Remote sensing classification method and system based on gradient neural network
By adopting a robust model based on gradient neural network in remote sensing image classification, the problem of low classification accuracy under noise conditions is solved, and higher classification accuracy and system stability are achieved.
Patent Information
- Application Number
- CN202510243802.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-03
- Publication Date
- 2025-06-03
AI Technical Summary
The prior art is difficult to effectively classify remote sensing images under noisy conditions, resulting in reduced classification accuracy and task failure.
Using a robust model based on gradient neural network, the objective function and robust gradient neural network model are constructed by constructing the L2 norm minimization objective function and the robust gradient neural network model, and iterative training is performed to obtain the coefficient vector of L2 norm minimization, and the regularization term is used to ensure the stability of the learning process.
It improves the accuracy of remote sensing image classification under noise conditions, enhances the tolerance for noise and outliers in remote sensing image data, and improves the stability and reliability of the system in complex environments.
Smart Images
Figure CN120088568A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of neural networks, and particularly relates to a remote sensing classification method and system based on a gradient neural network. Background Art
[0002] Image classification is a basic step in many applications such as face recognition, disease diagnosis, and remote sensing. In related technologies, a new collaborative representation method is proposed, which maps samples to a complex space to capture discriminative and non-linear information, and is superior to other traditional facial dataset methods. In addition, an improved deep neural network model with optimal feature selection is proposed for the classification of lung cancer, brain images, and Alzheimer's disease. In the field of remote sensing, related technologies use convolutional networks with classifiers to segment remote sensing images. Remote sensing images provide multi-spectral data of a vast area, and can be used to analyze in detail land cover, agricultural and fishery development, urban growth, etc. This natural characteristic makes it crucial for climate change research, biodiversity conservation, precision agriculture, and other applications. Due to the ability of remote sensing images to monitor large-scale environmental changes, land use, and disasters, remote sensing image classification has attracted much attention.
[0003] Classifiers based on L2 minimization are widely popular due to their closed-form solutions with low computational complexity. However, most of them classify under noise-free conditions, while noise is inevitable in computing systems. There are many sources of noise, such as hardware implementation, working environment noise, errors in data storage and signal transmission, etc. The inevitable noise will reduce the accuracy of the computational solution and even lead to the failure of the task. Therefore, it is necessary to design a robust solution to achieve remote sensing image classification under noisy conditions and improve the accuracy of remote sensing image classification under noisy conditions. Summary of the Invention
[0004] The purpose of the present invention is to provide a remote sensing classification method and system based on a gradient neural network, which can improve the accuracy of remote sensing image classification under noisy conditions.
[0005] In a first aspect, an embodiment of the present application provides a remote sensing classification method based on a gradient neural network, including the following steps:
[0006] Obtain a plurality of training samples generated based on remote sensing images, divide the plurality of training samples into multiple groups of training samples, and form a training sample matrix with the multiple groups of training samples; wherein, each group of training samples contains a plurality of training samples;
[0007] Construct an objective function for minimizing the L2 norm based on the CRC model, construct a robust gradient neural network model for solving the objective function, and iteratively train the robust gradient neural network model based on the remote sensing image to be classified and each class of training sample groups to obtain a coefficient vector with minimized L2 norm; wherein, the objective function is to solve the coefficient vector γ of the linear combination of the remote sensing image to be classified and each class of training sample groups, and has the minimum L2 norm.
[0008] Split the coefficient vector into multiple sub-vectors according to categories, calculate the expression errors between the remote sensing image to be classified and each sub-vector, obtain the expression errors of each category, and take the category with the minimum expression error as the classification result of the remote sensing image.
[0009] Preferably, constructing the objective function for minimizing the L2 norm based on the CRC model includes:
[0010] Establish a mathematical model for solving the coefficient vector γ based on the CRC model, and the expression is:
[0011]
[0012] where y is the test sample, y ∈ R m , λ CRC is the regularization parameter, γ is the coefficient vector; X is the training sample matrix containing remote sensing data, x i ∈ R m , n is the number of training samples, m is the dimension of the feature space or the number of spectral bands, x i is the i-th training sample; in the training sample matrix X, there are N classes, c represents the class index, N c represents the number of training samples belonging to the c-th class, X c is the c-th class of training sample group, is the N c -th training sample in the c-th class;
[0013] Transform the mathematical model into an objective function for minimizing the L2 norm, and the expression is:
[0014]
[0015] where γ(t) is a time-varying coefficient vector, and F(γ(t)) represents a function with respect to γ(t).
