Hyperspectral image classification method and system based on cat and mouse honeybee algorithm band selection

Abstract

The application discloses a hyperspectral image classification method and system based on cat-mouse honey horse algorithm band selection, and the method comprises the following steps: S1. Generating a band data set, and extracting the ground feature label corresponding to each band in the band data set to obtain a ground feature category label set; S2. Designing a cat-mouse honey horse algorithm based on the band data set, and sorting the fitness of cat-mouse honey horse individuals to extract representative band combinations in the band data set; S3. Training a 3DCNN-ViT neural network model based on the representative band combinations; S4. Collecting a to-be-processed hyperspectral image data set, and inputting the to-be-processed hyperspectral image data set into the 3DCNN-ViT neural network model for classification. The application selects the most representative band combination under the condition of ensuring high image classification accuracy, and increases the universality of band selection.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image processing, and particularly relates to a hyperspectral image classification method and system based on band selection of a cat-mouse honeybee algorithm. BACKGROUND

[0002] With the continuous leap of remote sensing technology, as an important part of remote sensing technology, hyperspectral remote sensing technology plays an important role in the development of remote sensing technology. Due to the characteristics of high resolution and rich spectral information, hyperspectral image classification has more extensive application. At the same time, hyperspectral image classification also faces some problems. Due to the fact that hyperspectral images will produce some redundant bands with less information and strong correlation, it is difficult to process the images and the classification accuracy is not high. Therefore, under the condition of ensuring high image classification accuracy, it is an important research content of hyperspectral to select more representative bands.

[0003] In recent years, the band selection method of hyperspectral image based on feature selection has developed rapidly, and the effective selection of band combination has been realized. Among them, the band selection method based on clustering is fast and simple to calculate, but it is very dependent on random seed initialization and has poor robustness. Although the band selection method based on sorting has the characteristics of high efficiency and simple data processing, this method only considers the condition of a certain band and does not optimize the whole band combination. The swarm intelligence algorithm has good robustness, considers the update of the whole band combination, and has the ability to handle high-dimensional complex engineering optimization problems, but it has the problems of low search ability, slow convergence speed and easy to fall into local optimum. To solve the above problems, the band selection method based on swarm intelligence and sorting can be combined, and the swarm intelligence algorithm can be improved, so that the selected band combination is more representative.

[0004] Through the search of the existing technical literature of the hyperspectral image band selection method based on swarm intelligence algorithm, it is found that although the "hyperspectral image band selection method based on evolutionary calculation" invented by Wang Zijun et al. in the invention (patent number: CN202210118319.9) and the "improved binary cuckoo algorithm for hyperspectral image band selection" published by Song Guangqin et al. in the Bulletin of Surveying and Mapping (2019(04):43-48) improve the swarm intelligence algorithm for band selection, they do not consider the selection of the most representative band combination according to the nature of the band. If the cat-mouse honeybee algorithm is designed considering multiple factors and setting them in the fitness function, the algorithm update formula is improved by introducing parameters, and the individual interaction mechanism is optimized by the sorting method to select the most representative band combination. Further, the band combination is used to train the neural network model, which is applied to the actual hyperspectral data set scene, and the universality of the method is improved. SUMMARY

[0005] The application aims to solve the problem that the search ability of existing swarm intelligence algorithm technology is low, the convergence speed is slow, and the locality is poor, resulting in that the wave band combination is not representative when selecting wave bands for hyperspectral images, and proposes a hyperspectral image classification method and system based on cat and mouse honey bee algorithm wave band selection, which provides the following scheme:

[0006] The hyperspectral image classification method based on cat and mouse honey bee algorithm wave band selection comprises the following steps:

[0007] S1. Generate a wave band data set, and extract the ground feature label corresponding to each wave band in the wave band data set to obtain a ground feature category label set;

[0008] S2. Design a cat and mouse honey bee algorithm based on the wave band data set, and sort the fitness of the cat and mouse honey bee individuals to extract a representative wave band combination in the wave band data set;

[0009] S3. Based on the representative wave band combination, train a 3DCNN-ViT neural network model;

[0010] S4. Collect a hyperspectral image data set to be processed, and input the hyperspectral image data set to be processed into the 3DCNN-ViT neural network model for classification.

