figure 3 and Figure 4 that E-FDPC, MVPCA and OCF perform poorly and are in the downstream. The effect of SOP-SRL and ASPS is unstable. The effect of GRSC is stable, but the performance is not outstanding. Taking the number of bands 15 as a dividing point, the RwSSR method proposed in the application gradually catches up with other comparison methods when the number of bands is less than 15, and RwSSR is far ahead of other comparison methods when the number of bands is greater than 15, and gradually tends to be stable. This shows that RwSSR shows good performance and can select a band subset that is helpful for hyperspectral image classification. In summary, the method of the application has excellent performance in the band selection process and stable performance.

[0124] The hyperspectral band selection method described in Embodiment 1 first uses the superpixel segmentation technology ERS to segment the hyperspectral image to form a series of superpixel blocks to capture the spectral feature difference of different ground object distributions, then constructs a local spectral-spatial hypergraph of each superpixel block to express the multivariate adjacency relationship between bands, and reasonably structures the band adjacency structure as much as possible, and finally integrates the band information of the superpixel blocks through the consensus matrix, so as to realize the purpose of representing the entire hyperspectral image with a unified band subset, and effectively enhances the local constraint of the pixel space and improves the quality of the band subset, thereby effectively solving the problem that the traditional sparse self-representation model ignores the pixel space information in the band selection process.

[0125] Embodiment 2

[0126] Embodiment 2 describes a local view assisted discriminant hyperspectral band selection system, which is based on the same inventive concept as the local view assisted discriminant hyperspectral band selection method in Embodiment 1.

[0127] Specifically, the local view assisted discriminant hyperspectral band selection system comprises:

[0128] a superpixel block segmentation module configured to segment the hyperspectral image using the superpixel segmentation technology ERS according to the distribution characteristics of the ground objects to form pixel-level superpixel blocks for capturing the spectral feature difference of different ground object distributions;

[0129] A local spectral-spatial supergraph construction module is configured to construct a local spectral-spatial supergraph of each superpixel block to express the multi-adjacency relationship between bands and rationalize the band adjacency structure by combining the spatial proximity and the band spectral correlation.

[0130] A hyperspectral band selection model construction module is configured to construct a local view-assisted discriminative hyperspectral band selection model RwSSR by combining the constructed local spectral-spatial supergraph of the superpixel block and the sparse self-representation model.

[0131] A band subset selection module is configured to obtain a consensus matrix and calculate the band priority and select a band subset by using an iterative update method to optimize and solve the local view-assisted discriminative hyperspectral band selection model RwSSR.

[0132] It should be noted that the functions and effects of the various functional modules in the local view-assisted discriminative hyperspectral band selection system are specifically described in the implementation process of the corresponding steps in the method of Embodiment 1 above, and will not be repeated here.

[0133] Embodiment 3

[0134] This embodiment 3 describes a computer device for implementing the steps of the local view-assisted discriminative hyperspectral band selection method described in Embodiment 1 above.

[0135] The computer device includes a memory and one or more processors. The executable code is stored in the memory, and when the processor executes the executable code, the steps of the local view-assisted discriminative hyperspectral band selection method are implemented.

[0136] The computer device in this embodiment is any device or apparatus with data processing capability, which will not be repeated here.

[0137] Embodiment 4

[0138] This embodiment 4 describes a computer readable storage medium for implementing the steps of the local view-assisted discriminative hyperspectral band selection method described in Embodiment 1 above.

[0139] The computer readable storage medium in this embodiment 4 has a program stored thereon, and when the program is executed by a processor, the steps of the local view-assisted discriminative hyperspectral band selection method are implemented.

[0140] The computer readable storage medium can be an internal storage unit of any data processing capable device or apparatus, such as a hard disk or a memory, or an external storage device of any data processing capable device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc.

[0141] Of course, the above description is merely preferred embodiments of the present application, and the present application is not limited to the above-described embodiments. It should be understood that any equivalent substitutions or obvious modifications made by those skilled in the art based on the teachings of the present specification fall within the scope of the present specification, and should be protected by the present application.

