Cross-scene hyperspectral image classification method based on semantic perception
Through the cross-scene hyperspectral remote sensing image classification method based on semantic perception, combined with visual and semantic features, the problem of insufficient feature fusion and semantic understanding ability in cross-scene classification of hyperspectral images is solved, the classification accuracy and efficiency are improved, and the model is enhanced robustness and adaptability.
Patent Information
- Application Number
- CN202411268501.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-12-13
AI Technical Summary
The lack of effective feature fusion mechanism and semantic understanding ability in cross-scene classification of hyperspectral images leads to the incompatibility of spectral offsets and scene changes, making it difficult to achieve accurate classification.
Using a cross-scene hyperspectral remote sensing image classification method based on semantic perception, a cross-scene hyperspectral remote sensing image classification network is constructed through the combination of visual feature extraction module and multi-level semantic fusion module, and the residual module and semantics are enhanced by principal component analysis, convolution and enhancement of residual modules and semantics. Feature extraction, build the total loss function for training, and improve the robustness and adaptability of the model.
It improves the classification accuracy and efficiency of hyperspectral images, enhances the model's fusion ability to spectral and spatial characteristics, improves the adaptability and semantic understanding of scene changes, and improves the performance and robustness of image classification.
Smart Images
Figure CN119131593A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hyperspectral image classification, and in particular to a cross-scene hyperspectral image classification method based on semantic perception. Background Art
[0002] As an advanced remote sensing technology, hyperspectral imaging technology has been widely used in agriculture, environmental monitoring, military reconnaissance, urban planning and other fields in recent years. Hyperspectral images can capture information in multiple bands such as visible light, near infrared and short-wave infrared, have extremely high spectral resolution, and can provide rich information on ground features, thereby achieving accurate classification and identification of ground objects.
[0003] The performance of hyperspectral image classification depends largely on the number and quality of labeled samples. However, the process of data labeling is usually expensive and time-consuming. At present, in the real-time cross-scene classification task of hyperspectral images, there are no available labeled samples for the target image. In this case, the label information of the existing image can only be used to assist the classification of the target image. For two hyperspectral images with similar distribution of objects obtained by the same sensor, their data characteristics are relatively similar. Therefore, using the label information of similar scenes to classify new hyperspectral images is a strategy to solve the problem of hyperspectral image classification with a small number of labeled samples or even no labeled samples. The problem faced by this strategy is that due to spectral shift, the labeled samples and the new image data come from different distributions, resulting in poor performance of the classification model learned directly using the labeled samples in the old image. In order to learn a classifier that is robust to spectral shift, domain adaptation is an effective method. In the field of domain adaptation, the image with previous labeled information is called the source domain, and the target image to be classified is called the target domain.
[0004] Existing methods often focus on the extraction of spectral features or spatial features, lacking an effective fusion mechanism, resulting in the classification results being greatly affected by a single feature, making it difficult to fully utilize the multi-dimensional information of hyperspectral images. Existing methods have poor adaptability to scene changes: factors such as lighting conditions and topography in different scenes will have a significant impact on hyperspectral images, and they often show low robustness and adaptability when applied across scenes. Lack of semantic understanding ability: Existing methods mainly rely on the physical features of images for classification, lack the ability to understand the semantics of image content, and find it difficult to achieve more complex classification tasks. Summary of the invention
[0005] In order to overcome the shortcomings of the prior art, the present invention provides a cross-scene hyperspectral remote sensing image classification method based on semantic perception, in order to achieve accurate classification of hyperspectral images and improve classification accuracy and efficiency, thereby promoting the widespread application of hyperspectral images in different fields.
