Endmember deep non-negative matrix factorization method for hyperspectral images under mixed noise

By reverse modeling and coupled observation modeling of the spectral depth linear mixing process of hyperspectral images, combining sparse constraints and spatial full variation constraints, the stability problem of depth non-negative matrix decomposition under mixed noise is solved, and high Hybrid cell decomposition with precision.

CN114861748BActive Publication Date: 2024-11-22WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210206935.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-04
Publication Date
2024-11-22
Estimated Expiration
2042-03-04

AI Technical Summary

Technical Problem

The existing deep non-negative matrix decomposition methods are susceptible to mixed noise pollution in hyperspectral images, resulting in uncontrollable models, non-convergence, and inunique solutions, which seriously threatens the stability of the hybrid cell decomposition process.

Method used

By reverse modeling the linear mixing process of spectral depth, a non-negative matrix decomposition model of end element depth of hyperspectral image is constructed, and the hyperspectral image observation model is coupled, and abundance L1/2 sparse constraints and spatial total variation constraints are introduced to construct The integrated observation model of noise removal-spectral decomposition is used to optimize the theoretical decomposition of non-convex cost functions using alternating direction multiplier method.

Benefits of technology

High-precision, high-efficiency, high-spectral image hybrid cell decomposition under mixed noise is achieved, which enhances the physical significance and stability of the decomposition process, and overcomes the problems of unclear physical significance and noise sensitivity in the intermediate process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114861748B_ABST
    Figure CN114861748B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise. The present invention couples the hyperspectral image observation model and the endmember deep non-negative matrix decomposition model to establish a noise removal-spectral decomposition integrated observation model, and based on the Bayesian inverse problem solving theory, establishes a theoretical framework for integrated processing to achieve a common description of the two processing tasks. At the same time, the present invention fully mines the hierarchical feature information of the image by reverse modeling the spectral deep linear mixed model, thereby enhancing the practicability and interpretability of the model. In addition, the model also applies L to the abundance matrix 1 / 2 The constraint and spatial total variation constraint take into account the sparsity and piecewise smoothness of the abundance matrix. The present invention fully exploits the deep feature information and spatial information of the image, provides technical support for high-precision mixed pixel decomposition, and effectively enhances the application potential of hyperspectral images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of mixed pixel decomposition of hyperspectral images, and in particular to an endmember deep non-negative matrix decomposition method obtained by reverse modeling a deep linear mixing process of a hyperspectral image, which learns the deep information between hyperspectral image pixels polluted by mixed noise to complete image spatial information extraction, thereby realizing mixed pixel decomposition of a hyperspectral image under mixed noise. Background Art

[0002] Hyperspectral images contain rich spatial and spectral information, and can perform fine analysis of ground objects with high spectral diagnostic capabilities. However, due to the mutual constraints between important parameters such as sensor spatial resolution, spectral resolution, temporal resolution and image width, hyperspectral images have low spatial resolution due to their high spectral resolution, and due to the complexity of ground object coverage types, mixed pixels are common in hyperspectral images. The existence of mixed pixels seriously restricts the development of hyperspectral images from qualitative analysis to quantitative description.

[0003] At present, mixed pixel decomposition methods can be classified into four categories:

[0004] The geometry-based method is based on convex geometry, assuming that the end members are located at the vertices of a simplex formed by data points in the feature space, and the mixed pixel decomposition is achieved by finding the vertices of the simplex as the end members. The advantages of this method are low computational complexity, clear concepts, and wide application in the decomposition of hyperspectral mixed pixels. The disadvantage is that it is not suitable for highly mixed situations.

[0005] The statistical method decomposes the hyperspectral mixed pixels into a statistically independent or uncorrelated Bayesian inference problem, regards the end members and abundance as random variables, and determines the end members and their corresponding abundance by estimating the parameters in the statistical model. The advantage of this method is that it can well overcome the situation of highly mixed data, and can perform end member extraction and abundance estimation at the same time. The disadvantage is that the computational complexity is high and the model is difficult to solve.

[0006] The sparse regression-based method uses the increasingly improved spectral library to decompose mixed pixels in a semi-supervised manner, taking the spectral library containing a large number of endmembers as a priori over-complete dictionary, and treating the endmembers as atoms. By effectively combining the linear mixed model and sparse expression, a small number of atoms are adaptively selected to sparsely reconstruct the target spectrum, thereby achieving robust abundance estimation. The advantage of this method is that it no longer relies on the performance of a certain endmember extraction algorithm when estimating abundance. The disadvantage is that it takes a lot of time and manpower to establish a spectral library of ground objects, and it cannot be guaranteed that the spectral library and the hyperspectral image used are under the same imaging conditions, which makes the endmembers in the spectral library different from the endmembers actually contained in the scene, limiting the development of sparse decomposition.

[0007] The deep learning-based method mainly uses an autoencoder network to map the input spectrum to abundance through the encoder, and the decoder uses linear mapping to reconstruct the original input spectrum, and then uses the linear mapping as the endmember information to finally achieve mixed pixel decomposition. The advantage of this method is that it can capture the hierarchical structure of hidden information in hyperspectral images, but the disadvantage is that the randomness of sample probability may lead to random local estimates of endmember extraction.

