Cross-domain hyperspectral band selection method and device, equipment and storage medium

Abstract

The invention is applicable to the technical field of remote sensing information processing, and provides a cross-domain hyperspectral band selection method, device and equipment and a storage medium, and the method comprises the steps: initializing a particle swarm according to the total number of bands of a target domain, and carrying out the grouping division of hyperspectral bands in the target domain according to the total number of bands and the number of target groups, calculating the fitness value of each particle in the particle swarm at the current position in a solution space formed by the target hyperspectral band group by using a neural network model obtained based on source domain training, judging whether an optimization termination condition is met or not, if yes, executing the step 2, and if not, executing the step 3; and determining the particle with the highest fitness value in the particle swarm as the optimal wave band combination of the target domain, otherwise, updating the position information and the speed information of each particle in the particle swarm according to the fitness value, and continuing to search and optimize in the search space, thereby reducing the hyperspectral data labeling cost and improving the hyperspectral data labeling efficiency. And the efficiency and the effect of selecting the most representative wave band from the hyperspectral data are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing information processing, and in particular relates to a cross-domain hyperspectral band selection method, device, equipment and storage medium. Background Art

[0002] Hyperspectral imaging (HSI) is a technology that combines spectral analysis and traditional imaging. It can simultaneously obtain spatial information and continuous, detailed spectral information of the target. This technology can not only provide two-dimensional image information of the object, but also reveal the chemical composition and physical properties of the object through spectral information. Unlike ordinary cameras (which only capture red, green, and blue channels) or multispectral imaging (a few discrete bands), hyperspectral imaging can cover dozens to hundreds of continuous narrow bands (such as visible light, near-infrared, or mid-infrared ranges), forming a "data cube" (containing spatial dimensions). and spectral dimensions (wavelength)), thereby revealing the unique spectral signature of a substance.

[0003] At present, hyperspectral imaging technology has been widely used in the field of remote sensing, especially in tasks such as surface analysis and classification. Traditional band selection methods are widely used to select the most representative bands from hyperspectral data, thereby reducing the dimension of the data and improving classification accuracy. However, existing band selection methods have several problems when processing hyperspectral data. First, the computational efficiency is low: existing band selection methods usually rely on fixed metrics such as information entropy or coefficient of variation. These methods have high computational costs when processing high-resolution, high-dimensional UAV hyperspectral data, especially when the number of bands and samples increases significantly, which may lead to a serious decline in computational efficiency; second, the high cost of hyperspectral data annotation: the hyperspectral data annotation process usually requires professional knowledge, and the annotation cost is higher in complex scenes. As the scale and dimension of UAV hyperspectral data continue to increase, traditional supervised learning methods face the problems of insufficient labeled data and excessively high annotation costs; third, insufficient use of prior knowledge: existing band selection methods usually ignore prior knowledge from similar areas. Although data from different regions may have similar ground object categories and spectral characteristics, traditional methods do not fully utilize these similarities for band selection, resulting in potential optimization opportunities not being effectively utilized. Summary of the invention

[0004] The purpose of the present invention is to provide a cross-domain hyperspectral band selection method, device, equipment and storage medium, aiming to solve the problems of low accuracy and efficiency and high labeling cost of cross-domain hyperspectral band selection caused by the prior art.

[0005] In one aspect, the present invention provides a cross-domain hyperspectral band selection method, the method comprising the following steps: Initialize the particle swarm according to the total number of bands in the target domain; According to the total number of bands and a preset number of target groups, the hyperspectral bands in the target domain are grouped and divided to obtain target hyperspectral band groups; Using a neural network model obtained through source domain training, calculating the fitness value of each particle in the particle swarm at a current position in a solution space, wherein the solution space is composed of the target hyperspectral band group; When a preset optimization termination condition is reached, according to the fitness value, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain; When the optimization termination condition is not met, the position information and speed information of each particle in the particle swarm are updated according to the fitness value, and the process jumps to the step of calculating the fitness value of each particle in the particle swarm at the current position in the solution space by using the neural network model obtained by source domain training.

[0006] Preferably, the step of grouping and dividing the hyperspectral bands in the target domain comprises: According to the segmentation point of each group, the hyperspectral bands in the target domain are divided into groups to obtain hyperspectral band groups corresponding to the target number of groups; According to the spectral similarity between the bands, the hyperspectral bands included in each of the hyperspectral band groups are adjusted to obtain the target hyperspectral band group.

