An Unsupervised Feature Selection Method and System Based on Multi-Source Remote Sensing Data

Abstract

The present invention discloses an unsupervised feature selection method and system based on multi-source remote sensing data. The method comprises: integrating the feature matrix of the acquired multi-source remote sensing data to obtain a data matrix in a preset format; extracting features from the data matrix using a principal component analysis method to obtain a principal component matrix; training a target deep prototype analysis network based on the principal component analysis matrix to obtain a coefficient matrix of the target data prototype; decomposing the coefficient matrix to obtain a target weight matrix; and screening key original features based on the target weight matrix. The present invention uses the principal component matrix obtained by the principal component analysis method as auxiliary information of the target deep prototype network, and then trains the target deep prototype analysis network to obtain a coefficient matrix of the target data prototype. By decomposing the coefficient matrix, a weight matrix can be obtained and key original features with the highest correlation with the prototype individual can be screened out. The method can be widely applied to the field of remote sensing image processing technology.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to an unsupervised feature selection method and system based on multi-source remote sensing data. Background Art

[0002] In recent years, multispectral sensors and other image sensors have been miniaturized and increasingly integrated. Furthermore, the payload and stability of the drone platforms that carry these sensors have significantly improved. These changes have made the combined, synchronized acquisition of data from multiple sensors an increasingly mainstream approach to remote sensing. While the combination of multimodal sensor data offers richer features, it also creates redundancy between these features. When applying this remote sensing image data for deeper intelligent analysis, the challenge of appropriately reducing its dimensionality has become a pressing issue.

[0003] Currently, the most popular dimensionality reduction method is principal component analysis (PCA). Its principle is to construct linear combinations of all features into principal components. The weights in these linear combinations are then selected to maximize the variance of the principal components and to ensure that they are orthogonal to all previous principal components. This significantly reduces the feature dimensionality of the data while capturing most of the variation present. However, the new features generated by PCA are not practically meaningful, and in the agricultural field, the focus is often on the feature information itself. Therefore, how to implement feature selection based on multi-source remote sensing data is an urgent problem that needs to be solved. Summary of the Invention

[0004] In view of this, an embodiment of the present invention provides an unsupervised feature selection method and system based on multi-source remote sensing data, which can screen out valuable key features.

[0005] In one aspect, an embodiment of the present invention provides an unsupervised feature selection method based on multi-source remote sensing data, comprising:

[0006] Integrate the feature matrix of multi-source remote sensing data to obtain a data matrix in a preset format;

[0007] Performing feature extraction on the data matrix by a principal component analysis method to obtain a principal component matrix;

[0008] According to the principal component matrix, a target deep prototype analysis network is trained to obtain a coefficient matrix of the target data prototype;

[0009] Decomposing the coefficient matrix to obtain a target weight matrix;

[0010] Based on the target weight matrix, key original features are screened and obtained.

[0011] Optionally, the step of integrating the feature matrices of the acquired multi-source remote sensing data to obtain a data matrix in a preset format includes:

[0012] Obtain all feature matrices of multi-source remote sensing data and determine the number of feature matrices and samples;

[0013] Determine the number of rows of the data matrix according to the number of feature matrices, and determine the number of columns of the data matrix according to the number of samples;

[0014] The characteristic matrix is integrated based on the number of rows of the data matrix and the number of columns of the data matrix to obtain a data matrix.

[0015] Optionally, extracting features from the data matrix using a principal component analysis method to obtain a principal component matrix includes:

[0016] Based on the target loss function of the principal component analysis method, the principal component variance is determined;

[0017] Based on the principal component variance, feature extraction is performed on the data matrix using a principal component analysis method to obtain a principal component matrix in which the cumulative contribution rate of the principal component variance reaches 90%.

[0018] Optionally, it also includes:

[0019] Based on the information bottleneck method and variational autoencoder method, a deep prototype analysis network is constructed;

[0020] Based on the deep prototype analysis network, a second deep prototype analysis network is obtained by combining a decoder branch;

[0021] Based on the second deep prototype analysis network and combined with the prototype loss, a target deep prototype analysis network is obtained.

