A remote sensing image active learning method based on class-level graph embedding representation

Abstract

The application relates to a remote sensing image active learning method based on a class-level graph embedding representation. The method comprises the following steps: randomly taking a labeled set and an unlabeled set from labeled data of a hyperspectral remote sensing image, taking the remaining samples as a test set, setting the iteration number and sample budget of active learning, training a classification model using the labeled set, obtaining spectral feature representation of the samples through network parameters, dividing the labeled set into K classes according to real labels, constructing graph nodes and an adjacency matrix for each class, training a class-level graph convolution network model, obtaining inter-class minimum uncertainty of the unlabeled samples, selecting B uncertain samples as a query set, giving the query set real labels and adding the query set to the labeled set, updating parameters of the classification model and the class-level graph convolution network model using the new labeled set, and when the iteration number is I, training the classification model using the updated labeled set and classifying the test set to obtain a classification result. Therefore, the accuracy of a deep learning model classification is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hyperspectral image processing, in particular to a remote sensing image active learning method based on class-level graph embedding representation. BACKGROUND

[0002] Hyperspectral remote sensing is one of the most commonly used technologies in remote sensing at present, and plays a crucial role in military exploration, ground surface monitoring, disaster prevention and other fields. Hyperspectral images can use the spectral curve characteristic information unique to various ground objects to identify many similar ground object categories that cannot be identified by spectral images, so that hyperspectral image classification has become a research hotspot today. Hyperspectral image classification has the following difficulties: 1. The spectral dimension of hyperspectral images is high and there is serious data redundancy, which increases the computational complexity of the classification task; 2. The data distribution is complex, and there is a nonlinear relationship and complex pattern between the spectral information of each pixel in the hyperspectral image and the corresponding ground object, so that traditional methods cannot accurately classify; 3. Insufficient samples, experts need to conduct field exploration and investigation, and sample labeling consumes manpower and material resources.

[0003] In recent years, with the development of deep learning, deep learning models have been gradually applied to remote sensing image classification and have shown good classification results. Some scholars have applied convolutional neural networks and generative adversarial networks to remote sensing image classification. However, the training process of these models often relies on a large amount of labeled data.

[0004] In view of the above difficulties, especially the high cost of expert labeling and the dependence of deep learning models on a large number of labeled samples, it is necessary to provide an active learning hyperspectral image classification method that can select high-quality and representative samples to improve the classification accuracy of deep learning models. SUMMARY

[0005] Therefore, it is necessary to provide a remote sensing image active learning method based on class-level graph embedding representation, which can select high-quality and representative samples.

[0006] A remote sensing image active learning method based on class-level graph embedding representation, the method comprising:

[0007] Step S1, from the labeled data of the hyperspectral remote sensing image, randomly take 10 samples of each class as a labeled set D l , randomly take 4000 samples as an unlabeled set D u , the remaining samples as a test set D t , set the number of iterations I of active learning and the sample budget B;

[0008] Step S2, use the labeled set D l to train a classification model f c, and the spectral feature of the sample is obtained through the network parameter;

[0009] Step S3, the labeled set D l is divided into K classes, and the graph node v and the adjacency matrix A are constructed for each class;

[0010] Step S4, the class-level graph convolution network model f g is trained, and the minimum inter-class uncertainty of the unlabeled sample is obtained;

[0011] Step S5, B uncertain samples are selected as the query set Q, the real label is given to the query set Q, and the query set Q is added to the labeled set D l ;

[0012] Step S6, steps S2-S5 are repeated, and the new labeled set is used to update the parameters of the classification model f c and the class-level graph convolution network model f g ;

[0013] Step S7, when the iteration number is I times, the updated labeled set D l is used to train the classification model f c , and the test set D t is classified to obtain the classification result.

[0014] In one of the embodiments, the method for obtaining the spectral feature is:

[0015] After the classification model f l is trained using the labeled set D c , the spectral feature f c (x) of the sample is generated using the trained parameters θ c :

[0016] f c (x) = x·θ c , θ c ∈R d×m

[0017] wherein f c (x) represents the spectral feature of the sample, f c (x) ∈R m×1 , m represents the feature dimension, x represents the original sample, x ∈R d×1 , d represents the number of bands of the hyperspectral remote sensing image, and R represents the domain.

