Remote sensing image ground object semi-supervised classification method, storage medium and equipment
By constructing a semi-supervised semantic segmentation model and utilizing the Swing Transformer and generative adversarial networks, the problems of insufficient remote sensing data and annotation difficulties were solved, achieving efficient remote sensing image land cover classification, improving classification accuracy and reducing costs.
Patent Information
- Application Number
- CN202411397523.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2025-02-18
AI Technical Summary
Due to insufficient remote sensing data and difficulties in pixel-level annotation of remote sensing data, existing semantic segmentation methods suffer from problems such as ambiguous category semantics and boundaries in remote sensing images, and are also costly.
A semi-supervised learning method is adopted to construct a semantic segmentation model for generator and discriminator. Using Swing Transformer and DeformableConv Block, combined with generative adversarial network, the model is trained with labeled and unlabeled data through cross-entropy, feature matching and self-training loss function, and high-quality pseudo-labels are generated for supervision.
It effectively solves the problems of insufficient remote sensing data and difficulty in annotation, improves the accuracy and efficiency of remote sensing image land cover classification, and reduces costs.
Smart Images
Figure CN119478490A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of remote sensing image classification, and in particular to a remote sensing image ground object semi-supervised classification method, storage medium and equipment. Background Art
[0002] Remote sensing image classification is a common method for remote sensing image interpretation. The basic processing unit of early classification methods is pixels, and the spectral information of pixels is used as the main identification feature. However, pixel-based methods only contain spectral features and cannot express features such as texture, space, and structure. With the improvement of spatial resolution of remote sensing images and the increase of spatial information of objects, object-level methods have emerged. The basic idea of this type of method is to obtain objects through segmentation technology, then calculate the underlying features of objects such as color, texture, and shape, and finally apply machine learning algorithms to achieve object-oriented classification. The rapid development of deep learning technology in recent years has provided a new solution for the intelligent interpretation of remote sensing images in the era of big data. Compared with traditional methods, deep learning methods are completely data-driven and learn hierarchical feature representations from labeled sample data, avoiding the pertinence and singleness of manually designed features. Semantic segmentation, also known as object classification in the field of remote sensing, is to assign a certain object category to each pixel in the image based on the texture, color, and scene information of the image itself.
[0003] Most existing semantic segmentation methods are based on fully supervised learning, which learns and understands the visual features and deep semantics of image data from a large amount of annotated data. Semantic segmentation methods based on fully convolutional networks or Vision Transformer (ViT: a deep learning model that applies the Transformer architecture to the field of computer vision) require a large amount of training data. When the samples are insufficient, the model parameters will overfit the training data, resulting in reduced model performance and decreased generalization ability. In addition, it is not easy to obtain finely annotated data for remote sensing images, which requires a lot of manpower and material resources. Complex ground scenes and sub-meter or even centimeter-level spatial resolutions pose great challenges to pixel-level annotation, and the cost is extremely high.
[0004] How to extract image features and understand visual information from limited labeled data is an important aspect of current research. Some scholars have tried to use unsupervised and weakly supervised methods, but it is difficult to distinguish the concept of category, and there are also great difficulties in fine boundary extraction. Some researchers have proposed a new framework for adversarial training - generative adversarial networks, which provides new ideas for the application of semi-supervised learning in semantic segmentation. Semi-supervised learning can use a small amount of labeled data and a large amount of unlabeled data, reducing the demand for labeled data and avoiding waste of resources, while solving the problem of model inaccuracy in unsupervised and weakly supervised learning. Researchers at home and abroad have proposed an adversarial-based semi-supervised semantic segmentation method, linking generative adversarial networks with semantic segmentation tasks.
[0005] Semi-supervised learning has made some progress in the field of natural image segmentation, but there are still few studies and applications in remote sensing images. There are problems of category semantics and boundary ambiguity in semi-supervised learning. Remote sensing data is expensive, the amount of remote sensing data is insufficient, and pixel-level annotation of remote sensing data is difficult. Summary of the invention
[0006] The purpose of the present invention is to solve the problem of insufficient remote sensing data and difficulty in pixel-level annotation of remote sensing data, and to propose a remote sensing image ground object semi-supervised classification method, comprising the following steps:
[0007] S1. Acquire remote sensing data of ground objects and divide the remote sensing data of ground objects into labeled data and unlabeled data;
[0008] S2. construct a semi-supervised semantic segmentation model, wherein the model includes a generator and a discriminator;
[0009] S3. When using labeled data, use a fully supervised approach to train the model. When using unlabeled data to train the generator, use the segmentation mask obtained by the generator with a probability greater than the threshold as a pseudo-label, and use the discriminator's discrimination result as a supervisory signal to implement a semi-supervised model mechanism.