[0022] Optionally, the expression of the prototype analysis network is:

[0023] min p(t|x) I(X;T)-λI(T;Y)

[0024] In the formula, p(t|x) represents the probability of t occurring under the condition that x occurs, min p(t|x) Indicates minimizing the value of the formula by controlling p(t|x), I represents mutual information, X represents the data matrix, T represents the latent variable, and λ represents the first Lagrange multiplier;

[0025] The parameterized form of the expression of the prototype analysis network is:

[0026]

[0027] Where Φ = p(t|x), θ = p(y|t), p(y|t) represents the probability of y occurring under the condition that t occurs. Indicates maximizing the value of the formula by controlling p(t|x) and p(y|t), t represents the t-th data point of the latent variable T, x represents the x-th data point of the data matrix X, y represents the y-th data point of the principal component matrix Y, I Φ (t; x) represents the encoder, I Φ，θ (t;y) represents the decoder.

[0028] Optionally, the obtaining of a second deep prototype analysis network based on the deep prototype analysis network and in combination with a decoder branch includes:

[0029] Adding a decoder branch and a second Lagrangian multiplier associated with the decoder branch to the deep prototype analysis network to obtain a second deep prototype analysis network;

[0030] Among them, the expression of the second deep prototype analysis network is:

[0031]

[0032] Where v represents the second Lagrange multiplier, Indicates that under the condition that t occurs The probability of occurrence, Represents the new data matrix obtained after the data matrix X is reconstructed The data points, Represents the decoder branch.

[0033] Optionally, obtaining a target deep prototype analysis network based on the second deep prototype analysis network and in combination with prototype loss includes:

[0034] Obtaining a nonlinear transformation of the data matrix by learning a first weight matrix and a second weight matrix;

[0035] Based on the nonlinear transformation of the data matrix, the structure of the latent space is driven to form a prototype, and the prototype is positioned to obtain an initialized prototype matrix;

[0036] Obtaining a prototype loss based on the first weight matrix, the second weight matrix, and the initialized prototype matrix;

[0037] Obtain a target deep prototype analysis network by combining the second deep prototype analysis network with the prototype loss;

[0038] Among them, the expression of the target deep prototype analysis network is:

[0039]

[0040] Where, l AT represents the prototype loss.

[0041] Optionally, decomposing the coefficient matrix to obtain a target weight matrix includes:

[0042] Based on the coefficient matrix, a target weight matrix for the connection between the data matrix and the coefficient matrix is obtained through matrix decomposition calculation combined with linear optimization.

[0043] Optionally, the screening to obtain key original features based on the target weight matrix includes:

[0044] Based on the target weight matrix, key original features are obtained by screening and calculating key original features;

[0045] The expression for the key original feature screening calculation is:

[0046] E g ={x i |C g (i)==max(C g )}

[0047] Where, E g represents the set of key original features, x i Represents the i-th column data in the data matrix X, C g Represents the g-th column data in the target weight matrix C, C g (i) represents the value of the i-th column and row of the g-th column of the target weight matrix C.

[0048] On the other hand, an embodiment of the present invention provides an unsupervised feature selection system based on multi-source remote sensing data, comprising:

[0049] The first module is used to integrate the feature matrix of the acquired multi-source remote sensing data to obtain a data matrix in a preset format;

[0050] The second module is used to extract features from the data matrix by using a principal component analysis method to obtain a principal component matrix;

[0051] The third module is used to train the target deep prototype analysis network according to the principal component matrix to obtain the coefficient matrix of the target data prototype;

[0052] The fourth module is used to decompose the coefficient matrix to obtain a target weight matrix;

[0053] The fifth module is used to screen and obtain key original features based on the target weight matrix.

[0054] On the other hand, an embodiment of the present invention provides an unsupervised feature selection system based on multi-source remote sensing data, comprising a processor and a memory;

[0055] The memory is used to store programs;

[0056] The processor executes the program to implement the method described above.

[0057] On the other hand, an embodiment of the present invention provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement the method described above.