[0007] Preferably, the step of grouping the hyperspectral bands in the target domain according to the segmentation point of each group comprises: According to the total number of bands and the target number of groups, a preset split point calculation formula is used to calculate the split point for grouping, wherein the split point calculation formula is: , represents the total number of bands, represents the target group number, , Indicates The split point of each group.

[0008] Preferably, the step of adjusting the hyperspectral bands included in each of the hyperspectral band groups according to the spectral similarity between the bands comprises: According to all the hyperspectral band groups, calculating the similarity distances of the boundary bands between adjacent groups; According to the similarity distance, a preset segmentation point adjustment strategy is adopted to adjust the segmentation points between the corresponding adjacent groups, and the hyperspectral bands in the corresponding hyperspectral band group are updated, and the process jumps to the step of calculating the similarity distances of the boundary bands between adjacent groups according to all the hyperspectral band groups until the preset band adjustment termination condition is reached.

[0009] Preferably, before the step of calculating the fitness value of each particle in the particle swarm at the current position in the solution space using the neural network model obtained based on source domain training, the method further comprises: Using a preset data set optimization strategy to construct a source domain data set for training the neural network model; The neural network model is iteratively trained using the source domain data set until a preset training termination condition is met.

[0010] Preferably, the step of using a preset data set optimization strategy to construct a source domain data set for training the neural network model includes: Using prior knowledge of historical hyperspectral data, an initial training dataset is constructed; According to a number of randomly generated groups of masks, the bands of the initial training data set are subjected to mask filtering to obtain corresponding band combinations, and the classification accuracy of each of the band combinations is calculated; Calculating the information entropy of each band in the initial training data set; The initial training data set is optimized according to the classification accuracy and the information entropy to obtain the source domain data set.

[0011] On the other hand, the present invention provides a cross-domain hyperspectral band selection device, the device comprising: A particle swarm initialization unit, used to initialize the particle swarm according to the total number of bands in the target domain; A band combination division unit, used for grouping and dividing the hyperspectral bands in the target domain according to the total number of bands and a preset number of target groups, to obtain a target hyperspectral band group; A fitness value calculation unit, used to calculate the fitness value of each particle in the particle swarm at a current position in a solution space using a neural network model obtained based on source domain training, wherein the solution space is composed of the target hyperspectral band group; An optimal combination determination unit, configured to determine, when a preset optimization termination condition is reached, the particle with the highest fitness value in the particle swarm as the optimal band combination of the target domain according to the fitness value; The particle information updating unit is used to update the position information and speed information of each particle in the particle swarm according to the fitness value when the optimization termination condition is not met, trigger the fitness value calculation unit, and execute the neural network model obtained by source domain training to calculate the fitness value of each particle in the particle swarm at the current position in the solution space.

[0012] Preferably, the band combination division unit comprises: A segmentation point division unit, used for grouping and dividing the hyperspectral bands in the target domain according to the segmentation point of each group, to obtain hyperspectral band groups corresponding to the target number of groups; The band combination adjustment unit is used to adjust the hyperspectral bands included in each of the hyperspectral band groups according to the spectral similarity between the bands to obtain the target hyperspectral band group.

[0013] On the other hand, the present invention also provides a computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps described in the above-mentioned cross-domain hyperspectral band selection method are implemented.

[0014] On the other hand, the present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps described in the above-mentioned cross-domain hyperspectral band selection method.

[0015] The present invention initializes a particle swarm according to the total number of bands in a target domain, groups and divides the hyperspectral bands in the target domain according to the total number of bands and the number of target groups, obtains a target hyperspectral band group, calculates the fitness value of each particle in the particle swarm at the current position in a solution space composed of the target hyperspectral band group by using a neural network model obtained based on source domain training, and determines whether an optimization termination condition is met. If so, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain. Otherwise, the position information and speed information of each particle in the particle swarm are updated according to the fitness value, and the search and optimization are continued in the search space, thereby reducing the hyperspectral data annotation cost while improving the efficiency and effect of selecting the most representative bands from the hyperspectral data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart for implementing the cross-domain hyperspectral band selection method provided in the first embodiment of the present invention; Figure 2 It is a structural schematic diagram of a cross-domain hyperspectral band selection device provided in Embodiment 2 of the present invention; Figure 3It is a schematic diagram of the structure of a computing device provided in Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] It should be understood that the terms "first", "second", "third", etc. in the embodiments of the present invention are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit a specific order or sequence. Unless otherwise noted, it should be understood that the terms used in this way can be interchangeable under appropriate circumstances, for example, they can be implemented in an order other than those given in the diagrams or descriptions of the embodiments of the present disclosure.

