An open set hyperspectral image transfer classification method based on JS distribution adaptive modeling

Abstract

The application discloses a kind of open set hyperspectral image transfer classification methods based on JS distribution self-adapting modeling, belong to hyperspectral image processing and application technical field. Including: the hyperspectral image of source domain and target domain is preprocessed and data set division is completed;Open set transfer classification model is constructed, and then the initial model with known class discrimination ability is obtained, preliminary open set discrimination is realized;The JS divergence distribution between the prediction distribution of target domain sample and the known class probability prototype distribution of source domain is self-adapting modeling, and known class and unknown class of target domain are dynamically distinguished, while cross-domain feature alignment is realized;After model optimization, target domain test sample is classified and output, and known class fine classification and unknown class unified identification are realized.The application avoids the problem of unknown class discrimination relying on fixed threshold, fully excavates the potential distribution structure of target domain sample, and improves the accuracy, stability and generalization ability of open set hyperspectral image transfer classification.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral image processing and application technology, specifically relating to an open-set hyperspectral image migration classification method, and more particularly to an open-set hyperspectral image migration classification method based on JS distribution adaptive modeling. Background Technology

[0002] Hyperspectral image classification is a fundamental step and key issue in various hyperspectral image applications. In practical hyperspectral image application scenarios, due to factors such as imaging environment, regional differences, and time span, the target domain hyperspectral data often contains private categories not seen in the source domain training data, thus giving rise to the open-set hyperspectral image transfer classification task. Open-set hyperspectral image transfer classification needs to simultaneously achieve accurate classification of common categories and effective identification of private categories. Its performance directly determines the application value of hyperspectral technology in complex real-world scenarios, and it is currently a research hotspot and challenge in the field of hyperspectral image processing.

[0003] Researchers have conducted extensive research on open-set hyperspectral image transfer classification methods and models, and have achieved excellent results: Li Zhaokui et al. (Reference: Li Zhaokui, Zhang Ziwei, Wang Yan, Fang Zhuoqun, Liu Cuiwei, Wang Chuanyun, Li Fei, Tian Jing, Qi Mingtai, Yuan Bo. Hyperspectral Image Open Set Classification Method Based on Inverse Prototype and MLP Network [P]. Chinese Invention Patent, Application No.: CN202411646540.7, 2025) proposed a hyperspectral image open set classification method based on inverse prototype and multilayer perceptron network. By constructing an inverse prototype to represent the feature space outside the common category boundary, and combining it with an MLP network to distinguish between common and private categories, the method effectively improves the classification ability of the model in open set scenarios.

[0004] Li Jie et al. (Reference: Li Jie, Lu Chengzheng, Lei Shengwei, Xu Tuo, Gao Xinbo. A hyperspectral open set classification method combining feature reconstruction and prototype constraints [P]. Chinese Invention Patent, Application Patent No.: CN202510182016.7, 2025) proposed a hyperspectral open set classification method combining feature reconstruction and prototype constraints. By introducing a feature reconstruction mechanism to enhance feature representation ability and using prototype constraints to guide the compact distribution of common category features, the classification accuracy and stability of hyperspectral images under open set conditions are improved.

[0005] Liu Xiaomin et al. (Reference: Liu Xiaomin, Qiao Zhenzhuang, Zhang Wei, Wang Haoyu, Xing Changda, Duan Tianyang, Yin Haojie, Guo Sifan. An Open Set Cross-Domain Hyperspectral Image Classification Method and Network Based on Critical Reflective Learning [P]. Chinese Invention Patent, Application Patent No.: CN202510085621.2, 2025) disclosed an open set cross-domain hyperspectral image classification method and network based on critical reflective learning. By constructing a cross-domain learning framework and introducing a reflective learning mechanism, the model can achieve effective transfer even when there are differences in the distribution between the source domain and the target domain, thus enhancing the model's adaptability in open set cross-domain scenarios.

[0006] Li Na et al. (Reference: Li Na, Song Xiaopeng, Quan Yinghui, Liu Yongxu, Zhu Wenxiang. Hyperspectral Image Open Set Classification Method Based on Private Class Feature Learning [P]. Chinese Invention Patent, Application Patent No.: CN202510434472.6, 2025) proposed a hyperspectral image open set classification method based on private class feature learning. By explicitly modeling the distribution of private class features, the method guides the model to learn the discriminative features that distinguish between public and private classes, thereby improving the model's ability to identify private class samples.

[0007] However, existing open-set hyperspectral image transfer classification methods still have certain limitations in practical applications: on the one hand, the discrimination of private categories often relies on preset thresholds or empirical judgment rules, and the selection of thresholds is highly sensitive to different datasets and application scenarios, limiting the model's generalization ability; on the other hand, existing methods fail to fully explore the potential clustering structure information within the target domain samples, making it difficult to effectively model private categories using the distribution characteristics of the target domain data itself. These shortcomings are key issues restricting the transition of hyperspectral image classification from theory to large-scale practical applications. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an open-set hyperspectral image transfer classification method based on JS distribution adaptive modeling. This method jointly models source and target domain samples from the perspective of consistency of classification probability distribution, introduces an adaptive discrimination mechanism based on distribution differences, and combines the guiding role of the potential clustering structure within the target domain samples. This solves the problems in existing open-set hyperspectral image transfer classification methods, such as relying on fixed thresholds for modeling private categories, failing to fully utilize the potential clustering structure within the target domain samples, and having difficulty dynamically distinguishing between common and private category samples during training. As a result, this method improves the accuracy, stability, and generalization ability of open-set hyperspectral image transfer classification.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An open-set hyperspectral image transfer classification method based on JS distribution adaptive modeling specifically includes the following steps: Step 1: Obtain the raw hyperspectral images and labels, and perform data preprocessing and dataset partitioning; specifically: Step 1.1: Acquire two sets of raw hyperspectral data from different sensors or collected at different times and locations. The raw hyperspectral data includes raw hyperspectral images and their corresponding category labels. Perform radiometric and geometric correction processing on the two sets of raw hyperspectral images respectively, and remove spectral bands that are greatly affected by noise to obtain two preprocessed sets of hyperspectral images and their corresponding category labels.

