Hyperspectral image band selection method and system based on three-dimensional convolution auto-encoder

Abstract

The invention discloses a hyperspectral image waveband selection method and system based on a three-dimensional convolution auto-encoder. The method comprises the following steps: carrying out block processing on an original hyperspectral image; designing a three-dimensional space-spectrum reconstruction network based on a global spectrum sensing module; constructing a combined loss function of a reconstruction error term and a sparse regularization constraint term, and performing iterative updating on the network by using an Adam optimizer; constructing a wave band comprehensive evaluation criterion based on triple constraints and a dynamic fusion mechanism; and successively selecting an optimal wave band by adopting a greedy algorithm according to a comprehensive evaluation criterion, and outputting a wave band selection result of the hyperspectral image. Through fusion of a three-dimensional spatial spectrum reconstruction module and a multi-head self-attention mechanism, local and global association of hyperspectral data is comprehensively mined, and deep joint representation of spectrum and spatial information is realized. Meanwhile, sparse regularization constraint is introduced into a loss function, and the network is guided to focus the most valuable key wave band.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for selecting hyperspectral image bands based on a three-dimensional convolutional autoencoder. Background Technology

[0002] With the rapid development of applications such as intelligent remote sensing monitoring, precision agriculture, and environmental perception, hyperspectral imaging technology, due to its ability to simultaneously provide rich spectral and spatial information, has become an important tool for accurate identification, target classification, and scene analysis. Hyperspectral images typically consist of hundreds of narrow and continuous spectral bands, significantly enhancing the ability to distinguish ground cover categories. However, with the dramatic increase in spectral dimensionality, hyperspectral data also exhibits problems such as severe data redundancy, heavy storage and computational burdens, and a significant "curse of dimensionality" (i.e., the Hughes phenomenon) during analysis and processing. These factors severely restrict the processing efficiency and analytical performance of hyperspectral images in practical applications. Therefore, how to effectively reduce the dimensionality of hyperspectral data and extract key bands has become a critical technical bottleneck that urgently needs to be addressed in current hyperspectral data processing.

[0003] In recent years, with the rapid development of deep learning technology, unsupervised deep learning methods for hyperspectral band selection have gradually emerged. These methods automatically learn the importance of spectral bands by training deep neural networks and capture complex spectral correlations using multi-layer nonlinear structures. Typical studies employ autoencoders and attention mechanisms, using hyperspectral reconstruction as the optimization objective for band selection. However, widely used fully connected autoencoders lack local receptive fields in their structure, making it impossible to effectively model local spectral correlations between adjacent bands; furthermore, when processing images, hyperspectral images need to be flattened into one-dimensional vectors, resulting in severe loss of spatial information.

[0004] To alleviate these problems, researchers have begun to introduce convolutional autoencoders to conduct unsupervised band selection research. One-dimensional convolution-based models can capture local spectral variations, while two-dimensional convolution-based models can extract spatial structure information. However, one-dimensional convolution can only slide along the spectral dimension, making it difficult to extract spatial features; two-dimensional convolution only slides along the spatial dimension, failing to fully characterize spectral variations between consecutive bands. Therefore, existing methods remain limited in fully integrating spectral and spatial information. Furthermore, most related studies focus on band representativeness while neglecting the redundancy relationships and inherent information content between bands, resulting in the final selected band subset potentially containing highly correlated redundant bands, which is detrimental to the performance improvement of subsequent classification and recognition tasks. In addition, representativeness, redundancy, and information content depend on different evaluation metrics, making it difficult for traditional methods to achieve adaptive joint optimization of these three metrics.

[0005] In summary, although current hyperspectral image band selection technology has made initial progress, it still has significant shortcomings in effectively modeling spectral-spatial joint features, handling complex nonlinear relationships, and achieving comprehensive optimization of representativeness, redundancy, and information content. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for selecting hyperspectral image bands based on a three-dimensional convolutional autoencoder. This method can comprehensively characterize the spectral-spatial structure and achieve multi-index adaptive optimization, thus meeting the data processing requirements of hyperspectral intelligent analysis for high precision, high efficiency, and strong robustness.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for selecting bands in hyperspectral images based on a three-dimensional convolutional autoencoder includes the following steps:

[0009] Step S1: The original hyperspectral image is divided into blocks to construct a hyperspectral three-dimensional cube sample set;

[0010] Step S2: Construct a space-spectral reconstruction network based on a global spectral sensing module and a three-dimensional spatial-spectral reconstruction module to learn a global spectral response that can effectively characterize the importance of bands while mining local space-spectral information in hyperspectral image patches.