[0018] In one of the embodiments, the step S3 includes:

[0019] Step 3.1: divide the labeled set D l into K classes according to the real label;

[0020] Step 3.2: Initialize graph node v as a single-class labeled empty spectral feature f c (c j ) and unlabeled spatial spectral features f c (D u The combination of ) in which, v i Let c represent the i-th node in the graph. j This indicates that the sample comes from a single-class label set of class j, and n j Let n be the number of labeled samples belonging to class j. u This represents the number of unlabeled samples.

[0021] Step 3.3: Calculate the L2 norm of graph node v;

[0022] Step 3.4: Calculate the vector product S of the graph nodes. Let S ab This represents the element in the vector product S of graph nodes located at row a and column b, where {a,b}∈n. j +n u S ab Represented as:

[0023]

[0024] Among them, v a Let v be the a-th node in the graph. Let v be the transpose of the a-th node in the graph. b Let b be the b-th node in the graph.

[0025] Step 3.5: Calculate the adjacency matrix A, the specific formula is as follows:

[0026] A = D -1 (SI)+I

[0027] In the formula, D is the degree matrix of S, and I is the identity matrix;

[0028] Step 3.6: Repeat steps 3.2-3.5 to construct graph node and adjacency matrices by category. The graph node and adjacency matrices for all categories are represented as follows: K is the maximum number of categories in the hyperspectral remote sensing image, and k is the category number.

[0029] In one embodiment, step S4 includes:

[0030] Step 4.1: Based on the graph nodes and adjacency matrix of all categories Construct K class-level graph convolutional network models f g In which, each class-level graph convolutional network model f gComprise: input layer, ReLU layer, hidden layer, Dropout layer, output layer and sigmoid layer;

[0031] Step 4.2: set the number of nodes of the input layer to 32, and the number of nodes of the output layer to 1;

[0032] Step 4.3: set the single class label sample and the unmarked sample label as 1 and 0 respectively;

[0033] Step 4.4: according to the artificial label, define the class level graph convolutional network model f g Loss function Loss function :

[0034]

[0035] Wherein, L (θ1, θ2) is the total loss of the network under the current parameters θ1, θ2, v is the graph node, A is the adjacency matrix, is the node score output by the network, λ represents the weighting parameter, the weighting parameter is used to adjust the proportion of marked samples and unmarked samples, θ1 is the parameter of the input layer, and θ2 is the parameter of the output layer;

[0036] Step 4.5: train K class level graph convolutional network models respectively;

[0037] Step 4.6: after training, the whole class level graph convolutional network model outputs K credible scores for each sample, since the interval range of the score is (0, 1), the uncertainty score can be obtained by subtracting the credible score from 1, and the minimum value is selected as the uncertainty of the sample, and the minimum uncertainty score u i Of the i-th unmarked sample to the class is:

[0038] u i = min (1-u i1 , 1-u i2 ,..., 1-u iK )

[0039] Wherein, u i1 is the credible score of the i-th sample to the first class marked sample, u i2 is the credible score of the i-th sample to the second class marked sample, and u iK is the credible score of the i-th sample to the Kth class marked sample.

[0040] In one embodiment, the step S5 comprises:

[0041] Step 5.1: obtaining the uncertainty score of all samples in the unmarked set by using the class level graph convolutional network model f g ​​

[0042] Step 5.2: Determine the threshold δ, and select the samples greater than δ in U as the set n δ The number of samples in the set.

[0043] Step 5.3: According to the relationship between the number of samples n δ and the sample budget B, select different strategies to determine the query set Q, as follows:

[0044]

[0045] When n δ ≤B, take C as the query set and add it to the labeled set; when n δ >B, introduce k-means clustering to divide C into B classes, and select the cluster centers of each class as the query set and add it to the labeled set.

[0046] The above active learning method for remote sensing images based on class-level graph embedding representation, by randomly taking 10 samples from each class in the labeled data of hyperspectral remote sensing images as the labeled set D l , randomly taking 4000 samples as the unlabeled set D u , and the remaining samples as the test set D t , setting the number of iterations I of active learning and the sample budget B, using the labeled set D l to train the classification model f c , obtaining the spectral feature representation of the samples through the network parameters, dividing the labeled set D l into K classes according to the real labels, constructing graph nodes v and adjacency matrix A for each class, training the class-level graph convolution network model f g , obtaining the inter-class minimum uncertainty of the unlabeled samples, selecting B uncertain samples as the query set Q, giving the query set Q real labels and adding it to the labeled set D l , updating the classification model f c and the parameters of the class-level graph convolution network model f g , when the number of iterations is I times, using the updated labeled set D l to train the classification model f c and classify the test set D t to obtain the classification result. By constructing a class-level graph convolution network to measure the inter-class uncertainty of unlabeled samples, high-quality and representative samples are selected for labeling to reduce the demand for labeled samples by deep learning models, thereby improving the accuracy of deep learning model classification. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flowchart of the active learning method for remote sensing images based on class-level graph embedding representation in one embodiment.