[0016] Preferably, the expression of the robust gradient neural network model is:
[0017] where, I is the identity matrix, β is a constant coefficient, e(t) represents the error function, and τ is the time variable.
[0018] Preferably, the expression of the error function is:
[0019] Preferably, iteratively training the robust gradient neural network model based on the remote sensing image to be classified and each class of training sample groups to obtain a coefficient vector with minimized L2 norm, including:
[0020] Inputting the training sample matrix, the remote sensing image to be classified, the regularization parameter, the constant coefficient, and the initial coefficient vector into the robust gradient neural network model, and solving to obtain
[0021] Passing through an integrator to integrate to obtain γ(t), updating the coefficient vector to γ(t), and then iteratively training the robust gradient neural network model; after reaching the condition for stopping iterative training, outputting the coefficient vector with minimized L2 norm.
[0022] Preferably, the expression of the expression error is where γ c is the c-th class sub-vector, and X c is the c-th class training sample.
[0023] In a second aspect, an embodiment of the present application provides a remote sensing classification system based on a gradient neural network, including:
[0024] At least one processor;
[0025] At least one memory for storing at least one program;
[0026] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.
[0027] The beneficial effects of the present invention are as follows: The present invention proposes a gradient neural network model for solving the L2 norm optimization problem that can resist noise, which improves the classification accuracy. At the same time, by introducing a robust gradient neural network model, the tolerance to noise and outliers in remote sensing image data is enhanced, thereby improving the stability and reliability of the system in complex environments. During the iterative training process, the robust gradient neural network model uses the backpropagation algorithm to continuously adjust the network parameters to minimize the loss of the objective function. In addition, the model also introduces a regularization term to ensure the stability of the learning process even when there is noise in the training samples. In this way, even in a complex remote sensing data environment, different types of ground objects can be effectively identified and classified, thereby significantly improving the classification accuracy and reducing misclassification phenomena. The present invention can improve the accuracy of remote sensing image classification under noise conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] For better understanding and implementation, the technical solutions of the present application will be described in detail below with reference to the drawings.
[0029] Figure 1 FIG. is a flowchart of the steps of a remote sensing classification method based on a gradient neural network provided by an embodiment of the present application;
[0030] Figure 2 FIG. is a structural diagram of a robust gradient neural network provided by an embodiment of the present application;
[0031] Figure 3 FIG. is a schematic structural diagram of a remote sensing classification system based on a gradient neural network provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0032] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the exemplary embodiments will be described in detail here, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of methods and systems consistent with some aspects of the present application as detailed in the appended claims.
[0033] The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "the", and "said" used in the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to any or all possible combinations of one or more of the associated listed items.
[0034] The following will, in conjunction with the accompanying drawings and preferred embodiments, elaborate in detail on the specific implementation manners, features, and effects of the present invention.
[0035] In the related art, deep learning models have been widely applied due to their excellent non-linear fitting and automatic feature extraction capabilities. Although deep learning models have achieved great success in image classification, there are still some deficiencies. For example, deep learning models require a large amount of labeled data to perform well, which is a time-consuming and labor-intensive task. In addition, deep learning models not only require high-performance hardware but also a large amount of training. These models also need to consider the risks of overfitting, lack of interpretability, and data quality sensitivity.
[0036] In addition to deep learning models, traditional machine learning models, such as support vector machines (SVMs), logistic regression, local binary patterns, etc., have also attracted the attention of researchers.
[0037] The SRC model (Sparse Representation Classifier) was initially designed for face classification and has been successfully applied to remote sensing images. The SRC model involves an L1 minimization problem, which has a large computational amount. Therefore, many classifiers based on L2 norm minimization have been developed, such as the CRC model (Collaborative Representation Classifier). Compared with the SRC model, the CRC model has been proven to be more efficient and has similar sparsity. Similar to the concept of CRC, the NRSC model (Nearest Regularized Subspace Classifier) is a classification method based on sparse representation and regularization techniques, mainly used in fields such as hyperspectral image classification. The NRSC model approximates test samples through non-collaborative methods and non-uniform regularization.
