Remote sensing image classification method and device based on unsupervised public subspace learning model

Abstract

The application discloses a kind of remote sensing image classification method and device based on unsupervised common subspace learning model.Each pixel in the remote sensing image to be classified is used as input data set;Using the unsupervised common subspace learning model solved in advance, the input data set is mapped to the common subspace, and the projection feature set to be classified of the input data set in the common subspace is determined;Through the pre-set classifier, the corresponding same mode projection feature solution of the remote sensing image to be classified in the common subspace and the multiple class labels corresponding to the projection feature solution are used to classify the projection feature set to be classified, and the classification result of each pixel is determined;According to the classification result of each pixel, the classification image of the remote sensing image to be classified is determined.The technical problem that classification is carried out depending on label information, resulting in poor classification effect is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image classification, and particularly relates to a remote sensing image classification method and device based on an unsupervised common subspace learning model. BACKGROUND

[0002] The purpose of remote sensing image classification is to assign meaningful attributes to each pixel of a remote sensing image according to the distribution of ground objects in the remote sensing image. Research on this problem has important practical significance and a wide range of application scenarios, such as mineral exploration, environmental monitoring, urban mapping, disaster detection, etc.

[0003] At present, most of the existing multi-modal or cross-modal feature learning methods, whether from the perspective of supervised modeling or semi-supervised modeling, are indispensable to rely on label information in the design. However, in the remote sensing image classification task, the pixel-level label itself not only needs experts to conduct on-site investigation and laboratory annotation, but also is easily damaged by noise caused by complex environmental factors. In addition, when the classifier uses the label represented by one-hot encoding to train the model, although the correct class can be used, the relationship information is easily missed. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a remote sensing image classification method and device based on an unsupervised common subspace learning model.

[0005] According to one aspect of the present application, there is provided a remote sensing image classification method based on an unsupervised common subspace learning model, comprising:

[0006] Each pixel in the remote sensing image to be classified is taken as an input data set, wherein the remote sensing image to be classified includes any one of the following modal remote sensing images: hyperspectral image, laser radar point cloud data, multispectral image and synthetic aperture radar image;

[0007] The input data set is mapped to a common subspace using a pre-solved unsupervised common subspace learning model, and a classification projection feature set of the input data set in the common subspace is determined;

[0008] The classification projection feature set is classified by a pre-set classifier using the corresponding same modal projection feature solution of the remote sensing image to be classified in the common subspace and the multiple class labels corresponding to the projection feature solution, and the classification result of each pixel is determined;

[0009] According to the classification result of each pixel, a classification image of the remote sensing image to be classified is determined.

[0010] Optionally, the input in the unsupervised common subspace learning model solving process is feature reorganization of the input data set of the multi-modal remote sensing image training set, the training set includes multiple modal data, and the multi-modal remote sensing image training set is obtained by

[0011] The unsupervised common subspace learning model is to regularize the feature reorganization data and its projection in the common subspace, and minimize the result and a preset regularization term.

[0012] Optionally, the process of solving the unsupervised common subspace learning model by the multi-modal remote sensing image is as follows:

[0013] Fixing the projection feature in the formula of the unsupervised common subspace learning model, a target function about the joint transformation matrix is determined;

[0014] The first-order derivative of the target function about the joint transformation matrix is calculated, and the joint transformation matrix optimal solution of the joint transformation matrix is determined by calculating the first-order derivative to be zero;

[0015] The joint transformation matrix optimal solution is brought into the formula of the unsupervised common subspace learning model, and a target function about the projection feature is determined;

[0016] The projection feature solution of the projection feature is determined by performing spectral decomposition on the target function of the projection feature;

[0017] According to the projection feature solution, the joint transformation matrix optimal solution of the joint transformation matrix is calculated.

[0018] Optionally, the input data set is mapped to the common subspace by using the unsupervised common subspace learning model solved in advance, and a set of projection features to be classified of the input data set in the common subspace is determined, including:

[0019] According to the target function about the joint transformation matrix, a mapping relationship formula of the projection feature and the joint transformation matrix is determined;

[0020] According to the corresponding same-modal joint transformation matrix optimal solution of the remote sensing image to be classified in the common subspace and the mapping relationship formula, the input data set is mapped into the set of projection features to be classified.

