A method for scene classification of remote sensing images based on continuous learning

Abstract

The present invention discloses a remote sensing image scene classification method based on continuous learning, which obtains a remote sensing image data set to be classified and a remote sensing image training data set at the current moment; wherein the image category in the remote sensing image data set to be classified is greater than the image category in the remote sensing image training data set; the remote sensing image classification model at the previous moment is trained by using the remote sensing image training data set; the remote sensing image classification model to be classified is classified by using the trained remote sensing image classification model; the present invention updates the remote sensing image classification model at the previous moment by using the remote sensing image training data set, and adds a second loss function to the loss function during the updating process, so that the remote sensing image classification model can retain the classification accuracy of the original remote sensing image, thereby improving the generalization ability of the remote sensing image classification model.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image classification, and in particular relates to a remote sensing image scene classification method based on continuous learning. Background Art

[0002] With the development of earth observation technology, computer technology, and artificial intelligence technology, the era of remote sensing big data has arrived. In particular, the innovation of communication technologies such as 5G has made the intelligent interconnection of all things gradually possible. In this case, developing remote sensing edge intelligence application technology, putting remote sensing data processing and analysis technology in the front, and applying it to scenarios such as satellite on-orbit processing and interpretation, real-time dynamic tracking of drones, large-scale urban environment reconstruction, and unmanned driving identification planning can undoubtedly save a lot of transmission bandwidth, processing time, and resource consumption.

[0003] Deep learning has been extensively studied in the field of intelligent interpretation of remote sensing data due to its powerful feature self-learning and application generalization capabilities. Many intelligent interpretation models of remote sensing data based on deep neural network models have been proposed and promoted.

[0004] However, in edge remote sensing data analysis, new tasks and new categories emerge in an endless stream. Existing methods often have poor generalization capabilities for continuously input data and suffer from catastrophic forgetting problems, that is, the model can only be applied to current tasks and remote sensing image samples and cannot adapt or expand over time. Summary of the invention

[0005] The purpose of the present invention is to provide a remote sensing image scene classification method based on continuous learning to solve the problem of poor generalization ability of the model in the continuous learning process.

[0006] The present invention adopts the following technical solution: a remote sensing image scene classification method based on continuous learning, comprising the following steps:

[0007] Obtaining a remote sensing image dataset to be classified and a remote sensing image training dataset at the current moment; wherein the image category in the remote sensing image dataset to be classified is greater than the image category in the remote sensing image training dataset;

[0008] The remote sensing image training data set is used to train the remote sensing image classification model at the previous moment;

[0009] The trained remote sensing image classification model is used to classify the remote sensing image dataset.

[0010] Furthermore, using the remote sensing image training data set to train the remote sensing image classification model at the previous moment includes:

[0011] use As the loss function of the remote sensing image classification model; where, is the loss function of the remote sensing image classification model, is the first loss function calculated based on the predicted category and the true category of each remote sensing image in the remote sensing image training dataset, It is a second loss function calculated based on the feature subspace of the remote sensing image training data set and the feature subspace at the previous moment, and the feature subspace is used to characterize the feature matrix of the remote sensing image training data set.

[0012] Furthermore, the second loss function calculation method is:

[0013] Calculate the distance between the feature subspace at the current moment and the feature subspace at the previous moment, and use the distance as the second loss function.

[0014] Furthermore, calculating the distance between the feature subspace at the current moment and the feature subspace at the previous moment includes:

[0015] Calculate the standard orthogonal basis of the characteristic subspace at the current moment and the standard orthogonal basis of the characteristic subspace at the previous moment;

[0016] The sum of the sine values ​​of the angle between the two standard orthogonal bases is calculated and used as the distance between the feature subspace at the current moment and the feature subspace at the previous moment.

[0017] Furthermore, the calculation of the standard orthogonal basis of the feature subspace at the current moment includes:

[0018] Extract the feature map of each remote sensing image in the remote sensing image training data set through the remote sensing image classification model at the previous moment;

[0019] Combine feature maps to get feature matrix;

[0020] Perform singular value decomposition on the characteristic matrix to obtain the standard orthogonal basis of the characteristic subspace at the current moment.

[0021] Furthermore, after performing singular value decomposition on the feature matrix, the following steps are also included:

[0022] The left singular matrix of the characteristic matrix is ​​used as the standard orthogonal basis of the characteristic subspace at the current moment.

