A Low-Redundancy Hyperspectral Band Selection Method and Device Considering Both Representativeness and Information Content

Abstract

The present invention relates to the field of hyperspectral remote sensing image processing, and discloses a low-redundancy hyperspectral band selection method and device that take into account representativeness and information content. The method includes: (1) constructing a 3D reconstruction network to measure the representativeness of candidate band subsets for the original hyperspectral image; (2) measuring the redundancy of candidate band subsets; (3) measuring the information content contained in candidate band subsets; (4) designing a band subset scoring function that takes into account band representativeness, redundancy, and information content to evaluate candidate band subsets; and (5) generating a number of candidate band subsets and selecting the candidate band subset with the highest score as the selected band subset. Starting from the characteristics of hyperspectral images, the present invention explores the inherent non-linear correlation relationships between the bands of hyperspectral images, fully utilizes the spatial information of hyperspectral images, combines advanced deep learning knowledge, and proposes a band selection method that takes into account band representativeness, redundancy, and information content, which can improve the accuracy of pixel classification of hyperspectral images.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a low-redundancy hyperspectral band selection method and device that takes into account both representativeness and information content. Background Art

[0002] Hyperspectral images consist of hundreds of continuous bands and contain rich spatial and spectral information. However, the high-dimensional data also brings some challenges, such as information redundancy, heavy computational burden, and the "Hughes phenomenon". Therefore, it is crucial to reduce the dimension of hyperspectral images in a reasonable way and retain effective information for subsequent processing of hyperspectral images. The methods of dimensionality reduction are usually divided into feature extraction and band selection. The former way of dimensionality reduction will lead to the loss of physical information of the original data, while the latter can retain the physical information of the original data. Therefore, the band selection method has received extensive attention from scholars.

[0003] The existing band selection methods at home and abroad are roughly divided into the following four categories: (1) band selection methods based on point-by-point search; (2) band selection methods based on grouped search; (3) band selection methods based on ranking; (4) band selection methods based on advanced machine learning algorithms. These methods directly select the subset of bands containing the most useful information from the original hyperspectral image. However, the existing band selection methods mainly face three problems:

[0004] (1) The existing band selection methods cannot take into account the representativeness, redundancy, and information content of the bands. For example, the band selection methods based on ranking usually only consider the information content or representativeness of the bands, while ignoring the correlation between the bands.

[0005] (2) Most of the existing band selection methods cannot well analyze the inherent non-linear correlation relationship between the bands, usually only simply considering the linear correlation between the bands or non-linear correlation based on predefined kernel functions.

[0006] (3) The existing band selection methods based on autoencoders and genetic algorithms cannot accurately reflect the representativeness of the bands for the original image, and there is a problem that the spatial information of the hyperspectral image cannot be utilized. Summary of the Invention

[0007] Aiming at the deficiencies in the prior art, the purpose of the present invention is to provide a low-redundancy hyperspectral band selection method and device that takes into account both representativeness and information content, reveals the inherent non-linear correlation relationship between the bands of the hyperspectral image, makes full use of the spatial information of the hyperspectral image, studies a band selection strategy that can take into account the representativeness, redundancy, and information content of the bands, improves the extraction effect of the subset of bands containing rich valuable information, and solves the problem of low pixel classification accuracy caused by the inability to select the subset of bands with the most valuable information.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The present invention provides a low-redundancy hyperspectral band selection method that takes into account both representativeness and information content, including the following steps:

[0010] Step 1) Construct and train a 3D reconstruction network:

[0011] Divide the original hyperspectral image into blocks, and each hyperspectral image block is used as a sample; use the hyperspectral image block as the input of the 3D reconstruction network and train it;

[0012] Step 2) Measure the band representativeness:

[0013] Pass the hyperspectral image data block through a sparse binary mask, and the calculation formula is expressed as:

[0014]