[0021] According to another aspect of the present application, a remote sensing image classification device based on an unsupervised common subspace learning model is provided, comprising:

[0022] As a module, each pixel in the remote sensing image to be classified is used as an input data set, wherein the remote sensing image to be classified includes any one of the following modal remote sensing images: hyperspectral image, laser radar point cloud data, multispectral image and synthetic aperture radar image;

[0023] The first determination module is configured to map the input data set to the common subspace by using the unsupervised common subspace learning model solved in advance, and determine a set of projection features to be classified of the input data set in the common subspace;

[0024] The second determining module is configured to classify the to-be-classified projection feature set by using a pre-set classifier, a corresponding same-mode projection feature solution of the to-be-classified remote sensing image in the common subspace, and a plurality of category labels corresponding to the projection feature solution, and determine the classification result of each pixel.

[0025] The third determining module is configured to determine a classification image of the to-be-classified remote sensing image according to the classification result of each pixel.

[0026] According to another aspect of the present application, a computer readable storage medium is provided, which stores a computer program for executing the method according to any one of the above aspects of the present application.

[0027] According to another aspect of the present application, an electronic device is provided, which comprises a processor, a memory for storing executable instructions of the processor, and the processor is configured to read the executable instructions from the memory and execute the instructions to implement the method according to any one of the above aspects of the present application.

[0028] The above at least one technical solution adopted by the embodiments of the present application can achieve the following beneficial effects:

[0029] The remote sensing image classification method based on the unsupervised common subspace learning model provided by the present application constructs a completely unsupervised cross-modal common subspace learning model to extract features of the to-be-classified remote sensing image, and finally performs classification calculation in a classifier to determine a classification image. Moreover, the unsupervised common subspace learning model directly learns cross-modal projection from the compactly reorganized multi-modal remote sensing image by implicitly introducing the to-be-learned projection feature in the implicit common subspace, without the need of marking target pixels, thereby reducing the need of costly pixel marking in image classification. In this way, the technical problem of poor classification effect caused by relying on label information for classification is solved. BRIEF DESCRIPTION OF DRAWINGS

[0030] The exemplary embodiments of the present application can be more completely understood by reference to the following drawings:

[0031] Figure 1 is a flowchart of the remote sensing image classification method based on the unsupervised common subspace learning model provided by the embodiments of the present application;

[0032] Figure 2 is a structural schematic diagram of the remote sensing image classification method based on the unsupervised common subspace learning model provided by the embodiments of the present application;

[0033] Figure 3 is a structural schematic diagram of the remote sensing image classification device based on the unsupervised common subspace learning model provided by the embodiments of the present application;

[0034] Figure 4 is a structure of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0035] For the purposes of the present application, the technical solutions and advantages will be more apparent, the technical solutions of the present application will be described in detail below with reference to the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0036] The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the drawings.

[0037] Figure 1 is a flowchart of a remote sensing image classification method based on an unsupervised common subspace learning model provided by an embodiment of the present application. The present embodiment can be applied on an electronic device, such as Figure 1 As shown in the figure, the remote sensing image classification method based on the unsupervised common subspace learning model comprises the following steps:

[0038] Step 101, taking each pixel in a to-be-classified remote sensing image as an input data set, wherein the to-be-classified remote sensing image comprises any one of the following modal remote sensing images: hyperspectral image, laser radar point cloud data, multispectral image and synthetic aperture radar image.

[0039] Wherein, the to-be-classified remote sensing image is not limited to the above-mentioned hyperspectral image, laser radar point cloud data, multispectral image and synthetic aperture radar image, but also can be other modal remote sensing image.

[0040] Step 102, using a pre-solved unsupervised common subspace learning model to map the input data set to a common subspace, and determining a to-be-classified projection feature set of the input data set in the common subspace.

