Self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification

Abstract

The present invention specifically discloses a self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification, the method comprising the following steps: S1, based on self-supervised learning and using remote sensing image data to train a preset deep neural network to obtain an embedded network model; S2, embedding the embedded network model into a meta-learning framework for training to obtain a small sample classification embedded network model; S3, iteratively optimizing the small sample classification embedded network model based on self-distillation training, thereby completing self-supervised self-distillation meta-learning for small sample remote sensing image classification and obtaining a trained SMSR model. The present invention applies the embedded network model with feature extraction capability obtained by self-supervised learning training to a meta-learning framework for small sample remote sensing image classification training, and then obtains a trained SMSR model through self-distillation simplification, thereby realizing self-supervised self-distillation meta-learning for small sample remote sensing image classification.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a self-supervised self-distillation meta-learning method for small-sample remote sensing image scene classification. Background Art

[0002] As more and more remote sensing satellites are launched, the acquisition of remote sensing images has become easier and easier, and the demand for faster and smarter remote sensing image understanding has also increased. With the improvement of remote sensing image resolution, the details of objects in remote sensing images have become richer and richer, and many new applications have emerged. Among them, remote sensing image scene classification is to assign a label to each remote sensing image. This label indicates the scene composed of the objects contained in the image, such as a commercial area, school or park. In the past, the method of manually designing features according to the characteristics of different objects to understand remote sensing images has been gradually replaced by machine (deep) learning methods because it is time-consuming, labor-intensive and inaccurate.

[0003] Deep learning technology has been increasingly widely used in the field of remote sensing image processing. By conducting supervised training on a large-scale labeled remote sensing image dataset, a deep neural network suitable for the dataset can be obtained. Remote sensing image scene classification is an important part of intelligent interpretation of remote sensing images. However, in certain specific scenarios, the remote sensing image classification data obtained is often very limited and lacks annotations, and cannot reflect the real data distribution. If it is used to directly train a deep learning network, the performance of the obtained network will be significantly reduced.

[0004] In view of this, how to solve the small sample problem in remote sensing image scene classification is a technical problem that needs to be solved urgently by people in this technical field. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings and deficiencies in the prior art and provide a self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification. An embedded network model with feature extraction capability is obtained through self-supervised learning training, and the embedded network model is applied to the meta-learning framework for small sample remote sensing image classification training. Self-distillation is then used for simplification to obtain a trained SMSR model. Finally, the trained SMSR model can be used to classify the remote sensing images acquired in real time.

[0006] In order to solve the above technical problems, the present invention provides a self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification, the method comprising the following steps:

[0007] S1. Based on self-supervised learning and using remote sensing image data, a preset deep neural network is trained to obtain an embedded network model that can extract feature information of remote sensing images;

[0008] S2, embedding the embedded network model obtained in step S1 into the framework of meta-learning for training, and enabling the embedded network model trained by meta-learning to classify new samples using existing prior knowledge, thereby obtaining a small sample classification embedded network model;

[0009] S3. Based on self-distillation training, the small sample classification embedding network model is iteratively optimized, and then the self-supervised self-distillation meta-learning for small sample remote sensing image classification is completed to obtain a trained SMSR model.

[0010] Preferably, the specific implementation of step S1 is as follows:

[0011] S11. Set contrastive learning as an auxiliary task of self-supervised learning, and select NCE loss as the measurement indicator of contrastive learning;

[0012] S12, using different data enhancement methods to perform data enhancement on the same remote sensing image, and calculating the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods;

[0013] S13. Using the stochastic gradient descent method, the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods is continuously reduced until the value of the NCE loss tends to be stable, thereby obtaining an embedded network model that can extract feature information of the remote sensing image.

[0014] Preferably, the different data enhancement methods in step S12 include scaling, random cropping, random center cropping, affine transformation, color jittering, random horizontal flipping and random grayscale transformation.

[0015] Preferably, the contrastive learning is performed by selecting feature maps generated by different convolutional layers in a preset deep neural network for comparison.