[0019] Unless otherwise specified, the term "several" in the embodiments of the present invention refers to two or more than two, and other quantifiers are similar.

[0020] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments: Embodiment 1: Figure 1 The implementation process of the cross-domain hyperspectral band selection method provided in the first embodiment of the present invention is shown. For the convenience of description, only the part related to the embodiment of the present invention is shown, which is described in detail as follows: In step S101 , a particle swarm is initialized according to the total number of bands in the target domain.

[0021] The embodiment of the present invention is applicable to computing devices, such as personal computers, servers, etc. In the embodiment of the present invention, the target domain is a hyperspectral image dataset to be processed, which includes multiple hyperspectral bands. The total number of hyperspectral bands (abbreviated as: bands) in the target domain is represented as , here, randomly generate encoded as a binary vector with length Each individual represents a possible band selection scheme (i.e., band combination), and each bit of the binary vector corresponds to the selected state of a band (1 means selected, 0 means unselected), and all individuals constitute a particle swarm.

[0022] In step S102, the hyperspectral bands in the target domain are grouped and divided according to the total number of bands and a preset number of target groups to obtain target hyperspectral band groups.

[0023] In the embodiment of the present invention, the preset target number of groups is expressed as , here, all hyperspectral bands in the target domain are divided into group, we get Group target hyperspectral bands into groups, where the bands within each group have similar spectral characteristics.

[0024] In a feasible embodiment, grouping is achieved through the following steps: (S102.1) grouping and dividing the hyperspectral bands in the target domain according to the segmentation point of each group, and obtaining hyperspectral band groups corresponding to the target number of groups; In the embodiment of the present invention, the segmentation point is a key parameter for determining the boundaries of each group. The hyperspectral band contained in each group is determined by the band range formed by two adjacent segmentation points, that is, Group has Here, according to the segmentation point of each group, the hyperspectral bands in the target domain are divided into Group Hyperspectral Band Group.

[0025] In a feasible embodiment, when the hyperspectral bands in the target domain are grouped and divided according to the segmentation points of each group, first, the total number of bands is and the number of target groups , calculate the split point for grouping , to evenly divide the bands, where , specifically, when When , the first split point is set to 1, recorded as ; when , the preset split point calculation formula is used to calculate the split point for grouping, where the split point calculation formula is: , Indicates the total number of bands, Indicates the number of target groups, Indicates Split points, It represents the remainder after B is divided by G. This item is used to ensure that the band can be evenly divided; when When the last split point is set , indicating that the group has covered all bands; Afterwards, according to the calculated segmentation point , the hyperspectral bands in the target domain are divided into continuous and non-overlapping intervals, each interval is a hyperspectral band group, and each hyperspectral band group Included band range All bands within This division ensures that all bands are fully allocated to In the groups, a coarse division of the hyperspectral bands in the target domain is achieved, thereby reducing the computational complexity and optimizing the subsequent band selection process, providing a reasonable initial grouping structure for subsequent steps.

[0026] (S102.2) According to the spectral similarity between the bands, the hyperspectral bands included in each hyperspectral band group are adjusted to obtain a target hyperspectral band group.

[0027] In an embodiment of the present invention, the hyperspectral bands contained in each hyperspectral band group are adjusted according to the spectral similarity between the bands, so that the number of bands in each group may no longer be uniform, but similar bands are classified into the same group, thereby achieving a refined division of the bands in the target domain and obtaining the final target hyperspectral band group.

[0028] In a feasible embodiment, the hyperspectral bands included in each hyperspectral band group are adjusted by the following steps: (S102.2.1) based on all hyperspectral band groups, calculating the similarity distances of boundary bands between adjacent groups; In the embodiment of the present invention, for two adjacent hyperspectral band groups , ( ), and its corresponding split point is , , ,definition For the previous group The last band, For the latter group For the first band of , the following four similarity distances are calculated: Using formula Calculation Group and The similarity distance between them; Using formula Calculation Group and The similarity distance between them; Using formula Calculation Group and The similarity distance between them; Using formula Calculation Group and The similarity distance between them; in, represents the Euclidean distance, For Group The last band , For Group The first band , For the bands.