[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:
[0007] The cross-scene hyperspectral remote sensing image classification method based on semantic perception of the present invention is characterized in that it is carried out according to the following steps:
[0008] Step 1: Obtain the source domain dataset and the target domain dataset, and select the hyperspectral images corresponding to the number of bands consistent with the target domain dataset from the source domain dataset through principal component analysis, and then perform data enhancement and normalization to obtain the preprocessed source domain dataset. ,in, After preprocessing, Hyperspectral images;
[0009] Let the label types of the preprocessed source domain dataset and the target domain dataset be the same; the category label corresponding to the preprocessed source domain dataset is recorded as ,in, express The category label of , represents the total number of categories, express The total number of mid- and high-spectral images;
[0010] Step 2: Construct a cross-scene hyperspectral remote sensing image classification network based on semantic perception, including: a visual feature extraction module, a multi-level semantic fusion module and a classifier;
[0011] Step 2.1: The visual feature extraction module Processing is performed to obtain the i-th visual feature ;
[0012] Step 2.2: The multi-level semantic fusion module Process and obtain Semantic features of ;
[0013] Step 2.3: The classifier uses the fully connected layer to After processing, we get The classification prediction probability ;
[0014] Step 3: Use formula (d) to construct the total loss function of the cross-scene hyperspectral remote sensing image classification network :
[0015] (d)
[0016] In formula (d), λ is a hyperparameter used to balance the alignment loss. represents the alignment loss, represents the classification loss;
[0017] Step 4: Use the gradient descent method to train the hyperspectral remote sensing image classification network across the scenes and calculate the total loss function To update the network parameters, when the number of training iterations reaches the set number or the total loss function When convergence occurs, the training stops, thereby obtaining a trained optimal hyperspectral image classification model, which is used to identify the hyperspectral image and obtain the image classification result.
[0018] The cross-scene hyperspectral remote sensing image classification method based on semantic perception described in the present invention is also characterized in that the visual feature extraction module in step 2.1 is composed of a cascade of The m-th level convolution enhanced residual module is composed of a convolution enhanced residual module and a three-dimensional convolution layer, wherein the m-th level convolution enhanced residual module is cascaded by the m-th level three-dimensional residual module and the m-th level deformable convolution attention module;
[0019] Step 2.1.1: The m-th level three-dimensional residual module is formed by cascading K deformable spatial convolution blocks, and two adjacent deformable spatial convolution blocks are connected using a batch normalization layer and an activation layer, and the output of the first deformable spatial convolution block after batch normalization and activation layer is residually connected with the output of the last deformable spatial convolution block;
[0020] When m=1, the hyperspectral remote sensing image spectral block After being input into the m-th level 3D residual module in the m-th level convolution enhanced residual module for processing, the m-th level 3D residual feature map output by the m-th level convolution enhanced residual module is obtained. ;
[0021] The m-th level deformable convolution attention module is composed of the m-th level channel attention module, the m-th level spatial attention module and the m-th level maximum pooling layer in sequence;
[0022] It is input to the m-th level deformable convolution attention module in the m-th level convolution enhanced residual module, and after the aggregation processing of the maximum pooling layer and the average pooling layer in the m-th level channel attention module, we get The spatial information is processed by the shared multi-layer perceptron in the m-th level channel attention module to generate a channel attention map, and the channel attention map is processed by the activation layer to output the m-th level channel enhanced feature map ;
[0023] Will and After element-by-element multiplication, it is input into the m-th level spatial attention module of the m-th level convolution enhanced residual module, and after being processed by the deformable convolution layer and the activation layer in turn, the m-th level spatial enhanced feature map is output. ;
[0024] Will and After element-by-element multiplication and processing through a maximum pooling layer, the m-th level intermediate feature map output by the m-th level convolution enhanced residual module is obtained. ;
[0025] Step 2.1.2: When When Input to the m-th level convolution enhanced residual module for processing, thus obtaining , and then the M-th level convolution enhanced residual module outputs the M-th level intermediate feature map ;
[0026] Step 2.1.3: After being processed by the three-dimensional convolutional layer, the i-th visual feature is obtained .
[0027] Furthermore, the multi-level semantic fusion module in step 2.2 includes: a semantic hint generation module, a semantic feature extraction module and a semantic fusion module;
[0028] Step 2.2.1: The semantic prompt generation module is composed of a prompt word generation module and a prompt sentence generation module;
[0029] The prompt word generation module uses the ConceptNet network to specify , obtain each initial prompt word and its score, and sort all the initial prompt words in descending order according to the score of each initial prompt word, so as to select the first Q initial prompt words as Prompt word collection ;
[0030] The prompt sentence generation module uses the ChatGPT4 network to specify After obtaining the query prompts in the query prompts, removing redundancy and adjusting the word order, we get Prompt sentence collection ;
[0031] Step 2.2.2: The semantic feature extraction module uses the CLIP text encoder network to extract and Processing is performed to obtain the first-level semantic hint features and secondary semantic cue features ;
[0032] Step 2.2.3: The semantic fusion module will and After splicing, we get Semantic features of .