[0010] S4. Input the remote sensing data of the objects to be classified into the trained semi-supervised semantic segmentation model to obtain the classification results of the remote sensing image objects.
[0011] Furthermore, the generator includes: a main branch consisting of a Swin Transformer, an auxiliary branch consisting of a Stem and three cascaded DeformableConv Blocks connected in series, and a dynamic decoder;
[0012] The input of the generator is input into the main branch and the auxiliary branch respectively. The results of the two branches are added and input into the dynamic decoder to obtain the output of the generator.
[0013] The discriminator consists of four convolutional layers, one average pooling layer, and two fully connected layers;
[0014] After the two inputs of the discriminator are concatenated, they pass through four convolutional layers, an average pooling layer, two fully connected layers, and finally are output through the Sigmoid activation function.
[0015] Furthermore, when training the model using labeled data, the cross entropy loss function of the generator is expressed as:
[0016] L ce =-∑ h，w，c [y l (h, w, c)logS(x l )(h,w,c)],
[0017] Among them, L ce represents the cross entropy loss function of the generator, represents the result of one-hot encoding of the real annotation of remote sensing data, S(x l )(h, w, c) represents the prediction result of the generator, x l represents the remote sensing data input to the generator, and (h, w, c) represents the classification probability map output by the generator.
[0018] Furthermore, when training the model using unlabeled data, the feature matching loss function of the generator is expressed as:
[0019]
[0020] Among them, L fm represents the feature matching loss function of the generator; E represents the expectation, (x l ,y l )~D l Indicates taking all labeled remote sensing data, D l represents a labeled remote sensing dataset, D k (·) represents the intermediate representation output by the k-th layer discriminator, represents the concatenation of vectors in the channel direction, x l represents the remote sensing data input to the generator, yl represents the label of the remote sensing data, (x U )~D U Indicates taking all unlabeled remote sensing data, S(x U ) indicates that the input is x U The output of the generator, x U Represents unlabeled data.
[0021] Furthermore, when training the model using unlabeled data, the self-training loss function of the generator is expressed as:
[0022]
[0023] Among them, L st represents the self-training loss function of the generator, y * represents a pseudo label, Indicates that the input is x U The output of the generator is D(S(x U )) indicates that the input is S(x U ) is the output of the discriminator, and γ represents the set threshold.
[0024] Furthermore, the total loss of the generator is expressed as:
[0025] L s =L ce +λ fm L fm +λ st L st ,
[0026] Among them, L s represents the total loss of the generator, λ fm and λ st is an adjustable hyperparameter.
[0027] Furthermore, when training the model using labeled data or unlabeled data, the cross entropy loss function of the discriminator is expressed as:
[0028] L D =-∑ h，w L bce (D(y l ), 1)+L bce (D(S(x l )),0),
[0029] L bce (D(y l ), 1) = -logD(y l ),
[0030] L bce (D(S(x l )), 0) = -log(1-D(S(x l ))),
[0031] Among them, L D represents the cross entropy loss function of the discriminator, L bce () represents the binary cross entropy loss function, D(y l ) indicates that the input of the discriminator is y l The output, y l Represents the label of remote sensing data, D(S(x l )) indicates that the input of the discriminator is S(xl ) output, S(x l ) indicates that the input of the generator is x l The output when x l Represents the remote sensing data that is input to the generator.
[0032] The present invention also proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned remote sensing image ground object semi-supervised classification method is implemented.
[0033] The present invention also proposes an electronic device, comprising a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned remote sensing image semi-supervised classification method.
