Incomplete hyperspectral image classification method based on geometric consistency learning

Abstract

The invention discloses an incomplete hyperspectral image classification method based on geometric consistency learning. The method comprises the following steps: acquiring a hyperspectral image to be classified; performing matrix decomposition on the hyperspectral image sample to obtain a consensus representation; complementing the data of each wave band and carrying out space geometric structure exploration; connecting the consensus representation and the space geometric structure exploration result of the data to a framework for joint optimization, thereby obtaining the consensus representation; inputting the consensus representation into a one-step classification strategy module to obtain a final classification matrix; using an iterative updating strategy to alternately update other variables while fixing one variable until convergence; and finally, obtaining an image classification result according to the indication matrix. According to the method, incomplete hyperspectral data are complemented through a tensor complementation technology, so that the relationship between the data is better captured, and a classification result is obtained by one step through spectral rotation, so that error accumulation caused by obtaining the result through two steps is avoided.

Description

Technical Field

[0001] The invention relates to the field of remote sensing image processing, and in particular to an incomplete hyperspectral image classification method based on geometric consistency learning. Background Art

[0002] Hyperspectral imaging works by simultaneously acquiring spectral data in multiple wavelength ranges of the target scene. Traditional color imaging systems can only acquire data in three bands: red, green, and blue, while hyperspectral imaging systems can acquire data in dozens to hundreds of bands. This technology not only provides color information for each pixel, but also provides detailed spectral information such as reflectance or transmittance in a continuous wavelength range at each pixel.

[0003] Hyperspectral image classification is the process of using the rich spectral information obtained by hyperspectral imaging to classify objects. Compared with traditional RGB images, hyperspectral images can provide continuous spectral information in multiple bands, which makes them have unique advantages and challenges in object classification.

[0004] Since the labeling cost of classification labels in the acquisition of hyperspectral image training samples is very time-consuming and labor-intensive, and data loss is often caused during the operation, how to perform unsupervised hyperspectral image classification and restore band data to improve the accuracy of image classification is a very challenging and practical problem. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an incomplete hyperspectral image classification method based on geometric consistency learning. The present invention completes the incomplete hyperspectral data through tensor completion technology, so that the relationship between the data can be better captured, and uses spectral rotation to obtain the classification result in one step, avoiding the error accumulation of the result obtained in two steps, thereby improving the image classification accuracy.

[0006] The technical means adopted by the present invention are as follows:

[0007] A method for classifying incomplete hyperspectral images based on geometric consistency learning includes the following steps:

[0008] S1. Acquire a hyperspectral image to be classified, wherein the hyperspectral image includes hyperspectral image samples extracted from different bands;

[0009] S2, perform matrix decomposition on the hyperspectral image samples of each band to construct a consensus representation of the hyperspectral image sample data of each band;

[0010] S3, complete the data of each band, introduce tensor constraints, and explore the spatial geometric structure of the data through tensor constraints;

[0011] S4. Through connectivity constraints, the consensus representation after tensor learning and the spatial geometric structure exploration results of the data are connected into a framework for joint optimization to obtain the consensus representation;

[0012] S5, inputting the consensus representation into a one-step classification framework to obtain a final classification indicator matrix;

[0013] S6. Based on the completed sample data, consensus representation and classification matrix, combined with auxiliary variables, an optimization target model is constructed;

[0014] S7, using an iterative update strategy, while fixing one variable, alternately updating other variables until the optimization target model converges;

[0015] S8. Obtain a classification indicator matrix after the model converges, and obtain an image classification result according to the classification indicator matrix.

[0016] Furthermore, matrix decomposition is performed on the hyperspectral image samples of each band according to the following formula:

[0017]

[0018] in, represents the basis matrix of the υth band, P∈R k×n The consensus is that represents the missing index matrix of the υth band, X υ represents the data of the υth band, V represents the number of bands, k represents the number of categories, m v The dimension of the υth band, n v The number of samples in the υth band.

[0019] Furthermore, the data of each band is completed by the following formula:

[0020]

[0021] sC Ω (Z v )=C Ω (S v ),0≤Z v ,Z v 1=1.

[0022] in represents a tensor, represents the similarity graph of the υth band, and represents the sample space of the similarity graph of the Vth band, N u ×N u represents the number of samples of the similarity graph of the υth band, and Ω represents the index space of non-missing data in the original data.

[0023] Furthermore, through connectivity constraints, the consensus representation and the spatial geometric structure exploration results of the data are connected into a framework for joint optimization, thereby obtaining the consensus representation, which also includes optimizing the consensus representation through the following formula:

[0024]

[0025] sP T P=I,Z v ≥0,Z v 1=1,RR T =I,y i,j ∈(0,1),y i 1=1.