[0041] Optionally, the input in the unsupervised common subspace learning model solving process is feature reorganization of the input data set of the multi-modal remote sensing image training set, the training set comprises multiple modal data, and

[0042] The unsupervised common subspace learning model is to regularize the feature reorganization data and its projection in the common subspace, and minimize the result with the pre-set regularization term.

[0043] Specifically, the unsupervised common subspace learning model formula is as follows:

[0044]

[0045] Wherein

[0046]

[0047]

[0048]

[0049] In the formula, Input dataset for training multimodal remote sensing images Characteristic recombination, For the k-th modal data, Indicates a dimension of d k A real matrix of size ×N, d k Let N be the number of channels, N be the number of pixels in the training data, K be the total number of modalities in the input dataset, and Θ = [Θ1,…,Θ2]. K ] is the joint transformation matrix. Let be the mapping matrix of mode k. Represents a dimension of d×d k The matrix is ​​a real number matrix, where d is the dimension after mapping for each mode. The regularization terms Φ(Θ) and Ψ(Z) are weighted using positive numbers α, β, and γ, respectively. for Projection characteristics in the public subspace For Frobenius regularization, L is the graph Laplacian matrix.

[0050] Optionally, the process of solving the unsupervised common subspace learning model using multimodal remote sensing images is as follows:

[0051] By fixing the projection features in the formula of the unsupervised common subspace learning model, the objective function with respect to the joint transformation matrix is ​​determined.

[0052] The optimal solution of the joint transformation matrix is ​​determined by taking the first derivative of the objective function with respect to the joint transformation matrix and finding that the first derivative is zero.

[0053] Substitute the optimal solution of the joint transformation matrix into the formula of the unsupervised common subspace learning model to determine the objective function with respect to the projected features;

[0054] The projection feature solution is determined by performing spectral decomposition on the objective function of the projection feature.

[0055] Based on the projection characteristic solution, calculate the optimal solution of the joint transformation matrix of the joint transformation matrix.

[0056] Specifically, by fixing the projection feature Z in the formula of the unsupervised common subspace learning model, the objective function with respect to the joint transformation matrix Θ is determined as follows:

[0057]

[0058] By taking the first derivative of the objective function with respect to the joint transformation matrix Θ, and by calculating that the first derivative is zero, the optimal solution of the joint transformation matrix Θ is determined. for:

[0059]

[0060] in,

[0061] The optimal solution of the joint transformation matrix Substituting into the formula for the unsupervised common subspace learning model, the objective function for the projected features Z is determined as follows:

[0062]

[0063] The projected characteristic solution of the projected characteristic Z is determined by performing spectral decomposition on the objective function of the projected characteristic Z. in, This represents the projection feature solution of modality k data in the common subspace; based on the projection feature solution... The optimal solution for the joint transformation matrix Θ is: in To find the optimal solution for the joint transformation matrix that maps modal k data to a common subspace,

[0064] In the above formula, Input dataset for multimodal remote sensing images Characteristic recombination, For the k-th modal data, Indicates a dimension of d k A real matrix of size ×N, d k Let N be the number of channels, K be the number of pixels, and Θ be the total number of modalities in the input dataset. Θ = [Θ1, ..., Θ2]. K ] is the joint transformation matrix. Let be the mapping matrix of mode k. Represents a dimension of d×d k The matrix is ​​a real number matrix, where d is the dimension after mapping for each mode. The regularization terms Φ(Θ) and Ψ(Z) are weighted using positive numbers α, β, and γ, respectively. for Projection characteristics in the public subspace For Frobenius regularization, L is the graph Laplacian matrix. This is the mapping matrix corresponding to the modality k data.

[0065] Optionally, the input data set is mapped to the common subspace using a pre-solved unsupervised common subspace learning model, and a set of to-be-classified projection features of the input data set in the common subspace is determined, comprising:

[0066] According to the objective function about the joint transformation matrix, a mapping relationship formula of the projection features and the joint transformation matrix is determined;

[0067] According to the optimal solution of the corresponding same-mode joint transformation matrix of the to-be-classified remote sensing image in the common subspace and the mapping relationship formula, the input data set is mapped to the set of to-be-classified projection features.