[0029] (S102.2.2) According to the similarity distance, the segmentation points between the corresponding adjacent groups are adjusted using the preset segmentation point adjustment strategy, and the hyperspectral bands in the corresponding hyperspectral band group are updated, and the process jumps to the step of calculating the similarity distance of the boundary bands between adjacent groups based on all the hyperspectral band groups until the preset band adjustment termination condition is reached.

[0030] In the embodiment of the present invention, according to the similarity distance, a preset segmentation point adjustment strategy is adopted to adjust the segmentation points between the corresponding adjacent groups, and the hyperspectral bands in the corresponding hyperspectral band group are updated. Specifically: when and , indicating the frequency band and is closer to the previous group, then the split point Move backwards so that Into the previous group At this time, the group The hyperspectral bands in have been updated; when and , indicating the frequency band and is closer to the latter group, then the split point Move forward so that Into the latter group At this time, the group The hyperspectral bands in have been updated; If the above conditions are not met, the current split point is considered reasonable and will not be adjusted; For the updated hyperspectral band group, repeat steps S102.2.1 to S102.2.2 until the segmentation point no longer changes or the preset maximum number of iterations is reached, and output the optimized final grouping result, i.e., the target hyperspectral band group .

[0031] Through the above steps S102.2.1 and S102.2.2, the band grouping is further refined based on similarity analysis, thereby reducing the redundancy between different groups, optimizing the grouping effect, and improving the effectiveness of data classification and the accuracy and stability of band selection.

[0032] In the above steps S102.1 to S102.2, bands with high similarity are divided into the same group through a uniform and fine grouping method, and bands are selected from each group, thereby reducing the selection of redundant bands, reducing the amount of calculation, and improving the efficiency and effect of band selection.

[0033] In step S103, the fitness value of each particle in the particle swarm at the current position in the solution space is calculated using the neural network model obtained based on the source domain training, and the solution space is composed of the target hyperspectral band group.

[0034] In the embodiment of the present invention, the particle swarm searches in the solution space composed of the target hyperspectral band group based on the grouping constraint of the target hyperspectral band group (ensuring that at least one band is selected for each group) to seek the optimal solution, and evaluates the fitness of each particle to judge the pros and cons of the current band selection scheme and guide the update of the particles. Here, for the band selection scheme represented by each particle, The classification performance is evaluated and the classification accuracy is calculated as the fitness value of the particle at the current position. Specifically, the fitness function is used Calculate the fitness value, where express The classification accuracy on the band combination selected by the particle, Represents particles The current position of the source domain is obtained by combining the neural network model trained based on the source domain, so that the particles can be evaluated based on the prior knowledge learned from the source domain during the search process, which improves the accuracy and efficiency of cross-domain band selection and makes the optimization process more efficient.

[0035] In a feasible embodiment, the neural network model obtained by source domain training is used to calculate the fitness value of each particle in the particle swarm at the current position in the solution space, and the training of the neural network model is implemented through the following steps: (S103.1) using a preset data set optimization strategy to construct a source domain data set for neural network model training; In an embodiment of the present invention, a preset data set optimization strategy is adopted to construct a source domain data set for neural network model training to improve data processing efficiency and reduce computational complexity.

[0036] In a feasible embodiment, the construction of the source domain dataset is achieved through the following steps: (S103.1.1) Use prior knowledge of historical hyperspectral data to construct an initial training dataset; In the embodiment of the present invention, prior knowledge is extracted from historical hyperspectral data, and this knowledge is used to construct an initial training data set ,in, is the number of samples in the initial training data set, that is, the initial training data set Depend on samples, each input sample Corresponding to a label .