[0038] Both the CRC model and the NRSC model need to estimate the coefficient vector of the representation by solving the L2 norm minimization, and this minimization has a closed-form solution. However, in real-world intelligent systems, many inevitable internal and external noises may reduce the effectiveness of the model, and few classifiers take this adverse factor into account. Therefore, the present invention proposes a gradient neural network model for solving the L2 norm optimization problem that can resist noise, and combines it into a noise-resistant remote sensing image classifier, which can improve the accuracy of remote sensing image classification under noisy conditions.
[0039] Please refer to Figure 1 , an embodiment of the present application provides a remote sensing classification method based on a gradient neural network, including the following steps:
[0040] S100. Obtain multiple training samples generated based on remote sensing images, divide the multiple training samples into multiple categories of training sample groups, and form a training sample matrix with the multiple categories of training sample groups; wherein, each training sample group contains multiple training samples.
[0041] In this embodiment, after obtaining the training samples, construct the training sample matrix X and the training samples X of each category. c Taking the publicly available Pavia University dataset as an example, this dataset is a part of the hyperspectral data of the image of Pavia City, Italy, taken by the airborne Reflective Optics Spectrographic Imaging System (ROSIS-03) in Germany. Retain the image formed by 103 spectral bands.
[0042] In this picture, the pixel points can be specifically divided into 9 categories, including trees, asphalt roads, bricks, pastures, etc. Each pixel point has more than 100 bands. When training, a certain number (n) of training samples are drawn from each category to form the training set X, and the training samples of each category are X. c .
[0043] S200. Based on the CRC model, construct an objective function for minimizing the L2 norm, construct a robust gradient neural network model for solving the objective function, and iteratively train the robust gradient neural network model based on the remote sensing image to be classified and each category of training sample groups to obtain a coefficient vector with minimized L2 norm; wherein, the objective function is to solve the coefficient vector γ of the linear combination of the remote sensing image to be classified and each category of training sample groups, and has the minimum L2 norm.
[0044] In this step, during the iterative training process, the robust gradient neural network model continuously adjusts the network parameters using the backpropagation algorithm to minimize the loss of the objective function. In addition, the model also introduces a regularization term to ensure the stability of the learning process even when there is noise in the training samples. In this way, even in a complex remote sensing data environment, different categories of ground objects can be effectively identified and classified, thus significantly improving the classification accuracy.
[0045] S300. Split the coefficient vector into multiple sub-vectors according to categories, calculate the expression errors between the remote sensing image to be classified and each category of sub-vectors respectively, obtain the expression errors of each category, and use the category with the minimum expression error as the classification result of the remote sensing image.
[0046] Specifically, γ is an N-dimensional vector, where the first to N 1 elements are related to the training samples of the first category, defined as γ 1 , the N 1From +1 to N 1 +N 2 elements are related to the second type of training samples, defined as γ 2 , and so on. Then, for the c-th type of samples, the expression error needs to be calculated. The class with the smallest expression error indicates that y belongs to that class. Finally, the classification is completed.
[0047] In an embodiment provided by the present application, the objective function for constructing the L2 norm minimization based on the CRC model includes:
[0048] Establish a mathematical model for solving the coefficient vector γ based on the CRC model, and the expression is:
[0049]
[0050] where y is the test sample, y ∈ R m , λ CRC is the regularization parameter, γ is the coefficient vector; X is the training sample matrix containing remote sensing data, x i ∈ R m , n is the number of training samples, m is the dimension of the feature space or the number of spectral bands, x i is the i-th training sample; in the training sample matrix X, there are N classes, c represents the class index, N c represents the number of training samples belonging to the c-th class, X c is the c-th class training sample group, is the N c -th training sample in the c-th class;
[0051] Transform the mathematical model into an objective function for L2 norm minimization, and the expression is:
[0052]
[0053] where γ(t) is a time-varying coefficient vector, and F(γ(t)) represents a function with respect to γ(t).
