Hyperspectral image increment classification method based on interactive graph distillation network
Through the interactive graph distillation network and prototype-assisted memory transfer technology, the problem of performance degradation of old class recognition in the hyperspectral image classification model in the incremental environment is solved, and efficient knowledge transfer and classification accuracy improvement are achieved.
Patent Information
- Application Number
- CN202510784985.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-12
- Publication Date
- 2025-09-23
AI Technical Summary
Traditional hyperspectral image classification models suffer from catastrophic forgetting in incremental environments, which leads to a sharp degradation in the recognition performance of old classes and they are unable to effectively cope with the emergence of new land object categories in dynamic data streams.
An incremental classification method for hyperspectral images based on an interactive graph distillation network is adopted. By constructing a feature extraction model and a dual-graph interaction strategy graph neural network, combined with interactive graph distillation and prototype-assisted memory transfer technology, knowledge solidification and cross-stage transfer are achieved, and model parameters are optimized to improve the old class recognition performance.
It effectively alleviates the catastrophic forgetting of old knowledge, reduces labeling costs, realizes the continuous learning ability of hyperspectral images, and improves classification accuracy and model adaptability.
Smart Images
Figure CN120689750A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hyperspectral image incremental classification, and in particular to a hyperspectral image incremental classification method based on an interactive graph distillation network. Background Art
[0002] Hyperspectral imagery consists of hundreds of continuous spectral bands, which facilitates target detection and provides richer information. Its nanometer-scale spectral resolution can resolve spectral differences between subclasses of surface materials, providing a unique advantage for accurate object identification. However, with the continuous improvement of remote sensing platform observation capabilities, new object categories are constantly emerging in dynamic data streams, posing a severe challenge to traditional hyperspectral image classification models. Due to their reliance on fixed class assumptions, catastrophic forgetting can lead to a sharp degradation in the recognition of old classes in incremental environments. Summary of the Invention
[0003] The present invention provides a hyperspectral image incremental classification method based on an interactive graph distillation network to overcome the problem of rapid degradation of old class knowledge recognition performance caused by catastrophic forgetting of old class knowledge.
[0004] In order to achieve the above object, the technical solution of the present invention is:
[0005] A hyperspectral image incremental classification method based on an interactive graph distillation network includes the following steps:
[0006] S1. Data preparation stage: Select hyperspectral scene images and randomly extract some samples from them as training sets; dynamically divide the categories in the training sets into old categories and new categories according to preset incremental learning, and the corresponding samples are called old class samples and new class samples respectively; randomly extract some samples from the old class samples and the new class samples as support sets and query sets respectively;
[0007] S2, initial model training stage: constructing a feature extraction model and a dual-image interaction strategy graph neural network model, training the feature extraction model and the dual-image interaction strategy graph neural network model through the old class samples, saving the model parameters after the training is completed to obtain an old class training model group, wherein the old class training model group includes the feature extraction model and the dual-image interaction strategy graph neural network model that are optimized by the old class sample training;
[0008] S3, incremental learning stage:
[0009] S31, inputting the new class sample into the untrained feature extraction model and dual-graph interaction strategy graph neural network model of step S2 for training to obtain a first similarity matrix between the support set and the query set and a first new class sample prototype, and saving the model parameters after training to obtain a new class training model group; the new class training model group includes the feature extraction model and dual-graph interaction strategy graph neural network model optimized by the new class sample training;
[0010] S32, inputting the new class sample into the old class training model group saved in step S2 for training to obtain a second similarity matrix between the support set and the query set and a second new class sample prototype;
[0011] The first similarity matrix and the second similarity matrix include a node graph similarity matrix and a distribution graph similarity matrix; the first new class sample prototype and the second new class sample prototype include a node graph new class sample prototype and a distribution graph new class sample prototype;
[0012] S4, interactive graph distillation stage: Based on the first similarity matrix and the second similarity matrix, the interactive graph distillation method is applied to perform knowledge distillation; and an adaptive temperature scaling mechanism is used to dynamically control the temperature parameter in the distillation loss function to generate the interactive graph distillation loss;
[0013] S5, prototype-assisted memory transfer stage: Based on the first new class sample prototype and the second new class sample prototype, the multi-level prototype constraint strategy of the prototype-assisted memory transfer method is applied to perform knowledge transfer and generate a new class sample prototype contrast loss;
[0014] S6, multi-loss collaborative optimization stage: combining the defined cross entropy loss, the interaction graph distillation loss, and the new class sample prototype comparison loss to construct a multi-loss collaborative optimization function; using the multi-loss collaborative optimization function to train the new class training model group to obtain an optimal model group, the optimal model group including the feature extraction model and the dual-graph interaction strategy graph neural network model under optimal parameters;
[0015] S7, model testing stage: selecting a test data set from the hyperspectral scene image, and randomly selecting a support set and a query set from the test data set, and obtaining the classification results of each class in the test data set and the classification results of the entire test data set through the optimal model group.