[0037] (S103.1.2) performing mask filtering on the bands of the initial training data set according to a number of randomly generated mask groups to obtain corresponding band combinations, and calculating the classification accuracy of each band combination; In the embodiment of the present invention, a number of different masks are randomly generated by binary coding, and the length of each mask is equal to the total number of bands in the initial training data set. , and the elements with a value of 1 in each group of masks indicate that the corresponding band is selected, and the elements with a value of 0 indicate that the band is not selected. Each group of masks is applied to the initial training data set, and the bands corresponding to the elements 1 in the mask are filtered out to form the corresponding band combinations. The band combinations corresponding to each group of masks are ,in, is the initial training data set. After that, for each group of band combinations generated by masks, the K-Nearest Neighbors (KNN) classifier is used to calculate the classification accuracy of the band combination. The classification accuracy is used to evaluate the advantages and disadvantages of different band combinations to screen out the optimal band combination. When using the KNN classifier to calculate the classification accuracy of the band combination, specifically, the Euclidean distance or cosine similarity is used to calculate the similarity between samples in the band combination, and the most similar k samples are selected based on the preset number of neighbors k. The category prediction label is determined by voting, and the matching ratio of the category prediction label to the true label is counted as the classification accuracy of the current band combination.

[0038] As an example, the initial training dataset is , the total number of bands is 10, and a set of masks of length 10 are randomly generated , the mask Application ,right The bands are masked and filtered, and the filtered bands are , , , , The obtained band combination is .

[0039] (S103.1.3) Calculate the information entropy of each band in the initial training data set; In the embodiment of the present invention, information entropy is a measurement method for measuring the amount of information in a band, which can reflect the information richness of a specific band in the overall data. Based on this, the calculation formula of information entropy is adopted: The information entropy of each band in the initial training data set is calculated to optimize the construction of the training data to make it more representative and effective. Indicates Bands The gray level, is the probability of the gray level, For band Information entropy.

[0040] (S103.1.4) Optimize the initial training data set based on the classification accuracy and information entropy to obtain the source domain data set.

[0041] In an embodiment of the present invention, the band combination whose classification accuracy meets the preset accuracy threshold is retained, and the bands with high information content and whose information entropy is higher than the preset entropy threshold are screened out, the retained band combination is fused with the screened bands to generate a final training data set, and the training data set is used as the source domain data set, thereby improving the quality of the training data.

[0042] The above steps S103.1.1 to S103.1.4 combine masking and information entropy to screen the bands, so that the samples finally used for neural network model training not only include the results of band selection, but also retain the band characteristics with high information content, thereby improving the adaptability of the model under different band selection schemes, enhancing the robustness of the model, and improving the final classification accuracy.

[0043] (S103.2) Iteratively train the neural network model using the source domain dataset until a preset training termination condition is met.

[0044] In the embodiment of the present invention, the source domain data set is used to construct and initialize the neural network model in advance. Iterative training is performed until the preset training termination conditions are met to ensure that the model can learn the effective features in the source domain data and ultimately be used for band selection optimization in the target domain. Specifically, first, the source domain data set is input into the neural network model for forward propagation calculation to obtain the predicted output. Then, the error rate between the predicted output and the true label is calculated by the loss function of the neural network model. The error rate calculation result is used to guide the parameter update of the model. After the error rate calculation is completed, the back propagation algorithm is used to calculate the gradient of the loss function relative to the model parameters, and the network weights and biases are adjusted based on the gradient descent strategy. The model parameters are continuously adjusted in each iteration. , in order to minimize the loss function value, thereby improving the prediction ability of the model, the training process continues until the preset training termination conditions are met. The preset termination conditions include the following two situations: First, when the error rate of the model decreases less than the decrease threshold after several consecutive iterations, that is, the error rate no longer decreases, indicating that the model has reached a convergence state, and further training is difficult to obtain performance improvement; second, when the training time reaches the maximum set value, that is, the number of training iterations reaches the upper limit, in order to avoid waste of computing resources and model overfitting caused by overtraining, after any training termination condition is met, the training process is terminated, and the final trained neural network model is obtained. .

[0045] The training of the neural network model is completed through the above steps S103.1 to S103.2, so that the model can optimize the band selection of the target domain unlabeled data by learning the prior knowledge of the source domain labeled hyperspectral data. This process uses transfer learning technology to reduce the dependence on the target domain labeled data, reduce the annotation cost, and effectively solve the cross-domain hyperspectral band selection problem.

[0046] In step S104, it is determined whether a preset optimization termination condition is met.

[0047] In an embodiment of the present invention, it is determined whether a preset optimization termination condition is met. Specifically, it is determined whether the current number of iterations reaches a preset optimization iteration threshold, or whether the improvement of the fitness value in several consecutive rounds of iterations is less than a preset improvement threshold. If so, the iteration is terminated and step S106 is executed. Otherwise, step S105 is executed and iteration continues.