[0023] Furthermore, the second loss function is specifically:

[0024]

[0025] in, Represents the feature subspace S at the current moment (t) and the characteristic subspace S at the previous moment (t-1) The distance, Θ is the characteristic subspace S (t) The orthonormal basis and characteristic subspace S of (t-1)The angle vector between the orthonormal basis of .

[0026] Furthermore, the angle vector includes several angles, and the calculation method of the several angles is:

[0027]

[0028]

[0029] in, is the first angle in the angle vector, is the characteristic subspace S (t-1) The first orthonormal basis of is the characteristic subspace S (t) The first orthonormal basis of is the second angle in the angle vector, is the characteristic subspace S (t-1) The second orthonormal basis of is the characteristic subspace S (t) The second orthonormal basis of is the bth angle in the angle vector, b is the number of angles in the angle vector, is the characteristic subspace S (t-1) The bth orthonormal basis of is the characteristic subspace S (t) The b-th orthonormal basis of .

[0030] Furthermore, the standard orthogonal basis is obtained by performing singular value decomposition on the characteristic matrix.

[0031] Another technical solution of the present invention: a remote sensing image scene classification device based on continuous learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned remote sensing image scene classification method based on continuous learning is implemented.

[0032] The beneficial effect of the present invention is: the present invention updates the remote sensing image classification model of the previous moment by using the remote sensing image training data set, and adds a second loss function to the loss function during the updating process, so that the remote sensing image classification model can retain the classification accuracy of the original remote sensing image, thereby improving the generalization ability of the remote sensing image classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of a network architecture of a remote sensing image classification model in an embodiment of the present invention;

[0034] Figure 2 A flowchart of a remote sensing image scene classification method based on continuous learning according to an embodiment of the present invention;

[0035] Figure 3 Schematic diagram of training of remote sensing image classification model at time t in an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram comparing the accuracy of continuous classification of remote sensing images of various methods in the verification embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] In edge remote sensing data analysis, when new data is available and the model needs to be updated, the model cannot maintain the performance requirements of the original task. For example, for the detection and recognition tasks of targets such as aircraft and ships based on satellite images, the current common method is to train models separately for different targets, or even train models separately for subdivided target models. This method is complex to implement and cumbersome to apply. Satellite images can be updated in large quantities every day, and the existing static models cannot be updated in time and utilize the newly added data. Therefore, how to conduct continuous learning to maintain the high accuracy and high performance of the model algorithm has become a challenge that needs to be solved urgently.

[0039] The difficulty in solving the continuous learning problem lies in learning new tasks while forgetting as little of the old tasks as possible. The existing technologies mainly include regularization-based methods, data reproduction-based methods, and model structure expansion-based methods.

[0040] Regularization-based methods: Additional regularization terms are introduced into the loss function to consolidate previous knowledge by limiting the degree of change of model parameters while learning new tasks. This avoids storing the original input data, prioritizes privacy protection, and reduces memory requirements.

[0041] Regularization-based methods reduce forgetting by constraining the degree of change of model parameters when facing continuous tasks, but they do not impose constraints on the higher level of data, which will lead to large differences in scale, semantics, etc. between the features of the previous task data and the features of the next round of tasks. As a result, in order to achieve new tasks, new features tend to cover old features, which is not effective in solving the problem of catastrophic forgetting.

[0042] Add a memory storage unit: Add a storage module to the model to save previous data and regularly replay the cross-data of previously learned knowledge and new samples. By storing samples of previous tasks or applying a generative model to generate pseudo samples, while learning new tasks, replaying the stored samples, and balancing the training data of different tasks, the catastrophic forgetting problem can be alleviated.

[0043] Adding a memory storage module A common disadvantage of continuous learning methods is that they need to explicitly store old task information, which will lead to larger working memory requirements and thus increase additional storage space. When the storage space is limited (such as remote sensing satellites), the model is not effective in solving the forgetting problem.

[0044] Parameter isolation-based methods: This method specifies different model parameters for each task to prevent any possible forgetting. When the architecture size is not limited, new branches can be grown for new tasks while freezing the previous task parameters or dedicating a copy of the model to each task. Alternatively, the architecture remains static with a fixed portion assigned to each task. The previous task portion is masked out during training of the new task, or imposed at the parameter level or unit level. Dividing the model into subsets dedicated to each task improves the task adaptability of the model and avoids catastrophic forgetting without relying on historical data.