[0015] where μ ∈ [0, 1] L represents the sparse binary mask, which is used to indicate whether each band is included in the candidate band subset, L represents the total number of bands, represents the multiplication symbol by band, X P represents the hyperspectral image block;

[0016] Rescale the data after passing through the mask; use the rescaled data as the input of the trained 3D reconstruction network to reconstruct the original hyperspectral data, and obtain the representativeness measure of the candidate band subset;

[0017] Step 3) Measure the band redundancy; the band redundancy measurement method uses the Pearson correlation coefficient or the vector subspace projection technique;

[0018] Step 4) Measure the band information content; the band information content measurement method uses the information divergence or the information entropy;

[0019] Step 5) Construct a comprehensive evaluation index that takes into account the band representativeness, redundancy, and information content, and the calculation formula is expressed as:

[0020]

[0021] where α and β represent the balance coefficients, R(X S ) represents the redundancy measurement value of the candidate band subset X S of, I(X S ) represents the information content measurement value of the candidate band subset X S of, represents the representativeness measure of the candidate band subset X S of, score(X S) represents the subset X of candidate wavelength bands S score of the comprehensive evaluation index;

[0022] Step 6) Search for the desired subset of wavelength bands:

[0023] Using the subset search strategy of the grouped search algorithm, calculate the score of the comprehensive evaluation index for each candidate subset of wavelength bands, and search for the subset of wavelength bands with the highest score as the selected subset; the grouped search algorithm uses the immune clone selection algorithm, genetic algorithm or particle swarm optimization algorithm.

[0024] Furthermore, the rescaling in step 2) refers to rescaling the data after masking according to the ratio between the total number of wavelength bands and the number of selected wavelength bands, and the calculation formula is expressed as:

[0025]

[0026] where k represents the number of selected wavelength bands, and Y P represents the rescaled data.

[0027] Furthermore, reconstruct the original hyperspectral data to obtain a representative measure of the candidate subset of wavelength bands, specifically:

[0028] Reconstruct the original hyperspectral data, and the formula is:

[0029]

[0030] where represents the reconstructed hyperspectral image block, F(·) represents the 3D reconstruction network, and θ represents the trainable parameters in the 3D reconstruction network;

[0031] Obtain a representative measure of the candidate subset of wavelength bands, and the formula is:

[0032]

[0033] where n represents the number of samples, represents the i-th input hyperspectral image block, represents the i-th reconstructed hyperspectral image block; ||·|| F represents the F-norm, represents the representative measure value.

[0034] Furthermore, the calculation formula for using the Pearson correlation coefficient to calculate the redundancy measure value of the candidate subset of wavelength bands is:

[0035]

[0036] where R(X S ) represents the candidate subset X of wavelength bands SRedundancy metric value, x s(i) represents the i-th band in the candidate wavelength subset, ||·|| represents the 2-norm, and the superscript T represents the transpose.

[0037] Furthermore, the formula for calculating the information measure value of the candidate wavelength subset using information divergence is:

[0038]

[0039]

[0040] where I(.) represents the information content, x s(i) represents the i-th band in the selected wavelength subset, X S represents the candidate wavelength subset, represents the i-th element after normalization of band x s(i) N represents the number of pixels, q i is the i-th element in the probability distribution obtained by normalizing the Gaussian distribution randomly initialized according to the mean and variance of band x s(i) and k represents the number of selected bands.