[0010] Step 1.2 involves sequentially performing spectral normalization, data standardization, and spatial cropping operations on the two sets of hyperspectral images obtained in Step 1.1. This divides the hyperspectral images into multiple fixed-size patches to construct two sets of hyperspectral image classification datasets. Each patch serves as a hyperspectral sample, and its class label is determined by the class label of the center pixel of the patch. The patch size is... Where P represents the spatial dimension and C represents the spectral dimension.

[0011] Step 1.3 involves dividing the two sets of hyperspectral image classification datasets and their corresponding labels obtained in Step 1.2: one set is selected as the source domain training dataset, and the other is selected as the target domain dataset. The source domain training dataset retains only the hyperspectral samples and labels corresponding to categories that exist in both the source and target domains (common categories); the target domain dataset contains hyperspectral samples corresponding to common categories and categories that exist only in the target domain (private categories). The target domain training set does not include any category labels, while the target domain test set includes the corresponding category labels for model performance evaluation.

[0012] Step 2: Construct an open-set transfer classification model based on JS distribution adaptation. This model includes a Mamba2-based two-branch feature extractor, a virtual class-based two-branch classifier, and a selective adversarial domain discriminator. Specifically: Step 2.1: Construct a Mamba2-based dual-branch feature extractor, which employs a spatial-spectral feature joint modeling structure to encode features from the hyperspectral samples obtained in Step 1. The dual-branch feature extractor consists of a spatial branch and a spectral branch. Specifically: Step 2.1.1: The spatial branch is constructed based on a multi-layer convolutional neural network, which includes convolutional layers, batch normalization layers, nonlinear activation layers, and downsampling operations. It is used to extract the local spatial structure features and texture information of land objects in the hyperspectral samples layer by layer, and finally outputs a spatial feature vector. Step 2.1.2: The spectral branch is constructed based on the Mamba2 state-space model to model the high-dimensional spectral information in the hyperspectral samples. This branch first performs weighted aggregation of the spectral vectors in the hyperspectral samples, generating representative spectral feature inputs based on the spectral similarity between the center pixel and its neighboring pixels, thus preserving key spectral discrimination information while suppressing noise interference. Subsequently, the spectral features are mapped to the input space of the Mamba2 model through linear mapping, and its state-space modeling capability is used to model long-range dependencies of the spectral features, obtaining spectral feature vectors with global spectral discriminative power. Finally, the spatial feature vector output from the spatial branch and the spectral feature vector output from the spectral branch are concatenated and fused along the feature dimension to form a joint spectral-spatial feature vector as the final output of the feature extractor. This constitutes the feature representation space supporting subsequent classification, clustering, and distribution modeling, achieving a complete representation of the spatial-spectral coupling information of the hyperspectral samples. Step 2.2: The virtual class-based dual-branch classifier includes a main classification branch and a virtual class classification branch. The main classification branch adopts a K+1 class output structure, where K is the number of common classes. It is used to classify the joint spectral-spatial feature representation obtained in Step 2.1, thereby achieving fine classification of common classes and unified identification of private classes. The virtual class classification branch constructs a virtual class representation space, guiding private class samples to form a clustered distribution in the feature representation space, thereby expanding the discrimination boundary of common classes and improving the separability of private classes.

[0013] Step 2.3, the selective adversarial domain discriminator is used to perform adversarial modeling on the joint spectral-spatial features of the source domain samples extracted in step 2.1 and the joint spectral-spatial features of the target domain samples obtained in step 2.1. Its selective adversarial mechanism only applies domain adversarial constraints to the target domain samples that are judged to be of the common category, thereby achieving effective alignment of the common category features of the source domain and the target domain, and avoiding negative transfer caused by private category samples participating in adversarial training.

[0014] Step 3: Input the source domain training dataset obtained in Step 1.3 into the open-set transfer classification model constructed in Step 2. Use the cross-entropy loss function as the optimization objective, and train only the main classification branch of the Mamba2-based two-branch feature extractor and the two-branch classifier. The loss function is defined as: (1) in, The number of samples in the source domain training dataset; The number of public categories; Indicates the first The true category label of each sample; This indicates that the main classification branch of the open set transfer classification model is related to the first... The sample belongs to the first The predicted probability of a class.

[0015] When the change in the loss function value within a preset number of consecutive rounds is less than a preset threshold, the main classification branch of the Mamba2-based dual-branch feature extractor and dual-branch classifier is considered to have reached convergence, and pre-training is terminated. This yields a feature extractor and main classifier with the ability to discriminate common category features of the source domain, providing a model foundation for subsequent adaptive target domain modeling and open-set transfer classification based on the JS distribution. The preset threshold is... .

[0016] Step 4: Based on the feature extractor pre-trained in Step 3, perform feature clustering on the hyperspectral samples from the source and target domains, and construct virtual classes based on distribution-level optimal transport matching. Specifically: Step 4.1: Using the Mamba2-based dual-branch feature extractor obtained in Step 2.1, extract features from the source domain training set samples and the target domain training set samples obtained in Step 1.3 respectively, to obtain the source domain sample feature vector set. and the set of feature vectors of the target domain samples ; Step 4.2: Based on the source domain sample feature vector set obtained in Step 4.1, calculate the feature center of each common category in the source domain to form the source domain category center set. Defined as: (2) in, For the source domain belonging to the first A collection of hyperspectral samples of this type; This represents a two-branch feature extractor based on Mamba2; Indicates that the source domain belongs to the first Hyperspectral samples of this type; Simultaneously, K-means clustering is performed on the target domain sample feature vector set obtained in step 4.1 to obtain the target domain cluster center set. The number of clusters is set to the sum of the number of common categories and the number of preset virtual classes. sum; Step 4.3: Based on the source domain cluster center set and the target domain cluster center set obtained in Step 4.2, calculate the feature distance between them. Furthermore, under the condition that the number of category centers in the source domain is less than the number of cluster centers in the target domain, an optimal transmission constraint is introduced. Based on the assumption of uniform distribution for both the source domain category center set and the target domain cluster center set, an optimal transmission problem is constructed and the transmission matrix T is solved. The matching weights of each common category center in the source domain and each cluster center in the target domain are determined to form the overall matching relationship between the two. The transmission matrix T is defined as follows: (3) in, The weights represent the transfer weights between the k-th cluster center in the source domain and the j-th cluster center in the target domain, satisfying... and ; This represents the feature distance between the k-th cluster center in the source domain and the j-th cluster center in the target domain; Indicates the number of common categories; Indicates the number of preset virtual classes; Step 4.4: For each common category, select the corresponding transmission weight in the transmission matrix T obtained in step 4.3. The largest target domain cluster center is designated as the target domain common category center, while the remaining target domain cluster centers are determined as virtual category centers, serving as feature centers for potential private categories.