[0048] Figure 2 Structure diagram of a class-level graph convolutional network model in an embodiment;

[0049] Figure 3 Pseudo-color image of an Indian Pines hyperspectral image in an embodiment;

[0050] Figure 4 Real object type of an Indian Pines hyperspectral image in an embodiment;

[0051] Figure 5 Classification results of the model under the samples selected by the Random of the comparative example of the present application;

[0052] Figure 6 Classification results of the model under the samples selected by the BT of the comparative example of the present application;

[0053] Figure 7 Classification results of the model under the samples selected by the Entropy of the comparative example of the present application;

[0054] Figure 8 Classification results of the model under the samples selected by the Cluster of the comparative example of the present application;

[0055] Figure 9 Classification results of the model under the samples selected by the GCN of the comparative example of the present application;

[0056] Figure 10 Classification results of the model under the samples selected by the CGCN of the comparative example of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0058] In an embodiment, as shown in Figure 1 , a remote sensing image active learning method based on class-level graph embedding representation is provided. The method is applied to a terminal as an example for illustration, which includes the following steps:

[0059] Step S1, from the labeled data of the hyperspectral remote sensing image, 10 samples of each class are randomly taken out as a labeled set D l , 4000 samples are randomly taken out as an unlabeled set D u , the remaining samples are taken out as a test set D t , and the number of iterations I and the sample budget B of active learning are set.

[0060] Step S2, using the label set D l training the classification model f c , obtaining the hyperspectral feature representation of the sample through the network parameters.

[0061] Step S3, dividing the label set D l into K classes according to the true labels, and constructing the graph nodes v and the adjacency matrix A for each class.

[0062] Step S4, training the class-level graph convolutional network model f g to obtain the inter-class minimum uncertainty of the unlabeled samples.

[0063] The class-level graph convolutional network model is shown in Figure 2 , and includes an input layer, a ReLU layer, a hidden layer, a Dropout layer, an output layer, and a sigmoid layer. The number of neurons of the input layer is equal to the feature dimension of the input sample, the number of neurons of the hidden layer is set to 32, and the number of neurons of the output layer is 1. The input of the class-level graph convolutional network model is the graph node and the adjacency matrix (the graph node is composed of single-class labeled samples and unlabeled samples in the set; the adjacency matrix is obtained by calculating the similarity between the graph nodes), and the output of the class-level graph convolutional network model is the credibility of the graph node (sample).

[0064] Step S5, selecting B uncertain samples as the query set Q, assigning true labels to the query set Q, and adding the query set Q to the label set D l .

[0065] Step S6, repeating steps S2-S5, and updating the parameters of the classification model f c and the class-level graph convolutional network model f g using the new label set.

[0066] Step S7, when the iteration number is I times, training the classification model f l using the updated label set D c and classifying the test set D t to obtain the classification result.

[0067] In one embodiment, the method for obtaining the hyperspectral feature representation is as follows:

[0068] After training the classification model f l using the label set D c , the hyperspectral feature f c (x) of the sample is generated using the trained parameters θ c , and is represented as:

[0069] f c (x) = x·θ c , θ c ∈R d×m

[0070] wherein f c (x) represents the spectral feature of the sample, f c (x)∈R m×1 , m represents the feature dimension, x represents the original sample, x∈R d×1 , d represents the number of bands of the hyperspectral remote sensing image, and R represents the domain.