[0041] Furthermore, the immune clonal selection algorithm is used to search for the expected wavelength subset, specifically:

[0042] (1) Construct the initial population:

[0043] All bands in the original hyperspectral image are divided into k groups according to the band index number. One band is randomly selected from each group of bands to form an initial antibody, and each antibody represents a candidate wavelength subset containing k bands; repeat m times to generate an initial population composed of m initial antibodies, and calculate the comprehensive evaluation index score of the candidate wavelength subset corresponding to each antibody in the initial population;

[0044] (2) Cloning operation:

[0045] Copy each antibody in the population, and the number of times n c (X S(i) ) depends on its comprehensive evaluation index score, that is:

[0046]

[0047] where Floor(·) represents rounding down, n c (X S(i) ) represents the number of copies of the i-th antibody X S(i) ; score(X S(i) ) represents the comprehensive evaluation index score of the i-th antibody X S(i) ;

[0048] (3) Mutation operation:

[0049] Randomly select some elements from each copied antibody and replace them with an equal number of other candidate bands;

[0050] (4) Selection operation:

[0051] Select the m antibodies with the highest affinity from all antibodies to form a new population;

[0052] (5) Repeat steps (2)-(4) until the change in the maximum value of the comprehensive evaluation index score of the candidate band subset in the new population is less than the threshold τ.

[0053] The present invention also provides a low-redundancy hyperspectral band selection device that takes into account representativeness and information content for implementing the above hyperspectral band selection method; the hyperspectral band selection device includes:

[0054] A 3D reconstruction network training module for constructing and training a 3D reconstruction network;

[0055] A band representativeness measurement module for measuring the representativeness of a candidate band subset for the original hyperspectral image;

[0056] A band redundancy measurement module for measuring the redundancy of a candidate band subset;

[0057] A band information content measurement module for measuring the information content contained in a candidate band subset;

[0058] A comprehensive evaluation index construction module for designing a band subset scoring function to evaluate a candidate band subset;

[0059] An optimal band subset search module for generating a certain number of candidate band subsets, scoring the candidate band subsets, and selecting the band subset with the highest score as the selected band subset;

[0060] A band selection result output module for outputting the result of the selected optimal band subset.

[0061] Furthermore, the band selection device further includes an application module that uses the band selection result for hyperspectral image classification or target detection.

[0062] Furthermore, the 3D reconstruction network training module includes:

[0063] An image block division module for dividing the hyperspectral image into blocks, with each hyperspectral image block serving as a sample;

[0064] A 3D reconstruction network module for constructing a 3D reconstruction network and training the constructed 3D reconstruction network.

[0065] Furthermore, the optimal band subset search module includes:

[0066] An initial population building module is used to obtain the initial population;

[0067] A cloning module, used to clone antibodies in the population;

[0068] The mutation module is used to perform mutation operations on antibodies in the population;

[0069] A selection module, used to perform selection operations on antibodies in the population;

[0070] Iteration stop condition judgment module, used for repeated cloning module, mutation module and selection module, to judge whether iteration should be stopped according to the situation of the generated new population;

[0071] The band selection module is used to sort the affinity of antibodies in the population and select the band subset corresponding to the antibody with the highest affinity as the selected band subset.

[0072] The beneficial effects of the present invention are:

[0073] 1) In view of the problem that the prior art cannot take into account the representativeness, information content and redundancy of the selected band subset, the present invention proposes a comprehensive evaluation criterion for band subsets that takes into account the band representativeness, information content and redundancy. It can select a band subset that can well represent the original hyperspectral image, contains rich information and has low redundancy, which is conducive to the implementation of downstream tasks.

[0074] 2) The present invention proposes to use a 3D reconstruction network to measure the representativeness of a band subset, which can solve the problem that the existing technology is usually unable to explore the inherent nonlinear relationship between bands, and further improve the implementation effect of downstream tasks.

[0075] 3) The present invention proposes to rescale the data after passing through the mask, and use the rescaled data as the input of the trained 3D reconstruction network to reconstruct the original hyperspectral image. This can solve the problem in the existing technology that the reconstruction error cannot accurately represent the representativeness of the band subset, which is beneficial to improving the pixel classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 The present invention is a flowchart of the basic steps of an embodiment of a method for selecting low-redundancy hyperspectral bands that takes into account both representativeness and information content.

[0077] Figure 2 It is a structural schematic diagram of the hyperspectral image band selection device of the present invention.