[0016] Preferably, the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods in step S12 is expressed by the formula:

[0017] L(x m ,x n )=L(f a (x m ),f5(x n ))+L(f a (x m ),f e (x n ))+L(f5(x m ),f5(x n ))+L(f a (x n ),f5(xm ))+L(f a (x n ),f e (x m ))+L(f e (x m ),f e (x n )) (1)

[0018] In formula (1), x m and x n They represent the processing results of the same remote sensing image after two different data enhancement methods, f a represents the global feature map extracted by the preset deep neural network, f5 represents the local feature map extracted by the fifth convolution layer in the preset deep neural network, and f e Represents the output feature map of the last layer of the network extracted by the last layer of convolution in the preset deep neural network, where:

[0019]

[0020]

[0021]

[0022]

[0023] N represents all remote sensing images, and d represents the square of the Euclidean distance.

[0024] Preferably, the specific implementation of step S2 is:

[0025] S21, embedding the embedded network model obtained in step S1 into the framework of meta-learning and using it as the embedded network model to be trained;

[0026] S22. Divide the samples in the data set consisting of all remote sensing images into several categories, randomly select K samples from each category as the support set, and the remaining samples as the query set. Input the samples in the support set into the embedded network model to be trained for measurement, and obtain the classification result w of the i-th sample in the support set. i , expressed as:

[0027]

[0028] In formula (2), f() represents the embedding network model to be trained, S represents the support set, and y i represents the i-th sample in the support set, z i Represents sample y iThe corresponding label of

[0029] S23, inputting the samples in the query set into the embedding network model to be trained to obtain the classification results of the samples in the query set, and then calculating the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, and then determining the types of the samples in the query set;

[0030] S24. Use the stochastic gradient descent method to continuously reduce the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, so that the embedding network model to be trained after meta-learning training can obtain the ability to classify new samples using existing prior knowledge, and then obtain a small sample classification embedding network model.

[0031] Preferably, the difference between the classification results of the samples in the query set and the classification results of the samples in the support set is expressed by the formula:

[0032] L meta =log∑ n exp(-D(f(q),w i )) (3)

[0033] In formula (3), q represents the sample in the query set, wi represents the classification result of the embedded network model to be trained for the i-th sample in the support set, Represents the difference between the classification results of samples in the query set and samples in the support set.

[0034] Preferably, the specific implementation method of step S3 is: first calculate the loss function in the self-distillation training process, that is, calculate the difference between the k-th generation predicted value and the k-th generation actual value and the difference between the k-th generation predicted value and the k-1-th generation predicted value in the self-distillation training; then use the induced network to reduce the two difference values, and continuously iterate until the sum of the two difference values ​​is reduced to a minimum value and stabilizes near the minimum value, thereby obtaining a trained SMSR model.

[0035] Preferably, the model parameterization when the loss function takes the minimum value in the kth iteration training of the self-distillation training is expressed by the formula:

[0036]

[0037] In formula (4), represents the parameterized model obtained from the kth iteration of self-distillation training, KL represents the Kullback-Leibler divergence, α represents the hyperparameter, Q represents the dataset for self-distillation training, and L ce is the cross entropy loss, represents a parameterized model, Representing a parameterized model Classification results for dataset Q.

[0038] Preferably, the method further includes: S4, inputting the remote sensing image acquired in real time into a trained SMSR model for classification, thereby achieving scene classification of the remote sensing image.

[0039] Compared with the prior art, the present invention uses self-supervised learning to train a preset deep neural network to obtain an embedded network model with feature extraction capability, that is, the ability to extract prior knowledge from a small amount of unlabeled data, and then applies the embedded network model to a meta-learning framework to perform small sample remote sensing image classification training to obtain a small sample classification embedded network model that uses known prior knowledge to classify new samples, and then the small sample classification embedded network model is streamlined and optimized through self-distillation to obtain a trained SMSR model, and finally the trained SMSR model can be used to classify the remote sensing images acquired in real time, thereby realizing scene classification of small sample remote sensing images, and the classification results of scene classification of remote sensing images based on the trained SMSR model are better than the classification results of traditional supervised learning remote sensing image scene classification methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of a self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification in the present invention.