[0011] Preferably, the S1 comprises:

[0012] The wave band data set X is:

[0013] X={X1,X2,…,X i ,…,X I}

[0014] Wherein, X i is the i-th wave band in X, i=1, 2,..., I, I is the number of wave bands;

[0015] The ground features corresponding to each wave band in the wave band data set are labeled to obtain the ground feature category label set Y:

[0016] Y={Y1,Y2,…,Y i ,…,Y I}

[0017] Y i ={y1,y2,…,y a ,y b， …,y k}

[0018] Wherein, Y i is the i-th wave band corresponding to the ground feature category label set in Y, y a and y b are respectively Y iwherein the number of the ground object species is a and b, a, b = 1, 2, …, k, and k is Y i wherein the number of the ground object species is a and b, a, b = 1, 2, …, k, and k is Y

[0019] Preferably, the S2 comprises:

[0020] The waveband dataset is taken as the initialization population of the cat-mouse honeybee algorithm, and the initialization population position information is calculated;

[0021] Based on the initialization population position information, the fitness function of the cat-mouse honeybee algorithm is calculated by using the Bhattacharyya distance;

[0022] Based on the fitness function, the initialization population position information is classified to obtain an alpha group and a candidate group;

[0023] The gas attraction factor and the development parameter are introduced to respectively update the positions of the alpha group and the candidate group and to sort the mixture to obtain a new alpha group and a new candidate group;

[0024] The new alpha group and the new candidate group are sorted, and the representative waveband combination is extracted based on the sorting result.

[0025] Preferably, the S3 comprises:

[0026] Based on the representative waveband combination, a training sample set is constructed;

[0027] An initial network model is constructed, and the initial network model is trained by using the training sample set to obtain the 3DCNN-ViT neural network model.

[0028] The application also provides a hyperspectral image classification system based on a cat-mouse honeybee algorithm waveband selection, comprising: a label extraction module, a calculation module, a training module, and a classification module;

[0029] The label extraction module is used to generate a waveband dataset and extract the ground object label corresponding to each waveband in the waveband dataset to obtain a ground object species label set;

[0030] The calculation module is used to design a cat-mouse honeybee algorithm based on the waveband dataset, to sort the fitness of the cat-mouse honeybee individuals, and to extract a representative waveband combination in the waveband dataset;

[0031] The training module is used to train a 3DCNN-ViT neural network model based on the representative waveband combination;

[0032] The classification module is used to collect a to-be-processed hyperspectral image dataset and input the to-be-processed hyperspectral image dataset into the 3DCNN-ViT neural network model for classification.

[0033] Preferably, the tag extraction module includes: a dataset generation unit and a first extraction unit;

[0034] The dataset generation unit is used to generate band dataset X:

[0035] X = {X1, X2, ..., X} i , ..., X I}

[0036] Among them, X i Let X be the i-th band, where i = 1, 2, ..., I, and I is the number of bands;

[0037] The first extraction unit is used to extract the land cover labels corresponding to each band in the band dataset, to obtain the land cover type label set Y:

[0038] Y = {Y1, Y2, ..., Y} i , ..., Y I}

[0039] Y i ={y1, 2, ..., y a y b , ..., y k}

[0040] Among them, Y i Let y be the set of land cover type labels corresponding to the i-th band in Y. a and y b Y i The a-th and b-th land cover types, a, b = 1, 2, ..., k, k is Y i The number of species of land cover in the region.

[0041] Preferably, the calculation module includes: a first calculation unit, a second calculation unit, a classification unit, an update unit, and a second extraction unit;

[0042] The first computing unit is used to use the band dataset as the initial population of the cat-mouse-honey horse algorithm and to calculate the initial population location information;

[0043] The second calculation unit is used to calculate the fitness function of the cat-mouse-honeyhorse algorithm based on the initial population location information and using Bach distance;

[0044] The classification unit is used to classify the initial population location information based on the fitness function to obtain alpha groups and candidate groups;

[0045] The updating unit is used to introduce gas attraction factors and development parameters to update the positions and perform mixed sorting of the alpha group and the candidate group respectively, so as to obtain a new alpha group and a new candidate group.