[0016] Furthermore, the support set and query set selection method adopts a meta-learning strategy, where the support set is a set containing labeled samples and the query set is a set without labeled samples. The steps for obtaining the support set and query set are as follows:
[0017] S11. Establish the support set of the training set. The corresponding expression is:
[0018] S={(x1,y1),(x2,y2),...,(x N×K ,y N×K )} (1)
[0019] Where x N×K Represents a sample of a hyperspectral image, y N×K Represents the corresponding category label, N represents the number of categories, and K represents the number of samples in each category;
[0020] S12. Create a query set for the training set. The corresponding expression is:
[0021] Q={(x N×K+1 ,y N×K+1 ),...,(x N×K+M ,y N×K+M )} (2)
[0022] Where, is the number of query set samples.
[0023] Furthermore, the construction method of the dual-graph interaction strategy graph neural network model is:
[0024] S21. Establishing the node graph of the dual-graph interaction strategy graph neural network model and distribution map
[0025] Among them, the node graph The expression is:
[0026]
[0027] Distribution Map The expression is:
[0028]
[0029] Where, Represents the encoding network, which contains two Conv2d-BN-ReLU blocks; The model function representing the encoding network; It is a node graph Node characteristics; It is a distribution map Node features; l is the number of layers of the dual-graph interaction strategy graph neural network;
[0030] S22, the node graph Aggregate to distribution graph The corresponding expression is:
[0031]
[0032] Where, It is an aggregation network consisting of a fully connected layer and a nonlinear activation function ReLU; represents the model function of the aggregate network; M is the number of support sets; || is the connection operation; δ(·) is the Kronecker delta function, which means that when y i =y j , output 1, otherwise output 0;
[0033] S23, the distribution map Aggregate to node graph The corresponding expression is:
[0034]
[0035] Where, It is an aggregation network consisting of two Conv2d-BN-ReLU blocks; The model function representing the aggregate network; T is the total number of nodes in the graph.
[0036] Furthermore, based on the first similarity matrix and the second similarity matrix, an interactive graph distillation method is applied to perform knowledge distillation. An adaptive temperature scaling mechanism is used to dynamically control the temperature parameter in the distillation loss function to generate an interactive graph distillation loss. The implementation steps are as follows:
[0037] S41. Define the adaptive temperature scaling mechanism. The corresponding expression is:
[0038]
[0039] Where, σ(·) refers to the SoftMax operation; T G,i represents the distillation temperature; N represents the total number of samples in the support set and query set;
[0040] S42, according to the distillation temperature T G,i , the interaction graph distillation process is performed on the first similarity matrix and the second similarity matrix, and its expression is:
[0041]
[0042] Where G1 is the node graph, G2 is the distribution graph, and graph G = {G1, G2}; is the knowledge distillation loss of graph G; L G is the weighted sum of the distillation loss of each layer in Figure G under temperature regulation; L KL is the total distillation loss of the node graph and the distribution graph; the first similarity matrix As the teacher similarity matrix, the second similarity matrix As the student similarity matrix; max(·) takes the maximum value of each node distribution; σ log(·) is the log_SoftMax operation; θ defaults to 3; Represents the proportion of the corresponding loss in the node graph G1 and the distribution graph G2.