[0071] In one embodiment, the step S3 comprises:

[0072] Step 3.1: dividing the labeled set D l into K classes according to the true labels;

[0073] Step 3.2: initializing the graph node v as the combination of the single-class labeled spectral feature f c (c j ) and the unlabeled spectral feature f c (D u ), wherein, v i represents the i-th node in the graph node, c j represents the single-class labeled set that the sample comes from the j-th class, n j is the number of labeled samples belonging to the j-th class, and n u is the number of unlabeled samples;

[0074] Step 3.3: calculating the two-norm of the graph node v;

[0075] Step 3.4: calculating the vector product S of the graph node, Let S ab represent the element located at the a-th row and b-th column in the vector product S of the graph node, {a,b}∈n j +n u , and S ab is represented as:

[0076]

[0077] wherein v a is the a-th node in the graph node v, is the transpose of the a-th node in the graph node v, and v b is the b-th node in the graph node;

[0078] Step 3.5: calculating the adjacency matrix A, and the specific formula is:

[0079] A=D -1 (S-I)+I

[0080] In the formula, D is the degree matrix of S, and I is the unit matrix.

[0081] Step 3.6: Repeat steps 3.2-3.5 to construct the graph nodes and adjacency matrix of all classes, and the graph nodes and adjacency matrix of all classes are represented as K is the maximum number of classes of the hyperspectral remote sensing image, and k is the class number.

[0082] In one embodiment, the step S4 comprises:

[0083] Step 4.1: According to the graph nodes and adjacency matrix of all classes K class-level graph convolution network models f are constructed g , wherein each class-level graph convolution network model f g comprises an input layer, a ReLU layer, a hidden layer, a Dropout layer, an output layer, and a sigmoid layer.

[0084] Step 4.2: Set the number of nodes of the input layer to 32 and the number of nodes of the output layer to 1;

[0085] Step 4.3: Set the labels of the single-class labeled samples and unlabeled samples to 1 and 0, respectively;

[0086] Step 4.4: According to the artificial labels, define the loss function of the class-level graph convolution network model f g The loss function of the class-level graph convolution network model f is:

[0087]

[0088] wherein, is the total loss of the network under the current parameters θ1, θ2, v is the graph node, A is the adjacency matrix, is the node score output by the network, λ represents a weighting parameter, the weighting parameter is used to adjust the proportion of the labeled samples and the unlabeled samples, θ1 is the parameter of the input layer, and θ2 is the parameter of the output layer.

[0089] Step 4.5: Train K class-level graph convolution network models respectively;

[0090] Step 4.6: After the training is completed, the entire class-level graph convolution network model outputs K credible scores for each sample. Since the interval range of the scores is (0, 1), the uncertainty score can be obtained by subtracting the credible score from 1. The minimum value of the uncertainty scores is selected as the uncertainty of the sample. The minimum uncertainty score u i of the i-th unlabeled sample in the class is represented as:

[0091] u i = min(1-u i1 , 1-u i2 ,..., 1-u iK ​)

[0092] where u i1 is the confidence score of the i-th sample for the first class labeled samples, u i2 is the confidence score of the i-th sample for the second class labeled samples, u iK is the confidence score of the i-th sample for the K-th class labeled samples.

[0093] In one of the embodiments, the step S5 comprises:

[0094] Step 5.1: using the class-level graph convolutional network model f g to obtain the uncertainty scores of all samples in the unlabeled set

[0095] Step 5.2: determine the threshold δ, and select the samples greater than δ in U as the set n δ is the number of samples in the set.

[0096] Step 5.3: according to the relationship between the number of samples n δ and the sample budget B, select different strategies to determine the query set Q, as follows:

[0097]

[0098] When n δ ≤ B, take C as the query set and add it to the labeled set; when n δ > B, introduce k-means clustering to divide C into B classes, and select the cluster centers of each class as the query set and add it to the labeled set.

[0099] The above remote sensing image active learning method based on class-level graph embedding representation, by taking 10 samples from each class in the labeled data of hyperspectral remote sensing image as the labeled set D l , randomly taking 4000 samples as the unlabeled set D u , and the remaining samples as the test set D t , setting the number of iterations I of active learning and the sample budget B, using the labeled set D l to train the classification model f c , obtaining the spectral feature representation of the sample through the network parameters, dividing the labeled set D l into K classes according to the real label, constructing the graph node v and the adjacency matrix A for each class, training the class-level graph convolutional network model f g , obtaining the minimum uncertainty between classes of the unlabeled samples, selecting B uncertain samples as the query set Q, giving the query set Q the real label and adding it to the labeled set D l , updating the classification model f c and the class-level graph convolutional network model fg parameters, when the number of iterations is I times, the updated label set D l training a classification model f c and classifying the test set D t to obtain a classification result. By constructing a class-level graph convolutional network to measure the inter-class uncertainty of unlabeled samples, high-quality and representative samples are selected for labeling to reduce the demand of deep learning models for labeled samples, thereby improving the classification accuracy of deep learning models.