[0011] Step S3: Construct a loss function that includes a reconstruction error term and a sparse regularization constraint term, and use the Adam optimizer to iteratively update the network.

[0012] Step S4: After the network training is completed, the learned global spectral response vector is extracted, and the inter-band correlation index and band information entropy index are introduced to construct a comprehensive evaluation criterion for bands based on triple constraints and dynamic fusion mechanism.

[0013] Step S5: Based on the comprehensive evaluation criteria, a greedy algorithm is used to select the best bands one by one until a subset of target bands containing the required number of bands is obtained, and finally the band selection results of the hyperspectral image are output.

[0014] Furthermore, in step S2, the three-dimensional spatial-spatial reconstruction module includes a three-dimensional convolutional encoder for extracting local spectral-spatial features and a three-dimensional deconvolutional decoder for recovering the original hyperspectral cube from latent features.

[0015] A 3D convolutional encoder includes 3D convolution, 3D batch normalization, and a nonlinear activation function, and its calculation formula is expressed as:

[0016]

[0017] In the formula, Represents a non-linear activation function; Represents three-dimensional batch normalization; symbol Represents a 3D convolution operation; Represents the convolution kernel; Represents bias; Representative encoding feature representation; This represents reweighted spectral data.

[0018] The 3D deconvolution decoder includes 3D deconvolution, 3D batch normalization, and a nonlinear activation function, and its calculation formula is expressed as:

[0019]

[0020] In the formula, Represents the convolution kernel; Represents bias; This represents the reconstruction of the original hyperspectral image patch.

[0021] Furthermore, in step S3, the loss function is expressed as:

[0022]

[0023] In the formula, Represents the loss function; These represent the parameters that the global spectral sensing module can learn; These represent the parameters that the 3D spatial spectrum reconstruction module can learn; Represents reconstruction error; This represents a sparse regularization constraint. This represents the penalty parameter.

[0024] Furthermore, in step S4, the formula for calculating the representative index between bands is expressed as follows:

[0025]

[0026] In the formula, As a representative indicator; Representative candidate bands; Represents the number of samples; Representing the In the nth sample Global spectral response in each band.

[0027] Furthermore, in step S4, the formula for calculating the inter-band correlation index is expressed as follows:

[0028]

[0029] In the formula, Representative candidate bands; Represents a subset of candidate bands; Represents a correlation index; This is the transpose operator.

[0030] Furthermore, in step S4, the formula for calculating the band information entropy index is expressed as follows:

[0031]

[0032] In the formula, Representative candidate bands; Represents the first candidate band in the normalized candidate band vector. One element; The first Gaussian distribution representing the correlation with the candidate band vector. One element; Indicators representing the amount of information contained; Represents the logarithmic function; It represents the number of elements contained in a band.

[0033] Furthermore, in step S4, the calculation formula for the band comprehensive evaluation criterion is expressed as follows:

[0034]

[0035]

[0036] In the formula, As a representative indicator; For correlation indicators; Information entropy is an indicator. Candidate bands; As a comprehensive evaluation criterion; This is the dynamic balance coefficient; This is the scaling factor; Total number of bands; This represents the number of selected bands. It is a constant balance coefficient.

[0037] The present invention also provides a hyperspectral image band selection system based on a three-dimensional convolutional autoencoder, the system comprising:

[0038] The image block extraction unit is used to divide the original hyperspectral image into blocks and construct a hyperspectral three-dimensional cube sample set.

[0039] A network establishment unit based on a global spectral sensing module and a three-dimensional spatial spectrum reconstruction module is used to establish a network architecture based on the global spectral sensing module and the three-dimensional spatial spectrum reconstruction module. The network based on the global spectral sensing module and the three-dimensional spatial spectrum reconstruction module includes a global spectral sensing module and a three-dimensional spatial spectrum reconstruction module. The global spectral sensing module is used to learn the global spectral response vector to characterize the contribution of each band to the hyperspectral image reconstruction. The three-dimensional spatial spectrum reconstruction module is used to compensate for the lack of local structural representation ability and realize the deep joint representation of spectral and spatial information.