[0043] Furthermore, in the prototype-assisted memory transfer stage, based on the first new class sample prototype and the second new class sample prototype, the multi-level prototype constraint strategy of the prototype-assisted memory transfer method is applied to perform knowledge transfer and generate the new class sample prototype contrast loss in the following specific steps:
[0044] S51, step S32 obtains the first new class sample prototype and the second new class sample prototype, and the corresponding expressions are:
[0045]
[0046] Where, It is the prototype of the first new type of sample; It is the prototype of the second new type of sample;
[0047] S52. Define a multi-level prototype constraint strategy for knowledge transfer, which is expressed as:
[0048]
[0049] Where, L align is the prototype contrast loss of new class samples; Represents the proportion of loss corresponding to distribution graph G2.
[0050] Furthermore, a multi-loss collaborative optimization function is constructed by combining the defined cross entropy loss, the interaction graph distillation loss, and the new class sample prototype comparison loss; the new class training model group is trained using the multi-loss collaborative optimization function to obtain an optimal model group, which includes a feature extraction model and a dual-graph interaction strategy graph neural network model under optimal parameters; the specific steps are:
[0051] S61. Define the cross entropy loss function as:
[0052]
[0053] Where, L CE is the cross entropy loss function; y j is the label of the jth sample in the support set; y i is the true label of the query set; Represents the edge features of the lth layer in the node graph G1; Represents the edge features of the lth layer in the distribution graph G2;
[0054] Define the cross entropy loss as:
[0055]
[0056] In the formula, λ1 and λ2 represent the proportion of the corresponding loss;
[0057] S62. Combined with cross entropy loss The interaction graph distillation loss L KL Compared with the prototype loss of the new class sample L align , construct a multi-loss collaborative optimization function, whose expression is:
[0058]
[0059] Where, is the total number of layers of the dual-graph interaction strategy graph neural network.
[0060] Beneficial effects of the present invention:
[0061] 1. The present invention uses an incremental classification method for hyperspectral images based on an interactive graph distillation network to perform incremental classification. By selecting a support set and a query set, it avoids the use of a large number of labeled samples and reduces the labeling cost, especially for hyperspectral scenes with scarce labels.
[0062] 2. The present invention uses interactive graph distillation and prototype memory auxiliary modules to achieve knowledge solidification, effectively realize cross-stage knowledge transfer, alleviate the catastrophic forgetting of old knowledge, and has important value for achieving continuous learning capabilities for incremental classification of hyperspectral images. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0064] Figure 1 This is a flow chart of the incremental classification method for hyperspectral images of the present invention;
[0065] Figure 2a This is a network structure diagram of the feature extraction model of the present invention;
[0066] Figure 2b This is a structural diagram of the dual-graph interaction strategy graph neural network model of the present invention;
[0067] Figure 3a This is a pseudo-color image of the test data set in an embodiment of the present invention;
[0068] Figure 3b This is a diagram of the classification results of the test data set in an embodiment of the present invention. DETAILED DESCRIPTION
[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0070] This embodiment provides a hyperspectral image incremental classification method based on an interactive graph distillation network. Figure 1 As shown, the steps of the method include:
[0071] S1. Data preparation stage: Select hyperspectral scene images and randomly extract some samples from them as training sets; dynamically divide the categories in the training sets into old categories and new categories according to preset incremental learning, and the corresponding samples are called old class samples and new class samples respectively; randomly extract some samples from the old class samples and the new class samples as support sets and query sets respectively;