[0045] The parameter isolation-based method will increase the number of network branches as the number of tasks increases, and the neural network architecture will continue to increase, which will directly reduce the scalability of the model. Since the number of tasks and sample size cannot be known in advance in continuous learning scenarios, it is inevitable to predefine sufficient storage resources without making strong assumptions about the probability distribution of input training samples, which increases storage costs. In addition, the available load of edge remote sensing devices themselves is limited, and there are major limitations on hardware devices such as storage and processing, making it impossible to store large amounts of historical remote sensing data.

[0046] Therefore, the present invention first defines the concept of feature subspace, explicitly models the description method of features, constrains different tasks to share feature subspaces through the way of self-knowledge distillation, reduces the feature gap between different tasks, alleviates catastrophic forgetting, and improves the classification accuracy of continuous learning without increasing additional storage costs. In the present invention, the feature subspace can be considered as the knowledge about the current remote sensing image learned by the network (i.e., the remote sensing image classification model), and the standard orthogonal basis can be considered as a description operator of the knowledge about the current remote sensing image learned by the network.

[0047] The invention discloses a remote sensing image scene classification method based on continuous learning, comprising the following steps: obtaining a remote sensing image data set to be classified and a remote sensing image training data set at the current moment; wherein the image category in the remote sensing image data set to be classified is greater than the image category in the remote sensing image training data set; using the remote sensing image training data set to train a remote sensing image classification model at the previous moment; and classifying the remote sensing image data set to be classified by using the trained remote sensing image classification model.

[0048] The present invention updates the remote sensing image classification model of the previous moment by using a remote sensing image training data set, and adds a second loss function to the loss function during the updating process, so that the remote sensing image classification model can retain the classification accuracy of the original remote sensing image, thereby improving the generalization ability of the remote sensing image classification model.

[0049] The present invention combines the idea of ​​feature subspace and knowledge distillation, which can help the model refine and preserve more complete knowledge about the data without increasing the storage cost, and help the model reduce catastrophic forgetting. Figure 1 As shown, the model consists of a feature extraction part, a subspace construction part and a classification part.

[0050] The method flow is as follows Figure 2 First, obtain the remote sensing image classification data D containing the new category at time t (t) , each sample in the data All have labels The entire sample set is represented by X (t) , the entire label set is represented as Y (t) , then D (t) =(X (t) , Y (t) ).

[0051] In this embodiment, the training phase of the invention is taken as an example. Using the newly acquired remote sensing image classification data D (t) Remote sensing image training dataset Train for model construction (t) With test data Test (t) ; Secondly, the training sample X at time t (t) Input to the model M of the previous round (t-1) In the above example, we obtain and record the feature subspace S of the current round of training samples in the previous round of model. (t-1) , which is the characteristic attribute of the feature map of the training sample data in this round; then, the remote sensing image classification data D (t) With the characteristic subspace S (t-1) For model M (t-1) Train and get the updated model M (t) ; Finally, save the training parameters of this round and update the model to M (t) , and use the model M after the parameters of this round are updated before the next round of training (t) The continuous classification task is performed on the remote sensing image dataset to be classified. The dataset construction and model training are described in detail below.

[0052] In the embodiment of the present invention, the training data Train of the model in the tth round of training (t) At the beginning of the tth round of training, the remote sensing image classification data D containing new categories is acquired(t) 75% of the data is augmented, i.e., random cutting, flipping, contrast and grayscale adjustment of D (t)* The model is trained as training data. After each round of training, the current training data is removed. That is, the training data of each round only contains the currently acquired remote sensing image training data set. (t) =D (t)* .

[0053] During the training phase of the model, the test data of the model in the tth round of testing is Test (t) At the end of the tth round of training, the overall sample 25% of the data is used as test data to evaluate the model.

[0054] The training process of the t-th round model is as follows: Figure 3 First, the t-th round model training sample X (t) Input the model M of the previous round (t-1) Secondly, use the feature extractor to partially extract the training data training sample X (t) Each sample The characteristic f i (t) ; Then, feature f i (t) The characteristic subspace S is obtained by the subspace construction part (t-1) ; At the same time, the feature f i (t) The prediction results are generated through the classification part.