[0040] The loss function definition and network training unit are used to define the network loss function that combines the reconstruction error term and the sparse regularization constraint term, and to optimize network training.

[0041] The band index calculation unit extracts the learned global spectral response vector and combines it with the inter-band correlation and band information entropy to construct a comprehensive band evaluation criterion based on triple constraints and dynamic fusion mechanism.

[0042] The band subset selection and output unit uses a greedy algorithm to select the best bands one by one according to the comprehensive evaluation criteria until a target band subset containing a predetermined number of bands is obtained, and finally outputs the band selection results of the hyperspectral image.

[0043] The present invention also provides a hyperspectral image band selection device based on a three-dimensional convolutional autoencoder, comprising:

[0044] Memory, used to store computer programs and data;

[0045] A processor is used to implement the steps of the above-described hyperspectral image band selection method based on a three-dimensional convolutional autoencoder when executing a computer program.

[0046] The beneficial effects of this invention are:

[0047] 1. In this invention, by integrating a three-dimensional convolutional autoencoder with a multi-head self-attention mechanism, the local and global correlations of hyperspectral data are fully explored, and a deep joint representation of spectral and spatial information is achieved. At the same time, sparse constraints are introduced into the loss function to guide the network to focus on the most valuable key bands.

[0048] 2. This invention constructs a triple index evaluation system composed of representativeness, information entropy and correlation, and establishes a dynamic fusion mechanism by combining adaptive balance coefficients, so that the three types of indicators can achieve adaptive collaborative optimization in the selection process, significantly reducing band redundancy and improving band effectiveness and complementarity.

[0049] 3. This invention can effectively improve the problems of insufficient utilization of band selection information and limited performance of downstream tasks caused by severe band redundancy and insufficient spectral-spatial feature coupling in hyperspectral image analysis.

[0050] 4. This invention enables efficient and robust unsupervised band selection without relying on prior knowledge, enhancing the performance of hyperspectral images in tasks such as classification and detection, meeting the urgent needs of modern hyperspectral intelligent analysis for high precision, high efficiency, and strong robustness, and providing an innovative and scalable band selection solution for hyperspectral data processing. Attached Figure Description

[0051] Figure 1 This is a flowchart of a hyperspectral image band selection method based on a three-dimensional convolutional autoencoder in this embodiment;

[0052] Figure 2 This is a structural block diagram of a hyperspectral image band selection system based on a three-dimensional convolutional autoencoder in this embodiment;

[0053] Figure 3 This is a structural block diagram of a hyperspectral image band selection device based on a three-dimensional convolutional autoencoder in this embodiment;

[0054] Figure 4 This is the ground truth image of the Indian Pines hyperspectral dataset in this embodiment.

[0055] Figure labels: 1. Image patch extraction unit; 2. Network establishment unit based on global spectral sensing module and three-dimensional spatial spectrum reconstruction module; 3. Loss function definition and network training unit; 4. Band index calculation unit; 5. Band subset selection and output unit; 6. Memory; 7. Processor. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example: A method for selecting bands in hyperspectral images based on a three-dimensional convolutional autoencoder, such as... Figure 1 As shown, it includes the following steps:

[0058] Step S1: Construct a hyperspectral 3D cube sample set. Obtain cube blocks from the original hyperspectral image using a sliding window sampling strategy, preserving the spatial neighborhood structure between pixels, and construct a training set for capturing the correlation between local texture and spectral density.

[0059] Step S2 involves constructing a spatial-spectral reconstruction network to simultaneously extract global spectral dependence and local spatial-spectral features from hyperspectral image patches. The network consists of a global spectral sensing module and a three-dimensional spatial-spectral reconstruction module.

[0060] Specifically, step S2 includes the following steps:

[0061] Step S21: Flatten each band of the hyperspectral image patch and input it into the constructed global spectral sensing module. Utilize a multi-head self-attention mechanism to learn global spectral dependencies in parallel across multiple feature subspaces, outputting a global spectral response vector characterizing the importance of each band. The calculation is expressed as follows:

[0062]

[0063] In the formula, Represents the global spectral response vector; Represents the functions of the global spectral sensing module; Represents a patch of the original hyperspectral image; This represents the trainable parameters of the global spectral sensing module.