[0058] The embodiments of the present invention further disclose a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the above method.

[0059] The embodiment of the present invention first integrates the feature matrix of the acquired multi-source remote sensing data to obtain a data matrix in a preset format; extracts features from the data matrix using a principal component analysis method to obtain a principal component matrix; trains a target deep prototype analysis network based on the principal component matrix to obtain a coefficient matrix of the target data prototype; decomposes the coefficient matrix to obtain a target weight matrix; and screens and obtains key original features based on the target weight matrix. The present invention uses the principal component matrix obtained by the principal component analysis method as auxiliary information for the target deep prototype network, and then trains the target deep prototype analysis network to obtain a coefficient matrix of the target data prototype. By decomposing the coefficient matrix, a weight matrix can be obtained and the key original features with the highest correlation with the prototype individual can be screened out. While removing redundant features to achieve a dimensionality reduction effect, the present invention allows the obtained features to retain the physical meaning of the original features. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0061] Figure 1 A flowchart of an unsupervised feature selection method based on multi-source remote sensing data provided by an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of a target deep prototype analysis network provided by an embodiment of the present invention;

[0063] Figure 3 A fitting graph for predicting sugarcane yield provided by an embodiment of the present invention;

[0064] Figure 4 This is a fitting diagram of the sugarcane yield prediction effect obtained using existing technology. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] On the one hand, referring to Figure 1 , an embodiment of the present invention provides an unsupervised feature selection method based on multi-source remote sensing data, comprising:

[0067] S100, integrating the feature matrix of the acquired multi-source remote sensing data to obtain a data matrix in a preset format;

[0068] It should be noted that in some embodiments, all feature matrices of multi-source remote sensing data are obtained to determine the number of feature matrices and the number of samples; the number of data matrix rows is determined based on the number of feature matrices, and the number of data matrix columns is determined based on the number of samples; based on the number of data matrix rows and the number of data matrix columns, the feature matrices are integrated to obtain the data matrix.

[0069] Specifically, all n feature matrices of multi-source remote sensing data of crops are obtained from sensors and integrated to construct a data matrix X with m rows as samples (that is, row m as pixels) and n columns as features. m×n .

[0070] S200, extracting features from the data matrix using a principal component analysis method to obtain a principal component matrix;

[0071] It should be noted that in some embodiments, the principal component variance is determined based on the target loss function of the principal component analysis method; based on the principal component variance, the data matrix is feature extracted by the principal component analysis method to obtain a principal component matrix with a cumulative contribution rate of 90% of the principal component variance.

[0072] Specifically, the principal component analysis method is used to analyze the data matrix X m×n Perform feature extraction to obtain the principal component matrix Y that satisfies the cumulative contribution rate of principal component variance to reach 90%. The objective loss function of the principal component analysis method is as follows:

[0073]

[0074] Among them, V(x) represents the variance of the principal component and all data, represents the number of samples of the data matrix X, n represents the number of features of the data matrix X, i represents the i-th quantization subinterval, j represents the j-th quantization subinterval, Represents the distance between the sample point projection and the principal axis of the principal component. By maximizing the variance V(x), we can obtain the corresponding multiple principal components and merge them into the principal component matrix Y.

[0075] S300, training a target deep prototype analysis network according to the principal component matrix to obtain a coefficient matrix of the target data prototype;

[0076] It should be noted that some embodiments also include: constructing a deep prototype analysis network based on the information bottleneck method and the variational autoencoder method; based on the deep prototype analysis network, combining the decoder branch to obtain a second deep prototype analysis network; based on the second deep prototype analysis network, combining the prototype loss to obtain a target deep prototype analysis network.