[0054] It should be noted that the objective function min γ F(γ(t)) represents solving the coefficient vector γ(t) that minimizes the L2 norm.
[0055] In an embodiment provided by the present application, the expression of the robust gradient neural network model is:
[0056] where, I is the identity matrix, β is a constant coefficient, e(t) represents the error function, and τ is the time variable.
[0057] Specifically, in the CRC model, the key step is to find a coefficient vector γ that is most suitable for the linear combination of the remote sensing image to be classified and each class of training sample groups, and has the smallest L2 norm. Through optimization, a mathematical model for solving the coefficient vector γ is established;
[0058] By using L2 norm regularization, overfitting of the CRC model can be avoided, and the computational complexity is low. The mathematical model for solving the coefficient vector γ is a convex optimization problem, and its solution is the point where its derivative is 0; therefore, its closed-form solution (theoretical solution) is: γ = (X T X + λ CRC I) -1 X T y;
[0059] where T is the transpose operation. After obtaining the solution of the coefficient vector γ, the difference between the linear representation of the c-th class of training samples and the test sample can be calculated, and this difference is expressed as γ c is the vector in the coefficient vector γ related to the c-th class of samples, and the test sample y is classified into the c-th class that obtains the minimum value .
[0060] It should be noted that classifiers based on L2 norm minimization are widely popular due to their closed-form solutions with low computational complexity. However, most of them perform classification under noise-free conditions, while in computing systems, noise is inevitable. There are many sources of noise, such as hardware implementation, environmental noise during operation, errors in data storage and signal transmission, etc. The inevitable noise will reduce the accuracy of the computational solution and even lead to the failure of the task. Therefore, it is necessary to design a robust algorithm to estimate the solution under noisy conditions. Iterative algorithms are the mainstream methods for solving mathematical problems and minimization problems, such as Newton iterative algorithms, GNN models, discrete-time reset neural networks, etc. These methods adjust the update step size of the solution according to different rules and can effectively handle the noise in the solution process. Based on this, in order to reduce or even eliminate the influence of noise, the present invention proposes a robust gradient neural network model (RGNN) model.
[0061] For the iterative algorithm, the mathematical model for solving the coefficient vector γ can be rewritten as an objective function for L2 norm minimization; the solution of this mathematical model satisfies
[0062] Define the error function as:
[0063]
[0064] Define a non-negative scalar a non-negative value
[0065] According to the concept of the GNN model, the traditional GNN model updates the solution along the negative gradient direction, which can be expressed as: Where I is the identity matrix; ideally, the error function e(t) should be equal to zero; however, due to the influence of noise, e(t) ≠ 0; that is to say, e(t) can reflect the influence of noise. In order to monitor and suppress the influence of noise, the update step size of the GNN model is adjusted using the historical residuals to obtain the expression of the robust gradient neural network model (RGNN model).
[0066] In an embodiment provided by the present application, iteratively training the robust gradient neural network model based on the remote sensing image to be classified and each class of training sample groups to obtain a coefficient vector with minimized L2 norm includes:
[0067] Inputting the training sample matrix, the remote sensing image to be classified, the regularization parameter, the constant coefficient, and the initial coefficient vector into the robust gradient neural network model, and solving to obtain
[0068] Integrating through an integrator to obtain γ(t), updating the coefficient vector to γ(t), and then iteratively training the robust gradient neural network model; after reaching the condition for stopping iterative training, outputting the coefficient vector with minimized L2 norm.
[0069] Specifically, the structure of the robust gradient neural network is as Figure 2 shown. First, it is necessary to input the training sample matrix X, the remote sensing image y to be classified, the regularization parameter λ CRC , the constant coefficient β, and the initial coefficient vector γ(0). Then, calculate B T , e(t) respectively; afterwards, e(t) passes through an integrator to obtain Adding e(t) and through an adder to obtain Putting and B T and -β into a multiplier to construct Finally passes through an integrator to obtain γ(t). During the execution of the algorithm, γ(t) is continuously updated, and γ(t) is used as the input to continuously update and train the entire network. Finally, when the running time arrives, γ(t) is input.