[0064] The multi-head self-attention mechanism layer is used to capture global dependencies, allowing each band to interact and compare with all other bands. Through this mechanism, the network can comprehensively analyze the entire spectral range. At the same time, the multi-head mechanism enables this process to be carried out in parallel in multiple different feature subspaces, thereby gaining a more comprehensive and robust understanding of the complex relationships between bands.

[0065] Furthermore, the calculation formula for the multi-head self-attention mechanism is expressed as follows:

[0066]

[0067] In the formula, This represents a splicing operation; Represents the number of attention heads; Represents the projection parameter matrix; Represents the multi-head self-attention mechanism function; Represents hyperspectral band embedding sequences;

[0068] For the multi-head self-attention mechanism, the first A person's attention Its calculation formula is expressed as:

[0069]

[0070] In the formula, , and Represents learnable parameters; , and Representing the first The query matrix, key matrix, and value matrix of each attention head; .

[0071] Step S22: Based on the global spectral response vector, perform channel-level feature recalibration on the original cube block to generate a reweighted cube, calculated using the following formula:

[0072]

[0073] In the formula, Represents reweighted spectral data; This represents multiplication by band.

[0074] Step S23: Construct a three-dimensional spatial spectrum reconstruction module to compensate for the lack of local structural expression ability; the three-dimensional spatial spectrum reconstruction module effectively extracts local spectral-spatial features and integrates modeling from global to local features to achieve deep joint representation of spectral and spatial information.

[0075] Furthermore, the 3D spatial spectrum reconstruction module includes a 3D convolutional encoder and a 3D deconvolutional decoder.

[0076] The 3D convolutional encoder includes 3D convolution, 3D batch normalization, and a nonlinear activation function to extract local spectral-spatial features. Its calculation formula is expressed as:

[0077]

[0078] In the formula, Represents a non-linear activation function; Represents three-dimensional batch normalization; symbol Represents a 3D convolution operation; Represents the convolution kernel; Represents bias; This represents the encoded features.

[0079] The 3D deconvolution decoder includes 3D deconvolution, 3D batch normalization, and a nonlinear activation function to recover the original hyperspectral cube from latent features. Its computational formula is expressed as follows:

[0080]

[0081] In the formula, Represents the convolution kernel; Represents bias; symbol Represents a 3D convolution operation; This represents the reconstruction of a patch from the original hyperspectral image; This represents the encoded feature representation, i.e., the output feature of the 3D convolutional encoder.

[0082] Step S3: Construct a loss function that includes a reconstruction error term and a sparse regularization constraint term, and use the Adam optimizer to iteratively update the network.

[0083] In step S3, to simultaneously optimize reconstruction performance and constrain the sparsity of the global spectral response vector, a loss function combining the reconstruction error term and the sparse regularization constraint term is designed to minimize the reconstruction mean square error and the global spectral response vector. With norm as the objective, the loss function is expressed as:

[0084]

[0085] In the formula, Represents the loss function; These represent the parameters that the global spectral sensing module can learn; These represent the parameters that the 3D spatial spectrum reconstruction module can learn; Represents reconstruction error; This represents a sparse regularization constraint. This represents the penalty parameter.

[0086] Furthermore, the reconstruction error can be selected as the reconstruction mean square error:

[0087]

[0088] In the formula, Represents the number of samples; Representing the rebuilt One hyperspectral image patch; Representing the One original hyperspectral image patch; This represents the F-norm.

[0089] Sparse regularization constraints can be selected from the global spectral response vector. Norm:

[0090]

[0091] In the formula, Representing the The global spectral response vector corresponding to each sample.

[0092] Step S4: After the network training is completed, the learned global spectral response vector is extracted, and the inter-band correlation index and band information entropy index are introduced to construct a comprehensive evaluation criterion for bands based on triple constraints and dynamic fusion mechanism.

[0093] The representative index for each spectral band is calculated using the following formula:

[0094]

[0095] In the formula, As a representative indicator; Representative candidate bands; Represents the number of samples; Representing the In the nth sample Global spectral response in each band.