[0046] The second extraction unit is configured to sort the new alpha group and the new candidate group, and extract the representative band combination based on a sorting result.

[0047] Preferably, the training module comprises a training set generation unit and a model training unit.

[0048] The training set generation unit is configured to construct a training sample set based on the representative band combination.

[0049] The model training unit is configured to construct an initial network model, and train the initial network model using the training sample set to obtain the 3DCNN-ViT neural network model.

[0050] Compared with the prior art, the application has the following beneficial effects:

[0051] The application introduces a development parameter, increases the spatial range of the cat-mouse honeybee population search, solves the problem of low population search capability, combines a gas attraction factor, adds an attraction degree attribute to the cat-mouse honeybee individual, accelerates the convergence speed of the cat-mouse honeybee algorithm, optimizes the individual interaction mechanism, and avoids the cat-mouse honeybee individual from falling into a local optimum. Meanwhile, the attention mechanism is combined, a new three-dimensional convolutional neural network is designed, the network model is trained using the image sample set corresponding to the most representative band combination, the most representative band combination is selected under the condition of ensuring high image classification accuracy, and the universality of band selection is increased. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the application, the following briefly introduces the drawings needed to be used in the embodiments. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, on the premise of not paying the creative labor, can also obtain other drawings according to these drawings.

[0053] Figure 1 The method flowchart of the embodiment of the application;

[0054] Figure 2 The cat-mouse honeybee algorithm schematic diagram of the band selection method of the embodiment of the application;

[0055] Figure 3 The 3DCNN-ViT neural network model structure schematic diagram of the embodiment of the application;

[0056] Figure 4 The system structure schematic diagram of the embodiment of the application. DETAILED DESCRIPTION

[0057] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0058] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0059] Embodiment one

[0060] In this embodiment, as shown in the following, the hyperspectral image classification method based on the cat-and-mouse bee algorithm band selection includes the following steps: Figure 1

[0061] S1. Generating a band data set, and extracting the ground feature label corresponding to each band in the band data set to obtain a ground feature category label set.

[0062] Step S1 includes: generating a band data set X:

[0063] X={X1,X2,…,X i ,…,X I}

[0064] Wherein, X i is the i-th band in X, i=1, 2,..., I, I is the number of bands;

[0065] Extracting the ground feature label corresponding to each band in the band data set to obtain a ground feature category label set Y:

[0066] Y={Y1,Y2,…,Y i ,…,Y I}

[0067] Y i ={y1,y2,…,y a ,y b ,…,y k}

[0068] Wherein, Y i is the i-th band corresponding to the ground feature category label set in Y, y a and y b are the a-th and b-th ground feature categories in Y i , respectively, a, b=1, 2,..., k, k is the number of ground feature categories in Y i .

[0069] ​S2. Design the Cat-Mouse-Honey Horse algorithm based on the band dataset, sort the individual Cat-Mouse-Honey Horses by fitness, and extract representative band combinations from the band dataset.

[0070] Step S2 includes: using the band dataset as the initial population for the Cat-Mouse-Mickey Mouse algorithm and calculating the initial population location information; calculating the fitness function of the Cat-Mouse-Mickey Mouse algorithm based on the initial population location information using the Bach distance; classifying the initial population location information based on the fitness function to obtain alpha groups and candidate groups; introducing gas attraction factors and development parameters to update the positions and perform mixed sorting on the alpha groups and candidate groups respectively to obtain new alpha groups and new candidate groups; sorting the new alpha groups and new candidate groups, and extracting representative band combinations based on the sorting results.

[0071] In this embodiment, as Figure 2 As shown, step S2 specifically includes:

[0072] S2.1. Initialize the population parameters and location information of the cat-mouse-honeyhorse algorithm using all bands generated in step S1. Set the population size I and the maximum number of iterations T. max The search's lower bound Lb and upper bound Ub are defined, and the population location information is initialized as follows:

[0073] P i (t)=rand×(Ub-Lb)+Lb

[0074] In the formula, P i (t) represents the position of the i-th cat-mouse-honey horse individual in the t-th iteration of the population, t = 1, 2, ..., T. max rand is a random number in the range [0, 1].