[0041] Figure 2 is a partial schematic diagram of the MASATI remote sensing image dataset in an embodiment of the present invention,

[0042] Figure 3 It is a flow chart of outputting feature maps of remote sensing images based on self-supervised learning networks in the present invention. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0044] like Figure 1-Figure 3 As shown, a self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification is provided, and the method comprises the following steps:

[0045] S1. Based on self-supervised learning and using remote sensing image data, a preset deep neural network is trained to obtain an embedded network model that can extract feature information of remote sensing images;

[0046] S2, embedding the embedded network model obtained in step S1 into the framework of meta-learning for training, and enabling the embedded network model trained by meta-learning to classify new samples using existing prior knowledge, thereby obtaining a small sample classification embedded network model;

[0047] S3. Iteratively optimize the small sample classification embedding network model based on self-distillation training, and then complete the self-supervised self-distillation meta-learning for small sample remote sensing image classification to obtain the trained SMSR (Self-supervised Metalearning and Self-distillation Remote Sensing Scene Classification) model;

[0048] S4. Input the remote sensing images acquired in real time into the trained SMSR model for classification, thereby realizing the scene classification of the remote sensing images.

[0049] In this embodiment, first, a preset deep neural network is trained using self-supervised learning to obtain an embedded network model with feature extraction capabilities; then, the embedded network model is applied to a meta-learning framework for small sample remote sensing image classification training and streamlined and optimized using self-distillation to obtain a trained SMSR model; finally, the trained SMSR model is used to classify the remote sensing images acquired in real time to achieve small sample remote sensing image scene classification. The classification results of scene classification of remote sensing images based on the trained SMSR model are better than the classification results of the traditional supervised learning remote sensing image scene classification method.

[0050] The specific implementation of step S1 is as follows:

[0051] S11, setting contrastive learning as an auxiliary task of self-supervised learning, and selecting NCE (Noise Contrastive Estimation) loss as a measurement indicator of contrastive learning. NCE loss is used to measure the degree of difference between target objects and the size of mutual information. The contrastive learning selects feature maps generated by different convolutional layers in a preset deep neural network for comparison, thereby further expanding the difference between contrast objects and improving the complexity of contrastive learning.

[0052] S12, using different data enhancement methods to perform data enhancement on the same remote sensing image, and calculating the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods, wherein the different data enhancement methods include scaling, random cropping, random center cropping, affine transformation, color jittering, random horizontal flipping and random grayscale transformation;

[0053] S13. Using the stochastic gradient descent method, the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods is continuously reduced until the value of the NCE loss tends to be stable, thereby obtaining an embedded network model that can extract feature information of the remote sensing image.

[0054] Among them, the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods is expressed by the formula:

[0055] L(x m ,x n )=L(f a (x m ),f5(x n ))+L(f a (x m ),f e (x n ))+L(f5(x m ),f5(x n ))+L(f a (x n ),f5(x m ))+L(f a (x n ),f e (x m ))+L(f e (x m ),f e (x n )) (1)

[0056] In formula (1), x m and x n They represent the processing results of the same remote sensing image after two different data enhancement methods, f a represents the global feature map extracted by the preset deep neural network, f5 represents the local feature map extracted by the fifth convolution layer in the preset deep neural network, and f e Represents the output feature map of the last layer of the network extracted by the last layer of convolution in the preset deep neural network, where:

[0057]

[0058]

[0059]

[0060]

[0061] N represents all remote sensing images, and d represents the square of the Euclidean distance.

[0062] In this embodiment, by using contrastive learning as an auxiliary task of self-supervised learning, the smaller the difference between the two contrasting targets, the higher their mutual information. By continuously expanding the mutual information between the two remote sensing image samples, the deep neural network is prompted to learn the feature information in the remote sensing image samples. At the same time, in order to expand the difference between different data enhancement results and avoid the collapse of model training, different data enhancement methods are used to perform data enhancement on the same remote sensing image, and then the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods is calculated. Finally, by continuously reducing the NCE loss until its value tends to be stable, an embedded network model that can extract the feature information of the remote sensing image is obtained.

[0063] The specific implementation of step S2 is as follows:

[0064] S21, embedding the embedded network model obtained in step S1 into the framework of meta-learning and using it as the embedded network model to be trained;

[0065] S22. Divide the samples in the data set consisting of all remote sensing images into several categories, randomly select K samples from each category as the support set, and the remaining samples as the query set. Input the samples in the support set into the embedded network model to be trained for measurement, and obtain the classification result w of the i-th sample in the support set. i , expressed as:

[0066]

[0067] In formula (2), f() represents the embedding network model to be trained, S represents the support set, and y i represents the i-th sample in the support set, z i Represents sample y i , wherein the query set is used to test the classification performance of a small sample classification embedding network model;