[0100] The hyperspectral remote sensing image dataset used in the embodiments of the present application is the Indian Pines hyperspectral dataset, which contains 200 effective spectral bands. As shown in Figure 3 , a pseudo-color image of the Indian Pines dataset is shown, with a spatial size of 145x145 pixels and a ground sampling distance of 20 meters. As shown in Figure 4 , the real ground object types are shown, including 16 ground object categories, of which 10249 pixel samples are available for experiments. The active learning comparison experiments respectively use random sampling (Random), boundary sampling (BT), maximum entropy sampling (Entropy), clustering sampling (Cluster), conventional graph embedding (GCN) and the sampling method of the present application (C GCN) to perform active learning and image classification on the labeled sample data of the embodiment Indian Pines dataset.

[0101] The active learning settings are as follows: 10 samples are selected from each of the 16 classes as the initial labeled set D l , 3000 samples are selected as the unlabeled set D u , and the remaining samples are used as the test set D t . The number of iterations I is set to 5, and the sample budget B is set to 30. Under these conditions, 5 repeated experiments are performed, and the classification accuracy is shown in Table 1, ± indicating the fluctuation range of the experimental results. The ResNet model is selected as the classification model, and the classification accuracy of the ResNet model under the Random sampling condition is used as the comparison benchmark.

[0102] Table 1 Comparison of image classification accuracy of Indian Pines dataset

[0103]

[0104]

[0105] From Table 1, it can be seen that Random, as a comparison benchmark of active learning method, achieves a classification accuracy of 92.31%, and the classification accuracies of the other four comparison methods are all higher than that of Random to varying degrees. BT sampling considers that the samples located at the class boundary have more abundant information, and determines the sample uncertainty by the direct distance difference between the maximum class and the second largest class, and BT achieves good classification results. Compared with the second, the present application achieves the best classification result. The present application mainly acquires the uncertainty of the sample by constructing a class-level graph embedding manner, effectively screens out samples with abundant information, and provides high-quality sample data guarantee for the deep learning classification model.

[0106] In order to visualize the classification results, Figures 5-9 respectively show the classification result graphs of Random, BT, Entropy, Cluster, GCN and CGCN, Figure 10 show the classification result graph of the present application. It can be seen directly that, with the help of the selected samples of the present application, the classification model can accurately identify the land cover category to which the sample belongs.

[0107] The above proves the feasibility of the present application in hyperspectral image classification. The present application greatly reduces the demand of deep learning model in the field of remote sensing image for labeled data, so that the deep learning model can also play an excellent performance under the condition of a small amount of labeled data, and effectively improves the classification accuracy of hyperspectral image.

[0108] It should be understood that, although Figure 1 the steps in the flowchart are shown in order according to the arrows, these steps are not necessarily executed in order according to the arrows. Unless otherwise specified in this article, the execution of these steps has no strict order limitation, and these steps can be executed in other order. Moreover, Figure 1 At least part of the steps in the above embodiment can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or sub-steps or stages of other steps.