[0055] In an embodiment of the present invention, using a remote sensing image training data set to train a remote sensing image classification model at a previous moment includes: using As the loss function of the remote sensing image classification model; where, It is the loss function of the remote sensing image classification model, that is, the first loss function. is the cross entropy loss function calculated based on the predicted category and the true category of each remote sensing image in the remote sensing image training data set, and is the training data Train (t) The first entropy loss function calculated by the first predicted category probability and the true label of each sample in . is the second loss function calculated based on the feature subspace of the remote sensing image training dataset and the feature subspace of the previous moment. The feature subspace is used to characterize the feature matrix of the remote sensing image training dataset. After determining the loss function, the gradient descent algorithm is used to optimize the loss function to train the model and obtain the model M after parameter update. (t) .

[0056] In an embodiment of the present invention, the second loss function calculation method is: calculating the distance between the feature subspace at the current moment and the feature subspace at the previous moment, and using the distance as the second loss function.

[0057] Specifically, calculating the distance between the feature subspace at the current moment and the feature subspace at the previous moment includes: calculating the standard orthogonal basis of the feature subspace at the current moment and the standard orthogonal basis of the feature subspace at the previous moment; calculating the sum of the sine values ​​of the angle between the two standard orthogonal bases, and using it as the distance between the feature subspace at the current moment and the feature subspace at the previous moment.

[0058] More specifically, calculating the standard orthogonal basis of the feature subspace at the current moment includes: extracting the feature map of each remote sensing image in the remote sensing image training data set through the remote sensing image classification model at the previous moment; combining the feature maps to obtain the feature matrix; performing singular value decomposition on the feature matrix to obtain the standard orthogonal basis of the feature subspace at the current moment.

[0059] Preferably, after performing singular value decomposition on the characteristic matrix, the method further includes: using the left singular matrix of the characteristic matrix as a standard orthogonal basis of the characteristic subspace at the current moment.

[0060] Specifically, in order to measure the distance between feature subspaces (Feature Subspace Distance, FSD), it is defined as the similarity of the standard orthogonal bases of two feature subspaces, that is, the sum of the sine values ​​of the angles between the standard orthogonal bases of the two feature subspaces. The second loss function is to construct the self-knowledge distillation loss function by minimizing the distance between the two feature subspaces, specifically:

[0061]

[0062] in, Represents the feature subspace S at the current moment (t) and the characteristic subspace S at the previous moment (t-1) The distance, Θ is the characteristic subspace S (t) The orthonormal basis and characteristic subspace S of (t-1) The angle vector between the standard orthogonal basis of ||·||1 is the 1-norm.

[0063] Preferably, the angle vector includes several angles, and the calculation method of the several angles is:

[0064]

[0065]

[0066] in, is corresponding to model M (t-1)The orthogonal basis of is corresponding to model M (t) An orthogonal basis. is the first angle in the angle vector, is the characteristic subspace S (t-1) The first orthonormal basis of is the characteristic subspace S (t) The first orthonormal basis of is the second angle in the angle vector, is the characteristic subspace S (t-1) The second orthonormal basis of is the characteristic subspace S (t) The second orthonormal basis of is the bth angle in the angle vector, b is the number of angles in the angle vector, is the characteristic subspace S (t-1) The bth orthonormal basis of is the characteristic subspace S (t) The b-th orthonormal basis of .

[0067] In this embodiment, the standard orthogonal basis is obtained by performing singular value decomposition on the feature matrix. First, the feature extractor part is used to extract each sample x in the training data. i The characteristic f i , a batch of samples pass through the feature extractor to form a feature matrix F = [f1...f b ], b is the number of samples in this batch. Next, the obtained feature matrix F is subjected to singular value decomposition (SVD):

[0068] F=U∑V,

[0069] Among them, U is a left singular matrix, V is a right singular matrix, and ∑ is a diagonal matrix. In this embodiment, the matrix U is used as an orthogonal basis in the model characteristic subspace S, and then the standard orthogonal basis of the characteristic subspace is used to characterize the subspace.

