A Hyperspectral Image Transfer Classification Method Based on Multivariate Mutual Learning Network

Abstract

The present invention provides a hyperspectral image transfer classification method based on a multi-source mutual learning network, belonging to the technical field of remote sensing image processing and applications. First, two hyperspectral images of different regions are collected respectively as the source domain and the target domain, and the dataset features are extracted; secondly, an adaptive neighborhood search method is used to find the neighbors most similar to the target domain samples, and a cross-domain manifold mixing 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 categories that the target samples may belong to; finally, an unknown-class classifier network is used to score the results output by the known-class classifier network. The present invention can solve the problem that existing hyperspectral image classification technologies are difficult to transfer across data and tasks; it can realize information sharing and transmission between hyperspectral images with different spatial-spectral resolutions, different spectral ranges, and different task coverage, and provide an effective solution for continuous hyperspectral image classification.

Description

Technical Field

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

[0002] Hyperspectral images have numerous imaging bands. The spectral imager has dozens or even hundreds of bands in the visible and infrared spectral regions, and can provide both spatial domain information and spectral domain information at the same time. Moreover, with high spectral resolution and rich detail information, it is conducive to fine ground object analysis. It has wide applications in the fields of military reconnaissance, vegetation research, agricultural monitoring, etc. However, the large amount of hyperspectral data causes redundancy and high correlation between bands, making classification require a large number of training samples. Most importantly, since the annotation of hyperspectral data sets depends on professional equipment and experts in related fields, it is unrealistic 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 the source domain with rich labels to the target domain with scarce labels, and have achieved remarkable results. However, the traditional domain adaptation assumes that the two domains share the same ground object types. Due to the different compositions of ground objects in different regions, there are very likely private categories in the source domain and the target domain. Therefore, it cannot be directly applied to cross-scene hyperspectral image classification. A more practical method is not to impose any prior knowledge on the target domain label set, and only assume that there is a common part between the source domain and the target domain labels, which we call general hyperspectral transfer classification.

[0004] Zhou Hao et al. (Chinese invention patent, application patent number: CN202310869709.4) provided a classification method, device and electronic device for hyperspectral images, and obtained a hyperspectral image data set. The dual-channel Siamese network based on convolutional residual blocks and spatial attention mechanism after network optimization was used for training the classification model, and the method of transfer learning was adopted to classify hyperspectral images in cross-domain scenarios, improving the classification accuracy of hyperspectral images. However, it was specifically optimized 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, application patent number: CN202311699520.1) disclosed a spectral classification method based on a multi-branch attention network. By using multi-branch attention to further learn the feature data output by the basic convolutional block, the feature extraction ability of the network is enhanced. Feature fusion is performed on the feature data output by different branches using branch weights, enabling the classification model to learn more rich global and local information of Raman characteristic peaks, and improving the accuracy of hyperspectral image classification. However, it does not consider the cross-domain class imbalance problem. Due to the possible different class distributions, the prediction accuracy of the minority classes is reduced, and the generalization ability of the model is insufficient.

[0006] These deficiencies have become the key bottlenecks restricting the transformation of hyperspectral image classification from theoretical research to large-scale practical applications. There is an urgent need to develop a cross-domain hyperspectral image recognition method that can handle different data class distributions and scales, so as to meet the new data and new tasks faced by the model 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 transfer classification method based on a multi-source mutual learning network to solve the problem that it is difficult for existing hyperspectral image classification technologies to achieve cross-data transfer. This method reduces intra-domain variations based on adaptive neighborhood selection to obtain a more generalized feature representation. This method uses a cross-domain manifold mixing method to simulate unknown class samples and expand the decision boundary. The present invention proposes a multi-source mutual learning network to solve the problem of separating common classes and private classes; proposes a cross-domain manifold mixing method based on the mutual learning network to solve the feature drift problem; proposes an adaptive neighborhood search method to solve the class imbalance problem. The present invention can realize information sharing and transfer between hyperspectral images with different spatial-spectral resolutions, different spectral ranges, and different task coverages, providing an effective solution for continuous hyperspectral image classification, and bringing the hyperspectral image classification theory closer to large-scale practical applications.