Claims

1. A method for selecting hyperspectral bands using partial view-assisted discrimination, characterized in that, Includes the following steps: Step 1. Based on the distribution characteristics of ground features, the hyperspectral image is segmented using the superpixel segmentation technique ERS to form pixel-level superpixel blocks for capturing the spectral feature differences of different ground feature distributions; Step 2. Combining spatial proximity and band spectral correlation, construct a local spectral-spatial hypergraph for each superpixel block to express the multi-dimensional adjacency relationship between bands and rationalize the band adjacency structure. Step 3. Combining the local spectral-empty hypergraph of the superpixel block constructed in Step 2 and the sparse self-representation model, construct the local view-assisted discrimination hyperspectral band selection model RwSSR; Step 3 specifically involves: Combining the local spectral-empty hypergraph of superpixel blocks and the sparse self-representation model, a local view-assisted discrimination hyperspectral band selection model RwSSR is constructed, with the objective function defined as: Where λ1 and λ2 are regularization parameters, L s It is every superpixel block X s The Laplace matrix, A s It is every superpixel block X s The coefficient matrix, Tr(·) denotes the trace operator, avoiding each band being represented by itself; w s It is an adaptive balancing parameter used to ensure Overall minimization; α represents the regularization parameter, ||A|| 2,1 Row sparsity was applied to consensus matrix A; ||·|| F Describing the F-norm, through The term refers to the coefficient matrix A of each superpixel block. s The data are merged into a consensus matrix A, which contains the key band information of all superpixel blocks. Step 4. The local view-assisted discrimination hyperspectral band selection model RwSSR is optimized and solved using the iterative update method to obtain the consensus matrix, calculate the band priority, and select a subset of bands. Step 4 specifically involves: An iterative update algorithm is used to solve the objective function to obtain A. s A, w s The process is as follows: Step 4.

1. First, fix A and w. s Update A s After fixing, the objective function becomes: For A s Taking the derivative, we get A. s The update equation is: Where n represents the current iteration number; A s (n+1) Let A represent the (n+1)th iteration. s The value of A (n) This represents the value of A in the nth iteration. w represents the nth iteration s The value of ; Step 4.

2. Next, fix A. s and w s After updating A and fixing it, the objective function becomes: Set U = diag(u1, u2, ..., u L ) is an L×L diagonal matrix; Among them, a i Let ||a| represent the i-th row in consensus matrix A. i ||2 represents a i The l2 norm, is u i To avoid numbers with zero denominators, formula (14) is rewritten as follows: Differentiate A with respect to A and let The update equation for A is: Among them, A (n+1) This represents the value of A in the (n+1)th iteration; Step 4.

3. Fix A again s And A, update w s , get w s The update equation is: in, w represents the (n+1)th iteration s The value of ; When the iteration update reaches the specified number of iterations, or ||A (n+1) -A (n) The process stops when the value is less than the set threshold. At this point, each row a in the resulting consensus matrix A... i This represents the contribution of the i-th band to the reconstruction of the original data X; Where, r i =||a i The larger the value of ||2, the more important the band is; Therefore, for r i Sort in descending order and select the top n. BS One band is used as the final band subset.

2. The method for selecting hyperspectral bands with local view assistance according to claim 1, characterized in that, Step 1 specifically involves: Hyperspectral images are segmented using the superpixel segmentation technique ERS, dividing them into a series of non-overlapping superpixel regions X with similar spectral features. s To represent different land cover, select the corresponding sensitive band set for different land cover, where S is defined as the number of superpixel blocks, 1≤s≤S; ERS maps the first principal component of a hyperspectral image to a graph G(V,E); where the vertex set V contains all pixels, and E is the set of edges connecting the pixels, with weights w(e). ij The similarity between adjacent pixels is represented by the following formula: Among them, v i v j Both represent pixels, v i v j ∈V;e ij Represents pixel v i and v j One of the connected edges, e ij ∈E; G s Let it be the s-th superpixel block, 1≤s≤S, that is The similarity between pixels within the same superpixel block is expressed as exp(-||v i -v j || / 2δ 2 The similarity between pixels in different superpixel blocks is calculated to be 0; where δ represents the kernel parameter. Define a hyperspectral image X, X = [b1, b2, ..., b L ], b l Let b represent the l-th band, l∈[1,L], where L represents the number of bands, and each band b l Contains N pixels, i.e., b l =(x1,x2,...,x N ) T x n This represents the nth pixel, where n∈[1,N]; The objective function of ERS is defined as shown in equation (2), which seeks a subset A from the edge set E such that the graph The image is segmented by removing certain edges from the original edge set E, which contains all Y connected subgraphs. Among them, Figure This image represents the hyperspectral image after ERS segmentation of the first principal component. H(A) is the graph The entropy rate of the random walk model ensures that the vertices in each cluster are compact and homogeneous; B(A) is the balance term describing the cluster distribution, ensuring that the size of each cluster is similar. μ is a variable weighting factor used to coordinate the proportional relationship between H(A) and B(A); N A Representation diagram The number of connected subgraphs in the graph, where Y represents the preset number of connected subgraphs, and V and A represent the graphs respectively. The set of vertices and a subset of edges; The objective function of ERS can eliminate the connectivity between clusters, thereby segmenting and generating S superpixel blocks. The corresponding hyperspectral image X is then re-represented as X = [X1, X2, ... X]. s ,...X S ],1≤s≤S,X s This represents a superpixel block.