[0068] S23, input the samples in the query set into the embedded network model to be trained to obtain the classification results of the samples in the query set, and then calculate the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, and then determine the type of the samples in the query set, wherein the type with the smallest difference between the classification results of the samples in the query set and the classification results of the samples in the support set is the type of the samples in the query set, and the classification result of each sample in the query set is the same as the classification result of the i-th sample in the support set, which will not be repeated here;

[0069] S24, using the stochastic gradient descent method to continuously reduce the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, so that the embedding network model to be trained after meta-learning training can obtain the ability to classify new samples using existing prior knowledge, thereby obtaining a small sample classification embedding network model, wherein the difference between the classification results of the samples in the query set and the classification results of the samples in the support set is expressed by the formula:

[0070] L meta =logΣ n exp(-D(f(q),w i )) (3)

[0071] In formula (3), q represents the sample in the query set, w i represents the classification result of the embedded network model to be trained for the i-th sample in the support set, It represents the difference between the classification results of the samples in the query set and the classification results of the samples in the support set.

[0072] In this embodiment, since there are multiple types of samples in the support set and the query set, when considering multiple types of samples comprehensively, the differences between the classification results of multiple samples need to be exponentially summed and then logarithmically calculated. Among them, the specific method of obtaining the small sample classification embedded network model is: the embedded network model obtained by self-supervised learning training, so that the embedded network model can extract prior knowledge from a small amount of unlabeled data, and then the embedded network model is applied to the meta-learning framework for training, which effectively deepens the network learning progress and obtains a small sample classification embedded network model that uses known prior knowledge to classify new samples.

[0073] It should be noted that in this embodiment, for the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, since there is more than one type of samples to be classified, it is necessary to use the exponential sum in formula (3) and then perform logarithmic calculation when considering multiple samples to be classified. Compared with directly adding the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, this processing can avoid a single sample from having too much influence on the whole and causing overfitting.

[0074] Among them, the specific implementation method of step S3 is: first calculate the loss function in the self-distillation training process, that is, calculate the difference between the k-th generation prediction value and the k-th generation actual value and the difference between the k-th generation prediction value and the k-1-th generation prediction value in the self-distillation training; then use the induced network to reduce the two difference values, and continuously iterate until the sum of the two difference values ​​is reduced to the minimum value and stabilizes near the minimum value, thereby obtaining a trained SMSR model. In this embodiment, the small sample classification embedding network model is optimized by self-distillation, so that the structure of the model is more streamlined and the accuracy of the model is greatly improved.

[0075] The model parameterization when the loss function takes the minimum value in the kth iteration training of the self-distillation training is expressed by the formula:

[0076]

[0077] In formula (4), represents the parameterized model obtained from the kth iteration of self-distillation training, KL represents the Kullback-Leibler divergence, α represents the hyperparameter, Q represents the dataset for self-distillation training, and L ce is the cross entropy loss, represents a parameterized model, Representing a parameterized model Classification results for dataset Q.

[0078] In order to further understand the working principle and technical effect of the present invention, a series of related experiments are carried out below to illustrate.

[0079] 1. Select the data set

[0080] like Figure 2 As shown in the figure, the remote sensing image data in the dataset comes from the MAritime SATellite Imagerydataset remote sensing image dataset (referred to as MASATI), which was compiled from different regions of Europe, Africa, Asia, the Mediterranean, and the Atlantic and Pacific Oceans between March 2016 and June 2019, providing optical remote sensing image scenes in the visible spectrum. The dataset contains seven types of samples, including land, coast, ocean, and ship, each with a resolution of 512x512, totaling about 7389 images.

[0081] 2. Deep neural network model training and results

[0082] like Figure 3 As shown in the figure, the preset deep neural network is trained using self-supervised learning. The self-supervised learning network extracts features from the input remote sensing image and generates corresponding feature maps for comparative learning. The self-supervised network obtained after the training is the embedded network model.

[0083] Table 1 shows the structural parameters of the self-supervised learning convolutional network, where each convolution block consists of multiple convolutional layers. Each convolution block from Conv3x to Conv7x contains 16 layers of convolution (Conv represents convolutional layer). 1x1 convolution is used to reduce the number of channels and reduce the complexity of the model. The same number of output channels is used in each convolution block. During the forward operation, the input and output of each block will be connected in the channel in turn to form a self-supervised learning network.