[0008] To achieve the above purpose, the technical solutions adopted by the present invention are:

[0009] A hyperspectral image transfer classification method based on a multi-source mutual learning network. In the hyperspectral image transfer classification method, first, two hyperspectral images of different regions are collected by a spectrometer, and each is used as the source domain and the target domain respectively. Second, a convolutional neural network based on channel attention is used to extract the features of the dataset. Second, an adaptive neighborhood search method is used to find the neighbors most similar to the target domain samples, and a cross-domain manifold mixing method is used to enhance the network's recognition ability for unknown classes. Third, a known class classifier network is used to output the label classes that the target samples may belong to. Finally, an unknown class classifier network is used to score the results output by the known class classifier network. Those with scores higher than the threshold can be classified as known classes, and those lower than the threshold are identified as unknown classes. The specific steps are as follows:

[0010] Step 1: Select two hyperspectral images containing the same label classes (hereinafter referred to as common classes). 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 classes contained in the source domain but not in the target domain are called source domain private classes, and the label classes not contained in the source domain but contained in the target domain are called target domain private classes. The source domain dataset will be used as the dataset for training the network, and the target domain dataset will be used as the dataset for testing. Read the data of the source domain and the target domain, and preprocess the data of the source domain and the target domain. The preprocessing includes dimension unification and data expansion.

[0011] Step 2: Input the preprocessed hyperspectral data of the source domain and the target domain obtained in Step 1 into the feature extraction network based on channel attention (hereinafter referred to as the feature extraction network) respectively to obtain the feature vectors of the source domain data and the target domain data.

[0012] Step 3: Use the spectral features extracted from the target domain in Step 2 to input into the similarity calculation network to calculate the neighborhood similarity of the target samples. First, maintain the target domain data memory bank, which is updated after each mini-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 neighborhood similarity between the input sample and the kth sample in the target domain is calculated by the inner product, and the calculation formula is: s j,k =m j ·m k where m k represents the L2-normalized feature of the kth sample.

[0013] Step 4. Meanwhile, based on the idea of manifold mixing, construct mixed samples between the source domain and the target domain to simulate potential unknown class samples (hereinafter referred to as cross-domain manifold mixing). Interpolate and mix the feature vectors of the source domain and the target domain extracted in Step 2 through a feature mixing network. Specifically: randomly sample an interpolation factor λ ∼ Beta(α, α) from the Beta distribution, where α is the hyperparameter of the Beta distribution, and the actual value is generally between [0.1, 10]. λ represents the interpolation factor; combined with the Beta distribution, use the features of the source domain data and the feature vectors of the target domain data in Step 2 to generate a mixed feature vector. The formula for generating the mixed feature vector is as follows: Where represents the mixed feature vector; λ represents the interpolation factor; represents the feature vector of the source domain data; represents the source domain data; represents the feature vector of 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 class with the maximum probability distribution of the training samples on all known classes. The known class classifier network consists of multiple binary classifier networks, and the number of binary classifier networks is the same as the number of label classes of the source domain hyperspectral image. The loss function used in the training process is the cross-entropy loss function.

[0015] Step 6. Input the input result of Step 5 into 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 decisions of the known class classifier network to guide the optimization of the unknown class classifier network. The unknown class classifier network also consists of multiple binary classifiers, and the number of binary classifier networks is the same as the number of label classes of the source domain hyperspectral image. Use the open-set entropy minimization loss and the cross-domain loss to train the unknown class classifier network. By minimizing the cross-domain loss function, the positive class score of the mixed samples is reduced.

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

[0017]

[0018] Where K represents the number of source domain classes; P o represents the probability output by the unknown class classifier network; l represents the label index of the known classes in the source domain; x t represents the unlabeled data in the target domain; P o (l|x t ) represents the positive class score of the unknown class classifier network for the l-th class.

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

[0020] where, represents the source - domain data; represents the source - domain data label; is the target - domain data; represents the mixed feature vector.

[0021] Step 7, test 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, input the features from Step 7.1 into the known - class classifier network to output the maximum - probability class label for each sample.

