Hyperspectral image migration classification method based on multivariate mutual learning network

Abstract

The invention provides a hyperspectral image migration classification method based on a multivariate mutual learning network, and belongs to the technical field of remote sensing image processing and application. The method comprises the following steps: firstly, acquiring two hyperspectral images in different areas as a source domain and a target domain respectively, and extracting data set features; secondly, searching a neighbor most similar to a target domain sample by using an adaptive neighborhood search method, and enhancing the recognition capability of the network on unknown categories by using a cross-domain manifold mixing method; thirdly, a known class classifier network is used for outputting a tag class which a target sample possibly belongs to; and finally, scoring results output by the known class classifier network by using the unknown class classifier network. According to the method, the problem that cross-data and cross-task migration is difficult in the existing hyperspectral image classification technology can be solved; information sharing and transmission among hyperspectral images with different spatial spectral resolutions, different spectral ranges and different task coverage can be realized, and an effective solution is provided for continuous hyperspectral image classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing and application, and relates to a hyperspectral image migration classification method based on a multivariate mutual learning network, specifically a hyperspectral image migration classification method based on a mutual learning network combining neighborhood invariance and cross-domain mixing. Background Art

[0002] Hyperspectral images have many imaging bands. Spectral imagers have dozens or even hundreds of bands in the visible and infrared spectral regions. They can provide spatial and spectral domain information at the same time, and have high spectral resolution and rich detail information, which is conducive to fine ground object analysis. They are widely used in military reconnaissance, vegetation surveys, agricultural monitoring and other fields. However, the large amount of hyperspectral data causes redundancy and high correlation between bands, which requires a large number of training samples for classification. The most important thing is that since the annotation of hyperspectral data sets depends on professional equipment and experts in related fields, it is impractical to obtain enough labeled data. Therefore, the current traditional supervised hyperspectral image classification methods face huge challenges.

[0003] In recent years, some researchers have proposed domain adaptation methods to transfer knowledge from a source domain with rich labels to a target domain with scarce labels, and have achieved remarkable results. However, traditional domain adaptation assumes that the two domains share the same type of objects. Due to the different composition of objects in different regions, there are likely to be private categories in the source and target domains, so it cannot be directly applied to cross-scene hyperspectral image classification. A more practical method is to not impose any prior knowledge on the target domain label set, but only assume that the source and target domain labels have common parts, which we call universal hyperspectral transfer classification.

[0004] Zhou Hao et al. (Chinese invention patent, patent application number: CN202310869709.4) provide a hyperspectral image classification method, device and electronic device to obtain a hyperspectral image dataset. The classification model is trained using a dual-channel twin network based on convolutional residual blocks and spatial attention mechanism after network optimization, and the hyperspectral image classification in cross-domain scenarios is classified using a transfer learning method, which improves the classification accuracy of hyperspectral images. However, it is optimized specifically for small sample scenarios. In practical applications, hyperspectral image data has a high degree of variability, and small samples may cause the model to be unable to fully capture the diversity and complexity of the data.

[0005] Zhu Shanshan et al. (Chinese invention patent, patent application number: CN202311699520.1) disclosed a spectral classification method based on a multi-branch attention network. The multi-branch attention is used to further learn the feature data output by the basic convolution block, which enhances the feature extraction ability of the network. The feature data output by different branches are fused using branch weights, so that the classification model can learn more abundant global and local information of Raman characteristic peaks, thereby improving the accuracy of hyperspectral image classification. However, it does not take into account the problem of cross-domain category imbalance. Since the category distribution may be different, the prediction accuracy of minority classes is reduced, and the generalization ability of the model is insufficient.