[0075] S2.2. Using the two land cover types a and b corresponding to the i-th individual cat-mouse-honey-horse in step S2.1, set the inter-class distance. The formula is as follows:

[0076]

[0077] In the formula, norm a and norm b The mean values ​​of radiation intensity for a and b are respectively, Cov a and Cov b Let be the covariance matrices of a and b, respectively. T and(·) -1 These represent the transpose function and the inverse function, respectively.

[0078] S2.3. Utilize the steps in S2.2 Calculate the Parseltongue distance J BH_i (t), the formula is as follows:

[0079]

[0080] S2.4. Use the J BH_i (t) Set the fitness function of the cat-mouse honeybee algorithm as follows:

[0081]

[0082] where fitness i (t) is the fitness value of the ith cat-mouse honeybee individual.

[0083] S2.5. Use the fitness i (t) in step S2.4 to set the grouping probability γ as follows:

[0084]

[0085] S2.6. Use the γ in step S2.5 to divide the initialized population position information P α (t) into an alpha group P β (t) and a candidate group P

[0086]

[0087] S2.7. Update the position of P α (t) in step S2.6 as follows:

[0088]

[0089] where P α (t+1) is the position of the t+1th iteration, is the nth cat-mouse honeybee individual in P α (t), R3, R4, and R5 are three different random numbers between 0 and 1, χ n is a gas attraction factor, δ is a convergence factor, and O is a direction transformation parameter.

[0090] The specific design of χ n is as follows:

[0091]

[0092] where is the position of a random cat-mouse honeybee individual in P α (t), is the position of the n+1th cat-mouse honeybee individual in P α (t); and

[0093] The specific design of δ is as follows:

[0094]

[0095] The specific design of O is as follows:

[0096]

[0097] In the formula, R2 is a random number in the range of [0, 1].

[0098] S2.8. Repeat steps S2.6 to S2.7 until P α (t) is changed or the number of iterations is greater than 30, execute steps S2.9 and S2.11.

[0099] S2.9. Replace P β (t) in step S2.6 with P

[0100]

[0101] In the formula, P β (t+1) is the position of the t+1 iteration, ω and λ are random numbers between [0, 1], and s β_j is a development parameter, is the direction vector of the cat-mouse-honeybee individual, and C is the movement ability parameter.

[0102] s β_j The specific design of s

[0103]

[0104] In the formula, w and v are two different cat-mouse-honeybee individuals in P β (t), j, w, v = 1, 2, …, ε, and ε is the number of cat-mouse-honeybee individuals in P β (t), and are the fitness values of w and v, respectively, max(·) and ∑(·) represent the maximum function and the summation function, respectively.

[0105] The specific design of s

[0106]

[0107] The specific design of C is as follows:

[0108]

[0109] S2.10. Replace P α (t) in step S2.7 with P β(t) are mixed and ranked from small to large using the fitness values in steps S2.2 to S2.4, as follows:

[0110]

[0111] wherein P i-new (t) is the i'th cat-and-mouse honeybee individual in P i-new (t), i' = 1, 2, …, I.

[0112] S2.11. Group P i-new (t) into a new alpha group P α-new (t) and a new candidate group P β-new (t), as follows:

[0113]

[0114] S2.12. Repeat steps S2.7 to S2.11 until the number of iterations equals the maximum number of iterations T max , and execute step S2.13.

[0115] S2.13. Rank all cat-and-mouse honeybee individuals in P α-new (t) at the end of the iterations from large to small using the fitness values in steps S2.2 to S2.4, and take the top q cat-and-mouse honeybee individuals as the final output of the most representative band combination:

[0116]

[0117] wherein X end is the final output of the most representative band combination, is the number'th band in X end , number = 1, 2, …, q, and q is the number of bands.

[0118] S3. Train the 3DCNN-ViT neural network model based on the representative band combination.

[0119] Step S3 includes: constructing a training sample set based on the representative band combination; constructing an initial network model and training the initial network model using the training sample set to obtain the 3DCNN-ViT neural network model.