[0096] Band correlation index is measured using orthogonal subspace projection technique, and its calculation formula is as follows:

[0097]

[0098] In the formula, Representative candidate bands; Represents a subset of candidate bands; Represents a correlation index; This is the transpose operator.

[0099] The band information content index is measured by spectral information divergence, and its calculation formula is expressed as:

[0100]

[0101] In the formula, Representative candidate bands; Represents the first in the candidate band vector elements, candidate band vector It was obtained by normalizing candidate bands; The first Gaussian distribution representing the correlation with the candidate band vector. One element; Indicators representing the amount of information contained; Represents the logarithmic function; It represents the number of elements contained in a band.

[0102] Furthermore, the three designed indicators—representative index, correlation index, and information entropy index—are integrated, and an adaptive balance coefficient is introduced to construct a comprehensive band evaluation criterion based on triple constraints and a dynamic balance mechanism. The calculation formula for the comprehensive band evaluation criterion based on triple constraints and a dynamic balance mechanism is expressed as follows:

[0103]

[0104] In the formula, As a representative indicator; For correlation indicators; Information entropy is an indicator. Candidate bands; As a comprehensive evaluation criterion; This is the dynamic balance coefficient; It is a constant balance coefficient.

[0105] Considering that the distance between candidate bands and the selected band subset gradually decreases as the number of selected bands increases, this in turn reduces the inter-band correlation metric. Therefore, the overall evaluation criteria for candidate bands become less sensitive to inter-band correlation and are primarily dominated by band representativeness and information entropy. To balance the contributions of redundancy with the other two metrics during band selection, a dynamic balance coefficient is used. The design value gradually increases with the number of selected bands, and its calculation formula is as follows:

[0106]

[0107] In the formula, This is the scaling factor; Total number of bands; This represents the number of selected bands.

[0108] Step S5: A greedy algorithm is used to select bands sequentially. In each selection step, the comprehensive evaluation of all unselected bands is calculated, and the band with the best comprehensive evaluation is added to the selected band subset until the selected band subset contains the required number of bands. Finally, the band selection result of the hyperspectral image is output.

[0109] In this embodiment, the above method first divides the original hyperspectral image into blocks using a sliding window approach to obtain hyperspectral image block samples for network training. Then, a spectral-spatial joint modeling network integrating a global spectral perception module and a three-dimensional spatial-spectral reconstruction module is constructed. The global spectral perception module uses a multi-head self-attention mechanism to model global spectral dependencies in multiple representation subspaces in parallel, generating a global spectral response vector to measure the importance of each spectral band. Based on the obtained global spectral response vector, the spectral bands of the input image block are weighted channel by channel to obtain reweighted spectral data that enhances the features of key bands. On this basis, a three-dimensional spatial-spectral reconstruction module is introduced to establish local spatial correlations and extract spectral-spatial joint features using three-dimensional convolution to compensate for the deficiencies in local structural representation and achieve deep encoding and reconstruction of the reweighted spectral data.

[0110] Furthermore, a joint loss function is constructed, incorporating both a reconstruction error term and a sparse regularization constraint term, to minimize the reconstruction mean square error and the global spectral response vector. The proposed network is trained using the Adam optimizer, which guides it to focus more attention on the most representative key spectral bands.

[0111] After network training, the learned global spectral response vector is extracted, and inter-band correlation and band information entropy indices are introduced to construct a comprehensive band evaluation criterion based on triple constraints and a dynamic fusion mechanism. Finally, a greedy algorithm is used to successively select the bands with the best comprehensive evaluation until a subset of target bands is obtained, thus achieving band selection for hyperspectral images.

[0112] This embodiment also provides a hyperspectral image band selection system based on a three-dimensional convolutional autoencoder, such as Figure 2 As shown, the system includes: an image patch extraction unit 1, a network establishment unit 2 based on a global spectral sensing module and a three-dimensional spatial spectrum reconstruction module, a loss function definition and network training unit 3, a band index calculation unit 4, and a band subset selection and output unit 5.

[0113] Among them, the image block extraction unit 1 is used to perform cube block processing on the original hyperspectral image to construct a hyperspectral three-dimensional cube sample set.