[0006] These deficiencies have become the key bottleneck restricting the transformation of hyperspectral image classification from theoretical research to large-scale practical applications. It is urgent to develop a cross-domain hyperspectral image recognition method that can handle data with different distributions and scales, so as to meet the needs of models facing new data and new tasks in practical applications. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the prior art and provide a hyperspectral image migration classification method based on a multivariate mutual learning network to solve the problem that the existing hyperspectral image classification technology is difficult to achieve cross-data migration. The method is based on adaptive neighborhood selection to reduce intra-domain changes and obtain a more generalized feature representation. The method uses a cross-domain manifold mixing method to simulate unknown class samples and expand the decision boundary. The present invention proposes a multivariate mutual learning network to solve the problem of separation of public and private classes; proposes a cross-domain manifold mixing method based on a mutual learning network to solve the problem of feature drift; and proposes an adaptive neighborhood search method to solve the problem of category imbalance. The present invention can realize information sharing and transmission between hyperspectral images with different spatial spectral resolutions, different spectral ranges, and different task coverage, providing an effective solution for continuous hyperspectral image classification, and bringing the hyperspectral image classification theory one step closer to large-scale practical use.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A hyperspectral image migration classification method based on a multivariate mutual learning network. The hyperspectral image migration classification method first uses a spectrometer to collect two hyperspectral images of different regions as the source domain and the target domain respectively. Secondly, a convolutional neural network based on channel attention is used to extract data set features. Then, an adaptive neighborhood search method is used to find the most similar neighbors of the target domain samples. A cross-domain manifold hybrid method is used to enhance the network's recognition ability for unknown categories. Thirdly, a known class classifier network is used to output the label category to which the target sample may belong. Finally, an unknown class classifier network is used to score the output of the known class classifier network. Scores above the threshold can be classified as known categories, and scores below the threshold can be identified as unknown categories. Specifically, the following steps are included:

[0010] Step 1, select two hyperspectral images containing the same label category (hereinafter referred to as the public category), select one of the hyperspectral images as the source data set (hereinafter referred to as the source domain), and select the other hyperspectral image as the target data set (hereinafter referred to as the target domain). It should be noted that the label category included in the source domain but not in the target domain is called the source domain private class, and the label category not included in the source domain but included in the target domain is called the target domain private class. The source domain data set will be used as the data set for training the network, and the target domain data set will be used as the test data set. Read the data in the source domain and the data in the target domain, and preprocess the source domain data and the target domain data, and the preprocessing includes dimension unification and data expansion.

[0011] In step 2, the hyperspectral data of the source domain and the target domain obtained after preprocessing in step 1 are respectively input into the feature extraction network based on channel attention (hereinafter referred to as the feature extraction network) to obtain the feature vectors of the source domain data and the target domain data respectively.

[0012] Step 3: Use the spectral features in the target domain extracted in step 2 to input into the similarity calculation network to calculate the neighborhood similarity of the target sample. First, maintain the target domain data memory library, update it after each small batch training, and perform L2 normalization on the features: Where m j represents the L2 normalized feature of the jth sample, represents the feature vector of the jth target domain sample, represents the L2 norm of the feature vector; then the input sample The neighborhood similarity between the kth sample in the target domain is calculated by the inner product, and the calculation formula is: j,k =m j ·m k , where m k Represents the L2 normalized feature of the kth sample.

[0013] Step 4. At the same time, based on the idea of ​​manifold mixing, a mixed sample is constructed between the source domain and the target domain to simulate potential unknown category samples (hereinafter referred to as cross-domain manifold mixing). The source domain and target domain feature vectors extracted in step 2 are interpolated and mixed through the feature mixing network. Specifically: an interpolation factor λ~Beta(α,α) is randomly sampled from the Beta distribution, where α is a hyperparameter of the Beta distribution, and the actual value is generally between [0.1,10], and λ represents the interpolation factor; combined with the Beta distribution, the features of the source domain data in step 2 and the feature vectors of the target domain data are used to generate a mixed feature vector. The formula for generating the mixed feature vector is as follows: in represents the mixed eigenvector; λ represents the interpolation factor; Feature vector representing source domain data; Represents source domain data; Feature vector representing the target domain data; Represents the target domain data.

[0014] Step 5, input the neighborhood similarity obtained in step 3 and the mixed feature vector in step 4 into the known class classifier network for training, and output the maximum probability distribution category of the training sample on all known categories. The known class classifier network is composed of multiple binary classifier networks, the number of which is the same as the number of label categories of the source domain hyperspectral image, and the loss function used in the training process is the cross entropy loss function.