[0026] in Represents a tensor, Z v represents the similarity matrix of the υth band, L z Represents Z v The Laplace matrix of , Y represents the classification indicator matrix, R represents the orthogonal rotation matrix, λ1 and λ2 represent the regularization parameters, I is the identity matrix, greater than 0 means non-negative, the value of the matrix Y is forced to 0 or 1, 0 means there is no data at the corresponding position, and 1 means there is data at the corresponding position.

[0027] Furthermore, the one-step classification strategy module is configured as follows:

[0028]

[0029] sR T =I,y i,j ∈(0,1),y i 1=1.

[0030] Where Y represents the final classification indicator matrix, and R represents the rotation matrix.

[0031] Furthermore, the optimization target model is:

[0032]

[0033] sC Ω (Z v )=C Ω (S v ), P T P=I,Z v ≥0,Z v 1=1,W v ≥0,W v 1=1.

[0034] Among them, W v and is an auxiliary variable introduced, and is the Lagrange multiplier, μ is the preset parameter.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] The present invention discloses an incomplete hyperspectral image classification method that obtains consensus representation through geometric consistency learning. First, consensus representation is obtained through matrix decomposition, and the potential representation of the original data can be learned. Then, the hyperspectral data is completed and the high-dimensional subspace structure of the hyperspectral data is explored through the subspace learning technology of tensor completion. The two modules are combined into a unified framework through connectivity constraints. Finally, the classification result is directly obtained through a one-step classification strategy to avoid suboptimal solutions. The present invention completes incomplete hyperspectral data through tensor completion technology, so that the relationship between the data is better captured, and spectral rotation is used to obtain the classification result in one step, avoiding the error accumulation of the result obtained by two steps. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 The present invention is a flowchart of an incomplete hyperspectral image classification method based on geometric consistency learning in an embodiment of the present invention.

[0039] Figure 2 It is a schematic diagram of the process of using the incomplete hyperspectral image classification method based on geometric consistency learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0041] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0042] like Figure 1 As shown, the present invention provides an incomplete hyperspectral image classification method based on geometric consistency learning, comprising the following steps:

[0043] Step 1: Given hyperspectral data X = {X 1 ,X 2 ,...,X υ}, where X υ Represents the data sample of the υth band;

[0044] Step 2: Matrix decomposition is performed on the hyperspectral data to capture the potential structure of the original data and obtain the consensus representation P of the data;

[0045] Zero-fill is performed to complete missing data, spatial combination structure is explored through tensor-related constraints, and the results after tensor learning are used to promote the learning of consensus representation to improve the results.

[0046] The consensus representation learning mentioned is as follows:

[0047]

[0048] in represents the basis matrix of the υth band, P∈R k×n The consensus is that represents the missing index matrix of the υth band. This formula fully explores the potential representation of the original data.

[0049] Step 3: Use tensor completion technology to complete the data and explore the spatial geometric structure of the data;

[0050] Data completion can be obtained through the following tensor completion formula:

[0051]

[0052] sCΩ (Z v )=C Ω (S v ),0≤Z v ,Z v 1=1.

[0053] in represents a tensor, represents the similarity graph of the υth band, and

[0054] Step 4: Through connectivity constraints, consensus representation learning and geometric consistency learning are connected into a unified framework for joint optimization to obtain the consensus representation P;

[0055] The specific formula can be obtained as follows:

[0056]

[0057] sP T P=I,Z v ≥0,Z v 1=1,C Ω (Z v )=C Ω (S v ).

[0058] Where P represents the consensus representation matrix.

[0059] Step 5: Input the above consensus representation matrix P into the one-step classification framework to obtain the final classification matrix;

[0060] The one-step classification strategy utilized is as follows:

[0061]

[0062] sR T =I,y i,j ∈(0,1),y i 1=1.

[0063] Where Y represents the final classification indicator matrix, and R represents the rotation matrix.

[0064] Step 6: Reference auxiliary variables, construct the optimization target equation, and optimize and solve the equation.

[0065] Step 6.1 The overall formula is as follows:

[0066]

[0067] Step 6.2 The optimization formula for adding variables is as follows:

[0068]

[0069] sC Ω (Z v )=C Ω (S v ), P T P=I,Z v ≥0,Z v 1=1,W v ≥0,W v 1=1.

[0070] Where W v and is an auxiliary variable introduced, and is the Lagrange multiplier, μ is the parameter.

[0071] Step 7: Use the iterative update strategy to solve the fixed variables.