[0120] In this embodiment, as shown in FIG. 3, step S3 specifically includes: Figure 3

[0121] ​​S3.1. Construct the corresponding hyperspectral image training sample set train using the most representative waveband combination generated in step S2:

[0122] train = (Z1, Y1), (Z2, Y2), …, (Z number , Y number ), …, (Z q , Y q )}

[0123] In the formula, Z number is the image corresponding to the number waveband in train, and Y number is the ground object class label set corresponding to the number waveband.

[0124] S3.2. Use train in step S3.1 as the first layer of the initial network model to preprocess the samples:

[0125]

[0126] In the formula, is the output after the first layer preprocessing operation, and neighborhood(·) is the preprocessing operation.

[0127] S3.3. Convolve the output of step S3.2 as the input of the second layer convolution to obtain the output of the second layer convolution:

[0128]

[0129] In the formula, “*” represents the convolution operation, represents the output feature of the second layer, and respectively represent the weight parameters and the offset of the second layer, the μth input data, μ = 1, 2, …, q, and g(·) represents the sigmoid activation function.

[0130] S3.4. Design two convolution layers same as step S3.3, and take the output of step S3.3 as the input of the convolution layer to obtain the output of the fourth layer convolution layer.

[0131] S3.5. Take the output of step S3.4 as the input of two fully connected layers to obtain the output feature of the sixth layer.

[0132]

[0133] S3.6. Take the output of step S3.5 as the input of ViT:

[0134]

[0135] F ViT (·) represents a four-layer ViT unit structure calculation function, Δ 3DCNN-ViT represents the output of the four-layer ViT structure after the features extracted by the front-end 3DCNN.

[0136] S3.7. Obtain the classification probability of the training sample set by using the last layer:

[0137]

[0138] wherein, is the classification probability of each sample set label, and softmax(·) is the activation function of the last layer of the network.

[0139] S3.8. Update the weights φ and bias coefficients σ of the deep learning network by using the back propagation gradient algorithm until the network learns all the samples, end the training of the network model, save the converged model parameters, and obtain the 3DCNN-ViT neural network model.

[0140] S4. Collect the to-be-processed hyperspectral image data set, and input the to-be-processed hyperspectral image data set into the 3DCNN-ViT neural network model for classification. In actual application, first, the received hyperspectral image band data set is subjected to band selection through step S2 to obtain the most representative band combination. Then, the image sample set corresponding to the band combination is input into the neural network model designed in step S3 for training. Finally, the hyperspectral image data set other than the band combination in the actual reception is input into the trained model to obtain accurate ground object category classification information, and the classification accuracy is improved.

[0141] Example Two

[0142] In this embodiment, as shown in Figure 4 , the hyperspectral image classification system based on the cat-and-mouse honeybee algorithm band selection includes a label extraction module, a calculation module, a training module, and a classification module.

[0143] The label extraction module is used to generate a band data set and extract the ground object label corresponding to each band in the band data set to obtain a ground object category label set.

[0144] The label extraction module includes a data set generation unit and a first extraction unit. The data generation unit is used to generate a band data set X:

[0145] X = {X1, X2, …, X i , …, X I}

[0146] wherein X i is the i-th band in X, i = 1, 2, …, I, I is the number of bands;

[0147] The first extraction unit is configured to extract the ground object label corresponding to each band in the band dataset to obtain a ground object category label set Y:

[0148] Y = {Y1, Y2, …, Y i , …, Y I}

[0149] Y i = {y1, y2, …, y a , y b , …, y k}

[0150] wherein Y i is the ground object category label set corresponding to the i-th band in Y, y a and y b are the a-th and b-th ground object categories in Y i , respectively, a, b = 1, 2, …, k, k is the number of ground object categories in Y i .

[0151] The calculation module includes a first calculation unit, a second calculation unit, a classification unit, an update unit, and a second extraction unit.