[0109] The technical features of the above embodiments can be combined in any way. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0110] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A remote sensing image active learning method based on class-level graph embedding representation, characterized in that, The method includes: Step S1: Randomly select 10 samples from each class from the labeled data of the hyperspectral remote sensing image as the label set D. l Then, 4000 samples are randomly selected as the unlabeled set D. u The remaining samples are used as the test set D. t Set the number of iterations I for active learning and the sample budget B; Step S2, using the tag set D l Training classification model f c The spatial spectral features of the samples are obtained through network parameters; Step S3, based on the real labels, the tag set D l Divide the data into K categories, and construct the graph node v and the adjacency matrix A for each category; Step S4, train the class-level graph convolutional network model f g This yields the minimum inter-class uncertainty for unlabeled samples; Step S5: Select B uncertain samples as the query set Q, assign true labels to the query set Q and add them to the label set D. l ; Step S6: Repeat steps S2-S5 to update the classification model f using the new label set. c With class-level graph convolutional network model f g Parameters; Step S7: When the iteration count is I, use the updated tag set D. l Training classification model f c And for test set D t The classification process yields the classification results. Step S4 includes: Step 4.1: Based on the graph nodes and adjacency matrix of all categories Construct K class-level graph convolutional network models f g In which, each class-level graph convolutional network model f g It includes: input layer, ReLU layer, hidden layer, Dropout layer, output layer, and sigmoid layer; Step 4.2: Set the number of nodes in the input layer to 32 and the number of nodes in the output layer to 1; Step 4.3: Set the labels for single-class labeled samples and unlabeled samples to 1 and 0, respectively; Step 4.4: Define the class-level graph convolutional network model f based on manual labels. g loss function loss function for: in, Let v be the total loss of the network under the current parameters θ1 and θ2, v be the graph node, and A be the adjacency matrix. λ represents the node score output by the network, λ represents the tradeoff parameter, which is used to adjust the ratio of labeled samples to unlabeled samples, θ1 is the parameter of the input layer, and θ2 is the parameter of the output layer. Step 4.5: Train K class-level graph convolutional network models respectively; Step 4.6: After training, the entire class-level graph convolutional network model outputs K confidence scores for each sample. Since the score range is (0,1), subtracting the confidence score from 1 yields the uncertainty score. The minimum value among these is selected as the uncertainty of the sample. The minimum uncertainty score u of the i-th unlabeled sample is... i Represented as: u i = min(1-u) i1 ,1-u i2 ,...,1-u iK ) Among them, u i1 Let u be the confidence score of the i-th sample against the first-class labeled samples. i2 Let u be the confidence score of the i-th sample relative to the second-class labeled samples. iK Let be the confidence score of the i-th sample relative to the K-th labeled sample.

2. The method according to claim 1, characterized in that, The method for representing the spatial spectral features is as follows: In the tag set D l Training classification model f c Then, using the trained parameters θ c Spatial spectral features f of generated samples c (x) represents: f c (x)=x·θ c ,the c ∈R d×m Among them, f c (x) represents the spatial spectral characteristics of the sample, f c (x)∈R m×1 m represents the feature dimension, x represents the original sample, and x∈R d ×1 d represents the number of bands in the hyperspectral remote sensing image, and R represents the domain.

3. The method according to claim 2, characterized in that, Step S3 includes: Step 3.1: Assign the tag set D based on the actual labels. l Divided into K categories; Step 3.2: Initialize graph node v as a single-class labeled empty spectral feature f c (c j ) and unlabeled spatial spectral features f c (D u The combination of ) in which, v i c represents the i-th node in the graph. j This indicates that the sample comes from a single-class label set of class j, and n j Let n be the number of labeled samples belonging to class j. u This represents the number of unlabeled samples. Step 3.3: Calculate the L2 norm of graph node v; Step 3.4: Calculate the vector product S of the graph nodes. Let S ab This represents the element in the vector product S of graph nodes located at row a and column b, where {a,b}∈n. j +n u S ab Represented as: Among them, v a Let v be the a-th node in the graph. Let v be the transpose of the a-th node in the graph. b Let b be the b-th node in the graph. Step 3.5: Calculate the adjacency matrix A, the specific formula is as follows: A=D -1 (S-I)+I In the formula, D is the degree matrix of S, and I is the identity matrix; Step 3.6: Repeat steps 3.2-3.5 to construct graph node and adjacency matrices by category. The graph node and adjacency matrices for all categories are represented as follows: K is the maximum number of categories in the hyperspectral remote sensing image, and k is the category number.

4. The method according to claim 3, characterized in that, The tradeoff parameter λ in step 4.4 is expressed by the formula: Where, n j Let n be the number of labeled samples belonging to class j. u This represents the number of unlabeled samples.

5. The method according to claim 4, characterized in that, Step S5 includes: Step 5.1: Utilize the class-level graph convolutional network model f g Obtain the uncertainty scores of all samples in the unlabeled set. Step 5.2: Determine the threshold δ and select samples in U that are greater than δ as a set. n δ The number of samples in the set; Step 5.3: Based on the sample size n δ The relationship with the sample budget B, and the selection of different strategies to determine the query set Q, are as follows: When n δ When n ≤ B, C is used as the query set and added to the tag set; when n ≤ B, C is used as the query set and added to the tag set. δ When the value is greater than B, k-means clustering is introduced to divide C into classes B. The cluster centers of each class are selected as query sets and added to the label set.

Images