[0077] In some embodiments, the expression of the prototype analysis network is:

[0078] min p(t|x) I(X;T)-λI(T;Y)

[0079] In the formula, p(t|x) represents the probability of t occurring under the condition that x occurs, min p(t|x) Indicates minimizing the value of the formula by controlling p(t|x), I represents mutual information, X represents the data matrix, T represents the latent variable, and λ represents the first Lagrange multiplier; the parameterized form of the expression of the prototype analysis network is:

[0080]

[0081] Where Φ = p(t|x), θ = p(y|t), p(y|t) represents the probability of y occurring under the condition that t occurs. Indicates maximizing the value of the formula by controlling p(t|x) and p(y|t), t represents the t-th data point of the latent variable T, x represents the x-th data point of the data matrix X, y represents the y-th data point of the principal component matrix Y, I Φ (t; x) represents the encoder, I Φ，θ (t;y) represents the decoder.

[0082] In some embodiments, a second deep prototype analysis network is obtained by combining a decoder branch based on the deep prototype analysis network, including: adding a decoder branch and a second Lagrangian multiplier associated with the decoder branch to the deep prototype analysis network to obtain the second deep prototype analysis network; wherein the expression of the second deep prototype analysis network is:

[0083]

[0084] Where v represents the second Lagrange multiplier, Indicates that under the condition that t occurs The probability of occurrence, Represents the new data matrix obtained after the data matrix X is reconstructed The data points, Represents the decoder branch.

[0085] In some embodiments, based on the second deep prototype analysis network, a target deep prototype analysis network is obtained in combination with the prototype loss, including:

[0086] By learning the first weight matrix and the second weight matrix, a nonlinear transformation of the data matrix is obtained;

[0087] Based on the nonlinear transformation of the data matrix, the structure of the latent space is driven to form a prototype, and the prototype is positioned to obtain the initialized prototype matrix;

[0088] Obtaining a prototype loss based on the first weight matrix, the second weight matrix, and the initialized prototype matrix;

[0089] The target deep prototype analysis network is obtained by combining the second deep prototype analysis network with the prototype loss;

[0090] Among them, the expression of the target deep prototype analysis network is:

[0091]

[0092] Where, l AT represents the prototype loss.

[0093] Specifically, by using the principal component as auxiliary information Y to guide the deep prototype analysis network for training, the target data prototype closest to the principal component and its coefficient matrix are obtained.

[0094] like Figure 2 As shown in Figure 1, the Deep Prototype Analysis Network combines the Information Bottleneck (IB) and Variational Autoencoder (VAE) methods. The goal of the IB method is to find a latent variable T that compresses the data X (i.e., the data matrix Xm×n) and performs a nonlinear transformation on X to obtain T in the high-dimensional space while preserving as much information as possible from the second auxiliary information matrix Y. Its objective function is as follows:

[0095] min p(t|x) I(X;T)-λI(T;Y)

[0096] In the formula, p(t|x) represents the probability of t occurring under the condition that x occurs, min p(t|x)represents the value of the formula minimized by controlling p(t|x), I represents the mutual information, X represents the data matrix, T represents the latent variable, and λ represents the first Lagrange multiplier. The parameterized form of the IB Markov chain TXY and the above formula can be expressed as:

[0097]

[0098] Where Φ = p(t|x), θ = p(y|t), p(y|t) represents the probability of y occurring under the condition that t occurs. Indicates maximizing the value of the formula by controlling p(t|x) and p(y|t), t represents the t-th data point of the latent variable T, x represents the x-th data point of the data matrix X, y represents the y-th data point of the principal component matrix Y, I Φ (t; x) represents the encoder, I Φ，θ (t; y) represents the decoder. The two terms in the above formula can be regarded as the encoder and decoder respectively. The specific formula is:

[0099]

[0100] Where D KL (*) represents relative entropy (i.e. KL divergence), p(*) represents the probability distribution of random variables, p Φ (t|x) represents the probability of t occurring under the condition that x occurs, p Φ (t|x,y) represents the probability of t occurring under the condition that the joint events x and y occur, p θ (y|t) represents the probability of y occurring under the condition that t occurs.

[0101] An additional decoder branch is then used to predict and reconstruct This extension requires adding an item and an additional Lagrange multiplier v. The specific objective function formula is as follows:

[0102]

[0103] Where, Represents the new data matrix obtained after the data matrix X is reconstructed The data points, Represents the mutual information estimate. Calculation method and formula I Φ，θ (t;y) are the same.