[0078] Figure 3 This is a hyperspectral image in the Indian Pines dataset.

[0079] Figure 4 Classification accuracy of different band selection methods when using the SVM classifier on the Indian Pines dataset.

[0080] Figure 5 Classification accuracy of different band selection methods when using the EPF-G-g classifier on the Indian Pines dataset. Detailed implementation manners

[0081] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0082] As Figure 1 shown, it is a basic step flowchart of an embodiment of the hyperspectral image band selection method of the present invention, which specifically includes the following steps.

[0083] Input: The original hyperspectral image X ∈ R W×H×L , where W×H is the number of pixels and L represents the number of bands; select the number of bands k.

[0084] Step 1: Train a 3D reconstruction network.

[0085] (1) Divide the original hyperspectral image into blocks, and represent the hyperspectral image block as X P ∈ R a×a×L , a×a is the number of pixels after blocking, and each hyperspectral image block is used as a sample, and a total of n samples are obtained;

[0086] (2) Use the hyperspectral image block as the input to construct a 3D reconstruction network and appropriately train the network.

[0087] In this embodiment, the specific implementation steps of training the 3D reconstruction network are to input the hyperspectral image block into the constructed convolutional autoencoder, use the mean square error between the input data and the output data obtained by reconstruction as the loss function, and train for 10 epochs.

[0088] The output channels of the first three two-dimensional convolutional layers of the convolutional autoencoder are 128, 64, and 32 respectively. The size of the convolutional kernel is 3×3, and after each convolutional layer, batch normalization processing and ReLU activation function layer processing are required. Subsequently, the obtained data passes through three two-dimensional transposed convolutional layers with a kernel size of 3×3, and the sizes of the output channels are 64, 128, and the original number of bands L respectively. After each transposed convolutional layer, batch normalization processing and activation function layer processing are required. Among them, except for the last activation function layer using the Sigmoid activation function, the rest of the activation function layers use the ReLU activation function.

[0089] Step 2: Measure the band representativeness based on the 3D reconstruction network.

[0090] The hyperspectral image patch X P Pass through the sparse binary mask, and the calculation formula is expressed as:

[0091]

[0092] where μ ∈ [0,1] L represents the sparse binary mask, which is used to indicate whether each band is included in the candidate band subset, and L represents the total number of bands. represents the multiplication symbol by band, and X P represents the hyperspectral image patch.

[0093] Rescale the data passed through the mask according to the ratio between the total number of bands and the number of selected bands:

[0094]

[0095] where k represents the number of selected bands, and Y P represents the rescaled data.

[0096] Use the rescaled data as the input of the trained 3D reconstruction network to reconstruct the original hyperspectral data, and the expression is:

[0097]

[0098] where represents the reconstructed hyperspectral image patch, F(·) represents the 3D reconstruction network, and θ represents the trainable parameters in the 3D reconstruction network.

[0099] By calculating the reconstruction error, obtain the representativeness measure of the candidate band subset, and the calculation formula is expressed as:

[0100]

[0101] where n represents the number of samples, represents the i-th input hyperspectral image patch. represents the i-th reconstructed hyperspectral image patch; ||·|| F represents the Frobenius norm, denotes the representative metric value.

[0102] Step 3: Measure the redundancy of the candidate wavelength subset.

[0103] In this embodiment, the Pearson correlation coefficient is used to calculate the redundancy of the candidate wavelength subset. Specifically, first calculate the Pearson correlation coefficient between every two wavelengths in the candidate wavelength subset, and then take the average to obtain the redundancy metric value of the candidate wavelength subset. The calculation formula is expressed as:

[0104]

[0105] where X S = [x s(1) , x s(2) ,..., x s(k) ∈ R N×k represents a candidate wavelength subset in the two-dimensional data obtained by unfolding the original hyperspectral image X along the spatial dimension. N represents the number of pixels, R(X S ) represents the redundancy metric value of the candidate wavelength subset X S , x s(i) represents the i-th wavelength in the candidate wavelength subset, ||·|| represents the 2-norm, and the superscript T represents the transpose.