[0152] The first calculation unit is configured to use the band dataset as the initialization population of the cat-mouse-honeybee algorithm and calculate the initialization population position information; the second calculation unit is configured to calculate the fitness function of the cat-mouse-honeybee algorithm based on the initialization population position information using the Bhattacharyya distance; the classification unit is configured to classify the initialization population position information based on the fitness function to obtain an alpha group and a candidate group; the update unit is configured to introduce a gas attraction factor and a development parameter to respectively update the position of the alpha group and the candidate group and mix the sorting to obtain a new alpha group and a new candidate group; and the second extraction unit is configured to sort the new alpha group and the new candidate group and extract a representative band combination based on the sorting result.

[0153] In this embodiment, the first calculation unit is configured to use all the bands generated in the step to initialize the population parameter and the position information of the cat-mouse-honeybee algorithm. The population size I, the maximum iteration number T max , the lower boundary of the search Lb, and the upper boundary of the search Ub are set, and the initialization population position information is:

[0154] P i (t) = rand × (Ub-Lb) + Lb

[0155] In the formula, P i(t) is the position of the i-th cat-and-mouse honeybee individual in the t-th iteration of the population, t = 1, 2, …, T max rand is a random number in the range of [0, 1].

[0156] The second calculation unit is configured to set the inter-class distance of the i-th cat-and-mouse honeybee individual corresponding to any two ground object classes a and b The formula is as follows:

[0157]

[0158] In the formula, norm a and norm b are the data mean of the radiation intensity of a and b respectively, Cov a and Cov b are the covariance matrix of a and b respectively, (·) T and (·) -1 represent the transpose function and the inverse function respectively.

[0159] Based on , the Bhattacharyya distance J BH_i (t) is calculated, and the formula is as follows:

[0160]

[0161] Based on J BH_i (t), the fitness function of the cat-and-mouse honeybee algorithm is set, and the formula is as follows:

[0162]

[0163] In the formula, fitness i (t) is the fitness value of the i-th cat-and-mouse honeybee individual.

[0164] The classification unit is configured to set the grouping probability γ using fitness i (t), and the formula is as follows:

[0165]

[0166] Based on the grouping probability γ, the initial population position information P i (t) is divided into an alpha group P α (t) and a candidate group P β (t), and the formula is as follows:

[0167]

[0168] The update unit is configured to update the position of P α (t), and the formula is as follows:

[0169]

[0170] where P α (t+1) is the position of the t+1th iteration, is P α the nth leech individual in P n is the gas attraction factor, δ is the convergence factor, and O is the direction transformation parameter.

[0171] χ n The specific design of χ

[0172]

[0173] where R2 is a random number in the range of [0, 1]. is P α the position of a random leech individual in P is P α the position of the n+1th leech individual in P

[0174] The specific design of δ is:

[0175]

[0176] The specific design of O is:

[0177]

[0178] where R2 is a random number in the range of [0, 1].

[0179] Repeat the above steps until P α (t) changes or the number of iterations is greater than 30.

[0180] The updating unit is also used to update the position of P β (t) according to the following formula:

[0181]

[0182] where P β (t+1) is the position of the t+1th iteration, ω and λ are random numbers in the range of [0, 1], and s β_j is the development parameter, is the direction vector of the leech individual, and C is the movement ability parameter.

[0183] s β_j The specific design of s

[0184]

[0185] where w and v are P β(t) two different cat-and-mouse honeybee individuals, j, w, v = 1, 2,..., ε, ε is P β (t) the number of cat-and-mouse honeybee individuals in (t), and are the fitness values of w and v respectively, max(·) and ∑(·) represent the maximum function and summation function respectively;

[0186] The specific design of C is as follows:

[0187]

[0188] The specific design of C is as follows:

[0189]

[0190] The obtained P α (t) and P β (t) are mixed, and all cat-and-mouse honeybee individuals are sorted in ascending order according to the fitness values, and the formula is as follows:

[0191]

[0192] In the formula, P i-new (t) is all populations after mixed sorting, is the i'th cat-and-mouse honeybee individual in P i-new (t), i' = 1, 2,..., I.

[0193] P i-new (t) is re-grouped to obtain a new alpha group P α-new (t) and a new candidate group P β-new (t), and the formula is as follows:

[0194]

[0195] The above steps are repeated until the iteration number is equal to the maximum iteration number T max .