[0072] S2, initial model training stage: constructing a feature extraction model and a dual-image interaction strategy graph neural network model, training the feature extraction model and the dual-image interaction strategy graph neural network model through the old class samples, saving the model parameters after the training is completed to obtain an old class training model group, wherein the old class training model group includes the feature extraction model and the dual-image interaction strategy graph neural network model that are optimized by the old class sample training;
[0073] S3, incremental learning stage:
[0074] S31, inputting the new class sample into the untrained feature extraction model and dual-graph interaction strategy graph neural network model of step S2 for training to obtain a first similarity matrix between the support set and the query set and a first new class sample prototype, and saving the model parameters after training to obtain a new class training model group; the new class training model group includes the feature extraction model and dual-graph interaction strategy graph neural network model optimized by the new class sample training;
[0075] S32, inputting the new class sample into the old class training model group saved in step S2 for training to obtain a second similarity matrix between the support set and the query set and a second new class sample prototype;
[0076] The first similarity matrix and the second similarity matrix include a node graph similarity matrix and a distribution graph similarity matrix; the first new class sample prototype and the second new class sample prototype include a node graph new class sample prototype and a distribution graph new class sample prototype;
[0077] S4, interactive graph distillation stage: Based on the first similarity matrix and the second similarity matrix, the interactive graph distillation method is applied to perform knowledge distillation; and an adaptive temperature scaling mechanism is used to dynamically control the temperature parameter in the distillation loss function to generate the interactive graph distillation loss;
[0078] S5, prototype-assisted memory transfer stage: Based on the first new class sample prototype and the second new class sample prototype, the multi-level prototype constraint strategy of the prototype-assisted memory transfer method is applied to perform knowledge transfer and generate a new class sample prototype contrast loss;
[0079] S6, multi-loss collaborative optimization stage: combining the defined cross entropy loss, the interaction graph distillation loss, and the new class sample prototype comparison loss to construct a multi-loss collaborative optimization function; using the multi-loss collaborative optimization function to train the new class training model group to obtain an optimal model group, the optimal model group including the feature extraction model and the dual-graph interaction strategy graph neural network model under optimal parameters;
[0080] S7, model testing stage: selecting a test data set from the hyperspectral scene image, and randomly selecting a support set and a query set from the test data set, and obtaining the classification results of each class in the test data set and the classification results of the entire test data set through the optimal model group.
[0081] Specifically, the hyperspectral scene images in this embodiment are derived from the Botswana dataset acquired by sensors on the EO-1 satellite. After processing, the Botswana dataset has 145 bands. To reduce computational complexity and preserve the maximum spatial-spectral characteristics, the dataset's spectral channels are reduced to 20 main bands. The Botswana dataset contains a total of 14 defined ground feature categories, as shown in Table 1:
[0082] Table 1. Number of samples of target object categories in the Botswana dataset:
[0083]
[0084] The support set and query set selection method adopts a meta-learning strategy, and the specific implementation steps are as follows:
[0085] S11, selecting 10% of the target object category samples in the data set from the hyperspectral scene image to form a training set; dividing the training set into old class samples and new class samples, randomly selecting 5 categories for each training, randomly selecting 5 samples from each category to form a support set, and randomly selecting 5 samples from the remaining samples to form a query set;
[0086] S12. Establish the support set of the training set. The corresponding expression is:
[0087] S={(x1,y1),(x2,y2),...,(x N×K ,y N×K )} (1)
[0088] Where x N×K Represents a sample of a hyperspectral image, y N×K Represents the corresponding category label, N represents the number of categories, and K represents the number of samples in each category;
[0089] S13. Establish a query set for the training set. The corresponding expression is:
[0090] Q={(x N×K+1 ,y N×K+1 ),...,(x N×K+M ,y N×K+M )} (2)
[0091] Where, is the number of query set samples.