[0015] Step 6, input the input result of step 5 to the unknown class classifier network for training. Jointly optimize the decision results of the known class classifier network and the unknown class classifier network, perform consistency constraints, and use the high confidence decision of the known class classifier network to guide the optimization of the unknown class classifier network. The unknown class classifier network is also composed of multiple binary classifiers, and the number of binary classifier networks is the same as the number of label categories of the source domain hyperspectral image. Use open set entropy minimization loss and cross-domain loss to train the unknown class classifier network, and by minimizing the cross-domain loss function, the positive class score of the mixed sample is reduced.

[0016] The formula for the open set entropy minimization loss is:

[0017]

[0018] Among them, K represents the number of source domain categories; P o represents the probability of the output of the unknown class classifier network; l represents the label index of the known category in the source domain; x t represents the unlabeled data in the target domain; P o (l|x t ) represents the positive score of the unknown class classifier network for the lth class.

[0019] The formula for the cross-domain loss is:

[0020] in, Represents source domain data; Represents the source domain data label; Target domain data; represents a mixed feature vector.

[0021] Step 7, testing phase, specifically:

[0022] Step 7.1, use the feature extraction network in step 2 to extract the target domain sample feature vector.

[0023] Step 7.2, use the features of step 7.1 to input into the known class classifier network, and output the maximum probability category label for each sample.

[0024] Step 7.3, select the corresponding unknown class classifier network according to the label prediction result of step 7.2, and record the positive score output by the unknown class classifier network.

[0025] Step 7.4, compare the positive score of the output of step 7.3 with the threshold, the threshold range is [0,1]. If the unknown class classifier network output score of the sample is lower than the threshold, it is marked as an unknown class and the unknown class label is output; otherwise, the known class of the preliminary prediction is retained and the corresponding known class label is output.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] First, in step 3, the present invention proposes an adaptive domain search algorithm to overcome the problem of imbalance of samples of different categories in hyperspectral remote sensing images while maximizing the feature similarity between the sample and its neighbors, enhancing feature generalization and reducing intra-domain variations.

[0028] Second, in step 4, the present invention is based on the idea of ​​manifold mixing and proposes a cross-domain manifold mixing method to simulate unknown class samples across domains in a smoother way, thereby reducing the phenomenon of cross-domain distribution deviation of hyperspectral remote sensing images.

[0029] Third, in step 5 and step 6, the present invention links the decisions of the known class classifier network and the unknown class classifier network through consistency constraints to innovatively identify unknown categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a principle flow chart of the present invention.

[0031] Figure 2 Acquisition scenario for source domain hyperspectral data; Figure 2(a) is the hyperspectral image of Pavia University; Figure 2 (b) in the figure is the true value graph.

[0032] Figure 3 Acquire scenes for target domain hyperspectral data; Figure 3 (a) is the Pavia Center hyperspectral image; Figure 3 (b) in the figure is the true value graph. DETAILED DESCRIPTION

[0033] The present invention is further described below in conjunction with specific embodiments.

[0034] A hyperspectral image migration classification method based on a multivariate mutual learning network comprises the following steps:

[0035] Step 1: Select two hyperspectral images containing the same label category (hereinafter referred to as the public category), select one of the hyperspectral images as the source dataset (hereinafter referred to as the source domain), and select the other hyperspectral image as the target dataset (hereinafter referred to as the target domain). It should be noted that the label category included in the source domain but not in the target domain is called the source domain private category, and the label category not included in the source domain but included in the target domain is called the target domain private category. The source domain dataset will be used as the dataset for training the network, and the target domain dataset will be used as the test dataset.

[0036] Step 1.1: Read the data from the source domain and target domain data If the spectral dimensions of the source domain and the target domain hyperspectral images are different, the low-dimensional hyperspectral image is padded with zeros to unify the dimensions with the high-dimensional hyperspectral image for dimension unification and data expansion. Zero matrices are added around the hyperspectral image, and mirror data blocks are obtained by splicing the expanded top, bottom, left, and right sides.