[0068] wherein, according to the formula: the mapping relationship formula of the projection features and the joint transformation matrix is Z=ΘX, and the set of to-be-classified projection features is wherein k' represents the mode of the to-be-classified remote sensing image.

[0069] Step 103, through a pre-set classifier, the set of to-be-classified projection features is classified by using the corresponding same-mode projection feature solution of the to-be-classified remote sensing image in the common subspace and a plurality of class labels corresponding to the projection feature solution, and a classification result of each pixel is determined.

[0070] wherein the plurality of class labels include: Compact High-Rise, Compact Mid-Rise, Compact Low-Rise, Open High-Rise, Open Mid-Rise, Open Low-Rise, Heavy Industry, etc., and the classifier can be a k-NN classifier.

[0071] Step 104, according to the classification result of each pixel, a classification image of the to-be-classified remote sensing image is determined.

[0072] Specifically, the same-mode projection feature solution and the plurality of class labels corresponding to the projection feature solution are input to the classifier k-NN, and for each pixel in the set of to-be-classified projection features, the k nearest pixels in the same-mode projection feature solution are selected to infer the class of each pixel in the set of to-be-classified projection features, and finally the classification image of the to-be-classified remote sensing image is output.

[0073] Further, the application proposes a method of unsupervised common subspace learning model for cross-modal image classification for remote sensing image classification. The method does not use labels as a regression target, and under more relaxed constraints, by implicitly setting a latent representation with the same multi-modal manifold structure as the original reorganized features, it better extracts the latent correlation between samples from multi-modal remote sensing images, further improving the classification performance.

[0074] Specifically, referring to Figure 2 the training process of the unsupervised common subspace learning model is as follows:

[0075] First, load and divide the multi-modal remote sensing data;

[0076] Common remote sensing data includes hyperspectral (HS), multispectral (MS), and synthetic aperture radar (SAR), etc. In cross-modal remote sensing image classification, only part of the data modalities used in the training stage are available in the test stage. Assuming is the kth modality data, where d k is the number of channels, and N is the number of pixels; then the training data can be represented as The test data is a subset of , where K is the total number of data modalities in the training stage. Further, the training, validation, and test sets are constructed.

[0077] First, load the multi-modal data under the same scene and perform super-pixel segmentation, by setting a relatively large super-pixel number N, a segment set is obtained, where the pixels within each segment have the same label. Then, by randomly selecting super-pixels the training set is iteratively constructed until a preset threshold N train is reached, and the final training set is obtained. The remaining segmented fragments are used as the test set or the validation set. In this application, only one modality is selected for the test set and the validation set, respectively, represented as where k' is the selected validation set and test set modality, N val and N test are the number of validation set and test set, respectively.

[0078] In the training of the application, N train is set to 30, N val and N test are set according to the number of samples in the data.

[0079] Second, construct the unsupervised common subspace learning model UCSL;

[0080] 1. Feature reorganization:

[0081] For convenience, the input dataset of the multi-modal remote sensing image training set is rewritten as The input is reorganized as follows:

[0082]

[0083] The output feature reorganization Simultaneously contains single-modal representation And multi-modal Representation, can enrich the data distribution of the original high-dimensional space.

[0084] 2. Construct joint prevalence regularization:

[0085] First, introduce the mapping matrix of modal k And concatenate them to get the joint transformation matrix Θ = [Θ1, …, Θ K ]. Then, map the feature reorganization To the d-dimensional implicit subspace using the joint transformation matrix Θ, and obtain the projected feature The invention assumes that the projected feature Z has orthogonality, which can reduce the ill-conditioning of unsupervised regression, and finally the model is formulated as follows:

[0086]

[0087] Where 1 / 2 is used to simplify the subsequent differential derivation, and the regularization terms Φ(Θ) and Ψ(Z) are weighted using positive numbers α, β and γ. Specifically, the invention uses the Frobenius norm to encourage the model to obtain small projections with better generalization ability. In addition, the invention uses manifold regularization to utilize the intrinsic geometric structure of multi-modal data, so that:

[0088]

[0089] Where, Is the Frobenius regularization, and about the second term on the right, we have:

[0090]

[0091] Where i, j are pixel indices, ||·||2 is the l2 norm, and L is called the graph Laplacian matrix derived from the adjacency matrix W of the graph. In this way, the neighborhood relationship on the manifold can also be strengthened in the latent common subspace, called manifold regularization. On the other hand, the invention suggests using the same graph Laplacian matrix L as described above to regularize the latent representation projected feature Z, as follows:

[0092]

[0093] This term doubles the alignment in the latent subspace through coupling with the first regression term.

[0094] In the training of the present application, the optimal combination of the values of α, β and γ is (10 -3 , 10 -2 , 10 0 ).

[0095] Further, an unsupervised graph is constructed for formula (3). For the model oriented to public subspace learning, the classification performance depends largely on L. Ideally, the accurate estimation of the graph adjacency matrix W is conducive to capturing the global manifold structure to help better feature extraction. However, in order to avoid heavy full graph calculation and its vulnerability to complex noise, the present application uses local calculation, that is, only considering a fixed size of neighbors for each pixel, denoted as N(·). The unsupervised graph adjacency matrix W Unsup of the unsupervised graph is constructed in the following formula:

[0096]

[0097] wherein k represents a modality, is the unsupervised adjacency matrix between the same modalities, i and j represent the subscripts of the elements in the matrix, respectively represent the kth modality feature of the reorganized feature corresponding to the i and j pixels, is a radial basis function, and the parameter σ controls its width, and a and b are two variables of the radial basis function. Considering that the reorganized multi-modal representation contains more information, the present application sets the intra-modality graph to be the graph calculated by , that is:

[0098]

[0099] wherein k1 and k2 represent two different modalities, , that is, the unsupervised adjacency matrix between any two different modalities, which are all set to , that is, calculated using the (K+1)th modality feature of the reorganized feature . The design of this formula provides more reliable guidance for joint manifold regularization.

[0100] Thus, by constructing the unsupervised graph to guide the public subspace learning, the cross-modal image classification is further completed.

[0101] Further, with the adjacency matrix W, the D ii =∑ j W​​i,j The degree matrix D can be calculated. Then, the normalized graph Laplacian matrix is L is expected to handle multi-modal features with regular and irregular graph structures well.

[0102] Finally, the equation (2) is the constructed model, and the projected feature Z and the joint transformation matrix Θ are to be solved.

[0103] Thirdly, the unsupervised common subspace learning model UCSL is solved.

[0104] Firstly, by fixing the projected feature Z, the sub-problem about the joint transformation matrix Θ is considered, and according to the equations (2), (3) and (5), the objective function of the joint transformation matrix Θ is as follows:

[0105]

[0106] The objective function is convex, and by calculating the first-order derivative to be zero, the closed-form optimal solution of it is Wherein,

[0107] Secondly, the optimal solution of the above joint transformation matrix is substituted into the equation (2), and the overall objective is rewritten as a function about the projected feature Z, as follows:

[0108]

[0109] Under the orthogonal constraint, the above optimization problem follows the form of spectral clustering, and the projected feature solution In general, the solution of minimizing the equation (9) can be obtained by the left eigenvector or the right eigenvector corresponding to the d smallest eigenvalues obtained by the eigen decomposition of According to the definition of H, M can be easily determined by H -1 = (H -1 ) T In order to avoid the numerical instability that may be caused by actually calculating the inverse of H, the present application suggests using eig((M+M T ) / 2) instead of eig(M), where eig(·) represents eigen decomposition. The eig((M+M T ) / 2) is also called spectral decomposition, which can more effectively generate orthogonal eigenvectors.