[0034] The beneficial effects brought by the technical solution provided by the present invention are:
[0035] The present invention proposes to use a dual-branch encoder embedded in Swin Transformer as a generative network to generate a prediction mask based on the input image. Its global modeling and detail extraction capabilities can effectively make up for the category semantics and boundary ambiguity problems in semi-supervised learning; the discriminant network includes four convolution blocks, an average pooling layer and two fully connected layers to distinguish between the prediction mask and the real annotation; in the training of unlabeled data, feature matching loss is introduced to reduce the distribution difference between the generated data and the real data; high-quality prediction results are selected as pseudo labels, and the self-training loss is calculated for supervised training. It solves the problems of insufficient remote sensing data, difficulty in pixel-level annotation of remote sensing data, and high cost of remote sensing data. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 is a flow chart of a method for semi-supervised classification of ground objects in remote sensing images according to an embodiment of the present invention;
[0037] Figure 2 This is the structure diagram of Swin Transformer;
[0038] Figure 3 is a structural diagram of a semi-supervised semantic segmentation model constructed in an embodiment of the present invention;
[0039] Figure 4 It is a block diagram of an electronic device in an exemplary embodiment of the present invention. DETAILED DESCRIPTION
[0040] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0041] The flowchart of the method for semi-supervised classification of ground objects in remote sensing images according to an embodiment of the present invention is as follows: Figure 1 , specifically including the following steps:
[0042] S1. Obtaining remote sensing data of ground objects, and dividing the remote sensing data of ground objects into labeled data and unlabeled data. The original data set of the remote sensing data of ground objects of the present invention comes from the ISPRS Vaihingen and Potsdam data sets. The Vaihingen data set contains 33 images with an image resolution range of 1996×1997 to 3816×2550 pixels, and the Potsdam data set contains 38 images with an image resolution of 6000×6000 pixels. The images and labels are cropped to a size of 512×512 pixels, and the data after cropping is shown in Table 1. Online data enhancement methods include horizontal flipping, vertical flipping, rotating 90°, random scaling and cropping, and mirror flipping. Obtain the ground object remote sensing data that can be used in the embodiment of the present invention, and divide the ground object remote sensing data into a training set and a test set, wherein the Vaihingen data set contains 16 of the 33 images for training and 17 for testing; the Potsdam data set contains 24 of the 38 images for training and 14 for testing.
[0043] Table 1
[0044] Dataset Size (pixels) Number of training sets Number of test sets Spatial resolution (cm) Vaihingen 512×512 344 398 9 Potsdam 512×512 3456 2016 5
[0045] S2. Build a semi-supervised semantic segmentation model, which includes a generator and a discriminator. The generator includes: a main branch consisting of a Swin Transformer, an auxiliary branch consisting of a Stem Block and three cascaded DeformableConv Blocks (deformable convolutions), and a dynamic decoder. The structure diagram of Swin Transformer is referenced in Figure 2 . The Swin Transformer structure comes from a paper: "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows". Swin Transformer is an innovative visual model. It successfully solves the problems faced by traditional Transformers when switching from NLP to CV tasks through innovative designs such as hierarchical multi-scale architecture, shifted window strategy, and local self-attention mechanism, achieving efficient image processing capabilities while reducing the consumption of computing resources. Stem Block is used for fast downsampling. This structure is designed to ensure strong feature expression capabilities while reducing a large number of parameters.
[0046] The input of the generator is input into the main branch and the auxiliary branch respectively. The results obtained from the two branches are added and input into the dynamic decoder to obtain the output of the generator.
[0047] Specifically, the generator inputs image data of H×W×3 dimensions and outputs a classification probability map (mask) of H×W×C. The one-dimensional vector at the (i, j) position in the probability map represents the probability value of each category, where C represents the number of categories to be identified in the data set.
[0048] The discriminator consists of four convolutional layers, one average pooling layer, and two fully connected layers.
[0049] After the two inputs of the discriminator are concatenated, they pass through four convolutional layers, an average pooling layer, two fully connected layers, and finally are output through the Sigmoid activation function.
[0050] The structure diagram of the semi-supervised semantic segmentation model constructed in the embodiment of the present invention is referenced Figure 3 .