[0033] Furthermore, the step 3 comprises:
[0034] Step 3.1: Use formula (a) to construct The cross entropy loss :
[0035] (a)
[0036] In formula (a), for The unique hot encoding of the label information of category c, for The predicted probability of category c in ;
[0037] Step 3.2: Use formula (b) to construct the classification loss :
[0038] (b)
[0039] Step 3.3: Use formula (c) to construct the alignment loss :
[0040] (c)
[0041] In formula (c), is a parameter that controls the logarithmic range in softmax, Represents X Semantic features of another image of the same category, Represents X The semantic features of another image of a different category, Represents X Visual features of another image of the same category, Represents X The semantic features of another image of a different category, Represents X The set of visual features for all images of the same category, Represents X The semantic feature set of all images of different categories, Represents X The semantic feature set of all images of the same category, Represents X The set of semantic features for all images of different categories.
[0042] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the cross-scene hyperspectral image classification method, and the processor is configured to execute the program stored in the memory.
[0043] The present invention provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, which is characterized in that when the computer program is executed by a processor, the steps of the cross-scene hyperspectral image classification method are executed.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] 1. The present invention extracts the spatial spectral features of hyperspectral images by using a convolution-enhanced residual network. The three-dimensional residual module can not only effectively fuse the spectral and spatial features of hyperspectral images by using deformable spatial convolution blocks, but also enhance the model's adaptability to geometric changes by introducing learnable offsets, thereby improving the richness and discrimination of features.
[0046] 2. The present invention uses the category label as the input of the multi-level semantic fusion module, avoiding the manual construction of semantic prior knowledge. Semantic prior knowledge is the shared knowledge between the source domain dataset and the target domain dataset. The semantic space unifies the visual modality and language modality containing prior knowledge. The present invention generates multi-level semantics and makes full use of the prior knowledge contained in the category label. It can not only provide rich information representation and layer-by-layer abstraction capabilities, but also improve the generalization ability and interpretability of the model.
[0047] 3. The present invention associates the extracted visual features of the hyperspectral image with the known semantic features by constructing an alignment loss to achieve semantic representation of the features, and introduces a language pattern with prior knowledge as a supervisory signal to improve visual representation learning. The model can better understand the content in the image, thereby improving the accuracy of classification.
[0048] 4. The present invention uses deformable convolution in multiple modules to improve the convolution operation. It enhances the model's adaptability to geometric changes by introducing learnable offsets, allowing the network to better handle objects or features with irregular shapes, thereby improving the performance and robustness of the image classification model. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 is a flow chart of the method of the present invention;
[0050] Figure 2 A framework diagram of a network of the method of the present invention;
[0051] Figure 3 It is a structural diagram of the visual feature extraction module of the method of the present invention;
[0052] Figure 4 This is a structural diagram of the multi-level semantic fusion module of the method of the present invention. DETAILED DESCRIPTION
[0053] In this embodiment, a cross-scene hyperspectral remote sensing image classification method based on semantic perception is described as follows: Figure 1 As shown, the following steps are performed:
[0054] Step 1: Obtain the source domain dataset and the target domain dataset, and select the hyperspectral images corresponding to the number of bands consistent with the target domain dataset from the source domain dataset through principal component analysis, and then perform data enhancement and normalization to obtain the preprocessed source domain dataset. ,in, After preprocessing, In this embodiment, the source domain dataset and the target domain dataset are Houston 2013 and Houston 2018, respectively. The Houston 2013 dataset has 144 spectral bands, a wavelength range of 380 ~ 1050 nm, and an image spatial resolution of 2.5 m. The Houston 2018 dataset has the same wavelength range, but only has 48 spectral bands, and the image spatial resolution is 1 m. 48 spectral bands are selected from the Houston 2013 dataset by principal component analysis, and then data enhancement and normalization are performed on the hyperspectral image.