[0072] Step 7.1 Optimize Z v , fixing other variables, we can get:

[0073]

[0074] According to the Lagrange multiplier method:

[0075]

[0076] in is the Lagrange multiplier.

[0077] Step 7.2 Optimization Fixing other variables yields:

[0078]

[0079]

[0080] Step 7.3 Optimize W v , fixing other variables, we get:

[0081]

[0082] Step 7.4 Optimize P and fix other variables to get:

[0083]

[0084] Step 7.5 Optimize R and fix other variables to get:

[0085]

[0086] sR T=I

[0087] Y T P performs SVD decomposition

[0088] Step 7.6 Optimize U and fix other variables to get:

[0089]

[0090] U v U vT =I

[0091]

[0092] Step 7.7 Optimize Y and fix other variables to get:

[0093]

[0094] sty i,j ∈(0,1),y i 1=1.

[0095]

[0096] where p i is the i-th column vector of P, r o is the 0th column vector of R.

[0097] Step 8. Finally, the classification result is obtained according to the indicator matrix Y. The value of the matrix Y is forced to be 0 or 1, 0 means there is no data at the corresponding position, and 1 means there is data at the corresponding position, that is, there is a class. Clustering is performed according to the knowledge matrix result to finally obtain the image classification result.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for incomplete hyperspectral image classification based on geometric consistency learning, characterized in that: The following steps are involved: S1. Acquire a hyperspectral image to be classified, wherein the hyperspectral image includes hyperspectral image samples extracted from different bands; S2, perform matrix decomposition on the hyperspectral image samples of each band to construct a consensus representation of the hyperspectral image sample data of each band; S3, complete the data of each band, introduce tensor constraints, and explore the spatial geometric structure of the data through tensor constraints; S4. Through connectivity constraints, the consensus representation after tensor learning and the spatial geometric structure exploration results of the data are connected into a framework for joint optimization to obtain the consensus representation; S5, inputting the consensus representation into a one-step classification framework to obtain a final classification indicator matrix; S6. Based on the completed sample data, consensus representation and classification matrix, combined with auxiliary variables, an optimization target model is constructed; S7, using an iterative update strategy, while fixing one variable, alternately updating other variables until the optimization target model converges; S8. Obtain a classification indicator matrix after the model converges, and obtain an image classification result according to the classification indicator matrix.

2. A method for incomplete hyperspectral image classification based on geometric consistency learning, characterized in that: The matrix decomposition is performed for the hyperspectral image samples of each band according to the following formula: in, represents the basis matrix of the υth band, P∈R k×n The consensus is that represents the missing index matrix of the υth band, X υ represents the data of the υth band, V represents the number of bands, k represents the number of categories, m v The dimension of the υth band, n v The number of samples in the υth band.

3. The incomplete hyperspectral image classification method based on geometric consistency learning according to claim 2 is characterized in that: The data of each band is completed by the following formula: s.t.C Ω (Z v )=C Ω (S v ),0≤Z v ,Z v 1=1. in represents a tensor, represents the similarity graph of the υth band, and represents the sample space of the similarity graph of the Vth band, N u ×N u represents the number of samples of the similarity graph of the υth band, and Ω represents the index space of non-missing data in the original data.

4. The incomplete hyperspectral image classification method based on geometric consistency learning according to claim 3 is characterized in that: Through connectivity constraints, the consensus representation and the spatial geometric structure exploration results of the data are connected into a framework for joint optimization to obtain the consensus representation, which also includes optimizing the consensus representation through the following formula: s.t.P T P=I,Z v ≥0,Z v 1=1,RR T =I,y i,j ∈(0,1),y i 1=1. in Represents a tensor, Z v represents the similarity matrix of the υth band, L z Represents Z v The Laplace matrix of , Y represents the classification indicator matrix, R represents the orthogonal rotation matrix, λ1 and λ2 represent the regularization parameters, I is the identity matrix, greater than 0 means non-negative, the value of the matrix Y is forced to 0 or 1, 0 means there is no data at the corresponding position, and 1 means there is data at the corresponding position.

5. The incomplete hyperspectral image classification method based on geometric consistency learning according to claim 4 is characterized in that: The one-step classification strategy module is configured as follows: s.t.RR T =I,y i,j ∈(0,1),y i 1=1. Where Y represents the final classification indicator matrix, and R represents the rotation matrix.

6. The incomplete hyperspectral image classification method based on geometric consistency learning according to claim 4 is characterized in that: The optimization target model is: s.t.C Ω (Z v )=C Ω (S v ),P T P=I,Z v ≥0,Z v 1=1,W v ≥0,W v 1=1. Among them, W v and is an auxiliary variable introduced, and is the Lagrange multiplier, μ is the preset parameter.

Images