[0104] By learning the first weight matrix A∈R m×k and the second weight matrix B∈R m×k, and using the nonlinear transformation T of the data X, it is possible to drive the structure of the latent space to form a prototype. The formula for nonlinear transformation of the data matrix is as follows:

[0105]

[0106] Where, t i represents the specific value of the latent variable T, x represents the specific value of the data matrix X, N represents the Gaussian distribution, Z represents the prototype matrix, a i Represents the specific parameters of the weight matrix A, μ i Indicates a i The average value of Indicates a i The variance of , I represents the mutual information. i The average value μ is given i and variance It is the nonlinear transformation of the specific data point x of the data matrix X learned by the encoder.

[0107] Because there is no absolute reference frame in the latent space, it is necessary to locate k prototypes at a simple coordinate point (k-1) and collect these fixed coordinates into the initial prototype matrix Z fixed This leads to an additional distance-dependent prototype loss, namely:

[0108] l AT =||Z fixed -BAZ fixed || 2 =||Z fixed -Z pred || 2

[0109] Where Z pred =BAZ fixed is the predicted prototype position given the learned weight matrices A and B. In order to make the loss function l AT To minimize, the objective function of the deep prototype analysis network can be refined as:

[0110]

[0111] By continuously learning the weight matrices A and B and finding the best result that satisfies the objective function, we can obtain the data prototype Z that is closest to the principal component. fixed and its weight matrix A (i.e. coefficient matrix).

[0112] S400, decomposing the coefficient matrix to obtain a target weight matrix;

[0113] It should be noted that, in some embodiments, based on the coefficient matrix, a target weight matrix that relates the data matrix to the coefficient matrix is obtained through matrix decomposition calculation combined with linear optimization.

[0114] Specifically, for the coefficient matrix A∈R m×k Decompose by the goal of A=XC, where X represents the data matrix. Initialize the weight matrix C n×k (ie, the target weight matrix), and let the coefficient matrix A m×k ≈X m×n C n×k Solve for C n× The k problem can be transformed into the following linear optimization problem:

[0115]

[0116] In the formula, C represents the target weight matrix, A represents the coefficient matrix, X represents the data matrix, m represents the number of samples, l represents the lth quantization subinterval, and a l represents the lth specific value of the coefficient matrix A, x l represents the lth specific value of the data matrix X, k represents the number of prototypes, i represents the i-th quantization subinterval, c j,i Represents the value of the j-th column and the i-th row in the C matrix.

[0117] By matrix decomposition calculation, we can get the matrix X that can reveal the original data m×n and the coefficient matrix A characterized by the prototype m ×k The weight matrix C of the connection between n×k .

[0118] S500: Based on the target weight matrix, key original features are screened and obtained.

[0119] It should be noted that in some embodiments, the key original features are obtained by screening and calculating the key original features based on the target weight matrix;

[0120] Among them, the expression for key original feature screening calculation is:

[0121] E g ={x i |C g (i)==max(C g )}

[0122] Where, E g represents the set of key original features, x i Represents the i-th column data in the data matrix X, C h Represents the g-th column data in the target weight matrix C, C g(i) represents the value of the i-th column and row of the g-th column of the target weight matrix C.

[0123] Specifically, from the weight matrix C∈R n×k In the above example, we can know which original features each of the k prototypes is mainly composed of, and finally select the key original features that have a greater contribution to the formation of the prototype. The formula is as follows

[0124] E g ={x i |C g (i)==max(C g )}

[0125] From the above formula, we can get the final set of key original features, namely E g .

[0126] In a specific embodiment, multi-source remote sensing data of sugarcane is used to process the data using different feature processing methods, and then applied to the yield prediction model for comparison. Figure 3 As shown in FIG, it is the yield prediction model performance obtained for sugarcane by the present invention. Figure 4 As shown in FIG, it is a fitting effect diagram of the yield prediction model performance obtained by the traditional method. Thus, according to the comparison, it can be seen that the yield prediction model obtained by the feature selection method proposed in the present invention has a better effect.