[0055] Let the label types of the preprocessed source domain dataset and the target domain dataset be the same; the category label corresponding to the preprocessed source domain dataset is recorded as ,in, express The category label of , represents the total number of categories, express The total number of medium and high spectral images; In this example, the source domain dataset and the target domain dataset have seven consistent classes, namely: Grass healthy, Grass stressed, Tress, Water, Residential buildings, Non-residential buildings, and Road.
[0056] Step 2: Construct a cross-scene hyperspectral remote sensing image classification network based on semantic perception. The network structure is as follows: Figure 2As shown, it includes: a visual feature extraction module, a multi-level semantic fusion module and a classifier;
[0057] Step 2.1: The visual feature extraction module consists of a cascade of The m-th level convolution enhancement residual module is composed of a convolution enhancement residual module and a three-dimensional convolution layer, wherein the m-th level convolution enhancement residual module is cascaded by the m-th level three-dimensional residual module and the m-th level deformable convolution attention module; in this embodiment, .
[0058] Step 2.1.1: The m-th level three-dimensional residual module is formed by cascading K deformable spatial convolution blocks, and two adjacent deformable spatial convolution blocks are connected using a batch normalization layer and an activation layer, and the output of the first deformable spatial convolution block after batch normalization and activation layer is residually connected with the output of the last deformable spatial convolution block. In this embodiment, , the m-th level three-dimensional residual module is as follows Figure 3 As shown, the deformable spatial convolution block has a convolution kernel of 3×3×3, a convolution step of 1, and is padded with zeros.
[0059] When m=1, the hyperspectral remote sensing image spectral block After being input into the m-th level 3D residual module in the m-th level convolution enhanced residual module for processing, the m-th level 3D residual feature map output by the m-th level convolution enhanced residual module is obtained. ;
[0060] The m-th level deformable convolutional attention module is composed of the m-th level channel attention module, the m-th level spatial attention module and the m-th level maximum pooling layer in sequence;
[0061] It is input to the m-th level deformable convolution attention module in the m-th level convolution enhanced residual module, and after the aggregation processing of the maximum pooling layer and the average pooling layer in the m-th level channel attention module, we get The spatial information is processed by the shared multi-layer perceptron in the m-th level channel attention module to generate a channel attention map. The channel attention map is then processed by the activation layer to output the m-th level channel enhanced feature map. In this embodiment, the m-th level deformable convolution attention module is as follows Figure 4 shown.
[0062] Will and After element-by-element multiplication, it is input into the m-th level spatial attention module of the m-th level convolution enhanced residual module, and after being processed by the deformable convolution layer and the activation layer in turn, the m-th level spatial enhanced feature map is output. ; In this embodiment, the convolution kernel of the deformable convolution layer is 3×3×3, the convolution step size is 1, the padding is zero, and the activation function of the activation layer uses the ReLu function;
[0063] Will and After element-by-element multiplication and processing through a maximum pooling layer, the m-th level intermediate feature map output by the m-th level convolution enhanced residual module is obtained. ; In this embodiment, the window size of the maximum pooling layer is set to 3×3, the step size is 1, and the padding is zero.
[0064] Step 2.1.2: When When Input to the m-th level convolution enhanced residual module for processing, thus obtaining , and then the M-th level convolution enhanced residual module outputs the M-th level intermediate feature map ;
[0065] Step 2.1.3: After being processed by the three-dimensional convolutional layer, the i-th visual feature is obtained .
[0066] Step 2.2: The multi-level semantic fusion module includes: a semantic hint generation module, a semantic feature extraction module and a semantic fusion module;
[0067] Step 2.2.1: The semantic prompt generation module consists of a prompt word generation module and a prompt sentence generation module;
[0068] The prompt word generation module uses the ConceptNet network to specify , obtain each initial prompt word and its score, and sort all the initial prompt words in descending order according to the score of each initial prompt word, so as to select the first Q initial prompt words as Prompt word collection In this embodiment, , specify The entity relationships in include: IsA, HasA, HasProperty, AtLocation, and RelatedTo.