[0106] Step 4: Measure the information content of the candidate wavelength subset.

[0107] In this embodiment, the information divergence is used as the information content of the candidate wavelength subset. Specifically, first calculate the information divergence of each wavelength in the candidate wavelength subset:

[0108]

[0109] where I(x) represents the information content of wavelength x, represents the i-th element after normalizing wavelength x. Normalizing wavelength x means dividing each element in wavelength x by the sum of all elements. q = [q 1 , q 2 ,..., q N T ∈ R N×1 is a probability distribution obtained by randomly initializing a Gaussian distribution based on the mean and variance of x and then normalizing it. q i is the i-th element in the probability distribution.

[0110] Take the average of the information contents of all wavelengths in the candidate wavelength subset, that is, obtain the information content metric value of the candidate wavelength subset. The calculation formula is expressed as:​

[0111]

[0112] Step Five: Construct a comprehensive evaluation index that takes into account band representativeness, redundancy, and information content. The calculation formula is expressed as:

[0113]

[0114] where α and β represent balance coefficients.

[0115] Step Six: Search for the desired subset of wavebands based on the immune clone selection algorithm

[0116] (1) Construct the initial population:

[0117] Assume that a population consists of m antibodies; an antibody represents a candidate subset of wavebands containing k wavebands, denoted as X S(i) (i = 1,..., m).

[0118] First, evenly divide all the wavebands (L wavebands) in the original hyperspectral image X into k groups according to the waveband index number. If the total number of wavebands L cannot be divided evenly by k, then each of the first k - 1 groups contains wavebands, and the remaining wavebands are assigned to the last group, where Round(·) represents the ceiling function. Subsequently, randomly select one waveband from each group of wavebands to form an initial antibody. Repeat the operation of generating the initial antibody m times to generate the initial population.

[0119] After the initial population is generated, use the comprehensive evaluation index of the candidate subset of wavebands as the affinity A(X S(i) ) of the antibody, and generate a new population through three operations: cloning, mutation, and selection.

[0120] (2) Cloning operation:

[0121] Copy each antibody in the population. The number of times n c (X S(i) ) that the antibody is copied depends on its affinity, that is:

[0122]

[0123] where Floor(·) represents the floor function, and n c (X S(i) ) represents the number of times the i-th antibody is copied.

[0124] (3) Mutation operation:

[0125] Randomly select some elements from each copied antibody and replace them with an equal number of other candidate wavebands.

[0126] (4) Selection operation:

[0127] Select the m antibodies with the highest affinity from all antibodies to form a new population.

[0128] (5) Repeat steps (2)-(4) until the change in the maximum affinity of the candidate wavelength subset in the new population is less than the threshold τ.

[0129] (6) Take the candidate wavelength subset with the highest affinity in the population as the finally selected wavelength subset.

[0130] Corresponding to the embodiment of the low-redundancy hyperspectral band selection method that takes into account both representativeness and information content described above, the present application also provides an embodiment of a low-redundancy hyperspectral band selection device that takes into account both representativeness and information content, which includes:

[0131] A 3D reconstruction network training module for constructing a 3D reconstruction network and properly training it;

[0132] A band representativeness measurement module for measuring the representativeness of the candidate wavelength subset for the original hyperspectral image;

[0133] A band redundancy measurement module for measuring the redundancy of the candidate wavelength subset;

[0134] A band information content measurement module for measuring the information content contained in the candidate wavelength subset;

[0135] A comprehensive evaluation index construction module for designing a band subset scoring function to evaluate the candidate wavelength subset;

[0136] A band subset search module based on the immune clonal selection algorithm for generating a large number of candidate wavelength subsets, scoring the candidate wavelength subsets, and selecting the wavelength subset with the highest score as the selected wavelength subset;

[0137] A band selection result output module for outputting the selected optimal wavelength subset result.