3. The method for selecting hyperspectral bands with local view assistance according to claim 2, characterized in that, Step 2 specifically involves: Combining spatial nearest neighbor and band spectral correlation, for each superpixel block X s Construct its corresponding local spectral-hyperspace graph To form a complete hypergraph structure for the image Each superpixel block X s Local spectral-hyperspace map Represented as in, Represents the local spectral-hypermap The set of vertices, Represents the local spectral-hypermap The set of superedges; Set band vector b i ∈V s Let V be the vertex set. s Composed of L band vectors, V s ={b1,b2,…b L }, i∈[1,L], e i Represents band vector b i A superedge consisting of K-nearest neighbors, e i ∈E s ; Super-edge e i weight w(e) i ) by superedge e i The connection relationships between all vertices within the interior are determined, and the calculation formula is as follows: Where θ represents the equilibrium parameter, its expression is as follows: Wherein, band vector b j ∈V s Let K be a vertex and K represent the hyperedge e. i The number of nearest neighbors of f(b); i ), f(b j ) represent vertices b in the hypergraph respectively. i b j Integral function; By b i and b i The superedge e formed around the center i The remaining vertices in the equation are determined, i.e. Because the other vertices in the hyperedge are related to b i The affinity varies depending on vertex b. i and vertex b j The affinity between them is weighted as follows: Wherein, band vector b k ∈V s For vertices; f ij The value of measures both the spectral correlation of the band and the spectral correlation. and spatial proximity Right now and As shown in formulas (4) and (5) respectively: in, For balance parameters; and These represent the band vectors b respectively. i and band vector b j Spatial index values; constructing each local spectral-space hypergraph The correlation matrix H s Weight matrix W s degree matrix of vertices and the degree matrix of the hyperedge W s =diag(w s (e1),w s (e2)…w s (e L (7) Among them, w s (e i ) represents the superedge e i Weights, i∈[1,L]; w s (e j ) represents the superedge e j The weight, h s (b i ,e j ) represents the band vector b i The incidence matrix, where diag(·) denotes a diagonal matrix; Thus, the local spectral-hyperspace map has been obtained. The corresponding Laplace matrix L s as follows: Where I is the identity matrix.

4. A local view-assisted hyperspectral band selection system for implementing the local view-assisted discrimination hyperspectral band selection method as described in any one of claims 1 to 3, characterized in that, The local view-assisted hyperspectral band selection system includes: The superpixel block segmentation module is used to segment hyperspectral images based on the distribution characteristics of ground features using the superpixel segmentation technology ERS, forming pixel-level superpixel blocks to capture the spectral feature differences of different ground feature distributions; The local spectral-space hypergraph construction module is used to combine spatial proximity and band spectral correlation to construct the local spectral-space hypergraph of each superpixel block, so as to express the multi-dimensional adjacency relationship between bands and rationalize the band adjacency structure. The hyperspectral band selection model construction module is used to combine the local spectral-space hypergraph of the constructed superpixel block with the sparse self-representation model to construct a local view to assist in the discrimination of the hyperspectral band selection model RwSSR. The module also includes a band subset selection module, which uses an iterative update method to optimize the local view-assisted discrimination hyperspectral band selection model RwSSR, obtain the consensus matrix, calculate the band priority, and select the band subset.

5. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements the steps of the local view-assisted discrimination hyperspectral band selection method as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the local view-assisted discrimination hyperspectral band selection method as described in any one of claims 1 to 3.

Images