[0092] Specifically, in this embodiment, Figure 2a As shown in the figure, the network structure used by the feature extraction model achieves differentiated feature learning in the spectral and spatial dimensions through a bimodal convolutional path: in the spectral dimension, a multi-level progressive 3D convolutional structure is used to construct a hierarchical receptive field, and multi-scale abstraction of spectral context is achieved through cascaded convolutional layers. A spectrally guided channel attention module (SE3D) is embedded in each stage to dynamically adjust the weights of each spectral channel; in the spatial dimension, after mapping the three-dimensional spectral feature tensor to the two-dimensional spatial domain, a 2D convolutional layer with channel attention (SE2D) is used for spatial downsampling, focusing on enhancing discriminative local spatial patterns. Table 2 shows the specific network structure;
[0093] Table 2 Feature extraction model structure
[0094]
[0095] like Figure 2b As shown, the network structure adopted by the dual-graph interaction strategy graph neural network model is through the node graph Distribution Map Perform three iterations to achieve multi-level feature optimization of the support set and query set. Edge features Direct impact distribution map Node characteristics Distribution Map Edge features Directly affects the node graph Node features in Node Graph and distribution map The edge features are determined by their respective node features, and the general process can be summarized as follows: The construction method of the dual-graph interaction strategy graph neural network model is as follows:
[0096] S21. Establishing the node graph of the dual-graph interaction strategy graph neural network model and distribution map
[0097] Among them, the node graph The expression is:
[0098]
[0099] Distribution Map The expression is:
[0100]
[0101] Where, Represents the encoding network, which contains two Conv2d-BN-ReLU blocks; The model function representing the encoding network; It is a node graph Node characteristics; It is a distribution map Node features; l is the number of layers of the dual-graph interaction strategy graph neural network;
[0102] S22, the node graph Aggregate to distribution graph The corresponding expression is:
[0103]
[0104] Where, It is an aggregation network consisting of a fully connected layer and a nonlinear activation function ReLU; represents the model function of the aggregate network; M is the number of support sets; || is the connection operation; δ(·) is the Kronecker delta function, which means that when y i =y j , output 1, otherwise output 0;
[0105] S23, the distribution map Aggregate to node graph The corresponding expression is:
[0106]
[0107] Where, It is an aggregation network consisting of two Conv2d-BN-ReLU blocks; The model function representing the aggregate network; T is the total number of nodes in the graph.
[0108] Specifically, based on the first similarity matrix and the second similarity matrix, an interactive graph distillation method is applied to perform knowledge distillation; and an adaptive temperature scaling mechanism is used to dynamically control the temperature parameter in the distillation loss function to generate an interactive graph distillation loss. The implementation steps are as follows:
[0109] S41. Define the adaptive temperature scaling mechanism. The corresponding expression is:
[0110]
[0111] Where, σ(·) refers to the SoftMax operation; T G,i represents the distillation temperature; N represents the total number of samples in the support set and query set;
[0112] S42, according to the distillation temperature T G,i , the interaction graph distillation process is performed on the first similarity matrix and the second similarity matrix, and its expression is:
[0113]
[0114] Where G1 is the node graph, G2 is the distribution graph, and graph G = {G1, G2}; is the knowledge distillation loss of graph G; L G is the weighted sum of the distillation loss of each layer in Figure G under temperature regulation; L KL is the total distillation loss of the node graph and the distribution graph; the first similarity matrix As the teacher similarity matrix, the second similarity matrix As the student similarity matrix; max(·) takes the maximum value of each node distribution; σ log (·) is the log_SoftMax operation; θ defaults to 3; Represents the proportion of the corresponding loss in the node graph G1 and the distribution graph G2.
[0115] Specifically, in the prototype-assisted memory transfer stage, based on the first new class sample prototype and the second new class sample prototype, the multi-level prototype constraint strategy of the prototype-assisted memory transfer method is applied to perform knowledge transfer and generate the new class sample prototype contrast loss in the following specific steps:
[0116] S51, step S32 obtains the first new class sample prototype and the second new class sample prototype, and the corresponding expressions are:
[0117]
[0118] Where, It is the prototype of the first new type of sample; It is the prototype of the second new type of sample;
[0119] S52. Define a multi-level prototype constraint strategy for knowledge transfer, which is expressed as:
[0120]
[0121] Where, L align is the prototype contrast loss of new class samples; Represents the proportion of loss corresponding to distribution graph G2.