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

[0025] Step 7.4, compare the positive scores output in Step 7.3 with the threshold, and the threshold range is [0, 1]. If the output score of the unknown - class classifier network for this 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 class is maintained and the corresponding known - class label is output.

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

[0027] First, in Step 3 of the present invention, an adaptive domain search algorithm is proposed, which overcomes the problem of unbalanced samples of different categories in hyperspectral remote - sensing images, maximizes the feature similarity between samples and their neighbors, enhances feature generalization, and reduces intra - domain variations.

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

[0029] Third, in Steps 5 and 6 of the present invention, the decisions of the known - class classifier network and the unknown - class classifier network are linked through consistency constraints to innovatively identify unknown classes. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the principle flowchart of the present invention.

[0031] Figure 2 is the source - domain hyperspectral data acquisition scenario; Figure 2Among them, (a) is the Pavia University hyperspectral image; Figure 2 Among them, (b) is the ground truth map.

[0032] Figure 3 is the hyperspectral data acquisition scenario of the target domain; Figure 3 Among them, (a) is the Pavia Center hyperspectral image; Figure 3 Among them, (b) is the ground truth map. Specific implementation manner

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

[0034] A hyperspectral image migration classification method based on a multi-source mutual learning network includes the following steps:

[0035] Step 1: Select two hyperspectral images containing the same label categories (hereinafter referred to as common categories), 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 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. Among them, 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 data set for testing.

[0036] Step 1.1: Read the data of the source domain and the data of the target domain If the spectral dimensions of the hyperspectral images in the source domain and the target domain are different, zero-padding is performed on the hyperspectral image with a lower dimension to make it dimensionally unified with the hyperspectral image with a higher dimension for dimensional unification and data expansion. Zero matrices are added around the hyperspectral image, and the mirror data blocks expanded up, down, left, and right are obtained through splicing.

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

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

[0039] Step 3.1: At the same time, use the target domain features extracted in Step 2 to maintain the target domain data memory bank, which is updated in place after each mini-batch training, and L2 normalization is performed on the features: Where represents the feature vector of the jth target domain sample, Denote the L2 norm of the feature vector, m j Denote the L2-normalized feature of the j-th sample.

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

[0041] Step 3.3, use the Softmax function to normalize the neighborhood similarities of all target domain samples, so that The probability of sharing the same class with other samples can be estimated as: where p jk Denote the probability that the target domain sample belongs to the same class, s j,k Denote the similarity score of the target domain sample, N t Denote 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 within-domain variability by maximizing the probability between the input sample and its neighbors, where the loss function is: where N j Denote the set of neighbors of the sample , w jk is the weight representing the neighbor confidence, p jk Denote the probability that the target domain sample belongs to the same class.

[0043] Step 4, interpolate and mix the source domain and target domain feature vectors extracted in Step 2 through a feature mixing network:

[0044] Step 4.1, randomly sample an interpolation factor λ ∼ Beta(α, α) from the Beta distribution, where λ represents the interpolation factor, and α is the hyperparameter of the Beta distribution, and α is set to 2.0 to enable the network to perform smooth interpolation in the intermediate states between different domains.

[0045] Step 4.2, use the source domain sample feature vector and the target domain sample feature vector in Step 2 to generate a cross-domain mixed feature vector where λ is the interpolation factor described in Step 4.1, where represents the mixed feature vector; λ represents the interpolation factor; represents the feature vector of the source domain data; represents the source domain data; represents the feature vector of the target domain data; Represents the target domain data.

[0046] Step 5: Input the neighborhood similarity obtained in Step 3 and the mixed feature vectors in Step 4 into the known class classifier network for training to obtain the output results. The known class classifier network consists of multiple binary classifier networks, and the number of binary classifier networks is the same as the number of label classes 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, and represents the predicted sample label value.