[0070] Originally, γ can be directly calculated according to the objective function, but because the computing system is easily affected by noise, the obtained γ is inaccurate. Therefore, the present invention randomly gives an initial coefficient vector γ(0) and uses the RGNN model to robustly solve the coefficient vector γ(t).
[0071] In an embodiment provided by the present application, the expression of the expression error is where γ c is the c-th class subvector, and X c is the c-th class training sample.
[0072] Please refer to Figure 3 , an embodiment of the present application provides a remote sensing classification system based on a gradient neural network, including:
[0073] At least one processor;
[0074] At least one memory for storing at least one program;
[0075] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.
[0076] It can be seen that the content in the above method embodiments is applicable to the embodiments of this system. The functions specifically implemented by the embodiments of this system are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above method embodiments.
[0077] In the above embodiments, the descriptions of the respective embodiments have their own focuses. For parts not detailed or recorded in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.
[0078] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above division of each functional unit and module is used as an example. In practical applications, the above functions can be allocated to different functional units and modules as needed, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here.
[0079] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.
[0080] As described above, the above are only preferred embodiments of the present invention, and there is no limitation to the present invention in any form. Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the above-disclosed technical content within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention still fall within the scope of the technical solution of the present invention.
Claims
1. A remote sensing classification method based on a gradient neural network, characterized in that: The steps include: Acquire a plurality of training samples generated based on remote sensing images, divide the plurality of training samples into a plurality of training sample groups, and form a training sample matrix with the plurality of training sample groups; wherein the training sample group includes a plurality of training samples; An objective function of minimizing the L2 norm is constructed based on the CRC model, a robust gradient neural network model for solving the objective function is constructed, and the robust gradient neural network model is iteratively trained based on the remote sensing image to be classified and each type of training sample group to obtain a coefficient vector that minimizes the L2 norm; wherein the objective function is a coefficient vector for solving a linear combination of the remote sensing image to be classified and each type of training sample group, and has a minimum L2 norm; The coefficient vector is split into multiple sub-vectors according to categories, and the expression errors between the remote sensing image to be classified and each sub-vector are calculated respectively to obtain the expression error of each category, and the category with the smallest expression error is taken as the classification result of the remote sensing image.
2. The method according to claim 1, characterized in that The objective function of minimizing the L2 norm based on the CRC model is constructed, including: Based on the CRC model, a mathematical model for solving the coefficient vector γ is established, and the expression is: Among them, y is the test sample, y∈R m ,λ CRC is the regularization parameter, γ is the coefficient vector; X is the training sample matrix containing remote sensing data, x i ∈R m , n is the number of training samples, m is the dimension of the feature space or the number of spectral bands, x i is the i-th training sample; in the training sample matrix X, there are N classes, c represents the index of the class, N c represents the number of training samples belonging to the cth class, X c is the c-th training sample group, is the Nth in the cth category c training samples; The mathematical model is transformed into an objective function that minimizes the L2 norm, and the expression is: Among them, γ(t) is a time-varying coefficient vector, and F(γ(t)) represents a function about γ(t).
3. The method according to claim 2, characterized in that The expression of the robust gradient neural network model is: in, I is the identity matrix, β is the constant coefficient, e(t) represents the error function, and τ is the time variable.
4. The method according to claim 3, characterized in that The expression of the error function is:
5. The method according to claim 4, characterized in that The iterative training of the robust gradient neural network model based on the remote sensing image to be classified and each type of training sample group to obtain a coefficient vector with a minimized L2 norm includes: The training sample matrix, the remote sensing image to be classified, the regularization parameter, the constant coefficient and the initial coefficient vector are input into the robust gradient neural network model to obtain Through the integrator Integrate to obtain γ(t), update the coefficient vector to γ(t), and iteratively train the robust gradient neural network model; after the condition for stopping the iterative training is met, output the coefficient vector with the minimized L2 norm.
6. The method according to claim 5, characterized in that: The expression of the expression error is: Among them, γ c is the c-th subvector, X c is the c-th type of training sample.
7. A remote sensing classification system based on a gradient neural network, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 6.