[0127] In summary, the present invention uses the data information obtained by the principal component analysis method as the auxiliary information input of the deep prototype network, which can guide the deep prototype network to find the representative individual of the input data closest to the principal component by learning appropriate matrix parameters in the latent space of the input data, and at the same time feed back the coefficient matrix result of the prototype, and then by decomposing the coefficient matrix, the weight matrix can be obtained and the key original features with the highest correlation with the prototype individual can be screened out. This is the processing result of the unsupervised feature selection model based on multi-source remote sensing data of crops proposed by the present invention. Compared with the existing technology, the beneficial effects and advantages of the present invention are: while removing redundant features to achieve the dimensionality reduction effect, the new features obtained retain the physical meaning of the original features, avoiding the problem that the features obtained by the traditional PCA dimensionality reduction method have no practical significance, which is of great help to the subsequent pursuit of interpretability in machine learning work.

[0128] On the other hand, an embodiment of the present invention provides an unsupervised feature selection system based on multi-source remote sensing data, including: a first module for integrating the feature matrix of the acquired multi-source remote sensing data to obtain a data matrix in a preset format; a second module for extracting features from the data matrix through a principal component analysis method to obtain a principal component matrix; a third module for training a target deep prototype analysis network based on the principal component matrix to obtain a coefficient matrix of the target data prototype; a fourth module for decomposing the coefficient matrix to obtain a target weight matrix; and a fifth module for screening key original features based on the target weight matrix.

[0129] The contents of the method embodiments of the present invention are all applicable to the system embodiments. The functions specifically implemented by the system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0130] Another aspect of the embodiment of the present invention further provides an unsupervised feature selection system based on multi-source remote sensing data, comprising a processor and a memory;

[0131] The memory is used to store programs;

[0132] The processor executes the program to implement the method described above.

[0133] The contents of the method embodiments of the present invention are all applicable to the system embodiments. The functions specifically implemented by the system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0134] Another aspect of the embodiments of the present invention further provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement the method described above.

[0135] The contents of the method embodiments of the present invention are all applicable to the computer-readable storage medium embodiments. The functions specifically implemented by the computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0136] The embodiments of the present invention further disclose a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the above method.

[0137] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0138] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0139] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0140] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution apparatus, device, or apparatus (e.g., a computer-based apparatus, a device including a processor, or other apparatus that can fetch instructions from and execute instructions on an instruction execution apparatus, device, or apparatus). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution apparatus, device, or apparatus.

[0141] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0142] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution device. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0143] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0144] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0145] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. An unsupervised feature selection method based on multi-source remote sensing data, characterized in that: include: Integrate the feature matrix of multi-source remote sensing data to obtain a data matrix in a preset format; Performing feature extraction on the data matrix by a principal component analysis method to obtain a principal component matrix; According to the principal component matrix, a target deep prototype analysis network is trained to obtain a coefficient matrix of the target data prototype; Decomposing the coefficient matrix to obtain a target weight matrix; Based on the target weight matrix, key original features are screened and obtained.

2. The unsupervised feature selection method based on multi-source remote sensing data according to claim 1, characterized in that: The step of integrating the feature matrices of the acquired multi-source remote sensing data to obtain a data matrix in a preset format includes: Obtain all feature matrices of multi-source remote sensing data and determine the number of feature matrices and samples; Determine the number of rows of the data matrix according to the number of feature matrices, and determine the number of columns of the data matrix according to the number of samples; The characteristic matrix is integrated based on the number of rows of the data matrix and the number of columns of the data matrix to obtain a data matrix.

3. The unsupervised feature selection method based on multi-source remote sensing data according to claim 1, characterized in that: The method of extracting features from the data matrix by a principal component analysis method to obtain a principal component matrix includes: Based on the target loss function of the principal component analysis method, the principal component variance is determined; Based on the principal component variance, feature extraction is performed on the data matrix using a principal component analysis method to obtain a principal component matrix in which the cumulative contribution rate of the principal component variance reaches 90%.