[0047] Step 6: Then, use the input results 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 represents the possibility that the sample belongs to a certain known class. Jointly optimize the decision results of the known class classifier network and the unknown class classifier network for consistency constraint, and use the high - confidence decisions of the known class classifier network to guide the optimization of the unknown class classifier network. The unknown class classifier network also consists of multiple binary classifiers, and the number of binary classifier networks is the same as the number of label classes of the source domain hyperspectral image. Use the open - set entropy minimization loss and the cross - domain mixing loss to train the unknown class classifier network, where the open - set entropy formula is as follows:

[0048]

[0049] where K represents the number of source domain classes. The cross - domain loss function is: where P o represents the probability output by the unknown class classifier network; l represents the index of the known class; x t represents the unlabeled data in the target domain; P o (l|x t ) represents the positive class score of the l - th class by the unknown class classifier network.

[0050] Step 7: Testing phase:

[0051] Step 7.1: Input the pre - processed target domain data in Step 1.1 into the feature extraction network in Step 2 to obtain the target domain sample feature vectors

[0052] Step 7.2: Further, use the features in Step 7.1 to input into the known class classifier network, and record the maximum - probability class label output for each sample.

[0053] Step 7.3: Further, according to the prediction results in Step 7.2, select the unknown class classifier network corresponding to the maximum - probability class, 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 a threshold value. The threshold value is taken as 0.5. If the output score of the unknown class classifier network for this sample is lower than the threshold value, it is marked as the unknown class, and the unknown class label is output. Otherwise, keep the initially predicted known class and output the corresponding known class label.

[0055] Simulation experiment to illustrate 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 processing unit, and 32 GB of memory.

[0058] The experimental data of the present invention are Pavia University hyperspectral data and Pavia Center hyperspectral data, captured by the ROSIS sensor in the northern region of Pavia, Italy. It consists of two scenes: Pavia University and Pavia Center. The spatial resolution is 1.3 m, and the spectral resolution is 4 nm. Figure 2 As shown in (a) in [reference], the source domain Pavia University hyperspectral data has a size of 610×340 pixels, 103 bands, and a total of 9 classes, including asphalt road, grassland, gravel, tree, metal plate, land, asphalt, brick, and shadow. Figure 2 As shown in (b) in [reference] is the ground truth map of Pavia University data. Figure 3 As shown in (a) in [reference] is the target domain Pavia Center hyperspectral data, with a size of 1096×715 pixels and a total of 9 classes, including water, tree, grassland, brick, land, asphalt road, asphalt, tile, and shadow. Figure 3 As shown in (b) in [reference] is the ground truth map of Pavia Center 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, analysis of simulation results. The classification results are shown in Table 1.

[0060] The comparative method adopted in the present invention is the 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 at present.

[0061] According to the simulation result data in Table 1, the method of the present invention significantly leads the existing best method in 5 out of 7 common categories. Especially in the bare soil category, the existing method only has an identification accuracy of 13.9% for it, while the present invention can reach an accuracy of 64.8%. Moreover, the average accuracy of all samples and the accuracy of each category are higher than those of the existing method. By using the hyperspectral image transfer classification method based on the multi-mutual learning network of the present invention, the classification model can be transferred to new data, and new categories can be identified without any labeled samples.