[0110] The projected feature solution of the projected feature Z obtained is Each pixel in the reorganized K+1 modal data is represented as a d-dimensional vector, and the implicit subspace alignment is realized. ​

[0111] Third step, obtaining the projection feature solution Then, the optimal solution of the joint transformation matrix is calculated Can be expressed as Wherein, Data of any modal k can be mapped to the subspace.

[0112] Therefore, by deeply studying the feature decomposition, the proposed unsupervised common subspace learning model USCL is solved by using almost one-step calculation.

[0113] Therefore, the remote sensing image classification method based on the unsupervised common subspace learning model provided by the present application constructs a completely unsupervised cross-modal common subspace learning model to extract features of the remote sensing image to be classified, and finally performs classification calculation in the classifier to determine the classified image. Moreover, the unsupervised common subspace learning model implicitly introduces the projection feature to be learned in the implicit common subspace, directly learns the cross-modal projection from the compactly reorganized multi-modal remote sensing image, and does not need to mark the target pixel, thereby reducing the need for costly pixel marking during image classification. Further, the technical problem of poor classification effect caused by relying on label information for classification is solved. Moreover, under the regularization of orthogonality and manifold structure prior, the proposed unsupervised common subspace learning model can be effectively transformed into feature decomposition, thereby effectively solving the trace norm optimization problem, and further reducing the need for costly pixel marking during image classification.

[0114] Figure 3 is a structural schematic diagram of the remote sensing image classification device based on the unsupervised common subspace learning model provided by the embodiment of the present application. As shown in Figure 3 The device comprises:

[0115] As the module 310, each pixel in the remote sensing image to be classified is used as the input data set, wherein the remote sensing image to be classified comprises any one of the following modal remote sensing images: hyperspectral image, laser radar point cloud data, multispectral image and synthetic aperture radar image;

[0116] The first determination module 320 is configured to map the input data set to the common subspace by using the pre-solved unsupervised common subspace learning model, and determine the classified projection feature set of the input data set in the common subspace;

[0117] The second determination module 330 is configured to classify the classified projection feature set by using the corresponding same-modal projection feature solution of the remote sensing image to be classified in the common subspace and the plurality of category labels corresponding to the projection feature solution through the pre-set classifier, and determine the classification result of each pixel.

[0118] The third determining module 340 is configured to determine a classification image of the remote sensing image to be classified according to the classification result of each pixel.

[0119] Optionally, input in the process of solving the unsupervised common subspace learning model is feature reconstruction of an input data set of a multi-modal remote sensing image training set, the training set includes multiple modal data, and

[0120] The unsupervised common subspace learning model is to regularize feature reconstruction data and its projection in the common subspace, and minimize the result and a preset regularization term.

[0121] Optionally, the process of solving the unsupervised common subspace learning model by the multi-modal remote sensing image is as follows:

[0122] The fourth determining module is configured to fix the projection feature in the formula of the unsupervised common subspace learning model, and determine a target function about the joint transformation matrix.

[0123] The fifth determining module is configured to perform first-order derivation of the target function about the joint transformation matrix, and determine a joint transformation matrix optimal solution of the joint transformation matrix by calculating the first-order derivative to be zero.

[0124] The sixth determining module is configured to bring the joint transformation matrix optimal solution into the formula of the unsupervised common subspace learning model, and determine a target function about the projection feature.

[0125] The seventh determining module is configured to determine a projection feature solution of the projection feature by performing spectral decomposition on the target function of the projection feature.

[0126] The calculating module is configured to calculate the joint transformation matrix optimal solution of the joint transformation matrix according to the projection feature solution.

[0127] Optionally, the first determining module includes:

[0128] The determining sub-module is configured to determine a mapping relationship formula of the projection feature and the joint transformation matrix according to the target function about the joint transformation matrix.

[0129] The mapping sub-module is configured to map the input data set into a to-be-classified projection feature set according to the corresponding same-modal joint transformation matrix optimal solution of the remote sensing image to be classified in the common subspace and the mapping relationship formula.

[0130] Figure 4 is the structure of the electronic device provided by the embodiment of the application. As shown in Figure 4 The electronic device 40 includes one or more processors 41 and a memory 42.