[0196] The second extraction unit sorts all cat-and-mouse honeybee individuals in P α-new (t) after the iteration ends in descending order according to the fitness values, and takes the first q cat-and-mouse honeybee individuals as the final output of the most representative waveband combination:

[0197]

[0198] In the formula, X end is the final output of the most representative waveband combination, is X endThe number-th wave band in the train set, number = 1, 2,..., q, q is the number of wave bands.

[0199] The training module is configured to train the 3DCNN-ViT neural network model based on the representative wave band combination.

[0200] The training module includes a training set generation unit and a model training unit. The training set generation unit is configured to construct a training sample set based on the representative wave band combination; and the model training unit is configured to construct an initial network model and train the initial network model using the training sample set to obtain the 3DCNN-ViT neural network model.

[0201] In this embodiment, the training set generation unit is configured to construct a corresponding hyperspectral image training sample set train based on the generated most representative wave band combination:

[0202] train = {(Z1, Y1), (Z2, Y2),..., (Z number , Y number ),..., (Z q , Y q )}

[0203] In the formula, Z number is an image corresponding to the number-th wave band in the train set, and Y nmber is a ground object class label set corresponding to the number-th wave band.

[0204] The model training unit is configured to use the train as the first layer of the initial network model to pre-process the sample:

[0205]

[0206] In the formula, is the output after the first layer pre-processing operation, and neighborhood(·) is the pre-processing operation.

[0207] The output is used as the input of the second layer convolution for convolution operation:

[0208]

[0209] In the formula, "*" represents the convolution operation, represents the output feature of the second layer, and respectively represent the weight parameters and the offset of the second layer and the μ-th input data, μ = 1, 2,..., q, and g(·) represents a sigmoid activation function.

[0210] Two layers of convolution layers identical to the above convolution layer are designed, and the output The output of the fourth layer convolutional layer is obtained as the input of the convolutional layer

[0211] The output of the fourth layer convolutional layer is obtained as the input of the convolutional layer The output of the sixth layer is obtained as the input of the two-layer fully connected layer

[0212]

[0213] The output of the fourth layer convolutional layer is obtained as the input of the convolutional layer As the input of the ViT:

[0214]

[0215] In the formula, F ViT (·) represents a four-layer ViT unit structure calculation function, Δ 3DCNN-ViT The output of the four-layer ViT structure is obtained after the features extracted by the front-end 3DCNN.

[0216] The classification probability of the training sample set is obtained by using the last layer:

[0217]

[0218] In the formula, is the classification probability of each sample set label, and softmax(·) is the activation function of the last layer of the network.

[0219] The weights φ and the bias coefficients σ of the deep learning network are updated by using the back propagation gradient algorithm until the network learns all the samples, the training of the network model is ended, the model parameters after convergence are saved, and the 3DCNN-ViT neural network model is obtained.

[0220] The classification module is used for collecting a to-be-processed hyperspectral image data set and inputting the to-be-processed hyperspectral image data set into the 3DCNN-ViT neural network model for classification. In actual application, first, the actual received hyperspectral image band data set is subjected to band selection through step S2 to obtain a most representative band combination. Then, the image sample set corresponding to the band combination is input into the neural network model designed in step S3 for training. Finally, the hyperspectral image data set other than the band combination in the actual reception is input into the trained model to obtain accurate ground object category classification information, and the classification accuracy is improved.

[0221] The embodiments described above are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A hyperspectral image classification method based on the cat-mouse-honey-horse algorithm for band selection, characterized in that, Includes the following steps: S1. Generate a band dataset and extract the land cover labels corresponding to each band in the band dataset to obtain a land cover type label set; S2. Based on the band dataset, design the Cat-Mouse-Honey Horse algorithm, sort the individual Cat-Mouse-Honey Horses by fitness, and extract representative band combinations from the band dataset; S3. Based on the representative band combination, train a 3DCNN-ViT neural network model; S4. Acquire the hyperspectral image dataset to be processed, and input the hyperspectral image dataset to be processed into the 3DCNN-ViT neural network model for classification; S2 includes: The band dataset is used as the initial population for the cat-mouse-honey horse algorithm, and the location information of the initial population is calculated. Based on the initial population location information, the fitness function of the cat-mouse-honey horse algorithm is calculated using Bach distance; The initial population location information is classified based on the fitness function to obtain the alpha group and the candidate group; By introducing a gas attraction factor and development parameters, the positions of the alpha group and the candidate group are updated and mixed sorted to obtain a new alpha group and a new candidate group. The new alpha group and the new candidate group are sorted, and the representative band combination is extracted based on the sorting results.