[0138] In a specific implementation of the present invention, the band selection device further includes an application module, and the application module uses the band selection result for hyperspectral image classification or target detection.

[0139] In a specific implementation of the present invention, the 3D reconstruction network training module includes:

[0140] An image block division module for dividing the hyperspectral image into blocks, and each hyperspectral image block is used as a sample;

[0141] A 3D reconstruction network module for constructing a 3D reconstruction network and properly training the constructed 3D reconstruction network.

[0142] In a specific implementation of the present invention, the wavelength band subset search module is implemented based on an immune clonal selection algorithm, including:

[0143] An initial population construction module, configured to obtain an initial population;

[0144] A cloning module, configured to perform a cloning operation on the antibodies in the population;

[0145] A mutation module, configured to perform a mutation operation on the antibodies in the population;

[0146] A selection module, configured to perform a selection operation on the antibodies in the population;

[0147] An iteration stop condition judgment module, configured to repeat the cloning module, the mutation module, and the selection module, and judge whether to stop the iteration according to the situation of the generated new population;

[0148] A wavelength band selection module, configured to sort the affinities of the antibodies in the population, and select the wavelength band subset corresponding to the antibody with the highest affinity as the selected wavelength band subset.

[0149] Regarding the device in the above embodiments, the specific manner in which each unit or module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated herein.

[0150] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial description of the method embodiments. The device embodiments described above are merely illustrative. The so-called 3D reconstruction network training module and the wavelength band subset search module based on the immune clonal selection algorithm may or may not be physically separated. In addition, each functional module in the present invention may be integrated in a processing unit, or each module may exist physically alone, or two or more modules may be integrated in a unit. The above integrated modules or units may be implemented in the form of hardware or in the form of software functional units, and some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present application.

[0151] To verify the effect of the present invention, experiments were carried out on real hyperspectral images. Taking the IndianPines dataset as an example, the specific implementation manner is as follows. The experiment is as follows:

[0152] The image used in this embodiment is the Indian Pines dataset, which is a hyperspectral image captured by the AVIRIS sensor. As Figure 3 shown, the size is 145×145 pixels. After removing the water vapor absorption bands and the bands with low signal-to-noise ratio, 185 bands are left to participate in the experiment.

[0153] To further verify the application effect of the present invention, the results obtained by the method of the present invention and other methods are used for pixel classification. In order to more intuitively compare the influence of various band selection methods on the accuracy of downstream classification tasks, Figure 4 shows the variation curve of the overall classification accuracy of different band selection methods with the number of bands when using the SVM classifier on the Indian Pines dataset. The abscissa is the number of selected bands, and the ordinate is the classification accuracy. From Figure 4 it can be seen that for the SVM classifier, specific examples of the present invention can achieve significantly better classification effects than other band selection methods when selecting different numbers of bands.

[0154] Figure 5 shows the variation curve of the overall classification accuracy of different band selection methods with the number of bands when using the EPF-G-g classifier on the Indian Pines dataset. The abscissa is the number of selected bands, and the ordinate is the classification accuracy. From Figure 5 it can be seen that when selecting different numbers of bands, for the EPF-G-g classifier, the classification accuracy obtained by specific examples of the present invention when selecting different numbers of bands is higher than that obtained by other band selection methods.

[0155] Table 1 Comparison of classification performance (%) of different band selection methods on the SVM classifier

[0156] Method OA AA 1. MVPCA 64.81 50.83 2. LCMVBCC 58.95 49.74 3. LCMVBCM 66.90 60.98 4. ECA 75.16 65.25 5. MR 78.42 71.24 6. BS-Net-Conv 78.91 72.27 7. The present invention 80.36 74.63

[0157] Taking k = 15 as an example, Table 1 gives the classification accuracy obtained by different band selection methods when using the SVM classifier on the Indian Pines dataset. OA (Overall Accuracy) in the table refers to the ratio of correctly classified samples to the total samples, and AA (Average Accuracy) refers to the average value of the accuracy rates obtained for each category. The results in Table 1 show that the low redundancy hyperspectral band selection method of the present invention that takes into account representativeness and information content can improve the classification effect of band selection.