[0122] Specifically, a multi-loss collaborative optimization function is constructed by combining the defined cross entropy loss, the interaction graph distillation loss, and the new class sample prototype comparison loss; the multi-loss collaborative optimization function is used to train the new class training model group to obtain an optimal model group, which includes a feature extraction model and a dual-graph interaction strategy graph neural network model under optimal parameters; the specific steps are:
[0123] S61. Define the cross entropy loss function as:
[0124]
[0125] Where, L CE is the cross entropy loss function; y j is the label of the jth sample in the support set; y i is the true label of the query set; Represents the edge features of the lth layer in the node graph G1; Represents the edge features of the lth layer in the distribution graph G2;
[0126] Define the cross entropy loss as:
[0127]
[0128] In the formula, λ1 and λ2 represent the proportion of the corresponding loss;
[0129] S62. Combined with cross entropy loss The interaction graph distillation loss L KL Compared with the prototype loss of the new class sample L align , construct a multi-loss collaborative optimization function, whose expression is:
[0130]
[0131] Where, is the total number of layers of the dual-graph interaction strategy graph neural network.
[0132] Specifically, in this embodiment, an experiment was conducted on the Botswana dataset using the hyperspectral image incremental classification method based on the interactive graph distillation network of the present invention. The experimental results are shown in Table 3:
[0133] Table 3 Botswana classification accuracy (%)
[0134] Classification accuracy Category 1 96.22±2.59 Category 2 98.81±1.29 Category 3 99.20±0.80 Category 4 99.44±1.01 Category 5 87.36±5.17 Category 6 81.79±11.42 Category 7 97.60±3.92 Category 8 87.88±13.47 Category 9 98.79±1.30 Category 10 92.58±6.79 Category 11 94.03±4.41 Category 12 95.80±3.80 Category 13 98.21±1.10 Category 14 94.95±4.79 OA 94.27±0.63 AA 94.48±2.31 Kappa 93.8±0.68
[0135] Here, OA (Overall Accuracy) represents the overall classification accuracy, AA (Average accuracy) represents the average classification accuracy, and Kappa represents the Kappa coefficient. The Kappa coefficient refers to a multivariate discrete method for evaluating the classification accuracy and error matrix of remote sensing images. It also takes into account various missed and misclassified pixels outside the diagonal, and can penalize the bias of the model, thereby more comprehensively evaluating the classification effect. Figure 3a and Figure 3b Pseudo-color images and classification results of the test dataset are shown.
[0136] In order to more objectively evaluate the role of each step in the model of the hyperspectral image incremental classification method based on the interactive graph distillation network in the present invention, the existing ablation experiment is added for illustration. On the basis of the feature extraction model and the dual-graph exchange strategy graph neural network model, a single module or a combination of different modules is added to compare the experimental results. The specific experimental results are shown in Table 4:
[0137] Table 4 Classification accuracy of different modules (%)
[0138]
[0139] The following conclusions can be drawn from the above experiments:
[0140] (1) The experimental results in Table 3 show that the proposed hyperspectral image incremental classification method based on interactive graph distillation network has good classification effect, which proves that this method has excellent performance in reducing forgetting in incremental classification.
[0141] (2) The ablation experiment data in Table 4 show that the classification results of adding the prototype auxiliary memory module are significantly better than those of using only the ordinary basic network, which proves that the prototype auxiliary memory module plays an important role in solidifying old knowledge, and is more conducive to improving the classification effect of the model and alleviating catastrophic forgetting.