[0051] Each convolutional layer of the discriminator includes a two-dimensional convolution with a kernel size of 4, a stride of 2, and a padding of 1, batch normalization, and a Leaky ReLU activation function. After each convolutional layer, the width and height of the feature map are reduced to half of the original, and the number of channels is doubled. The number of output channels of the four convolutional layers is {32, 64, 128, 256} respectively. Average pooling can reduce the shift of internal variables and reduce the sensitivity of the network model to parameter changes. Leaky ReLU is used as the activation function of this discriminator, so that when the input data is less than 0, a very small gradient can be maintained to avoid the gradient of 0. After four convolutions, global average pooling is used to refine and simplify the information, and then the final classification is performed through two fully connected layers.
[0052] S3. When using labeled data, use a fully supervised approach to train the model. When using unlabeled data to train the generator, in order to ensure a balance between the generator and the discriminator, use the segmentation mask obtained by the generator with a probability greater than the threshold as a pseudo-label, and use the discriminator’s judgment result as a supervisory signal to implement a semi-supervised model mechanism.
[0053] When using labeled data, the input of the discriminator is the concatenation of the remote sensing data of the object and the labeled data input by the generator in the channel direction; when using unlabeled data, the input of the discriminator is the concatenation of the remote sensing data of the object and the generated mask input by the generator in the channel direction.
[0054] (1) Loss function of the generator
[0055] When using labeled data to train the model, the input image data is Given an image, the true annotation is one-hot encoded as The prediction result of the generator is The generator is constructed by minimizing and The cross entropy loss function of the generator is expressed as follows:
[0056] L ce =-∑ h,w,c [y 1 (h,w,c)logS(x l )(h,w,c)],
[0057] Among them, L ce represents the cross entropy loss function of the generator, represents the result of one-hot encoding of the real annotation of remote sensing data, S(x l )(h,w,c) represents the prediction result of the generator, x l represents the remote sensing data input to the generator, and (h,w,c) represents the probability distribution map of the generator output.
[0058] When training a model using unlabeled data, in order to make the generated mask as consistent as possible with the actual distribution of the label, a feature matching loss is designed based on the difference in feature distribution. The purpose of the feature matching loss is to minimize the generator prediction result. With the true label The average difference in feature statistics between them makes the data generated by the generator more consistent with the real distribution. The feature matching loss function of the generator is expressed as:
[0059]
[0060] Among them, L fm represents the feature matching loss function of the generator; E represents the expectation, (x l ,y l )~D l Indicates taking all labeled remote sensing data, D l represents a labeled remote sensing dataset, D k (·) represents the intermediate representation output by the k-th layer discriminator, represents the concatenation of vectors in the channel direction, x l Represents the remote sensing data input to the generator, y l Represents the label of remote sensing data, (x U )~D U Indicates taking all unlabeled remote sensing data, S(x U ) indicates that the input is x U The output of the generator, x U Represents unlabeled data.
[0061] In the training process of the generative adversarial network, it is necessary to maintain a balance between the generator and the discriminator. In order to promote the balance between the two, self-training loss is introduced. When training the model with unlabeled data, the generator's better prediction results are selected as a pseudo label for supervised training. The cross entropy is calculated based on the prediction results and pseudo labels as the self-training loss. The self-training loss function of the generator is expressed as:
[0062]
[0063] Among them, L st represents the self-training loss function of the generator, y * represents a pseudo label, Indicates that the input is x U The output of the generator is D(S(x U )) indicates that the input is S(x U ) is the output of the discriminator, and γ represents the set threshold. Represents the output value of the discriminator and the prediction result, ranging from 0 to 1. The output value is used as a confidence indicator. If the confidence is greater than the set threshold γ, it is considered a high-quality prediction result; * For high-quality prediction results Generated pseudo labels.
[0064] Combining the above three loss functions, the total loss of the generator is expressed as:
[0065] L S =L ce +λ fm L fm +λ st L st ,
[0066] Among them, L S represents the total loss of the generator, λ fm and λ st is an adjustable hyperparameter.
[0067] The generated network is optimized using the SGD optimizer, with an initial learning rate of 2.5e. -4 , momentum is 0.9, weight decay is 5e -4 .