2. The hyperspectral image classification method based on the cat-mouse-honey-horse algorithm for band selection according to claim 1, characterized in that, S1 includes: The band dataset X is: X={X1,X2,...,X i ,...,X I } Among them, X i Let X be the i-th band, where i = 1, 2, ..., I, and I is the number of bands; Extract the land cover labels corresponding to each band in the band dataset to obtain the land cover type label set Y: Y={Y1, Y2,…, Y i ,…,Y I } Y i ={y1,y2,…,y a ,y b ,…,y k } Among them, Y i Let y be the set of land cover type labels corresponding to the i-th band in Y. a and y b Y i The a-th and b-th land cover types, a, b = 1, 2, ..., k, k is Y i The number of species of land cover in the region.

3. The hyperspectral image classification method based on the cat-mouse-honey-horse algorithm for band selection according to claim 1, characterized in that, S3 includes: A training sample set is constructed based on the representative band combinations; An initial network model is constructed, and the initial network model is trained using the training sample set to obtain the 3DCNN-ViT neural network model.

4. A hyperspectral image classification system based on the cat-mouse-honey-horse algorithm for band selection, characterized in that, include: The module includes a label extraction module, a calculation module, a training module, and a classification module. The label extraction module is used to generate a band dataset and extract the land cover labels corresponding to each band in the band dataset to obtain a land cover type label set. The computing module is used to design the Cat-Mouse-Honey Horse algorithm based on the band dataset, sort the individual Cat-Mouse-Honey Horses by fitness, and extract representative band combinations from the band dataset. The training module is used to train a 3DCNN-ViT neural network model based on the representative band combination; The classification module is used to collect a dataset of hyperspectral images to be processed and input the dataset of hyperspectral images to be processed into the 3DCNN-ViT neural network model for classification. The calculation module includes: a first calculation unit, a second calculation unit, a classification unit, an update unit, and a second extraction unit; The first computing unit is used to use the band dataset as the initial population of the cat-mouse-honey horse algorithm and to calculate the initial population location information; The second calculation unit is used to calculate the fitness function of the cat-mouse-honeyhorse algorithm based on the initial population location information and using Bach distance; The classification unit is used to classify the initial population location information based on the fitness function to obtain alpha groups and candidate groups; The updating unit is used to introduce gas attraction factors and development parameters to update the positions and perform mixed sorting of the alpha group and the candidate group respectively, so as to obtain a new alpha group and a new candidate group. The second extraction unit is used to sort the new alpha group and the new candidate group, and extract the representative band combination based on the sorting result.

5. The hyperspectral image classification system based on the cat-mouse-honey-horse algorithm for band selection according to claim 4, characterized in that, The label extraction module includes: a dataset generation unit and a first extraction unit; The dataset generation unit is used to generate band dataset X: X={X1,X2,…,X i ,...,X I } Among them, X i Let X be the i-th band, where i = 1, 2, ..., I, and I is the number of bands; The first extraction unit is used to extract the land cover labels corresponding to each band in the band dataset, to obtain the land cover type label set Y: Y={Y1,Y2,…,Y i ,…,Y I } Y i ={y1,y2,…,y a ,y b ,…,y k } Among them, Y i Let y be the set of land cover type labels corresponding to the i-th band in Y. a and y b Y i The a-th and b-th land cover types, where a, b = 1, 2, ..., k, and k is Y. i The number of species of land cover in the region.

6. The hyperspectral image classification system based on the cat-mouse-honey-horse algorithm for band selection according to claim 4, characterized in that, The training module includes: a training set generation unit and a model training unit; The training set generation unit is used to construct a training sample set based on the representative band combination; The model training unit is used to construct an initial network model and train the initial network model using the training sample set to obtain the 3DCNN-ViT neural network model.

Images