[0142] (3) The ablation experiment data in Table 4 show that the interactive graph distillation module has a great impact on the experimental results. The classification results of using this module alone are significantly better than those of the ordinary prototype network. After adding the interactive graph distillation module on the basis of the prototype-assisted memory transfer module, the classification effect of the Botswana dataset has been improved to a certain extent. This proves that the addition of the interactive graph distillation module reduces the forgetting of old classes and plays an important role in improving the classification accuracy.
[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A hyperspectral image incremental classification method based on an interactive graph distillation network, characterized by: The following steps are involved: S1, data preparation stage: select hyperspectral scene images and randomly extract some samples from them as training sets; according to the preset incremental learning, the categories in the training sets are dynamically divided into old categories and new categories, and the corresponding samples are called old class samples and new class samples respectively; Randomly extracting some samples from the old class samples and the new class samples as a support set and a query set respectively; S2, initial model training stage: constructing a feature extraction model and a dual-image interaction strategy graph neural network model, training the feature extraction model and the dual-image interaction strategy graph neural network model through the old class samples, saving the model parameters after the training is completed to obtain an old class training model group, wherein the old class training model group includes the feature extraction model and the dual-image interaction strategy graph neural network model that are optimized by the old class sample training; S3, incremental learning stage: S31, inputting the new class sample into the untrained feature extraction model and dual-graph interaction strategy graph neural network model of step S2 for training to obtain a first similarity matrix between the support set and the query set and a first new class sample prototype, and saving the model parameters after training to obtain a new class training model group; the new class training model group includes the feature extraction model and dual-graph interaction strategy graph neural network model optimized by the new class sample training; S32, inputting the new class sample into the old class training model group saved in step S2 for training to obtain a second similarity matrix between the support set and the query set and a second new class sample prototype; The first similarity matrix and the second similarity matrix include a node graph similarity matrix and a distribution graph similarity matrix; the first new class sample prototype and the second new class sample prototype include a node graph new class sample prototype and a distribution graph new class sample prototype; S4, interactive graph distillation stage: Based on the first similarity matrix and the second similarity matrix, the interactive graph distillation method is applied to perform knowledge distillation; and an adaptive temperature scaling mechanism is used to dynamically control the temperature parameter in the distillation loss function to generate the interactive graph distillation loss; S5, prototype-assisted memory transfer stage: Based on the first new class sample prototype and the second new class sample prototype, the multi-level prototype constraint strategy of the prototype-assisted memory transfer method is applied to perform knowledge transfer and generate a new class sample prototype contrast loss; S6, multi-loss collaborative optimization stage: combining the defined cross entropy loss, the interaction graph distillation loss, and the new class sample prototype comparison loss to construct a multi-loss collaborative optimization function; using the multi-loss collaborative optimization function to train the new class training model group to obtain an optimal model group, the optimal model group including the feature extraction model and the dual-graph interaction strategy graph neural network model under optimal parameters; S7, model testing stage: selecting a test data set from the hyperspectral scene image, and randomly selecting a support set and a query set from the test data set, and obtaining the classification results of each class in the test data set and the classification results of the entire test data set through the optimal model group.
2. The method for incremental hyperspectral image classification based on an interactive graph distillation network according to claim 1, characterized in that: The support set and query set selection method adopts a meta-learning strategy. The support set is a set containing labeled samples, and the query set is a set without labeled samples. The steps for obtaining the support set and query set are as follows: S11. Establish the support set of the training set. The corresponding expression is: S={(x1,y1),(x2,y2),...,(x N×K ,y N×K )} (1) Where x N×K Represents a sample of a hyperspectral image, y N×K Represents the corresponding category label, N represents the number of categories, and K represents the number of samples in each category; S12. Create a query set for the training set. The corresponding expression is: Where, is the number of query set samples.