[0017] Step 5, Open-set adaptive training guided by virtual classes: Step 5.1: Introduce the virtual class centers obtained in Step 4.4 into the virtual class-based dual-branch classifier in Step 2 to construct the initial discrimination space of the target domain private categories; Step 5.2: Input the source domain training dataset and the target domain training dataset obtained in Step 1.3 into the open set transfer classification model based on JS distribution adaptive constructed in Step 2 for Steps 5.3 and 5.4. Step 5.3: For the source domain training dataset, the cross-entropy loss function is used to constrain the common category output of the main classification branch, so that the open set transfer classification model can maintain its ability to distinguish the common category of the source domain. This loss is denoted as the source domain classification loss. Step 5.4: Simultaneously, for the target domain training dataset, a virtual class loss function is calculated based on the virtual class classification branch. This guides samples in the target domain whose class center distribution differs significantly from that in the source domain to cluster towards the virtual class centers, thereby forming an initial clustering structure of potential private classes in the feature space. The virtual class loss function is: (4) in, V represents the number of samples in the target domain training dataset; V represents the number of predefined virtual classes. This represents the pseudo-label of the i-th sample in the target domain corresponding to the v-th virtual class; This represents the predicted probability of the open-set transfer classification model that the i-th sample in the target domain belongs to the v-th virtual class; Step 5.5: When the virtual class loss calculated in Step 5.4 and the source domain classification loss calculated in Step 5.3 tend to stabilize within a preset number of consecutive rounds, the first stage of training ends, and a feature representation space with preliminary private category discrimination capability is obtained. The stabilization refers to the sum of the virtual class loss and the source domain classification loss changing by less than a preset threshold within a preset number of consecutive rounds.

[0018] Step 6, Boundary optimization training based on JS distribution adaptive modeling: Step 6.1: Based on the open-set transfer classification model trained in Step 5, predict the samples in the target domain training dataset to obtain the predicted probability distribution of each target domain sample in the common category. Simultaneously, based on the statistical results of the samples in the classification probability space of the source domain training dataset, a probability prototype distribution corresponding to each common category in the source domain is constructed. Defined as: (5) in, This represents the predicted probability vector that the i-th sample in the source domain belongs to the k-th common category; For the source domain belonging to the first A collection of hyperspectral samples of this type; Indicates that the source domain belongs to the first Hyperspectral samples of this type.

[0019] Step 6.2: Use the Jensen-Shannon divergence measure to predict the probability distribution from step 6.1. With probability prototype distribution The degree of difference between them, and using this degree of difference as an adaptive metric for distinguishing the public and private categories of the target domain samples, is defined as: (6) in, This represents the predicted probability distribution of the i-th sample in the target domain; This represents the prototype probability distribution of the k-th common category in the source domain; This represents the average distribution of the predicted probability distribution of samples in the target domain and the prototype probability distribution of the common category in the source domain. JS divergence represents the Kullback-Leibler divergence, which measures the difference between two probability distributions. The JS divergence value ranges from [0,1]. The smaller the divergence value, the closer the two probability distributions are, and the more likely the target domain sample belongs to the common category. The larger the divergence value, the greater the difference between the two probability distributions, and the more likely the target domain sample belongs to the private category.

[0020] Step 6.3: Based on the statistical distribution characteristics of the Jensen-Shannon divergence values ​​in the target domain samples calculated in Step 6.2, a Laplace mixture probability model is used to model the Jensen-Shannon divergence distribution. The EM algorithm is used to optimize the model parameters to achieve accurate fitting. After fitting, high and low divergence components are adaptively distinguished based on the posterior probability of each component without setting a fixed threshold: components with small location parameters and narrow scale parameters are low divergence components (corresponding to common category samples), and components with large location parameters and wide scale parameters are high divergence components (corresponding to private category samples). Thus, the hyperspectral samples of the target domain are adaptively divided into subsets of common and private category samples to achieve unsupervised class discrimination.

[0021] Step 6.4: For samples that were determined to be common categories of the target domain in Step 6.3, an adversarial discriminator based on selective adversarial training is introduced to perform adversarial training. Cross-domain feature alignment is achieved by minimizing the distribution difference of sample features of common categories between the source domain and the target domain.

[0022] Step 6.5: For samples identified as private categories of the target domain in Step 6.3, no domain adversarial constraints are imposed. Instead, the virtual class loss in Step 5.4 is continued through the virtual class classification branch to optimize the clustering of private category samples in the virtual class space. This further widens the feature discrimination boundary between private and public categories and improves the accuracy of private category identification.

[0023] Step 6.6: By jointly optimizing the source domain supervised classification loss, virtual class loss, and selective adversarial loss, the open set discrimination boundary of the open set transfer classification model is gradually optimized, improving the classification accuracy of common categories and the stability of private category recognition in the target domain. (Joint loss...) Defined as: (7) in, This represents the source domain classification loss; Indicates virtual class loss; This indicates selective resistance to loss; Step 7: Testing the open-set transfer classification model and outputting the transfer classification results. Specifically: Step 7.1: Input the target domain test set samples obtained in Step 1.3 into the open set transfer classification model based on JS distribution adaptive training trained in Step 6, and output the final classification result of the target domain samples through the main classification branch.

[0024] Step 7.2: When a sample is determined to be in a public category, output the corresponding public category label; when a sample is determined to be in a private category, output the private category identifier.

[0025] Step 7.3 evaluates the performance of the open-set hyperspectral image transfer classification method based on JS distribution adaptive modeling by calculating four evaluation metrics: overall classification accuracy, common category classification accuracy, private category recognition rate, and harmonic open set index. The calculation formulas for each index are as follows: Overall classification accuracy: refers to the proportion of correctly classified samples in the target domain test set to the total number of samples in the test set. The calculation formula is: (8) in, This represents the number of correctly classified samples in the target domain test dataset. This represents the total number of samples in the target domain test dataset.

[0026] Common category classification accuracy: refers to the average classification accuracy of all common categories, calculated using the following formula: (9) Where K is the number of common categories, This represents the number of correctly classified samples in the c-th common category. This represents the total number of samples in the c-th common category.