[0114] Network establishment unit 2 based on global spectral sensing module and three-dimensional spatial spectrum reconstruction module is used to establish a network architecture based on global spectral sensing module and three-dimensional spatial spectrum reconstruction module; the network based on global spectral sensing module and three-dimensional spatial spectrum reconstruction module includes global spectral sensing module and three-dimensional spatial spectrum reconstruction module.

[0115] The global spectral sensing module is used to learn the global spectral response of each spectral band to characterize the contribution of each band to hyperspectral image reconstruction.

[0116] The three-dimensional spatial spectrum reconstruction module is used to compensate for the lack of local structural representation ability and realize the deep joint representation of spectral and spatial information.

[0117] The loss function definition and network training unit 3 are used to define the network loss function that combines the reconstruction error term and the sparse regularization constraint term, and to optimize network training.

[0118] Band index calculation unit 4 extracts the learned global spectral response vector and, combined with inter-band correlation and band information entropy, constructs a comprehensive band evaluation criterion based on triple constraints and dynamic fusion mechanism.

[0119] The band subset selection and output unit 5 uses a greedy algorithm to select the best bands one by one according to the comprehensive evaluation criteria until the target band subset containing the required number of bands is obtained, and finally outputs the band selection results of the hyperspectral image.

[0120] This embodiment also provides a hyperspectral image band selection device based on a three-dimensional convolutional autoencoder, such as... Figure 3 As shown, it includes:

[0121] Memory 6 is used to store computer programs and data;

[0122] Processor 7 is used to implement the steps of the above-described hyperspectral image band selection method based on a three-dimensional convolutional autoencoder when executing a computer program.

[0123] To verify the effectiveness of the proposed method in hyperspectral band selection, experiments were conducted on real hyperspectral images. The specific implementation is illustrated below using the Indian Pines dataset as an example; the experiments are as follows:

[0124] Indian Pines dataset, such as Figure 4 As shown, excluding the water vapor absorption band and the low signal-to-noise ratio band, the remaining 185 bands participated in the experiment.

[0125] Table 1 shows the classification accuracy obtained by using SVM and EPF classifiers respectively when selecting 15 bands on the Indian Pines dataset using different band selection methods. The table presents the results of two quantitative evaluation metrics for pixel classification: Overall Accuracy (OA) and Average Accuracy (AA). The former refers to the ratio of correctly classified samples to the total number of samples, while the latter refers to the average classification accuracy for each land cover category. As can be seen from Table 1, the hyperspectral image band selection method based on a 3D convolutional autoencoder proposed in this embodiment can improve the classification performance of band selection.

[0126] Table 1. Comparison of classification performance of different band selection methods on SVM and EPF classifiers.

[0127]

[0128] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for selecting bands in hyperspectral images based on a three-dimensional convolutional autoencoder, characterized in that, Includes the following steps: Step S1: The original hyperspectral image is divided into blocks to construct a hyperspectral three-dimensional cube sample set; Step S2: Construct a space-spectral reconstruction network based on a global spectral sensing module and a three-dimensional spatial-spectral reconstruction module to learn a global spectral response that can effectively characterize the importance of bands while mining local space-spectral information in hyperspectral image patches. Step S3: Construct a loss function that includes a reconstruction error term and a sparse regularization constraint term, and use the Adam optimizer to iteratively update the network. Step S4: After the network training is completed, the learned global spectral response vector is extracted, and the inter-band correlation index and band information entropy index are introduced to construct a comprehensive evaluation criterion for bands based on triple constraints and dynamic fusion mechanism. Step S5: Based on the comprehensive evaluation criteria, a greedy algorithm is used to select the best bands one by one until a subset of target bands containing the required number of bands is obtained, and finally the band selection results of the hyperspectral image are output.