3. The method for incremental hyperspectral image classification based on an interactive graph distillation network according to claim 1, characterized in that: The construction method of the dual-graph interaction strategy graph neural network model is: S21. Establishing the node graph of the dual-graph interaction strategy graph neural network model and distribution map Among them, the node graph The expression is: Distribution Map The expression is: Where, Represents the encoding network, which contains two Conv2d-BN-ReLU blocks; The model function representing the encoding network; It is a node graph Node characteristics; It is a distribution map Node features; l is the number of layers of the dual-graph interaction strategy graph neural network; S22, the node graph Aggregate to distribution graph The corresponding expression is: Where, It is an aggregation network consisting of a fully connected layer and a nonlinear activation function ReLU; represents the model function of the aggregate network; M is the number of support sets; || is the connection operation; δ(·) is the Kronecker delta function, which means that when y i =y j , output 1, otherwise output 0; S23, the distribution map Aggregate to node graph The corresponding expression is: Where, It is an aggregation network consisting of two Conv2d-BN-ReLU blocks; The model function representing the aggregate network; T is the total number of nodes in the graph.
4. The method for incremental hyperspectral image classification based on an interactive graph distillation network according to claim 1, characterized in that: Based on the first similarity matrix and the second similarity matrix, an interactive graph distillation method is applied to perform knowledge distillation; and an adaptive temperature scaling mechanism is used to dynamically control the temperature parameter in the distillation loss function to generate an interactive graph distillation loss; The implementation steps are as follows: S41. Define the adaptive temperature scaling mechanism. The corresponding expression is: Where, σ() refers to the SoftMax operation; T G,i represents the distillation temperature; N represents the total number of samples in the support set and query set; S42, according to the distillation temperature T G,i , the interaction graph distillation process is performed on the first similarity matrix and the second similarity matrix, and its expression is: Where G1 is the node graph, G2 is the distribution graph, and graph G = {G1, G2}; is the knowledge distillation loss of graph G; L G is the weighted sum of the distillation loss of each layer in Figure G under temperature regulation; L KL is the total distillation loss of the node graph and the distribution graph; the first similarity matrix As the teacher similarity matrix, the second similarity matrix As the student similarity matrix; max(·) takes the maximum value of each node distribution; σ log () is the log_SoftMax operation; θ defaults to 3; Represents the proportion of the corresponding loss in the node graph G1 and the distribution graph G2.
5. The method for incremental hyperspectral image classification based on an interactive graph distillation network according to claim 1, characterized in that: Based on the first new class sample prototype and the second new class sample prototype, and applying the multi-level prototype constraint strategy of the prototype-assisted memory transfer method to perform knowledge transfer, the specific steps of generating the new class sample prototype contrast loss are as follows: S51, step S32 obtains the first new class sample prototype and the second new class sample prototype, and the corresponding expressions are: Where, It is the prototype of the first new type of sample; It is the prototype of the second new type of sample; S52. Define a multi-level prototype constraint strategy for knowledge transfer, which is expressed as: Where, L align is the prototype contrast loss of new class samples; Represents the proportion of loss corresponding to distribution graph G2.
6. The method for incremental hyperspectral image classification based on an interactive graph distillation network according to claim 1, characterized in that: Combining the defined cross entropy loss, the interaction graph distillation loss, and the new class sample prototype contrast loss, a multi-loss collaborative optimization function is constructed; the new class training model group is trained using the multi-loss collaborative optimization function to obtain an optimal model group, which includes a feature extraction model and a dual-graph interaction strategy graph neural network model under optimal parameters; the specific steps are: S61. Define the cross entropy loss function as: Where, L CE is the cross entropy loss function; y j is the label of the jth sample in the support set; y i is the true label of the query set; Represents the edge features of the lth layer in the node graph G1; Represents the edge features of the lth layer in the distribution graph G2; Define the cross entropy loss as: In the formula, λ1 and λ2 represent the proportion of the corresponding loss; S62. Combined with cross entropy loss The interaction graph distillation loss L KL = Contrast loss L with the prototype of the new class sample align , construct a multi-loss collaborative optimization function, whose expression is: Where, is the total number of layers of the dual-graph interaction strategy graph neural network.