[0069] The prompt sentence generation module uses the ChatGPT4 network to specify After obtaining the query prompts in the query prompts, removing redundancy and adjusting the word order, we get Prompt sentence collection In this embodiment, specify The query prompts in are: describe the shape, size and distribution pattern of the target object, and describe the spectral band characteristics of the target object.
[0070] Step 2.2.2: The semantic feature extraction module uses the CLIP text encoder network to and Processing is performed to obtain the first-level semantic hint features and secondary semantic cue features ; In this example, the basic size of the text encoder is 33M parameters, three layers, 512 width, 8 attention heads, the text is represented by the converter using lowercase byte pair encoding, and the vocabulary size is 49,152. In order to improve computational efficiency, the maximum sequence length is limited to 76.
[0071] Step 2.2.3: The semantic fusion module will and After splicing, we get Semantic features of .
[0072] Step 2.3: The classifier uses the fully connected layer to After processing, we get The classification prediction probability ;
[0073] Step 3: Use formula (d) to construct the total loss function of the cross-scene hyperspectral remote sensing image classification network :
[0074] (d)
[0075] In formula (d), λ is a hyperparameter used to balance the alignment loss. represents the alignment loss, represents the classification loss;
[0076] Step 3.1: Use formula (a) to construct The cross entropy loss :
[0077] (a)
[0078] In formula (a), for The unique hot encoding of the label information of category c, for The predicted probability of category c in ;
[0079] Step 3.2: Use formula (b) to construct the classification loss :
[0080] (b)
[0081] Step 3.3: Use formula (c) to construct the alignment loss :
[0082] (c)
[0083] In formula (c), is a parameter that controls the logarithmic range in softmax, Represents X Semantic features of another image of the same category, Represents X The semantic features of another image of a different category, Represents X Visual features of another image of the same category, Represents X The semantic features of another image of a different category, Represents X The set of visual features for all images of the same category, Represents X The semantic feature set of all images of different categories, Represents X The semantic feature set of all images of the same category, Represents X The set of semantic features for all images of different categories.
[0084] Step 4: Use the gradient descent method to train the hyperspectral remote sensing image classification network across the scenes and calculate the total loss function To update the network parameters, when the number of training iterations reaches the set number or the total loss function When convergence occurs, the training stops, thereby obtaining a trained optimal hyperspectral image classification model, which is used to identify the hyperspectral image and obtain the image classification result. In this embodiment, the Adam optimizer is used, the learning rate is halved every 50 iterations, and the training is iterated 200 times.
[0085] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.
[0086] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.
Claims
1. A cross-scene hyperspectral remote sensing image classification method based on semantic perception, characterized in that Follow these steps: Step 1: Obtain the source domain dataset and the target domain dataset, and select the hyperspectral images corresponding to the number of bands consistent with the target domain dataset from the source domain dataset through principal component analysis, and then perform data enhancement and normalization to obtain the preprocessed source domain dataset. ,in, After preprocessing, Hyperspectral images; Let the label types of the preprocessed source domain dataset and the target domain dataset be the same; the category label corresponding to the preprocessed source domain dataset is recorded as ,in, express The category label of , represents the total number of categories, express The total number of mid- and high-spectral images; Step 2: Construct a cross-scene hyperspectral remote sensing image classification network based on semantic perception, including: a visual feature extraction module, a multi-level semantic fusion module and a classifier; Step 2.1: The visual feature extraction module Processing is performed to obtain the i-th visual feature ; Step 2.2: The multi-level semantic fusion module Process and obtain Semantic features of ; Step 2.3: The classifier uses the fully connected layer to After processing, we get The classification prediction probability ; Step 3: Use formula (d) to construct the total loss function of the cross-scene hyperspectral remote sensing image classification network : (d) In formula (d), λ is a hyperparameter used to balance the alignment loss. represents the alignment loss, represents the classification loss; Step 4: Use the gradient descent method to train the hyperspectral remote sensing image classification network across the scenes and calculate the total loss function To update the network parameters, when the number of training iterations reaches the set number or the total loss function When convergence occurs, the training stops, thereby obtaining a trained optimal hyperspectral image classification model, which is used to identify the hyperspectral image and obtain the image classification result.