[0131] The processor 41 can be a central processing unit (CPU) or other form of processing unit that has data processing and / or instruction execution capabilities and can control other components in the electronic device to perform desired functions.

[0132] The memory 42 can include one or more computer program products that can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory, for example, can include random access memory (RAM), cache memory, and / or the like. The non-volatile memory, for example, can include read only memory (ROM), hard disk, flash memory, and / or the like. One or more computer program instructions can be stored on the computer-readable storage media, which the processor 41 can execute to implement the methods of the software programs of the various embodiments of the present application described above and / or other desired functions. In one example, the electronic device can further include an input device 43 and an output device 44, which are interconnected through a bus system and / or other form of connection mechanism (not shown).

[0133] In addition, the input device 43 can further include, for example, a keyboard, a mouse, and / or the like.

[0134] The output device 44 can output various information to the outside. The output device 44 can include, for example, a display, a speaker, a printer, a communication network and a remote output device connected thereto, and / or the like.

[0135] Of course, in order to simplify, Figure 4 Only some of the components of the electronic device related to the present application are shown in FIG. 1, and components such as a bus, an input / output interface, and the like are omitted. In addition to this, the electronic device can further include any other appropriate components according to a specific application.

[0136] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and / or the like) embodying computer readable program code thereon.

[0137] Embodiments of the present application can be applied to terminal devices, computer systems, servers, and the like electronic devices, which can operate with many other general-purpose or special-purpose computing system environments or configurations. Examples of well-known terminal devices, computing systems, environments, and / or configurations suitable for use with terminal devices, computer systems, servers, and the like electronic devices include, but are not limited to, personal computers, server computers, thin clients, thick clients, hand-held or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network personal computers, minicomputers, mainframe computers, and distributed cloud computing technology environments that include any of the above systems, and the like.

[0138] Terminal devices, computer systems, servers, and the like electronic devices can be described in the general context of computer system-executable instructions, such as program modules, being executed by a computer system. Generally, program modules can include routines, programs, objects, components, logic, data structures, and the like, which perform particular tasks or implement particular abstract data types. Computer systems / servers can be implemented in a distributed cloud computing environment, where tasks are performed by remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules can be located in local or remote computer system storage media including memory storage devices.

[0139] The computer program product can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, and the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device or entirely on the remote computing device or server.

[0140] In addition, embodiments of the present application can also be a computer readable storage medium, which stores computer program instructions, when the computer program instructions are run by a processor, the processor executes the steps in the method of various embodiments described in the specification.

[0141] The computer readable storage medium can be any combination of one or more computer readable medium(s). The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, or apparatus or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0142] It should also be noted that the terms "comprising," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus including the recited element.

[0143] The above description is merely illustrative of the application, and not restrictive. Various modifications and changes can become apparent to those skilled in the art. The scope of the application is defined by the appended claims, along with the full scope of equivalents to which such claims are entitled. The disclosures of each patent, patent application, and publication cited herein are hereby incorporated herein by reference, each in its entirety.

Claims

1. A method for remote sensing image classification based on unsupervised common subspace learning model, characterized in that, The method comprises the following steps: each pixel in a remote sensing image to be classified is taken as an input data set, wherein the remote sensing image to be classified comprises any one of the following modal remote sensing images: a hyperspectral image, laser radar point cloud data, a multispectral image, and a synthetic aperture radar image; an unsupervised common subspace learning model is solved by using multi-modal remote sensing images, which comprises the following steps: fixing a projection feature in a formula of the unsupervised common subspace learning model, determining a target function about a joint transformation matrix; performing first-order derivation about the joint transformation matrix on the target function, and determining a joint transformation matrix optimal solution of the joint transformation matrix by calculating the first-order derivative to be zero; bringing the joint transformation matrix optimal solution into the formula of the unsupervised common subspace learning model, determining a target function about the projection feature; determining a projection feature solution of the projection feature by performing spectral decomposition on the target function of the projection feature; and calculating a joint transformation matrix optimal solution of the joint transformation matrix according to the projection feature solution; the input data set is mapped to a common subspace by using a pre-solved unsupervised common subspace learning model, and a to-be-classified projection feature set of the input data set in the common subspace is determined; the to-be-classified projection feature set is classified by using corresponding same-modal projection feature solutions of the remote sensing image to be classified in the common subspace and a plurality of class labels corresponding to the projection feature solutions by using a pre-set classifier, and a classification result of each pixel is determined; a classification image of the remote sensing image to be classified is determined according to the classification result of each pixel.