[0037] Step 2: The hyperspectral data of the source domain and the target domain obtained after preprocessing in step 1 are respectively input into the feature extraction network based on channel attention (hereinafter referred to as the feature extraction network) to obtain the feature vectors of the source domain data. and the feature vector of the target domain data

[0038] Step 3: Use the similarity calculation network to calculate the neighborhood similarity of the target sample:

[0039] Step 3.1: Use the target domain features extracted in step 2 to maintain the target domain data memory library, update it in place after each small batch training, and perform L2 normalization on the features: in represents the feature vector of the jth target domain sample, represents the L2 norm of the feature vector, m j Represents the L2 normalized feature of the jth sample.

[0040] Step 3.2, input sample The neighborhood similarity between the kth sample in the target domain is calculated by the inner product: s j,k =m j ·m k , m k Represents the L2 normalized feature of the kth sample.

[0041] Step 3.3, use the Softmax function to normalize the neighborhood similarity of all target domain samples so that The probability of sharing the same class with other samples can be estimated as: where p jk represents the probability that the target domain sample belongs to the same category, s j,k Represents the similarity score of the target domain sample, N t represents the number of target samples, and τ is a scaling factor that adjusts the smoothness of the probability distribution.

[0042] Step 3.4, reduce the intra-domain variability by maximizing the probability between the input sample and its neighbors, where the loss function is: Where N j Representation sample The neighbor set of w jk is the weight representing the neighbor confidence, p jk Represents the probability that the target domain sample belongs to the same category.

[0043] Step 4: The source domain and target domain feature vectors extracted in step 2 are interpolated and mixed through a feature mixing network:

[0044] In step 4.1, an interpolation factor λ~Beta(α,α) is randomly sampled from the Beta distribution, where λ represents the interpolation factor and α is a hyperparameter of the Beta distribution. α is set to 2.0 so that the network can smoothly interpolate in the intermediate states between different domains.

[0045] Step 4.2: Use the source domain sample feature vector in step 2 and the target domain sample feature vector Generating cross-domain mixed feature vectors where λ is the interpolation factor described in step 4.1, where represents the mixed eigenvector; λ represents the interpolation factor; Feature vector representing source domain data; Represents source domain data; Feature vector representing the target domain data; Represents the target domain data.

[0046] Step 5: Add the neighborhood similarity obtained in step 3 and the mixed feature vector obtained in step 4 Input to the known class classifier network for training to obtain the output result. The known class classifier network is composed of multiple binary classifier networks. The number of binary classifier networks is the same as the number of label categories of the source domain hyperspectral image. The loss function used in the training process is the cross entropy loss function, and the formula is as follows: where y i represents the true label of the sample, Represents the predicted sample label value.

[0047] Step 6. Then, use the input result of step 5 to input into the unknown class classifier network for training, and output the positive score corresponding to the sample. The positive score indicates the possibility that the sample belongs to a known category. Jointly optimize the decision results of the known class classifier network and the unknown class classifier network, perform consistency constraints, and use the high confidence decision of the known class classifier network to guide the optimization of the unknown class classifier network. The unknown class classifier network is also composed of multiple binary classifiers, and the number of binary classifier networks is the same as the number of label categories of the source domain hyperspectral image. Use open set entropy minimization loss and cross-domain hybrid loss to train the unknown class classifier network, where the open set entropy formula is as follows:

[0048]

[0049] Among them, K represents the number of source domain categories. The cross-domain loss function is: Where P o represents the probability of the output of the unknown class classifier network; l represents the index of the known category; x t represents the unlabeled data in the target domain; P o (l|x t ) represents the positive score of the unknown class classifier network for the lth class.

[0050] Step 7, testing phase:

[0051] Step 7.1: Input the preprocessed target domain data in step 1.1 into the feature extraction network in step 2 to obtain the target domain sample feature vector

[0052] Step 7.2, further, use the features of step 7.1 to input into the known class classifier network, and record the maximum probability category label output by each sample.

[0053] Step 7.3, further, according to the prediction result of step 7.2, select the unknown class classifier network corresponding to the maximum probability category, and record the positive score output by the unknown class classifier network.

[0054] Step 7.4: Further, compare the output score of step 7.3 with the threshold value, and the threshold value is 0.5. If the output score of the unknown class classifier network of the sample is lower than the threshold value, it is marked as an unknown class and the unknown class label is output. Otherwise, the initially predicted known class is retained and the corresponding known class label is output.