[0062] Table 1. Analysis of simulation results and classification results

[0063] Related task category Proportion of comparative example (%) Classification result of the present invention (%) Tree 85.5 89.84 Asphalt road 82.0 90.32 Self-locking brick 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 embodiments only represent the implementation manners of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A hyperspectral image transfer classification method based on a multi-source mutual learning network, characterized in that The described hyperspectral image transfer classification method is as follows: First, use a spectrometer to collect two hyperspectral images of different regions respectively as the source domain and the target domain. Second, use a convolutional neural network based on channel attention to extract the features of the dataset. Second, use the adaptive neighborhood search method to find the neighbors most similar to the target domain samples, and use the cross-domain manifold mixing method to enhance the network's recognition ability for unknown classes. Third, use the known class classifier network to output the label classes that the target samples may belong to. Finally, use the unknown class classifier network to score the results output by the known class classifier network. Samples with scores higher than the threshold are classified as known classes, and those with scores lower than the threshold are identified as unknown classes; The described hyperspectral image transfer classification method specifically includes the following steps: Step 1: Select two hyperspectral images containing the same label classes; select one of the hyperspectral images as the source dataset, defined as the source domain; select the other hyperspectral image as the target dataset, defined as the target domain; the source domain dataset is used as the dataset for training the network, and the target domain dataset is used as the dataset for testing; read the data of the source domain and the target domain, and preprocess the source domain data and the target domain data; Step 2: Input the preprocessed hyperspectral data of the source domain and the target domain obtained in Step 1 into the feature extraction network based on channel attention respectively to obtain the feature vectors of the source domain data and the target domain data; Step 3: Using the spectral features in the target domain extracted in Step 2, input them into the similarity calculation network to calculate the neighborhood similarity of the target sample. First, maintain the target domain data memory bank, which is updated after each batch training, and perform L2 normalization on the features. Input the sample and the th sample in the target domain, and calculate the neighborhood similarity through the inner product; Step 4: At the same time, based on the cross-domain manifold mixing method, construct mixed samples between the source domain and the target domain to simulate potential unknown class samples; interpolate and mix the feature vectors of the source domain and the target domain extracted in Step 2 through the feature mixing network; combined with the Beta distribution, generate mixed feature vectors using the features of the source domain data and the feature vectors of the target domain data in Step 2. The formula for generating the mixed feature vectors is: , where represents the mixed feature vector; represents the interpolation factor; represents the feature vector of the source domain data; represents the source domain data; represents the feature vector of the target domain data; represents the target domain data; Step 5: Input the neighborhood similarity obtained in Step 3 and the mixed feature vectors in Step 4 into the known class classifier network for training, and output the maximum probability distribution class of the training samples on all known classes; Step 6: Input the input result of Step 5 into 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 decisions 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 classes of the source domain hyperspectral images; use the open-set entropy minimization loss and the cross-domain loss to train the unknown class classifier network; by minimizing the cross-domain loss function, reduce the positive class scores of the mixed samples; Step 7: In the test phase, specifically: Step 7.1: Use the feature extraction network in Step 2 to extract the feature vectors of the target domain samples; Step 7.2: Use the feature vectors of the target domain samples in Step 7.1 to input into the known class classifier network to 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 in Step 7.2, and record the positive scores output by the unknown class classifier network; Step 7.4: Compare the positive fraction output in Step 7.3 with the threshold. If the output score of the unknown class classifier network for this sample is lower than the threshold, mark it as the unknown class and output the unknown class label; otherwise, keep the initially predicted known class and output the corresponding known class label. In the said Step 6: The formula for the open-set entropy minimization loss is: , where K represents the number of source domain classes; represents the probability output by the unknown class classifier network; represents the label index of the known classes in the source domain; represents the unlabeled data in the target domain; represents the positive class score of the unknown class classifier network for the -th class; The formula for the cross-domain loss is as follows: , where represents the source domain data; represents the source domain data label; the target domain data; represents the mixed feature vector.

2. The hyperspectral image transfer classification method based on a multi-source mutual learning network according to claim 1, wherein, In the said Step 1, the label categories that are included in the source domain but not in the target domain are called source domain private classes, and the label categories that are not included in the source domain but are included in the target domain are called target domain private classes.

3. A hyperspectral image transfer classification method based on a multi-source mutual learning network according to claim 1, wherein In the said Step 1, the preprocessing includes dimension unification and data augmentation.

4. A hyperspectral image transfer classification method based on a multi-source mutual learning network according to claim 1, characterized in that In the said Step 3, L2 normalization adopts the following formula: , where represents the feature vector of the th target domain sample, represents the L2-normalized feature of the th sample; The inner product calculation formula is as follows: , where represents the L2-normalized feature of the th sample.

5. A hyperspectral image transfer classification method based on a multi-source mutual learning network according to claim 1, characterized in that In the aforementioned step 4, the Beta distribution is: randomly sample an interpolation factor from the Beta distribution , where are the hyperparameters of the Beta distribution; represents the interpolation factor.

6. A hyperspectral image migration classification method based on a multi-source mutual learning network according to claim 5, characterized in that The value ranges from to 7. A hyperspectral image transfer classification method based on a multi-source mutual learning network according to claim 1, characterized in that In the said Step 5, the known class classifier network consists 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.

8. A hyperspectral image transfer classification method based on a multi-source mutual learning network according to claim 1, characterized in that, The threshold range in step 7.4 described above is .

Images