[0158] The present invention describes specific embodiments to simplify the present invention. However, it should be recognized that the present invention is not limited to the illustrated embodiments, and various modifications of the present invention are possible without departing from the basic principles, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

Claims

1. A low-redundancy hyperspectral band selection method that takes into account both representativeness and information content, characterized in that, it includes the following steps: Step 1) Construct and train a 3D reconstruction network: Divide the original hyperspectral image into blocks, and each hyperspectral image block is used as a sample; use the hyperspectral image block as the input of the 3D reconstruction network and train it; Step 2) Measure the band representativeness: Pass the hyperspectral image data block through a sparse binary mask, and the calculation formula is expressed as: where μ ∈ [0, 1] L represents a sparse binary mask for indicating whether each band is included in the candidate band subset, L represents the total number of bands, represents the multiplication symbol by band, X P represents a hyperspectral image patch; Rescale the data after passing through the mask; use the rescaled data as the input of the trained 3D reconstruction network to reconstruct the original hyperspectral data, and obtain the representativeness measure of the candidate band subset; Step 3) Measure the band redundancy; the band redundancy measurement method uses the Pearson correlation coefficient or the vector subspace projection technique; Step 4) Measure the band information content; the band information content measurement method uses the information divergence or the information entropy; Step 5) Construct a comprehensive evaluation index that takes into account the band representativeness, redundancy, and information content, and the calculation formula is expressed as: score(X S ) = -l - αR(X S ) + βI(X S ) Among them, α and β represent balance coefficients, R(X S ) represents the redundancy metric value of the candidate wavelength subset X S , I(X S ) represents the information metric value of the candidate wavelength subset X S , l represents the representativeness metric of the candidate wavelength subset X S , score(X S ) represents the score of the comprehensive evaluation index of the candidate wavelength subset X S ; Step 6) Search for the desired band subset: Use the subset search strategy of the grouped search algorithm to calculate the comprehensive evaluation index score of each candidate band subset, and search for the band subset with the highest score as the selected band subset; the grouped search algorithm uses the immune clonal selection algorithm, the genetic algorithm, or the particle swarm optimization algorithm.

2. The low-redundancy hyperspectral band selection method that takes into account both representativeness and information content according to claim 1, characterized in that, the rescaling in step 2) refers to rescaling the data after passing through the mask according to the ratio between the total number of bands and the number of selected bands, and the calculation formula is expressed as: Among them, k represents the number of selected bands, and Y P represents the rescaled data.

3. The low-redundancy hyperspectral band selection method that takes into account both representativeness and information content according to claim 1, characterized in that, the reconstruction of the original hyperspectral data to obtain the representativeness measure of the candidate band subset is specifically: Reconstruct the original hyperspectral data, and the formula is: Among them, represents the reconstructed hyperspectral image patch, F(·) represents the 3D reconstruction network, and θ represents the trainable parameters in the 3D reconstruction network; Obtain the representativeness measure of the candidate band subset, and the formula is: where n represents the number of samples, represents the i-th input hyperspectral image patch, represents the i-th reconstructed hyperspectral image patch; ||·|| F represents the Frobenius norm, and l represents the representative metric value.

4. The low-redundancy hyperspectral band selection method that takes into account both representativeness and information content according to claim 1, characterized in that, the calculation formula for calculating the redundancy measure value of the candidate band subset using the Pearson correlation coefficient is: where, R(X S ) represents the redundancy metric value of the candidate waveband subset X S , x s(i) represents the i-th waveband in the candidate waveband subset, ||·|| represents the 2-norm, and the superscript T represents the transpose.