4. The unsupervised feature selection method based on multi-source remote sensing data according to claim 1, characterized in that: Also includes: Based on the information bottleneck method and variational autoencoder method, a deep prototype analysis network is constructed; Based on the deep prototype analysis network, a second deep prototype analysis network is obtained by combining a decoder branch; Based on the second deep prototype analysis network and combined with the prototype loss, a target deep prototype analysis network is obtained.

5. The unsupervised feature selection method based on multi-source remote sensing data according to claim 4, characterized in that: The expression of the prototype analysis network is: min p(t|x) I(X;T)-λI(T;Y) In the formula, p(t|x) represents the probability of t occurring under the condition that x occurs, min p(t|x) Indicates minimizing the value of the formula by controlling p(t|x), I represents mutual information, X represents the data matrix, T represents the latent variable, and λ represents the first Lagrange multiplier; The parameterized form of the expression of the prototype analysis network is: Where Φ = p(t|x), θ = p(y|t), p(y|t) represents the probability of y occurring under the condition that t occurs. Indicates maximizing the value of the formula by controlling p(t|x) and p(y|t), t represents the t-th data point of the latent variable T, x represents the x-th data point of the data matrix X, y represents the y-th data point of the principal component matrix Y, I Φ (t; x) represents the encoder, I Φ，θ (t;y) represents the decoder.

6. The unsupervised feature selection method based on multi-source remote sensing data according to claim 5, characterized in that: The second deep prototype analysis network is obtained by combining the decoder branch based on the deep prototype analysis network, including: Adding a decoder branch and a second Lagrangian multiplier associated with the decoder branch to the deep prototype analysis network to obtain a second deep prototype analysis network; Among them, the expression of the second deep prototype analysis network is: Where v represents the second Lagrange multiplier, Indicates that under the condition that t occurs The probability of occurrence, Represents the new data matrix obtained after the data matrix X is reconstructed The data points, Represents the decoder branch.

7. The unsupervised feature selection method based on multi-source remote sensing data according to claim 6, characterized in that: The step of obtaining a target deep prototype analysis network based on the second deep prototype analysis network and combining the prototype loss includes: Obtaining a nonlinear transformation of the data matrix by learning a first weight matrix and a second weight matrix; Based on the nonlinear transformation of the data matrix, the structure of the latent space is driven to form a prototype, and the prototype is positioned to obtain an initialized prototype matrix; Obtaining a prototype loss based on the first weight matrix, the second weight matrix, and the initialized prototype matrix; Obtain a target deep prototype analysis network by combining the second deep prototype analysis network with the prototype loss; Among them, the expression of the target deep prototype analysis network is: Where, l AT represents the prototype loss.

8. The unsupervised feature selection method based on multi-source remote sensing data according to claim 1, characterized in that: Decomposing the coefficient matrix to obtain a target weight matrix includes: Based on the coefficient matrix, a target weight matrix for the connection between the data matrix and the coefficient matrix is obtained through matrix decomposition calculation combined with linear optimization.

9. The unsupervised feature selection method based on multi-source remote sensing data according to claim 1, characterized in that: The key original features obtained by screening based on the target weight matrix include: Based on the target weight matrix, key original features are obtained by screening and calculating key original features; The expression for the key original feature screening calculation is: E g ={x i |C g (i)==max(C g )} Where, E g represents the set of key original features, x i Represents the i-th column data in the data matrix X, C g Represents the g-th column data in the target weight matrix C, C g (i) represents the value of the i-th column and row of the g-th column of the target weight matrix C.

10. An unsupervised feature selection system based on multi-source remote sensing data, characterized in that: include: The first module is used to integrate the feature matrix of the acquired multi-source remote sensing data to obtain a data matrix in a preset format; The second module is used to extract features from the data matrix by using a principal component analysis method to obtain a principal component matrix; The third module is used to train the target deep prototype analysis network according to the principal component matrix to obtain the coefficient matrix of the target data prototype; The fourth module is used to decompose the coefficient matrix to obtain a target weight matrix; The fifth module is used to screen and obtain key original features based on the target weight matrix.

Images