[0055] The simulation experiment illustrates the technical effect of this embodiment:

[0056] First, simulation experiment conditions:

[0057] The experimental platform of the present invention is a desktop computer with the following configuration: Intel Core i7, 4.0 GHz central processing unit, GeForce GTX 1080Ti graphics processor, and 32 GB memory.

[0058] The experimental data of this invention are Pavia University hyperspectral data and Pavia Center hyperspectral data, which are captured by ROSIS sensor in the northern area of ​​Pavia, Italy. It consists of two scenes: University of Pavia and Pavia Center. The spatial resolution is 1.3m and the spectral resolution is 4nm. Figure 2 (a) shows the source domain Pavia University hyperspectral data, which is 610×340 pixels in size, 103 bands, and 9 categories, including asphalt, grass, gravel, tree, metal plate, land, asphalt, brick and shadow. Figure 2 (b) in the figure is the true value map of Pavia University data. Figure 3 (a) shows the target domain Pavia Center hyperspectral data, which is 1096×715 pixels in size and has 9 categories, including water, tree, grass, brick, land, asphalt, asphalt, tile and shadow. Figure 3 (b) in the figure is the truth map of PaviaCenter data. There are 7 common classes between the source domain and the target domain, and each domain has two private classes. The source domain consists of 42776 samples, and the target domain consists of 148152 samples.

[0059] Second, the simulation results are analyzed and the classification results are shown in Table 1.

[0060] The comparative method adopted in the present invention is cross-domain adaptation based on the optimal transport framework [Chang W, Shi Y, Tuan H, et al. Unified optimal transport framework for universal domain adaptation [J]. Advances in Neural Information Processing Systems, 2022, 35: 29512-29524], which is one of the most advanced cross-domain transfer learning methods currently available.

[0061] According to the simulation result data in Table 1, the method of the present invention is significantly ahead of the best existing methods in 5 of the 7 common categories, especially in the bare soil category, where the existing method has only 13.9% recognition accuracy, while the present invention can achieve 64.8% accuracy, and the average accuracy of all samples and the accuracy of each category are higher than the existing methods. The hyperspectral image migration classification method based on the multivariate mutual learning network of the present invention can be used to migrate the classification model to new data and identify new categories in the absence of any labeled samples.

[0062] Table 1, simulation results analysis and classification results

[0063] Related Task Categories Comparative ratio (%) Classification results of the present invention (%) Trees 85.5 89.84 Asphalt Road 82.0 90.32 Self-locking bricks 80.1 78.46 asphalt 39.2 76.47 shadow 90.2 78.26 grassland 68.8 73.03 Bare soil 13.9 64.80 Target domain private class 90.7 88.17 Average accuracy 75.8 83.2 Average accuracy per class 68.8 79.92

[0064] The above-described embodiments merely express the implementation methods of the present invention, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A hyperspectral image migration classification method based on a multivariate mutual learning network, characterized in that: The hyperspectral image migration classification method first uses a spectrometer to collect two hyperspectral images of different regions as the source domain and the target domain respectively. Secondly, a convolutional neural network based on channel attention is used to extract data set features. Then, an adaptive neighborhood search method is used to find the most similar neighbors of the target domain samples. A cross-domain manifold hybrid method is used to enhance the network's recognition ability of unknown categories. Thirdly, a known class classifier network is used to output the label category to which the target sample may belong. Finally, an unknown class classifier network is used to score the results output by the known class classifier network. Scores above the threshold are classified as known categories, and scores below the threshold are identified as unknown categories.

2. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 1 is characterized in that: The hyperspectral image migration classification method specifically includes the following steps: Step 1, select two hyperspectral images containing the same label category; select one of the hyperspectral images as the source data set, defined as the source domain; select the other hyperspectral image as the target data set, defined as the target domain; the source domain data set is used as the data set for training the network, and the target domain data set is used as the test data set; read the data of the source domain and the data of the target domain, and preprocess the source domain data and the target domain data; Step 2, inputting the hyperspectral data of the source domain and the target domain obtained after preprocessing in step 1 into the feature extraction network based on channel attention respectively, and obtaining the feature vectors of the source domain data and the target domain data respectively; Step 3: Use the spectral features in the target domain extracted in step 2 to input the similarity calculation network to calculate the neighborhood similarity of the target sample. First, maintain the target domain data memory library, update it after each batch training, and perform L2 normalization on the features. Input the sample The neighborhood similarity between the kth sample in the target domain is calculated by the inner product; Step 4: At the same time, based on the cross-domain manifold mixing method, a mixed sample is constructed between the source domain and the target domain to simulate potential unknown category samples; the source domain and target domain feature vectors extracted in step 2 are interpolated and mixed through a feature mixing network; combined with Beta distribution, the features of the source domain data in step 2 and the feature vectors of the target domain data are used to generate a mixed feature vector, and the formula for generating the mixed feature vector is: in represents the mixed eigenvector; λ represents the interpolation factor; Feature vector representing source domain data; Represents source domain data; Feature vector representing the target domain data; Represents the target domain data; Step 5, input the neighborhood similarity obtained in step 3 and the mixed feature vector in step 4 into the known class classifier network for training, and output the maximum probability distribution category of the training sample on all known categories; Step 6, inputting the input result of step 5 into the unknown class classifier network for training; jointly optimizing the decision results of the known class classifier network and the unknown class classifier network, performing consistency constraints, and using the high confidence decision of the known class classifier network to guide the optimization of the unknown class classifier network; the unknown class classifier network is also composed of multiple binary classifiers, and the number of binary classifier networks is the same as the number of label categories of the source domain hyperspectral image; using open set entropy minimization loss and cross-domain loss to train the unknown class classifier network; by minimizing the cross-domain loss function, the positive class score of the mixed sample is reduced; Step 7, testing phase, specifically: Step 7.1, using the feature extraction network in step 2, extract the target domain sample feature vector; Step 7.2, use the target domain sample feature vector of step 7.1 to the known class classifier network, and output the maximum probability class label output by each sample; Step 7.3, select the corresponding unknown class classifier network according to the label prediction result of step 7.2, and record the positive score output by the unknown class classifier network; In step 7.4, the positive score of the output of step 7.3 is compared with the threshold; if the unknown class classifier network output score of the sample is lower than the threshold, it is marked as an unknown class and the unknown class label is output; otherwise, the initially predicted known category is kept and the corresponding known class label is output.

3. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2 is characterized in that: In step 1, the label categories included in the source domain but not in the target domain are called source domain private categories, and the label categories not included in the source domain but included in the target domain are called target domain private categories.

4. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2 is characterized in that: In the step 1, the preprocessing includes dimension unification and data expansion.

5. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2 is characterized in that: In step 3, L2 normalization uses the following formula: in, represents the feature vector of the jth target domain sample, m j represents the L2 normalized feature of the jth sample; the inner product calculation formula is: s j,k =m j ·m k , where m k Represents the L2 normalized feature of the kth sample.

6. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2 is characterized in that: The Beta distribution in step 4 is as follows: randomly sampling an interpolation factor λ~Beta(α,α) from the Beta distribution, where α is a hyperparameter of the Beta distribution; λ represents the interpolation factor.

7. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 6 is characterized in that: The value of α is between [0.1,10].

8. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2 is characterized in that: In step 5, the known class classifier network is composed of multiple binary classifier networks, the number of binary classifier networks is the same as the number of label categories of the source domain hyperspectral image, and the loss function used in the training process is the cross entropy loss function.

9. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2 is characterized in that: In step 6: The formula for the open set entropy minimization loss is: Among them, K represents the number of source domain categories; P o represents the probability of the output of the unknown class classifier network; l represents the label index of the known category in the source domain; x t represents the unlabeled data in the target domain; P o (l|x t ) represents the positive score of the unknown class classifier network for the lth class; The formula for the cross-domain loss is: in, Represents source domain data; Represents the source domain data label; Target domain data; represents a mixed feature vector.

10. The hyperspectral image migration classification method based on a multivariate mutual learning network according to claim 2, characterized in that: The threshold range in step 7.4 is [0, 1].

Images