2. The method of claim 1, wherein, In the solving process of the unsupervised common subspace learning model, the input is feature reconstruction of an input data set of a multi-modal remote sensing training set, the training set comprises multiple modal data, and the unsupervised common subspace learning model is: regularization feature reconstruction data and its projection in a common subspace, and the result is minimized with a pre-set regularization term.

3. The method of claim 1, wherein, the input data set is mapped to a common subspace by using a pre-solved unsupervised common subspace learning model, and a to-be-classified projection feature set of the input data set in the common subspace is determined, which comprises the following steps: a mapping relationship formula of the projection feature and the joint transformation matrix is determined according to the target function about the joint transformation matrix; the input data set is mapped to the to-be-classified projection feature set according to the corresponding same-modal joint transformation matrix optimal solution of the remote sensing image to be classified in the common subspace and the mapping relationship formula.

4. A remote sensing image classification device based on an unsupervised common subspace learning model, characterized in that, The method comprises the following steps: each pixel in a remote sensing image to be classified is taken as an input data set, wherein the remote sensing image to be classified comprises any one of the following modal remote sensing images: a hyperspectral image, laser radar point cloud data, a multispectral image, and a synthetic aperture radar image; a first determination module is configured to map the input data set to a common subspace by using a pre-solved unsupervised common subspace learning model, and determine a to-be-classified projection feature set of the input data set in the common subspace; The second determining module is configured to classify the to-be-classified projection feature set by using a preset classifier, using the corresponding same-modality projection feature solution of the to-be-classified remote sensing image in the common subspace and a plurality of category labels corresponding to the projection feature solution, and determining a classification result of each pixel. The third determining module is configured to determine a classification image of the to-be-classified remote sensing image according to the classification result of each pixel. The fourth determining module is configured to fix the projection feature in the unsupervised common subspace learning model formula, and determine a target function about the joint transformation matrix. The fifth determining module is configured to perform first-order derivation about the joint transformation matrix on the target function, and determine a joint transformation matrix optimal solution of the joint transformation matrix by calculating a first-order derivative to be zero. The sixth determining module is configured to bring the joint transformation matrix optimal solution into the unsupervised common subspace learning model formula, and determine a target function about the projection feature. The seventh determining module is configured to determine a projection feature solution of the projection feature by performing spectral decomposition on the target function of the projection feature. The computing module is configured to calculate the joint transformation matrix optimal solution of the joint transformation matrix according to the projection feature solution.

5. The apparatus of claim 4, wherein, The input in the unsupervised common subspace learning model solving process is feature reorganization of an input data set of a multi-modality remote sensing image training set, the training set includes a plurality of modal data, and The unsupervised common subspace learning model is: regularized feature reorganization data and its projection in a common subspace, and the result is minimized with a preset regularization term.

6. The apparatus of claim 5, wherein, The first determining module comprises: The determining sub-module is configured to determine a mapping relationship formula of the projection feature and the joint transformation matrix according to the target function about the joint transformation matrix. The mapping sub-module is configured to map the input data set into the to-be-classified projection feature set according to the corresponding same-modality joint transformation matrix optimal solution of the to-be-classified remote sensing image in the common subspace and the mapping relationship formula.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is used to execute the method in any one of claims 1-3.

8. An electronic device, comprising: The electronic device comprises: A processor; A memory for storing executable instructions of the processor; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method in any one of claims 1-3. The electronic device comprises: A processor; A memory for storing executable instructions of the processor; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method in any one of claims 1-3.

Images