[0048] In step S105, the position information and speed information of each particle in the particle swarm are updated according to the fitness value.

[0049] In the embodiment of the present invention, the result of fitness evaluation is used to guide the update of particles. Here, when the preset optimization termination condition is not reached, each particle dynamically adjusts its position according to its current fitness value and jumps to step S103 to seek a better solution. Specifically, the update of the position information and speed information of each particle is achieved through the following steps: (S105.1) updating the individual optimal position corresponding to each particle in the particle swarm and the global optimal position of the particle swarm according to the fitness value; In the embodiment of the present invention, for the convenience of description, the number of iterations in this round is expressed as ,particle So far (i.e. The optimal position (i.e., the individual optimal position) found before the round of iteration is expressed as , the optimal position (i.e., the global optimal position) found by the particle swarm so far is expressed as ,particle In the The current position information and speed information of the round iteration are expressed as and Here, according to the fitness value, the individual optimal position of each particle in the particle swarm and the global optimal position of the particle swarm are updated. Specifically, when the particle The individual optimal position The corresponding fitness value is not higher than the particle's current position When the fitness value is The individual optimal position is updated to the current position ,Right now Otherwise, the individual optimal position remains unchanged. When there is a particle swarm with a fitness value higher than the global optimal position When the corresponding fitness value is Update to the position information corresponding to the particle with the highest fitness value in the particle swarm. When the fitness values ​​of all particles in the particle swarm are not higher than the global optimal position When the corresponding fitness value is , the global optimal position remains unchanged.

[0050] (S105.2) According to the updated individual optimal position and global optimal position, the position information and speed information of each particle in the particle swarm are updated.

[0051] In the embodiment of the present invention, first, the speed update formula is adopted Calculate the new velocity information of the particle , then, according to the updated speed, the position update formula is used Calculate the new position information of the particle ,in, is the inertia weight, and is the acceleration constant, and is a random number between [0,1].

[0052] Through the above steps S105.1 to S105.2, the position information and speed information of each particle in the particle swarm are updated, thereby ensuring that the particle can continuously adjust its position in the search space and gradually approach the optimal solution.

[0053] In step S106, according to the fitness value, the particle with the highest fitness value in the particle group is determined as the optimal band combination of the target domain.

[0054] In the embodiment of the present invention, when the preset optimization termination condition is reached, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain.

[0055] In an embodiment of the present invention, a particle swarm is initialized according to the total number of bands in a target domain, and the hyperspectral bands in the target domain are grouped and divided according to the total number of bands and the target number of groups to obtain a target hyperspectral band group. A neural network model obtained based on source domain training is used to calculate the fitness value of each particle in the particle swarm at the current position in the solution space composed of the target hyperspectral band group, and it is determined whether the optimization termination condition is met. If so, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain. Otherwise, the position information and speed information of each particle in the particle swarm are updated according to the fitness value, and the search and optimization are continued in the search space, thereby reducing the hyperspectral data annotation cost while improving the efficiency and effect of selecting the most representative bands from the hyperspectral data.

[0056] Embodiment 2: Figure 2 The structure of the cross-domain hyperspectral band selection device provided by the second embodiment of the present invention is shown. For the convenience of description, only the part related to the embodiment of the present invention is shown, including: A particle swarm initialization unit 21 is used to initialize the particle swarm according to the total number of bands in the target domain; The band combination division unit 22 is used to group and divide the hyperspectral bands in the target domain according to the total number of bands and the preset number of target groups to obtain a target hyperspectral band group; A fitness value calculation unit 23 is used to calculate the fitness value of each particle in the particle swarm at the current position in the solution space using the neural network model obtained based on the source domain training, and the solution space is composed of the target hyperspectral band group; The optimal combination determination unit 24 is used to determine the particle with the highest fitness value in the particle group as the optimal band combination of the target domain according to the fitness value when the preset optimization termination condition is reached; The particle information updating unit 25 is used to update the position information and speed information of each particle in the particle swarm according to the fitness value when the optimization termination condition is not met, trigger the fitness value calculation unit 23, and execute the neural network model obtained based on the source domain training to calculate the fitness value of each particle in the particle swarm at the current position in the solution space.