2. According to the semantic perception-based cross-scene hyperspectral remote sensing image classification method of claim 1, it is characterized in that: The visual feature extraction module in step 2.1 consists of a cascade of The m-th level convolution enhanced residual module is composed of a convolution enhanced residual module and a three-dimensional convolution layer, wherein the m-th level convolution enhanced residual module is cascaded by the m-th level three-dimensional residual module and the m-th level deformable convolution attention module; Step 2.1.1: The m-th level three-dimensional residual module is formed by cascading K deformable spatial convolution blocks, and two adjacent deformable spatial convolution blocks are connected using a batch normalization layer and an activation layer, and the output of the first deformable spatial convolution block after batch normalization and activation layer is residually connected with the output of the last deformable spatial convolution block; When m=1, the hyperspectral remote sensing image spectral block After being input into the m-th level 3D residual module in the m-th level convolution enhanced residual module for processing, the m-th level 3D residual feature map output by the m-th level convolution enhanced residual module is obtained. ; The m-th level deformable convolution attention module is composed of the m-th level channel attention module, the m-th level spatial attention module and the m-th level maximum pooling layer in sequence; It is input to the m-th level deformable convolution attention module in the m-th level convolution enhanced residual module, and after the aggregation processing of the maximum pooling layer and the average pooling layer in the m-th level channel attention module, we get The spatial information is processed by the shared multi-layer perceptron in the m-th level channel attention module to generate a channel attention map, and the channel attention map is processed by the activation layer to output the m-th level channel enhanced feature map ; Will and After element-by-element multiplication, it is input into the m-th level spatial attention module of the m-th level convolution enhanced residual module, and after being processed by the deformable convolution layer and the activation layer in turn, the m-th level spatial enhanced feature map is output. ; Will and After element-by-element multiplication and processing through a maximum pooling layer, the m-th level intermediate feature map output by the m-th level convolution enhanced residual module is obtained. ; Step 2.1.2: When When Input to the m-th level convolution enhanced residual module for processing, thus obtaining , and then the M-th level convolution enhanced residual module outputs the M-th level intermediate feature map ; Step 2.1.3: After being processed by the three-dimensional convolutional layer, the i-th visual feature is obtained .
3. According to the semantic perception-based cross-scene hyperspectral remote sensing image classification method of claim 2, it is characterized in that: The multi-level semantic fusion module in step 2.2 includes: a semantic hint generation module, a semantic feature extraction module and a semantic fusion module; Step 2.2.1: The semantic prompt generation module is composed of a prompt word generation module and a prompt sentence generation module; The prompt word generation module uses the ConceptNet network to specify , obtain each initial prompt word and its score, and sort all the initial prompt words in descending order according to the score of each initial prompt word, so as to select the first Q initial prompt words as Prompt word collection ; The prompt sentence generation module uses the ChatGPT4 network to specify After obtaining the query prompts in the query prompts, removing redundancy and adjusting the word order, we get Prompt sentence collection ; Step 2.2.2: The semantic feature extraction module uses the CLIP text encoder network to extract and Processing is performed to obtain the first-level semantic hint features and secondary semantic cue features ; Step 2.2.3: The semantic fusion module will and After splicing, we get Semantic features of .
4. The method for cross-scene hyperspectral remote sensing image classification based on semantic perception according to claim 3 is characterized in that: The step 3 comprises: Step 3.1: Use formula (a) to construct The cross entropy loss : (a) In formula (a), for The unique hot encoding of the label information of category c, for The predicted probability of category c in ; Step 3.2: Use formula (b) to construct the classification loss : (b) Step 3.3: Use formula (c) to construct the alignment loss : (c) In formula (c), is a parameter that controls the logarithmic range in softmax, Represents X The semantic features of another image of the same category, Represents X The semantic features of another image of a different category, Represents X Visual features of another image of the same category, Represents X The semantic features of another image of a different category, Represents X The set of visual features for all images of the same category, Represents X The semantic feature set of all images of different categories, Represents X The semantic feature set of all images of the same category, Represents X The set of semantic features for all images of different categories.
5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the cross-scene hyperspectral image classification method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the cross-scene hyperspectral image classification method according to any one of claims 1 to 4 are performed.