[0070] That is, the training sample X at time t (t) Input to the model M of the previous round (t-1) In the above example, we obtain and record the feature subspace S of the current round of training samples in the previous round of model. (t-1) The orthonormal basis U (t-1) , the feature subspace updated at the training time is S (t) .

[0071] In one embodiment, the model uses a Resnet34 network. The subspace construction part uses SVD decomposition without parameter update. The model is optimized using the Adam optimizer, and the parameters are set to default.

[0072] In addition, in order to prove the feasibility of the method of the present invention, the following verification examples are specifically carried out:

[0073] First, determine the experimental dataset. The validated Continual Learning Benchmark for Remote Sensing (CLRS) dataset consists of 15,000 remote sensing images, which are divided into 25 scene classes, namely, airport, bare land, beach, bridge, business, desert, farmland, forest, golf course, highway, industry, grassland, mountain, overpass, park, parking lot, playground, port, railway, railway station, residential, river, runway, stadium, and storage tank. Each class has 600 images, and the image size is 256×256. The resolution of the images ranges from 0.26m to 8.85m. These 15,000 photos come from more than 100 countries and regions around the world.

[0074] Next, we set up the experiment. The 25 scene classes are divided into 5 non-overlapping parts, each containing 5 types of images. Each part is set as a task, with 75% used for training and 25% used for testing. That is, the training data for each task contains 5 scene classes and 2250 images, and the test data also contains 5 scene classes and 750 images.

[0075] However, the test data after each round of training includes the test data set for each previous task, that is, t*5 scene classes, t*750 images, and t is the number of tasks. In this task setting, the test set contains the scene classes that appeared in the training set in the previous task. The classifier not only needs to learn to distinguish the scene classes currently seen, but also needs to distinguish all scene classes seen so far. Therefore, the model must be able to quickly learn new scene categories and not forget the previously learned category knowledge. The model is required to accurately predict new scene categories without losing the prediction accuracy of the categories that have already been learned.

[0076] In order to verify the efficiency of the method, the present invention is compared with the current mainstream continuous learning methods, including EWC and LWF. The results show that on the test set, the average classification accuracy of the present invention in each stage of continuous tasks is at least 4% higher than that of all other methods. Figure 4 As shown in Table 1, this fully demonstrates the high efficiency of the present invention in edge intelligent remote sensing scene classification.

[0077] Table 1 Comparison of methods for continuous classification of remote sensing images

[0078] Methods Task1 Task2 Task3 Task4 Task5 LqCy 98.59％ 50.59％ 34.00％ 23.52％ 14.32％ EWC 98.06％ 47.33％ 31.25％ 22.56％ 12.78％ Fine-tuning 98.89％ 41.51％ 32.93％ 21.41％ 10.88％ Ours 98.71％ 67.89％ 41.85％ 27.93％ 22.94％

[0079] In summary, the present invention proposes the concept of feature subspace in continuous learning tasks for the first time. By explicitly performing singular value decomposition on the feature matrix of the data extracted by the model, the left singular matrix of the feature matrix is ​​obtained, which is the standard orthogonal basis of the feature subspace. The feature subspace is characterized by the standard orthogonal basis. At the same time, the similarity of the feature subspace can be measured using the standard orthogonal basis, thereby constraining the feature subspace. Moreover, the invention proposes to limit the consistency of feature subspaces of different tasks in a self-knowledge distillation manner. The feature subspaces of different task data at different times in the same model are limited by the knowledge distillation loss function, so as to achieve self-knowledge distillation of the model while reducing catastrophic forgetting.

[0080] That is to say, the present invention uses feature subspaces to characterize different data features of tasks at different stages, explicitly extracts the features of the data through the subspace modeling network, unifies the feature description method, and reduces the semantic drift problem caused by data differences. It is proposed to limit the consistency of feature subspaces of different tasks in the form of self-knowledge distillation. By retaining the feature subspace instead of directly retaining the features or samples, the requirements for storage resources are reduced, and at the same time, the model always maintains the stability of the feature space during the optimization process, alleviates the catastrophic forgetting phenomenon, and improves the continuous classification accuracy of edge remote sensing images without the need for additional storage resources.

[0081] The present invention also discloses a remote sensing image scene classification device based on continuous learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned remote sensing image scene classification method based on continuous learning is implemented.