[0027] Private category recognition rate: refers to the proportion of private category samples correctly identified as "private category" in the target domain test set out out of the total number of private category samples. The calculation formula is: (10) in, This represents the number of samples in the private category that were correctly identified as "private category". This represents the total number of samples in the private category.

[0028] Harmonic Open Set Metric: This metric measures both the classification accuracy of common categories and the recognition rate of private categories. It better reflects the open set capability than either the accuracy of common categories or the recognition rate of private categories alone. The calculation formula is as follows: (11) Compared with the prior art, the present invention has the following main advantages: First, by introducing an open set discrimination mechanism based on JS distribution adaptive modeling, this invention avoids the problem of relying on manually setting fixed thresholds to distinguish between public and private categories in existing technologies.

[0029] Second, by performing optimal transmission matching between the source domain category center and the target domain cluster center, and combining K-means clustering to mine the inherent distribution pattern of the target domain samples, this invention can fully mine the potential clustering structure inside the target domain samples and explicitly model the private categories by constructing a virtual class representation space.

[0030] Third, this invention constructs a spatial-spectral dual-branch feature extraction structure based on Mamba2. By combining the local spatial modeling capability of convolutional neural networks with the long-range spectral dependency modeling capability of Mamba2 state space models, it achieves collaborative modeling of spatial and spectral information of hyperspectral data, thereby improving the hyperspectral feature modeling capability. Attached Figure Description

[0031] Figure 1 This is a diagram of the open set migration classification model based on JS distribution adaptive model of the present invention.

[0032] Figure 2 This is a flowchart illustrating the implementation of the present invention.

[0033] Figure 3 This is a hyperspectral image from the Pavia Centre.

[0034] Figure 4 This is a hyperspectral image from Pavia University. Detailed Implementation

[0035] The following section, in conjunction with the technical solution and accompanying drawings, uses the Pavia Centre hyperspectral dataset and the Pavia University hyperspectral dataset for open-set hyperspectral image transfer classification as examples to further illustrate the specific implementation of the present invention.

[0036] Step 1: Obtain the raw hyperspectral images and labels, and perform data preprocessing and dataset partitioning: Step 1.1: This example uses hyperspectral data from Pavia University and Pavia Centre, and the image data is as follows: Figures 3-4 As shown. Specifically, the Pavia University hyperspectral image contains 610×340 pixels, with a spectral range of 430nm-860nm, a total of 103 spectral bands, and includes 9 land cover types, named PU, as shown. Figure 3 As shown; the hyperspectral image of Pavia Centre contains 1096×715 pixels, with a spectral range of 430nm-860nm, a total of 102 spectral bands, and includes 9 land cover types, named PC, as shown. Figure 4 As shown.

[0037] Step 1.2 involves sequentially performing spectral normalization, data standardization, and spatial cropping operations on the two sets of hyperspectral images obtained in Step 1.1. This divides the hyperspectral images into multiple fixed-size patches to construct two sets of hyperspectral image classification datasets. Each patch serves as a sample, and the class label of a sample is determined by the class label of its central pixel. The patch size is... .

[0038] Step 1.3: Divide the two sets of hyperspectral image classification datasets and their corresponding labels after processing in Step 1.2: Select PU as the source domain training dataset, and retain only the image samples and labels corresponding to the 7 common categories; PC as the target domain dataset, which contains image samples corresponding to the 7 common categories and 1 private category. The target domain training set does not have any category labels, while the target domain test set contains the corresponding category labels for model performance evaluation.

[0039] Step 2: Construct an open-set transfer classification model based on JS distribution adaptation. This model includes a Mamba2-based two-branch feature extractor, a virtual class-based two-branch classifier, and a selective adversarial domain discriminator. Specifically: Step 2.1: Construct a Mamba2-based dual-branch feature extractor, which employs a spatial-spectral feature joint modeling structure to encode features from the hyperspectral samples obtained in Step 1. The dual-branch feature extractor consists of a spatial branch and a spectral branch. Specifically: Step 2.1.1: The spatial branch is constructed based on a multi-layer convolutional neural network, which includes convolutional layers, batch normalization layers, nonlinear activation layers, and downsampling operations. It is used to extract the local spatial structure features and texture information of land objects in the hyperspectral samples layer by layer, and finally outputs a spatial feature vector. Step 2.1.2: The spectral branch is constructed based on the Mamba2 state-space model to model the high-dimensional spectral information in the hyperspectral samples. This branch first performs weighted aggregation of the spectral vectors in the hyperspectral samples, generating representative spectral feature inputs based on the spectral similarity between the center pixel and its neighboring pixels, thus preserving key spectral discrimination information while suppressing noise interference. Subsequently, the spectral features are mapped to the input space of the Mamba2 model through linear mapping, and its state-space modeling capability is used to model long-range dependencies of the spectral features, obtaining spectral feature vectors with global spectral discriminative power. Finally, the spatial feature vector output from the spatial branch and the spectral feature vector output from the spectral branch are concatenated and fused along the feature dimension to form a joint spectral-spatial feature vector as the final output of the feature extractor. This constitutes the feature representation space supporting subsequent classification, clustering, and distribution modeling, achieving a complete representation of the spatial-spectral coupling information of the hyperspectral samples. Step 2.2: The virtual class-based dual-branch classifier includes a main classification branch and a virtual class classification branch. The main classification branch adopts a 7+1 class output structure, where 7 is the number of common classes. It is used to classify the joint spectral-spatial feature representation obtained in Step 2.1, thereby achieving fine classification of common classes and unified identification of private classes. The virtual class classification branch constructs a virtual class representation space, guiding private class samples to form a clustered distribution in the feature representation space, thereby expanding the discrimination boundary of common classes and improving the separability of private classes.

[0040] Step 2.3, the selective adversarial domain discriminator is used to perform adversarial modeling on the joint spectral-spatial features of the source domain samples extracted in step 2.1 and the joint spectral-spatial features of the target domain samples obtained in step 2.1. Its selective adversarial mechanism only applies domain adversarial constraints to the target domain samples that are judged to be of the common category, thereby achieving effective alignment of the common category features of the source domain and the target domain, and avoiding negative transfer caused by private category samples participating in adversarial training.