[0084] Table 1. Network structure parameters for self-supervised learning

[0085]

[0086]

[0087] During the self-supervised meta-learning training, the selected remote sensing image small sample scene classification task is 1-shot 5-way. Before training, the remote sensing image pictures are pre-processed and cropped to 256×256 and input into the network. The optimizer uses SGD (stochastic gradient descent method), the learning rate is 0.0001, the dropout (a parameter in deep neural network) is 0.5, and the batch size is 32. A total of 300 epochs are trained (one epoch means the process of training all samples once).

[0088] When the model is self-distilled, the model obtained by the self-supervised meta-learning training phase is used on the same dataset for self-distillation. The hyperparameter α is set to 0.5, and a total of 10 epochs of self-distillation training are performed. The model obtained after self-distillation is named SMSR model, which is a small sample remote sensing image scene classification method model based on self-supervised meta-learning and self-distillation.

[0089] For comparison, this experiment selected several classic supervised small sample learning algorithms for training in 1-shot 5-way (where 5-Way means 5 ways or 5 categories, and 1-Shot means 1 time or 1) tasks, and tested them in the dataset. The test results are shown in Table 2:

[0090] Table 2 Comparison of MASATI 1-shot 5-way accuracy, with a confidence level of 95%

[0091] Algorithm Name Embedded Network 1-shot 5-way accuracy MAML Convnet 53.70±0.82％ RelationNet Convnet 54.21±0.14％ SNAIL ResNet50 56.25±0.29％ ProtoNet ResNet50 57.63±0.66％ Self-jig ResNet50 58.16±0.84％ SMSR Self-supervised learning network 59.62±0.16％

[0092] It can be seen from Table 2 that the SMSR model classification method proposed in the present invention has certain advantages in accuracy over the classic supervised small sample learning algorithm in the 1-shot 5-way task performed in the MASATI remote sensing image dataset (remote sensing image dataset in deep learning), and its accuracy is improved by about 1.5% over Self-jig.

[0093] 3 Ablation experiments

[0094] In order to verify the effectiveness of the self-supervised meta-learning and self-distillation used in the SMSR model classification method proposed in the present invention, a series of ablation experiments were further performed.

[0095] 3.1 Ablation Experiments on Self-Supervised Learning

[0096] In order to verify the effectiveness of self-supervised learning, supervised training is used instead of self-supervised training in SMSR. The embedding model directly uses ResNet50 (a neural network) instead of the self-supervised learning network for meta-learning. The other parts remain the same as the SMSR model, and the algorithm is named MSR model. The MSR model is trained and tested in the same training environment as the SMSR model. The experimental results are shown in Table 3.

[0097] Table 3 Comparison of MASATI 1-shot 5-way accuracy, with a confidence level of 95%

[0098]

[0099]

[0100] As shown in Table 3, when supervised learning uses only a small amount of labeled data to train the embedding network, overfitting will occur due to too few samples, which further leads to insufficient generalization ability of the meta-learning model. Its performance in the 1-shot 5-way task is not as good as that of the SMSR model using self-supervised learning. Therefore, self-supervised learning can be used to obtain an embedding network model with good performance and improve the performance of the meta-learning model.

[0101] 3.2 Self-distillation ablation experiment

[0102] In order to verify the effectiveness of self-distillation, the self-distillation part in the SMSR model is removed, and the other structures are consistent with the SMSR model. Training and testing are carried out in the same training environment as the SMSR model. The experimental results are shown in Table 4.

[0103] Table 4 Comparison of MASATI 1-shot 5-way accuracy, with a confidence level of 95%

[0104] Algorithm Name Basic Network 1-shot 5-way accuracy SMSR- Self-supervised learning network 62.83±0.91％ SMSR Self-supervised learning network 64.54±0.16％

[0105] Among them, the SMSR-model represents the SMSR model without self-distillation. It can be seen from Table 4 that the SMSR model trained by self-distillation is about 1.7% more accurate than SMSR-. This shows that self-distillation training can make the model more refined.