[0057] Preferably, the band combination division unit 22 includes: A segmentation point division unit is used to divide the hyperspectral bands in the target domain into groups according to the segmentation points of each group, so as to obtain hyperspectral band groups corresponding to the target number of groups; The band combination adjustment unit is used to adjust the hyperspectral bands contained in each hyperspectral band group according to the spectral similarity between the bands to obtain a target hyperspectral band group.

[0058] Preferably, the cross-domain hyperspectral band selection device of the embodiment of the present invention further includes: A data set construction unit, used to construct a source domain data set for neural network model training by adopting a preset data set optimization strategy; The model training unit is used to iteratively train the neural network model using the source domain data set until a preset training termination condition is met.

[0059] Preferably, the data set construction unit includes: An initial data construction unit, used to construct an initial training data set using prior knowledge of historical hyperspectral data; The band mask filtering unit is used to perform mask filtering on the bands of the initial training data set according to a number of randomly generated mask groups to obtain corresponding band combinations and calculate the classification accuracy of each band combination; An information entropy calculation unit, used to calculate the information entropy of each band in the initial training data set; The data set optimization unit is used to optimize the initial training data set according to the classification accuracy and information entropy to obtain the source domain data set.

[0060] In the embodiment of the present invention, each unit of the cross-domain hyperspectral band selection device can be implemented by a corresponding hardware or software unit, and each unit can be an independent software or hardware unit, or can be integrated into a software or hardware unit, which is not intended to limit the present invention. Specifically, the implementation of each unit can refer to the description of the aforementioned embodiment 1, which will not be repeated here.

[0061] Embodiment three: Figure 3 The structure of a computing device provided in the third embodiment of the present invention is shown. For the convenience of description, only the part related to the embodiment of the present invention is shown.

[0062] The computing device 3 of the embodiment of the present invention includes a processor 30, a memory 31, and a computer program 32 stored in the memory 31 and executable on the processor 30. When the processor 30 executes the computer program 32, the steps in the above cross-domain hyperspectral band selection method embodiment are implemented, such as Figure 1 Alternatively, when the processor 30 executes the computer program 32, the functions of each unit in the above-mentioned device embodiments are implemented, for example Figure 2 Function of the unit shown.

[0063] In an embodiment of the present invention, a particle swarm is initialized according to the total number of bands in a target domain, and the hyperspectral bands in the target domain are grouped and divided according to the total number of bands and the target number of groups to obtain a target hyperspectral band group. A neural network model obtained based on source domain training is used to calculate the fitness value of each particle in the particle swarm at the current position in the solution space composed of the target hyperspectral band group, and it is determined whether the optimization termination condition is met. If so, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain. Otherwise, the position information and speed information of each particle in the particle swarm are updated according to the fitness value, and the search and optimization are continued in the search space, thereby reducing the hyperspectral data annotation cost while improving the efficiency and effect of selecting the most representative bands from the hyperspectral data.

[0064] The computing device of the embodiment of the present invention may be a personal computer. The steps implemented when the processor 30 in the computing device 3 executes the computer program 32 to implement the cross-domain hyperspectral band selection method can refer to the description of the aforementioned method embodiment, which will not be repeated here.

[0065] Embodiment 4: In an embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps in the above cross-domain hyperspectral band selection method embodiment are implemented, for example, Figure 1 Alternatively, when the computer program is executed by a processor, the functions of each unit in the above-mentioned device embodiments are realized, for example Figure 2 Function of the unit shown.

[0066] In an embodiment of the present invention, a particle swarm is initialized according to the total number of bands in a target domain, and the hyperspectral bands in the target domain are grouped and divided according to the total number of bands and the target number of groups to obtain a target hyperspectral band group. A neural network model obtained based on source domain training is used to calculate the fitness value of each particle in the particle swarm at the current position in the solution space composed of the target hyperspectral band group, and it is determined whether the optimization termination condition is met. If so, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain. Otherwise, the position information and speed information of each particle in the particle swarm are updated according to the fitness value, and the search and optimization are continued in the search space, thereby reducing the hyperspectral data annotation cost while improving the efficiency and effect of selecting the most representative bands from the hyperspectral data.