[0041] Step 3: Input the source domain training dataset obtained in Step 1.3 into the open-set transfer classification model constructed in Step 2. Use the cross-entropy loss function as the optimization objective, and train only the main classification branch of the Mamba2-based two-branch feature extractor and the two-branch classifier. The loss function is defined as: (12) in, The number of samples in the source domain training dataset; Indicates the first The true category label of each sample; This indicates that the main classification branch of the open set transfer classification model is related to the first... The sample belongs to the first The predicted probability of the class. When the change in the loss function value over 5 consecutive rounds is less than... When the model is considered to have reached convergence, pre-training is terminated, resulting in a feature extractor and a main classifier capable of discriminating common category features in the source domain. The SGD algorithm is used for optimization, with a learning rate set to... The batch size is set to 64.

[0042] Step 4: Based on the feature extractor pre-trained in Step 3, perform feature clustering on the hyperspectral samples from the source and target domains, and construct virtual classes based on distribution-level optimal transport matching. Specifically: Step 4.1: Using the Mamba2-based dual-branch feature extractor obtained in Step 2.1, extract features from the source domain training set samples and the target domain training set samples obtained in Step 1.3 respectively, to obtain the source domain sample feature vector set. and the set of feature vectors of the target domain samples ; Step 4.2: Based on the source domain sample feature vector set obtained in Step 4.1, calculate the feature center of each common category in the source domain to form the source domain category center set. Defined as: (13) in, For the source domain belonging to the first A collection of hyperspectral samples of this type; This represents a two-branch feature extractor based on Mamba2; Indicates that the source domain belongs to the first Hyperspectral samples of this type; Simultaneously, K-means clustering is performed on the target domain sample feature vector set obtained in step 4.1 to obtain the target domain cluster center set. The number of clusters is set to the number of common categories. The number of pre-defined virtual classes sum; Step 4.3: Based on the source domain cluster center set and the target domain cluster center set obtained in Step 4.2, calculate the feature distance between them. Furthermore, under the condition that the number of category centers in the source domain is less than the number of cluster centers in the target domain, an optimal transmission constraint is introduced. Based on the assumption of uniform distribution for both the source domain category center set and the target domain cluster center set, an optimal transmission problem is constructed and the transmission matrix T is solved. The matching weights of each common category center in the source domain and each cluster center in the target domain are determined to form the overall matching relationship between the two. The transmission matrix T is defined as follows: (14) in, The weights represent the transfer weights between the k-th cluster center in the source domain and the j-th cluster center in the target domain, satisfying... and ; This represents the feature distance between the k-th cluster center in the source domain and the j-th cluster center in the target domain; Indicates the number of common categories; Indicates the number of preset virtual classes; Step 4.4: For each common category, select the corresponding transmission weight in the transmission matrix T obtained in step 4.3. The largest target domain cluster center is designated as the target domain common category center, while the remaining target domain cluster centers are determined as virtual category centers, serving as feature centers for potential private categories.

[0043] Step 5, Open-set adaptive training guided by virtual classes: Step 5.1: Introduce the virtual class centers obtained in Step 4.4 into the virtual class-based dual-branch classifier in Step 2 to construct the initial discrimination space of the target domain private categories; Step 5.2: Input the source domain training dataset and the target domain training dataset obtained in Step 1.3 into the open set transfer classification model based on JS distribution adaptive constructed in Step 2 for Steps 5.3 and 5.4. Step 5.3: For the source domain training dataset, the cross-entropy loss function is used to constrain the common category output of the main classification branch, so that the open set transfer classification model can maintain its ability to distinguish the common category of the source domain. This loss is denoted as the source domain classification loss. Step 5.4: Simultaneously, for the target domain training dataset, a virtual class loss function is calculated based on the virtual class classification branch. This guides samples in the target domain whose class center distribution differs significantly from that in the source domain to cluster towards the virtual class centers, thereby forming an initial clustering structure of potential private classes in the feature space. The virtual class loss function is: (15) in, Indicates the number of samples in the target domain training dataset; This represents the pseudo-label of the i-th sample in the target domain corresponding to the v-th virtual class; This represents the predicted probability of the open-set transfer classification model that the i-th sample in the target domain belongs to the v-th virtual class; Step 5.5: When the virtual class loss calculated in Step 5.4 and the source domain classification loss calculated in Step 5.3 tend to stabilize within a preset number of consecutive rounds, the first stage of training ends, and a feature representation space with preliminary private category discrimination capability is obtained. The stabilization refers to the sum of the virtual class loss and the source domain classification loss changing by less than a preset threshold within a preset number of consecutive rounds. The total loss is: (16) Step 6, Boundary optimization training based on JS distribution adaptive modeling: Step 6.1: Based on the open-set transfer classification model trained in Step 5, predict the samples in the target domain training dataset to obtain the predicted probability distribution of each target domain sample in the common category. Simultaneously, based on the statistical results of the samples in the classification probability space of the source domain training dataset, a probability prototype distribution corresponding to each common category in the source domain is constructed. Defined as: (17) in, This represents the predicted probability vector that the i-th sample in the source domain belongs to the k-th common category; For the source domain belonging to the first A collection of hyperspectral samples of this type; Indicates that the source domain belongs to the first Hyperspectral samples of this type.

[0044] Step 6.2: Use the Jensen-Shannon divergence measure to predict the probability distribution from step 6.1. With probability prototype distribution The degree of difference between them, and using this degree of difference as an adaptive metric for distinguishing the public and private categories of the target domain samples, is defined as: (18) in, This represents the predicted probability distribution of the i-th sample in the target domain; This represents the prototype probability distribution of the k-th common category in the source domain; This represents the average distribution of the predicted probability distribution of samples in the target domain and the prototype probability distribution of the common category in the source domain. JS divergence represents the Kullback-Leibler divergence, which measures the difference between two probability distributions. The JS divergence value ranges from [0,1]. The smaller the divergence value, the closer the two probability distributions are, and the more likely the target domain sample belongs to the common category. The larger the divergence value, the greater the difference between the two probability distributions, and the more likely the target domain sample belongs to the private category.