5. The low-redundancy hyperspectral band selection method that takes into account both representativeness and information content according to claim 1, characterized in that, the calculation formula for calculating the information content measure value of the candidate band subset using the information divergence is: where, I(.) represents the amount of information, and x s(i) represents the i-th band in the selected band subset, X S represents the candidate band subset, represents the i-th element after normalization of band x s(i) N represents the number of pixels, and q i is the i-th element in the probability distribution obtained by normalizing the Gaussian distribution randomly initialized according to the mean and variance of band x s(i) and k represents the number of selected bands.

6. The low-redundancy hyperspectral band selection method that takes into account both representativeness and information content according to claim 1, characterized in that, using the immune clonal selection algorithm to search for the desired band subset, specifically: (1) Construct an initial population: Divide all the bands in the original hyperspectral image into k groups according to the band index number, randomly select one band from each group of bands to form an initial antibody, and each antibody represents a candidate band subset containing k bands; repeat m times to generate an initial population composed of m initial antibodies, and calculate the comprehensive evaluation index score of the candidate band subset corresponding to each antibody in the initial population; (2) Cloning operation: Copy each antibody in the population, and the number of times n for antibody replication c (X S(i) ) depends on its comprehensive evaluation index score, that is: Among them, Floor(·) represents rounding down, and n c (X S(i) ) represents the replication times of the i-th antibody X S(i) ; score(X S(i) ) represents the comprehensive evaluation index score of the i-th antibody X S(i) ; (3) Mutation operation: Randomly select some elements from each replicated antibody and replace them with an equal amount of other candidate bands; (4) Selection operation: Select the m antibodies with the highest affinity from all antibodies to form a new population; (5) Repeat steps (2)-(4) until the change in the maximum value of the comprehensive evaluation index score of the candidate band subset in the new population is less than the threshold τ.

7. A low-redundancy hyperspectral band selection device that takes into account both representativeness and information content, characterized in that, it is used to implement the hyperspectral band selection method described in claim 1; the hyperspectral band selection device includes: A 3D reconstruction network training module for constructing and training a 3D reconstruction network; A band representativeness measurement module for measuring the representativeness of a candidate band subset for the original hyperspectral image; A band redundancy measurement module for measuring the redundancy of a candidate band subset; A band information content measurement module for measuring the information content contained in a candidate band subset; A comprehensive evaluation index construction module for designing a band subset scoring function to evaluate a candidate band subset; An optimal band subset search module for generating a certain number of candidate band subsets, scoring the candidate band subsets, and selecting the band subset with the highest score as the selected band subset; A band selection result output module for outputting the result of the selected optimal band subset.

8. The low-redundancy hyperspectral band selection device according to claim 7 that takes into account both representativeness and information content, characterized in that, the band selection device further includes an application module, and the application module uses the band selection result for hyperspectral image classification or target detection.

9. The low-redundancy hyperspectral band selection device according to claim 7 that takes into account both representativeness and information content, characterized in that, the 3D reconstruction network training module includes: An image block division module for dividing a hyperspectral image into blocks, and each hyperspectral image block is used as a sample; A 3D reconstruction network module for constructing a 3D reconstruction network and training the constructed 3D reconstruction network.

10. The low-redundancy hyperspectral band selection device according to claim 7 that takes into account both representativeness and information content, characterized in that, the optimal band subset search module includes: An initial population construction module for obtaining an initial population; A cloning module for performing a cloning operation on the antibodies in the population; A mutation module for performing a mutation operation on the antibodies in the population; A selection module for performing a selection operation on the antibodies in the population; An iteration stop condition judgment module for repeating the cloning module, the mutation module, and the selection module, and judging whether to stop the iteration according to the situation of the generated new population; A band selection module for sorting the affinities of the antibodies in the population and selecting the band subset corresponding to the antibody with the highest affinity as the selected band subset.

Images