[0082] The above-mentioned device can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The device may include but is not limited to a processor and a memory. Those skilled in the art will appreciate that the device may include more or fewer components, or a combination of certain components, or different components, for example, it may also include input and output devices, network access devices, etc.

[0083] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0084] In some embodiments, the memory may be an internal storage unit of the device, such as a hard disk or memory of the device. In other embodiments, the memory may also be an external storage device of the device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the device. Further, the memory may also include both an internal storage unit of the device and an external storage device. The memory is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory may also be used to temporarily store data that has been output or is to be output.

[0085] It should be noted that the specific content of the above-mentioned device is based on the same concept as the embodiment of the method of the present invention. Its specific functions and technical effects can be found in the method embodiment part, and will not be repeated here.

[0086] In the above embodiments, the description of each embodiment has its own emphasis. For the parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. Those of ordinary skill in the art can appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented with electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

Claims

1. A remote sensing image scene classification method based on continuous learning, characterized in that: The following steps are involved: Acquire a remote sensing image dataset to be classified and a remote sensing image training dataset at the current moment; wherein the image category in the remote sensing image dataset to be classified is greater than the image category in the remote sensing image training dataset; Using the remote sensing image training data set to train the remote sensing image classification model at the previous moment; Classifying the remote sensing image data set to be classified by using the trained remote sensing image classification model; Using the remote sensing image training data set to train the remote sensing image classification model at the previous moment includes: use As the loss function of the remote sensing image classification model; wherein, is the loss function of the remote sensing image classification model, is a first loss function calculated according to the predicted category and the true category of each remote sensing image in the remote sensing image training data set, is a second loss function calculated based on a feature subspace of the remote sensing image training data set and a feature subspace at a previous moment, wherein the feature subspace is used to characterize a feature matrix of the remote sensing image training data set; The second loss function calculation method is: Calculating the distance between the feature subspace at the current moment and the feature subspace at the previous moment, and using the distance as the second loss function; Calculating the distance between the feature subspace at the current moment and the feature subspace at the previous moment includes: Calculate the standard orthogonal basis of the characteristic subspace at the current moment and the standard orthogonal basis of the characteristic subspace at the previous moment; Calculate the sum of the sine values ​​of the angle between the two standard orthogonal bases, and use it as the distance between the feature subspace at the current moment and the feature subspace at the previous moment; The second loss function is specifically: in, Represents the feature subspace S at the current moment (t) and the characteristic subspace S at the previous moment (t-1) The distance, Θ is the characteristic subspace S (t) The orthonormal basis and characteristic subspace S of (t-1) The angle vector between the orthonormal basis of .

2. A remote sensing image scene classification method based on continuous learning as claimed in claim 1, characterized in that: The standard orthogonal basis for calculating the characteristic subspace at the current moment includes: Extracting a feature map of each remote sensing image in the remote sensing image training data set through a remote sensing image classification model at a previous moment; Combining the feature maps to obtain a feature matrix; Perform singular value decomposition on the characteristic matrix to obtain a standard orthogonal basis of the characteristic subspace at the current moment.

3. A remote sensing image scene classification method based on continuous learning as claimed in claim 2, characterized in that: After performing singular value decomposition on the feature matrix, the method further includes: The left singular matrix of the characteristic matrix is ​​used as the standard orthogonal basis of the characteristic subspace at the current moment.

4. A remote sensing image scene classification method based on continuous learning as claimed in claim 3, characterized in that: The angle vector includes a plurality of angles, and the calculation method of the plurality of angles is: in, is the first angle in the angle vector, is the characteristic subspace S (t-1) The first orthonormal basis of is the characteristic subspace S (t) The first orthonormal basis of is the second angle in the angle vector, is the characteristic subspace S (t-1) The second orthonormal basis of is the characteristic subspace S (t) The second orthonormal basis of is the bth angle in the angle vector, b is the number of angles in the angle vector, is the characteristic subspace S (t-1) The bth orthonormal basis of is the characteristic subspace S (t) The b-th orthonormal basis of .

5. A remote sensing image scene classification method based on continuous learning as claimed in claim 4, characterized in that: The standard orthogonal basis is obtained by performing singular value decomposition on the characteristic matrix.

6. A remote sensing image scene classification device based on continuous learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the remote sensing image scene classification method based on continuous learning as described in any one of claims 1 to 5 is implemented.

Images