[0045] Step 6.3: Based on the statistical distribution characteristics of the Jensen-Shannon divergence values ​​in the target domain samples calculated in Step 6.2, a Laplace mixture probability model is used to model the Jensen-Shannon divergence distribution. The EM algorithm is used to optimize the model parameters to achieve accurate fitting. After fitting, high and low divergence components are adaptively distinguished based on the posterior probability of each component without setting a fixed threshold: components with small location parameters and narrow scale parameters are low divergence components (corresponding to common category samples), and components with large location parameters and wide scale parameters are high divergence components (corresponding to private category samples). Thus, the hyperspectral samples of the target domain are adaptively divided into subsets of common and private category samples to achieve unsupervised class discrimination.

[0046] Step 6.4: For samples that were determined to be common categories of the target domain in Step 6.3, an adversarial discriminator based on selective adversarial training is introduced to perform adversarial training. Cross-domain feature alignment is achieved by minimizing the distribution difference of sample features of common categories between the source domain and the target domain.

[0047] Step 6.5: For samples identified as private categories of the target domain in Step 6.3, no domain adversarial constraints are imposed. Instead, the virtual class loss in Step 5.4 is continued through the virtual class classification branch to optimize the clustering of private category samples in the virtual class space. This further widens the feature discrimination boundary between private and public categories and improves the accuracy of private category identification.

[0048] Step 6.6: By jointly optimizing the source domain supervised classification loss, virtual class loss, and selective adversarial loss, the open set discrimination boundary of the open set transfer classification model is gradually optimized, improving the classification accuracy of common categories and the stability of private category recognition in the target domain. (Joint loss...) Defined as: (19) in, This represents the source domain classification loss; Indicates virtual class loss; This indicates selective resistance to loss; Step 7: Testing the open-set transfer classification model and outputting the transfer classification results. Specifically: Step 7.1: Input the target domain test set samples obtained in Step 1.3 into the open set transfer classification model based on JS distribution adaptive training trained in Step 6, and output the final classification result of the target domain samples through the main classification branch.

[0049] Step 7.2: When a sample is determined to be in a public category, output the corresponding public category label; when a sample is determined to be in a private category, output the private category identifier.

[0050] Step 7.3 evaluates the performance of the open-set hyperspectral image transfer classification method based on JS distribution adaptive modeling by calculating four evaluation metrics: overall classification accuracy, common category classification accuracy, private category recognition rate, and harmonic open set index. The calculation formulas for each index are as follows: Overall classification accuracy: refers to the proportion of correctly classified samples in the target domain test set to the total number of samples in the test set. The calculation formula is: (20) in, This represents the number of correctly classified samples in the target domain test dataset. This represents the total number of samples in the target domain test dataset.

[0051] Common category classification accuracy: refers to the average classification accuracy of all common categories, calculated using the following formula: (twenty one) Where K is the number of common categories, This represents the number of correctly classified samples in the c-th common category. This represents the total number of samples in the c-th common category.

[0052] Private category recognition rate: refers to the proportion of private category samples correctly identified as "private category" in the target domain test set out out of the total number of private category samples. The calculation formula is: (twenty two) in, This represents the number of samples in the private category that were correctly identified as "private category". This represents the total number of samples in the private category.

[0053] Harmonic Open Set Metric: This metric measures both the classification accuracy of common categories and the recognition rate of private categories. It better reflects the open set capability than either the accuracy of common categories or the recognition rate of private categories alone. The calculation formula is as follows: (twenty three) The above model is used to calculate four commonly used evaluation metrics for open-set hyperspectral image transfer classification: overall classification accuracy (OS), common category classification accuracy (OS), and common category classification accuracy (OS). The four metrics are: Unique Class Recognition Rate (UNK), Harmonic Open Set (HOS), and Private Class Recognition Rate (UNK). Higher values ​​for these four metrics indicate better classification results.

[0054] The technical effects of the embodiments of the present invention are illustrated below through simulation experiments: First, the simulation conditions are as follows: The experimental platform for this invention embodiment is a desktop computer with the following configuration: Intel Core i9-10900X, 3.70GHz CPU, NVIDIA TITAN RTX GPU, and 24 GB of RAM. Python 3.11 and the PyTorch toolkit were used in the experiment.

[0055] Second, simulation results analysis: In the open set hyperspectral image transfer classification task with Pavia University (PU) as the source domain and Pavia Centre (PC) as the target domain, the method proposed in this invention was used for simulation testing, and the results are shown in Table 1.

[0056] Table 1. Simulation results data

[0057] As shown in Table 1, this invention achieves high overall open-set accuracy and harmonic open-set index in the target domain, indicating that the model achieves a good balance between common-class classification performance and private-class recognition ability. Specifically, the private-class recognition rate reaches 99.44%, demonstrating that by introducing virtual class modeling and an adaptive discrimination mechanism based on Jensen-Shannon divergence, the model can effectively guide private-class samples in the target domain to form stable clusters in the feature space, thereby significantly improving the recognition and rejection capabilities of private classes. Meanwhile, the average accuracy of the common class reaches 85.07%, indicating that the proposed Mamba2-based spectral-spatial joint feature modeling and selective adversarial alignment strategy can maintain effective transfer of common-class discrimination information from the source domain while avoiding negative transfer of private classes. Therefore, the open-set hyperspectral image transfer classification method based on JS distribution adaptive modeling proposed in this invention can simultaneously improve common-class classification accuracy and private-class recognition performance under the condition of no target domain labeling, exhibiting good open-set discrimination stability and cross-domain generalization ability, verifying the effectiveness and practical value of the technical solution of this invention.

[0058] The above-described embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling, characterized in that, The open-set hyperspectral image migration classification method includes the following steps: Step 1: Obtain the original hyperspectral images and labels, and perform data preprocessing and dataset partitioning to obtain the source domain training dataset and the target domain dataset; Step 2: Construct an open-set transfer classification model based on JS distribution adaptation. This model includes a Mamba2-based two-branch feature extractor, a virtual class-based two-branch classifier, and a selective adversarial domain discriminator. Specifically: Step 2.1: Construct a Mamba2-based dual-branch feature extractor, which adopts a spatial-spectral feature joint modeling structure to encode the features of the hyperspectral samples obtained in Step 1; the dual-branch feature extractor consists of two parts: a spatial branch and a spectral branch. Step 2.2: The dual-branch classifier based on virtual classes includes a main classification branch and a virtual class classification branch. The main classification branch adopts a K+1 class output structure, where K is the number of common classes. It is used to classify the joint spectral-spatial feature representation obtained in Step 2.1, achieving fine classification of common classes and unified identification of private classes. The virtual class classification branch constructs a virtual class representation space, guiding private class samples to form a clustered distribution in the feature representation space, expanding the discrimination boundary of common classes and improving the separability of private classes. Step 2.3: The selective adversarial domain discriminator is used to perform adversarial modeling on the joint spectral-spatial features of the source domain samples extracted in step 2.1 and the joint spectral-spatial features of the target domain samples obtained in step 2.