[0068] (2) Loss function of the discriminator
[0069] When using labeled data, the image is input into the generator to obtain a classification probability map, and the labeled data is One-hot encoded to obtain a feature map, which is input into the discriminator to obtain the discrimination result. When using unlabeled data images, they are input into the segmentation network to obtain an initial prediction or segmentation mask, and then this prediction is input into the discrimination network to calculate the discrimination result, and the discrimination result is used as a supervision signal to achieve semi-supervised training.
[0070] Training the discriminant network is training a binary classifier to determine whether the input data comes from generated samples or real data. When the discriminator outputs 1, it means that the input data is a real label; when the discriminator outputs 0, it means that the input data is a generated sample. When using labeled data or unlabeled data to train the model, the cross entropy loss function of the discriminator is expressed as:
[0071] L D =-∑ h,w L bce (D(y 1 ),1)+L bce (D(S(x 1 )),0),
[0072] L bce (D(y 1 ),1)=-logD(y 1 ),
[0073] L bce (D(S(x 1 )),0)=-log(1-D(S(x 1 ))),
[0074] Among them, L D represents the cross entropy loss function of the discriminator, L bce () represents the binary cross entropy loss function, D(y l ) indicates that the input of the discriminator is y l The output, y l Represents the label of remote sensing data, D(S(x l )) indicates that the input of the discriminator is S(x l ) output, S(x l ) indicates that the input of the generator is x l The output when x 1 Represents the remote sensing data that is input to the generator.
[0075] The discriminant network is optimized using the Adam optimizer with an initial learning rate of 1e -4 The exponential decay rates of the first-order moment and second-order moment estimates are 0.9 and 0.999, respectively. The batch size is set to 6 and the number of iterations is 20,000.
[0076] S4. Input the remote sensing data of the objects to be classified into the trained semi-supervised semantic segmentation model to obtain the classification results of the remote sensing image objects.
[0077] Both the Vaihingen dataset and the Potsdam dataset contain six types of objects, namely, impervious surfaces, buildings, low vegetation, trees, cars, and miscellaneous objects / background. The evaluation indicators are the intersection over union (IoU) of each category and the average intersection over union (mIoU) (Note: the calculation of mIoU does not include the background category).
[0078] Through the full supervision and semi-supervised experimental results under different data ratios in Table 2 (%), the following conclusions can be drawn:
[0079] (1) As the amount of labeled data increases, the overall evaluation index (average intersection-over-union ratio) of the test set in both fully supervised and semi-supervised cases increases, indicating that an increase in data volume will improve network performance.
[0080] (2) Under the same data ratio, the semi-supervised method achieves better experimental results than the fully supervised method. Under the condition of 1 / 2 labeled data, the semi-supervised method achieves 81.70% mIoU on the Vaihingen dataset and 85.09% mIoU on the Potsdam dataset, which is close to the extraction results of the full-supervised data.
[0081] (3) When the amount of labeled data is small, the segmentation effect of the semi-supervised method is much better than that of the fully supervised method. For example, when there is 1 / 8 labeled data, the semi-supervised method improves by 3.03% on the Vaihingen dataset and 3.92% on the Potsdam dataset compared to the fully supervised method.
[0082] Table 2
[0083]
[0084]
[0085] In summary, the present invention aims at the time-consuming, labor-intensive and costly pixel-level annotation of remote sensing data, and introduces generative adversarial networks into semantic segmentation. The idea of generator and discriminator confronting each other is adopted to train a small amount of labeled data and a large amount of unlabeled data. It is able to perform effective global modeling and local adjustment extraction, and tap into potential semantic information, with better results than common semi-supervised algorithms. Experiments on the Vaihinge and Potsdam datasets have shown that under the same annotation ratio, the semi-supervised method achieves better experimental results than the fully supervised method, and semi-supervised training with 1 / 2 of the labeled data can achieve a land object classification effect equivalent to that of full supervision; the less labeled data used, the more obvious the advantages of the semi-supervised algorithm.
[0086] In an exemplary embodiment, a computer-readable storage medium is included, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned remote sensing image ground object semi-supervised classification method is implemented.
[0087] See also Figure 4 In an exemplary embodiment, an electronic device is also included, including at least one processor, at least one memory, and at least one communication bus.
[0088] Wherein, a computer program is stored in the memory, and the computer program includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-mentioned remote sensing image ground object semi-supervised classification method.