2. The hyperspectral image band selection method based on a three-dimensional convolutional autoencoder according to claim 1, characterized in that, In step S2, the three-dimensional spatial-spatial reconstruction module includes a three-dimensional convolutional encoder for extracting local spectral-spatial features and a three-dimensional deconvolutional decoder for recovering the original hyperspectral cube from latent features; A 3D convolutional encoder includes 3D convolution, 3D batch normalization, and a nonlinear activation function, and its calculation formula is expressed as: In the formula, Represents a non-linear activation function; Represents three-dimensional batch normalization; symbol Represents a 3D convolution operation; Represents the convolution kernel; Represents bias; Representative encoding feature representation; Represents reweighted spectral data; The 3D deconvolution decoder includes 3D deconvolution, 3D batch normalization, and a nonlinear activation function, and its calculation formula is expressed as: In the formula, Represents the convolution kernel; Represents bias; This represents the reconstruction of the original hyperspectral image patch.

3. The hyperspectral image band selection method based on a three-dimensional convolutional autoencoder according to claim 1, characterized in that, In step S3, the loss function is expressed as: In the formula, Represents the loss function; These represent the parameters that the global spectral sensing module can learn; These represent the parameters that the 3D spatial spectrum reconstruction module can learn; Represents reconstruction error; This represents a sparse regularization constraint. This represents the penalty parameter.

4. The hyperspectral image band selection method based on a three-dimensional convolutional autoencoder according to claim 1, characterized in that, In step S4, the formula for calculating the representative index between bands is expressed as follows: In the formula, As a representative indicator; Representative candidate bands; Represents the number of samples; Representing the In the nth sample Global spectral response in each band.

5. The hyperspectral image band selection method based on a three-dimensional convolutional autoencoder according to claim 1, characterized in that, In step S4, the formula for calculating the inter-band correlation index is expressed as follows: In the formula, Representative candidate bands; Represents a subset of candidate bands; Represents a correlation index; This is the transpose operator.

6. The hyperspectral image band selection method based on a three-dimensional convolutional autoencoder according to claim 1, characterized in that, In step S4, the formula for calculating the band information entropy index is as follows: In the formula, Representative candidate bands; Represents the first candidate band in the normalized candidate band vector. One element; The first Gaussian distribution representing the correlation with the candidate band vector. One element; Indicators representing the amount of information contained; Represents the logarithmic function; It represents the number of elements contained in a band.

7. The hyperspectral image band selection method based on a three-dimensional convolutional autoencoder according to claim 1, characterized in that, In step S4, the calculation formula for the band comprehensive evaluation criterion is expressed as follows: In the formula, As a representative indicator; For correlation indicators; Information entropy is an indicator. Candidate bands; As a comprehensive evaluation criterion; This is the dynamic balance coefficient; This is the scaling factor; Total number of bands; This represents the number of selected bands. It is a constant balance coefficient.

8. A hyperspectral image band selection system based on a three-dimensional convolutional autoencoder for implementing the method of claim 1, characterized in that, The system includes: Image block extraction unit (1) is used to divide the original hyperspectral image into blocks and construct a hyperspectral three-dimensional cube sample set; The network establishment unit (2) based on the global spectral sensing module and the three-dimensional spatial spectrum reconstruction module is used to establish a network architecture based on the global spectral sensing module and the three-dimensional spatial spectrum reconstruction module. The network based on the global spectral sensing module and the three-dimensional spatial spectrum reconstruction module includes the global spectral sensing module and the three-dimensional spatial spectrum reconstruction module. The global spectral sensing module is used to learn the global spectral response vector to characterize the contribution of each band to the reconstruction of the hyperspectral image. The three-dimensional spatial spectrum reconstruction module is used to make up for the lack of local structural expression ability and realize the deep joint characterization of spectral and spatial information. The loss function definition and network training unit (3) are used to define the network loss function that combines the reconstruction error term and the sparse regularization constraint term, and to optimize the network training. The band index calculation unit (4) extracts the learned global spectral response vector and combines the inter-band correlation and band information entropy to construct a comprehensive evaluation criterion for bands based on triple constraints and dynamic fusion mechanism. The band subset selection and output unit (5) selects the best bands one by one using a greedy algorithm based on the comprehensive evaluation criteria until a target band subset containing a predetermined number of bands is obtained, and finally outputs the band selection results of the hyperspectral image.

9. A hyperspectral image band selection device based on a three-dimensional convolutional autoencoder, characterized in that, include: Memory (6) is used to store computer programs and data; The processor (7) is configured to implement the steps of the hyperspectral image band selection method based on any one of claims 1-7 when executing a computer program.

Images