[0067] The computer-readable storage medium of the embodiment of the present invention may include any entity or device or recording medium capable of carrying computer program code, for example, ROM / RAM, magnetic disk, optical disk, flash memory and other memories.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A cross-domain hyperspectral band selection method, characterized in that: The method comprises the following steps: Initialize the particle swarm according to the total number of bands in the target domain; According to the total number of bands and a preset number of target groups, the hyperspectral bands in the target domain are grouped and divided to obtain target hyperspectral band groups; Using a neural network model obtained through source domain training, calculating the fitness value of each particle in the particle swarm at a current position in a solution space, wherein the solution space is composed of the target hyperspectral band group; When a preset optimization termination condition is reached, according to the fitness value, the particle with the highest fitness value in the particle swarm is determined as the optimal band combination of the target domain; When the optimization termination condition is not met, the position information and speed information of each particle in the particle swarm are updated according to the fitness value, and the process jumps to the step of calculating the fitness value of each particle in the particle swarm at the current position in the solution space by using the neural network model obtained by training based on the source domain.

2. The method according to claim 1, characterized in that The step of grouping and dividing the hyperspectral bands in the target domain comprises: According to the segmentation point of each group, the hyperspectral bands in the target domain are divided into groups to obtain hyperspectral band groups corresponding to the target number of groups; According to the spectral similarity between the bands, the hyperspectral bands included in each of the hyperspectral band groups are adjusted to obtain the target hyperspectral band group.

3. The method according to claim 2, characterized in that The step of grouping the hyperspectral bands in the target domain according to the segmentation point of each group includes: According to the total number of bands and the target number of groups, a preset split point calculation formula is used to calculate the split point for grouping, wherein the split point calculation formula is: , represents the total number of bands, represents the target group number, , Indicates A split point.

4. The method according to claim 2, characterized in that The step of adjusting the hyperspectral bands included in each of the hyperspectral band groups according to the spectral similarity between the bands comprises: According to all the hyperspectral band groups, calculating the similarity distances of the boundary bands between adjacent groups; According to the similarity distance, a preset segmentation point adjustment strategy is adopted to adjust the segmentation points between the corresponding adjacent groups, and the hyperspectral bands in the corresponding hyperspectral band group are updated, and the process jumps to the step of calculating the similarity distances of the boundary bands between adjacent groups according to all the hyperspectral band groups until the preset band adjustment termination condition is reached.

5. The method according to claim 1, characterized in that Before the step of calculating the fitness value of each particle in the particle swarm at the current position in the solution space using the neural network model obtained based on source domain training, the method further includes: Using a preset data set optimization strategy to construct a source domain data set for training the neural network model; The neural network model is iteratively trained using the source domain data set until a preset training termination condition is met.

6. The method according to claim 5, characterized in that The step of using a preset data set optimization strategy to construct a source domain data set for training the neural network model includes: Using prior knowledge of historical hyperspectral data, an initial training dataset is constructed; According to a number of randomly generated groups of masks, the bands of the initial training data set are subjected to mask filtering to obtain corresponding band combinations, and the classification accuracy of each of the band combinations is calculated; Calculating the information entropy of each band in the initial training data set; The initial training data set is optimized according to the classification accuracy and the information entropy to obtain the source domain data set.

7. A cross-domain hyperspectral band selection device, characterized in that: The device comprises: A particle swarm initialization unit, used to initialize the particle swarm according to the total number of bands in the target domain; A band combination division unit, used for grouping and dividing the hyperspectral bands in the target domain according to the total number of bands and a preset number of target groups, to obtain a target hyperspectral band group; A fitness value calculation unit, used to calculate the fitness value of each particle in the particle swarm at a current position in a solution space using a neural network model obtained based on source domain training, wherein the solution space is composed of the target hyperspectral band group; An optimal combination determination unit, configured to determine, when a preset optimization termination condition is reached, the particle with the highest fitness value in the particle swarm as the optimal band combination of the target domain according to the fitness value; The particle information updating unit is used to update the position information and speed information of each particle in the particle swarm according to the fitness value when the optimization termination condition is not met, trigger the fitness value calculation unit, and execute the neural network model obtained by source domain training to calculate the fitness value of each particle in the particle swarm at the current position in the solution space.

8. The device according to claim 7, characterized in that The band combination division unit comprises: A segmentation point division unit, used for grouping and dividing the hyperspectral bands in the target domain according to the segmentation point of each group, to obtain hyperspectral band groups corresponding to the target number of groups; The band combination adjustment unit is used to adjust the hyperspectral bands included in each of the hyperspectral band groups according to the spectral similarity between the bands to obtain the target hyperspectral band group.

9. A computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images