[0089] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for semi-supervised classification of objects in remote sensing images, characterized in that: The following steps are involved: S1. Acquire remote sensing data of ground objects and divide the remote sensing data of ground objects into labeled data and unlabeled data; S2. construct a semi-supervised semantic segmentation model, wherein the model includes a generator and a discriminator; S3. When using labeled data, use a fully supervised approach to train the model. When using unlabeled data to train the generator, use the segmentation mask obtained by the generator with a probability greater than the threshold as a pseudo-label, and use the discriminator's discrimination result as a supervisory signal to implement a semi-supervised model mechanism. S4. Input the remote sensing data of the objects to be classified into the trained semi-supervised semantic segmentation model to obtain the classification results of the remote sensing image objects.
2. A remote sensing image object semi-supervised classification method according to claim 1, characterized in that: The generator includes: The main branch consists of a Swin Transformer, an auxiliary branch consisting of a Stem and three cascaded DeformableConv Blocks, and a dynamic decoder; The input of the generator is input into the main branch and the auxiliary branch respectively. The results of the two branches are added and input into the dynamic decoder to obtain the output of the generator. The discriminator consists of four convolutional layers, one average pooling layer, and two fully connected layers; After the two inputs of the discriminator are concatenated, they pass through four convolutional layers, an average pooling layer, two fully connected layers, and finally are output through the Sigmoid activation function.
3. The method for semi-supervised classification of objects in remote sensing images according to claim 1, characterized in that: When training the model using labeled data, the cross entropy loss function of the generator is expressed as: L ce =-∑ h,w,c [y 1 (h,w,c)logS(x 1 )(h,w,c)], Among them, L ce represents the cross entropy loss function of the generator, represents the result of one-hot encoding of the real annotation of remote sensing data, S(x 1 )(h,w,c) represents the prediction result of the generator, x l represents the remote sensing data input to the generator, and (h,w,c) represents the classification probability map output by the generator.
4. A method for semi-supervised classification of objects in remote sensing images according to claim 3, characterized in that: When training the model using unlabeled data, the feature matching loss function of the generator is expressed as: Among them, L fm represents the feature matching loss function of the generator; E represents the expectation, (x l ,y l )~D 1 Indicates taking all labeled remote sensing data, D 1 represents a labeled remote sensing dataset, D k (·) represents the intermediate representation output by the k-th layer discriminator, represents the concatenation of vectors in the channel direction, x 1 Represents the remote sensing data input to the generator, y l Represents the label of remote sensing data, (x U )~D U Indicates taking all unlabeled remote sensing data, S(x U ) indicates that the input is x U The output of the generator, x U Represents unlabeled data.
5. A method for semi-supervised classification of objects in remote sensing images according to claim 4, characterized in that: When training the model using unlabeled data, the self-training loss function of the generator is expressed as: Among them, L st represents the self-training loss function of the generator, y * represents a pseudo label, Indicates that the input is x U The output of the generator is D(S(x U )) indicates that the input is S(x U ) is the output of the discriminator, and γ represents the set threshold.
6. A method for semi-supervised classification of objects in remote sensing images according to claim 5, characterized in that: During model training, the total loss of the generator is expressed as: L S =L ce +λ fm L fm +λ st L st , Among them, L S represents the total loss of the generator, λ fm and λ st is an adjustable hyperparameter.
7. The method for semi-supervised classification of objects in remote sensing images according to claim 1, characterized in that: When training a model using labeled or unlabeled data, the cross entropy loss function of the discriminator is expressed as: L D =-∑ h,w L bce (D(y l ),1)+L bce (D(S(x l )),0), L bce (D(y l ),1)=-logD(y l ), L bce (D(S(x 1 )),0)=-log(1-D(S(x 1 ))), Among them, L D represents the cross entropy loss function of the discriminator, L bce () represents the binary cross entropy loss function, D(y l ) indicates that the input of the discriminator is y l The output, y l Represents the label of remote sensing data, D(S(x l )) indicates that the input of the discriminator is S(x l ) output, S(x l ) indicates that the input of the generator is x l The output when x l Represents the remote sensing data that is input to the generator.
8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.
9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1 to 7.