1. Its selective adversarial mechanism applies domain adversarial constraints only to the target domain samples that are judged to be of the common category, thereby achieving effective alignment of the common category features of the source domain and the target domain and avoiding negative transfer caused by private category samples participating in adversarial training. Step 3: Input the source domain training dataset from Step 1 into the open set transfer classification model constructed in Step 2, and use the cross-entropy loss function as the optimization objective to train only the main classification branch of the Mamba2-based two-branch feature extractor and the two-branch classifier. Step 4: Based on the feature extractor pre-trained in Step 3, perform feature clustering on the hyperspectral samples of the source and target domains, and construct virtual classes based on the optimal transmission matching at the distribution level; Step 5: Open set adaptive training guided by virtual classes; Step 6: Boundary optimization training based on JS distribution adaptive modeling; Step 7: Test the open set transfer classification model and output the transfer classification results.

2. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Acquire two sets of raw hyperspectral data from different sensors or collected at different times and locations. The raw hyperspectral data includes raw hyperspectral images and their corresponding category labels. Perform radiometric correction and geometric correction on the two sets of raw hyperspectral images respectively, and remove noise bands to obtain two preprocessed sets of hyperspectral images and their corresponding category labels. Step 1.2 involves sequentially performing spectral normalization, data standardization, and spatial cropping operations on the two sets of hyperspectral images obtained in Step 1.

1. This divides the hyperspectral images into multiple fixed-size patches to construct two sets of hyperspectral image classification datasets. Each patch serves as a hyperspectral sample, and its class label is determined by the class label of the center pixel of the patch. The size of each patch is [missing information]. Where P represents the spatial dimension and C represents the spectral dimension; Step 1.3: Divide the two sets of hyperspectral image classification datasets and their corresponding labels processed in Step 1.2: Select one set as the source domain training dataset and the other set as the target domain dataset; The source domain training dataset retains only the hyperspectral samples and labels corresponding to the common categories that exist in both the source and target domains; The target domain dataset contains hyperspectral samples corresponding to the common categories and the private categories that exist only in the target domain. The target domain training set does not have any category labels, while the target domain test set contains the corresponding category labels for model performance evaluation.

3. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 2, characterized in that, Specifically, step 2.1 is as follows: Step 2.1.1: The spatial branch is constructed based on a multi-layer convolutional neural network, which includes convolutional layers, batch normalization layers, nonlinear activation layers, and downsampling operations. It is used to extract the local spatial structure features and texture information of land objects in the hyperspectral samples layer by layer, and finally outputs a spatial feature vector. Step 2.1.2: The spectral branch is constructed based on the Mamba2 state-space model and is used to model the high-dimensional spectral information in the hyperspectral samples. First, the spectral vectors in the hyperspectral samples are weighted and aggregated, and representative spectral feature inputs are generated based on the spectral similarity between the center pixel and its neighboring pixels. Subsequently, spectral features are mapped to the input space of the Mamba2 model through linear mapping, and its state space modeling capability is used to model the long-range dependency of spectral features to obtain spectral feature vectors with global spectral discriminativeness. Finally, the spatial feature vectors output by the spatial branch and the spectral feature vectors output by the spectral branch are concatenated and fused in the feature dimension to form a joint spectral-spatial feature vector as the final output of the feature extractor. This constitutes the feature representation space that supports subsequent classification, clustering and distribution modeling, and realizes a complete representation of the spatial-spectral coupling information of hyperspectral samples.

4. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 3, characterized in that, In step 3, the loss function is defined as: （1） in, The number of samples in the source domain training dataset; The number of public categories; Indicates the first The true category label of each sample; This indicates that the main classification branch of the open set transfer classification model is related to the first... The sample belongs to the first The predicted probability of a class; When the change in the loss function value within a preset number of consecutive rounds is less than a preset threshold, the main classification branch of the Mamba2-based dual-branch feature extractor and dual-branch classifier is considered to have reached convergence, and pre-training is terminated. This yields a feature extractor and main classifier with the ability to discriminate common category features of the source domain, providing a model foundation for subsequent adaptive target domain modeling and open-set transfer classification based on the JS distribution. The preset threshold is... .

5. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 4, characterized in that, Step 4 specifically involves: Step 4.1: Using the Mamba2-based dual-branch feature extractor obtained in Step 2.1, extract features from the source domain training set samples and the target domain training set samples obtained in Step 1.3 respectively, to obtain the source domain sample feature vector set. and the set of feature vectors of the target domain samples ; Step 4.2: Based on the source domain sample feature vector set obtained in Step 4.1, calculate the feature center of each common category in the source domain to form the source domain category center set. Defined as: （2） in, For the source domain belonging to the first A collection of hyperspectral samples of this type; This represents a two-branch feature extractor based on Mamba2; Indicates that the source domain belongs to the first Hyperspectral samples of this type; Simultaneously, K-means clustering is performed on the target domain sample feature vector set obtained in step 4.1 to obtain the target domain cluster center set. The number of clusters is set to the sum of the number of common categories and the number of preset virtual classes. sum; Step 4.3: Based on the source domain cluster center set and the target domain cluster center set obtained in Step 4.2, calculate the feature distance between them. Furthermore, under the condition that the number of source domain category centers is less than the number of target domain cluster centers, an optimal transmission constraint is introduced. Based on the assumption of uniform distribution for the source domain category center set and the target domain cluster center set, an optimal transmission problem is constructed and the transmission matrix T is solved. The matching weights of each source domain common category center and each target domain cluster center are determined to form the overall matching relationship between the two. Step 4.4: For each common category, select the corresponding transmission weight in the transmission matrix T obtained in step 4.

3. The largest target domain cluster center is designated as the target domain common category center, while the remaining target domain cluster centers are determined as virtual category centers, serving as feature centers for potential private categories.

6. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 5, characterized in that, In step 4.3, the transmission matrix T is defined as: （3） in, The weights represent the transfer weights between the k-th cluster center in the source domain and the j-th cluster center in the target domain, satisfying... and ; This represents the feature distance between the k-th cluster center in the source domain and the j-th cluster center in the target domain; Indicates the number of common categories; This indicates the number of predefined virtual classes.

7. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 6, characterized in that, Step 5 specifically involves: Step 5.1: Introduce the virtual class centers obtained in Step 4.4 into the virtual class-based dual-branch classifier in Step 2 to construct the initial discrimination space of the target domain private categories; Step 5.2: Input the source domain training dataset and the target domain training dataset obtained in Step 1.3 into the open set transfer classification model based on JS distribution adaptive constructed in Step 2 for Steps 5.3 and 5.

4. Step 5.3: For the source domain training dataset, the cross-entropy loss function is used to constrain the common category output of the main classification branch, so that the open set transfer classification model can maintain its ability to distinguish the common category of the source domain. This loss is denoted as the source domain classification loss. Step 5.4 Simultaneously, for the target domain training dataset, calculate the virtual class loss function based on the virtual class classification branch, and guide the samples in the target domain that have a large difference from the distribution of the class centers in the source domain to gather towards the virtual class centers, thereby forming an initial clustering structure of potential private classes in the feature space; The virtual class loss function is: （4） in, V represents the number of samples in the target domain training dataset; V represents the number of predefined virtual classes. This represents the pseudo-label of the i-th sample in the target domain corresponding to the v-th virtual class; This represents the predicted probability of the open-set transfer classification model that the i-th sample in the target domain belongs to the v-th virtual class; Step 5.5: When the virtual class loss calculated in step 5.4 and the source domain classification loss calculated in step 5.3 tend to stabilize within a preset number of consecutive rounds, the first stage of training ends, and a feature representation space with preliminary private category discrimination capability is obtained; the stabilization means that the sum of the virtual class loss and the source domain classification loss changes by less than a preset threshold within a preset number of consecutive rounds.

8. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 7, characterized in that, Step 6 specifically involves: Step 6.1: Based on the open-set transfer classification model trained in Step 5, predict the samples in the target domain training dataset to obtain the predicted probability distribution of each target domain sample in the common category. Simultaneously, based on the statistical results of the samples in the classification probability space of the source domain training dataset, a probability prototype distribution corresponding to each common category in the source domain is constructed. Defined as: （5） in, This represents the predicted probability vector that the i-th sample in the source domain belongs to the k-th common category; For the source domain belonging to the first A collection of hyperspectral samples of this type; Indicates that the source domain belongs to the first Hyperspectral samples of this type; Step 6.2: Use the Jensen-Shannon divergence measure to predict the probability distribution from step 6.

1. With probability prototype distribution The degree of difference between them, and using this degree of difference as an adaptive metric for distinguishing the public and private categories of the target domain samples, is defined as: （6） in, This represents the predicted probability distribution of the i-th sample in the target domain; This represents the prototype probability distribution of the k-th common category in the source domain; This represents the average distribution of the predicted probability distribution of samples in the target domain and the prototype probability distribution of the common category in the source domain. This represents the Kullback-Leibler divergence, used to measure the difference between two probability distributions; the JS divergence value ranges from [0,1]. Step 6.3: Based on the statistical distribution characteristics of the Jensen-Shannon divergence values ​​in the target domain samples calculated in Step 6.2, a Laplace mixture probability model is used to model the Jensen-Shannon divergence distribution. The EM algorithm is used to optimize the model parameters to achieve accurate fitting. After fitting, the high and low divergence components are adaptively distinguished based on the posterior probability of each component without setting a fixed threshold. Thus, the hyperspectral samples in the target domain are adaptively divided into subsets of public and private category samples, achieving unsupervised class discrimination. Step 6.4: For samples that were determined to be common categories of the target domain in Step 6.3, an adversarial discriminator based on selective adversarial training is introduced to perform adversarial training. Cross-domain feature alignment is achieved by minimizing the distribution difference of sample features of common categories between the source domain and the target domain. Step 6.5: For samples identified as private categories of the target domain in Step 6.3, no domain adversarial constraints are imposed. Instead, the virtual class loss in Step 5.4 is continued through the virtual class classification branch to optimize the clustering of private category samples in the virtual class space and improve the accuracy of private category identification. Step 6.6: By jointly optimizing the source domain supervised classification loss, virtual class loss, and selective adversarial loss, the open set discrimination boundary of the open set transfer classification model is gradually optimized, improving the classification accuracy of common categories and the stability of private category recognition in the target domain. (Joint loss...) Defined as: （7） in, This represents the source domain classification loss; Indicates virtual class loss; This indicates selective resistance to loss.

9. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 8, characterized in that, Step 7 specifically involves: Step 7.1: Input the target domain test set samples obtained in Step 1.3 into the open set transfer classification model based on JS distribution adaptive training trained in Step 6, and output the final classification result of the target domain samples through the main classification branch; Step 7.2: When a sample is determined to be in a common category, output the corresponding common category label; when a sample is determined to be in a private category, output the private category identifier. Step 7.3 evaluates the performance of the open-set hyperspectral image transfer classification method based on JS distribution adaptive modeling by calculating four evaluation metrics: overall classification accuracy, common category classification accuracy, private category recognition rate, and harmonic open set index.

10. The method for open-set hyperspectral image transfer classification based on JS distribution adaptive modeling according to claim 9, characterized in that, In step 7.3, the calculation formulas for each indicator are as follows: Overall classification accuracy: refers to the proportion of correctly classified samples in the target domain test set to the total number of samples in the test set. The calculation formula is: （8） in, This represents the number of correctly classified samples in the target domain test dataset. This represents the total number of samples in the target domain test dataset; Common category classification accuracy: refers to the average classification accuracy of all common categories, calculated using the following formula: （9） Where K is the number of common categories, This represents the number of correctly classified samples in the c-th common category. This represents the total number of samples in the c-th common category; Private category recognition rate: refers to the proportion of private category samples correctly identified as "private category" in the target domain test set out out of the total number of private category samples. The calculation formula is: （10） in, This represents the number of samples in the private category that were correctly identified as "private category". This represents the total number of samples in the private category; Harmonic Open Set Metric: This metric measures both the classification accuracy of common categories and the recognition rate of private categories. It better reflects the open set capability than either the accuracy of common categories or the recognition rate of private categories alone. The calculation formula is as follows: （11）。

Images