[0106] The above is a detailed introduction to the self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification, characterized by: The method comprises the following steps: S1. Based on self-supervised learning and using remote sensing image data, a preset deep neural network is trained to obtain an embedded network model that can extract feature information of remote sensing images; S2, embedding the embedded network model obtained in step S1 into the framework of meta-learning for training, and enabling the embedded network model trained by meta-learning to classify new samples using existing prior knowledge, thereby obtaining a small sample classification embedded network model; S3. Iteratively optimize the small sample classification embedding network model based on self-distillation training, and then complete the self-supervised self-distillation meta-learning for small sample remote sensing image classification to obtain a trained SMSR model; The specific implementation of step S1 is as follows: S11. Set contrastive learning as an auxiliary task of self-supervised learning, and select NCE loss as the measurement indicator of contrastive learning; S12, using different data enhancement methods to perform data enhancement on the same remote sensing image, and calculating the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods; S13, using a stochastic gradient descent method to continuously reduce the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods until the value of the NCE loss tends to be stable, thereby obtaining an embedded network model capable of extracting feature information of the remote sensing image; The specific implementation of step S2 is: S21, embedding the embedded network model obtained in step S1 into the framework of meta-learning and using it as the embedded network model to be trained; S22. Divide the samples in the data set consisting of all remote sensing images into several categories, randomly select K samples from each category as the support set, and the remaining samples as the query set. Input the samples in the support set into the embedded network model to be trained for measurement, and obtain the classification result w of the i-th sample in the support set. i , expressed as: In formula (2), f() represents the embedding network model to be trained, S represents the support set, and y i represents the i-th sample in the support set, and zi represents the corresponding label of sample yi; S23, inputting the samples in the query set into the embedding network model to be trained to obtain the classification results of the samples in the query set, and then calculating the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, and then determining the types of the samples in the query set; S24, using the stochastic gradient descent method to continuously reduce the difference between the classification results of the samples in the query set and the classification results of the samples in the support set, so that the embedding network model to be trained after meta-learning training can obtain the ability to classify new samples using the existing prior knowledge, thereby obtaining a small sample classification embedding network model; The specific implementation method of step S3 is: first calculate the loss function in the self-distillation training process, that is, calculate the difference between the predicted value of the kth iteration and the actual value of the kth iteration and the difference between the predicted value of the kth iteration and the predicted value of the k-1th iteration in the self-distillation training; then use the induced network to reduce the two difference values, and continuously iterate until the sum of the two difference values ​​is reduced to a minimum value and stabilizes near the minimum value, thereby obtaining a trained SMSR model.

2. The self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification according to claim 1, characterized in that: The different data enhancement methods in step S12 include scaling, random cropping, random center cropping, affine transformation, color jittering, random horizontal flipping and random grayscale transformation.

3. The self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification according to claim 2, characterized in that: The contrastive learning is performed by selecting feature maps generated by different convolutional layers in a preset deep neural network for comparison.

4. The self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification according to claim 3, characterized in that: In step S12, the NCE loss between the data enhancement processing results of the remote sensing image under different data enhancement methods is expressed by the formula: L(x m ,x n )=L(f a (x m ),f5(x n ))+L(f a (x m ),f e (x n ))+L(f5(x m ),f5(x n ))+L(f a (x n ),f5(x m )) +L(f a (x n ),f e (x m ))+L(f e (x m ),f e (x n ))(1) In formula (1), x m and x n They represent the processing results of the same remote sensing image after two different data enhancement methods, f a represents the global feature map extracted by the preset deep neural network, f5 represents the local feature map extracted by the fifth convolution layer in the preset deep neural network, and f e Represents the output feature map of the last layer of the network extracted by the last layer of convolution in the preset deep neural network, where: N represents all remote sensing images, and d represents the square of the Euclidean distance.

5. The self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification according to claim 4, characterized in that: The difference between the classification results of the samples in the query set and the classification results of the samples in the support set is expressed as follows: L meta =log∑ n exp(-D(f(q),w i )) (3) In formula (3), q represents the sample in the query set, w i It represents the classification result of the embedded network model to be trained for the i-th sample in the support set. Represents the difference between the classification results of samples in the query set and samples in the support set.

6. The self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification according to claim 5, characterized in that: The model parameterization when the loss function takes the minimum value in the kth iteration training of the self-distillation training is expressed by the formula: In formula (4), represents the parameterized model obtained from the kth iteration of self-distillation training, KL represents the Kullback-Leibler divergence, α represents the hyperparameter, Q represents the dataset for self-distillation training, and L ce is the cross entropy loss, represents a parameterized model, Representing a parameterized model Classification results for dataset Q.

7. The self-supervised self-distillation meta-learning method for small sample remote sensing image scene classification according to any one of claims 1 to 6, characterized in that: Also includes: S4. Input the remote sensing images acquired in real time into the trained SMSR model for classification, thereby achieving scene classification of the remote sensing images.

Images