[0008] The traditional hyperspectral mixed pixel decomposition method optimizes and solves by establishing a mathematical model to obtain endmember and abundance information. In recent years, deep learning has developed rapidly and has been widely used in various fields, and has achieved good results. The essence of hyperspectral mixed pixel decomposition is a function approximation problem, that is, using the endmember matrix and the abundance matrix to infinitely approximate the original observation data. The powerful fitting ability of deep learning is very suitable for mixed pixel decomposition. Therefore, the deep non-negative matrix decomposition method combined with the concept of deep learning has entered the field of vision of researchers. However, the existing deep non-negative matrix decomposition method is based on mathematical theory, which deeply decomposes the abundance, and the physical meaning of the intermediate process is unclear.

[0009] However, due to the failure of sensor electronic components, differences in band energy distribution and the influence of imaging environment, the hyperspectral remote sensing imaging process is very susceptible to contamination by multiple types of mixed noise such as Gaussian, pulse, stripe and bad pixel. Non-negative matrix decomposition is a typical highly ill-conditioned non-convex inverse problem, which is extremely sensitive to noise. Non-Gaussian noise in the image can easily become a singular point of the decomposition model, causing the model to be uncontrollable, non-convergent and non-unique, which seriously threatens the stability of the mixed pixel decomposition process.

[0010] In summary, how to provide a method for deep non-negative matrix decomposition of hyperspectral images that starts from the physical process of spectral imaging and takes into account the influence of mixed noise is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0011] In view of the shortcomings of existing deep non-negative matrix decomposition methods, the present invention proposes a method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise. By reversely modeling the physical process of spectral depth linear mixing, a hyperspectral image endmember deep non-negative matrix decomposition model is constructed, and the hyperspectral image observation model and the endmember deep non-negative matrix decomposition model are coupled to establish a noise removal-spectral decomposition integrated observation model. In addition, in order to solve the problem of non-convex non-negative matrix decomposition cost function, the abundance L is introduced to make full use of the sparsity and spatial smoothness of the abundance distribution of hyperspectral images. 1 / 2Sparse constraints and spatial total variation constraints are used to mine the content information of hyperspectral images in the spatial dimension, overcome the shortcomings of unclear physical meaning of the intermediate process of abundance depth non-negative matrix decomposition, and achieve high-precision and high-efficiency hyperspectral image mixed pixel decomposition under mixed noise.

[0012] The technical solution of the present invention is to provide a method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise, comprising the following steps:

[0013] Step 1: Based on the given hyperspectral image, the spectral depth linear mixing process is reversely modeled to construct a hyperspectral image endmember depth non-negative matrix decomposition model;

[0014] Step 2, coupling the hyperspectral image observation model and the endmember deep non-negative matrix decomposition model to construct a noise removal-spectral decomposition integrated observation model;

[0015] Step 3, construct the cost function of the noise removal-spectral decomposition integrated observation model, and based on the alternating direction multiplier method optimization theory, introduce auxiliary variables and perform variable separation to decompose the complex non-convex integrated observation model cost function into several convex sub-objective functions to achieve mixed pixel decomposition.

[0016] Furthermore, in step 1, the spectral depth linear mixing process is as follows;

[0017] The given hyperspectral image is denoted as Among them, each represents a pixel point with B bands, N is the number of pixels, ignoring the influence of noise, is the pure and noise-free data matrix of the hyperspectral image, is the pure endmember spectrum matrix, p is the number of endmembers, The two are linearly mixed for the first time to obtain matrix A2, which is regarded as the data matrix formed after the electromagnetic wave is reflected once on the surface; then the electromagnetic wave is reflected for the second time on the surface, and the end member matrix A2 at this time is linearly mixed with its corresponding abundance matrix S2 again to generate a new data matrix A3; this is repeated, when the electromagnetic wave is reflected l times on the surface, the end member matrix A at this time is l The corresponding abundance matrix S l Perform linear mixing to obtain the matrix A l+1 ; Until the electromagnetic wave reaches the hyperspectral sensor after L times of linear mixing, the noise-free hyperspectral image data matrix Y after the final deep linear mixing will be obtained. Then the spectrum depth linear mixing model obtained after the electromagnetic wave is deeply linearly mixed after multiple reflections is:

[0018] Y=A1S1S2…S L (1).

[0019] Furthermore, in step 1, the spectral depth linear mixing process is reversely simulated. In the first layer, the noise-free hyperspectral image data matrix Y is subjected to a non-negative matrix decomposition to obtain the endmember spectral matrix A1 after the first decomposition and its corresponding abundance matrix S1; due to the spectral depth linear mixing, the endmember spectral matrix obtained after the first decomposition is not pure enough, so a second decomposition is performed to obtain the endmember spectral matrix A2 after the second decomposition and its corresponding abundance matrix S2; this is repeated, and when the lth decomposition is performed, the endmember spectral matrix A obtained by the previous decomposition is l-1 Decomposed into a purer endmember spectrum matrix A l and its corresponding abundance matrix S l ; After L decompositions, it is considered that a pure end member spectrum matrix A is obtained. L and its corresponding abundance matrix S L , then the endmember deep non-negative matrix decomposition model of hyperspectral image is obtained as:

[0020] Y=A L S L S L-1 …S2S1 (2).

[0021] Furthermore, in step 2, the hyperspectral image observation model is:

[0022] X=Y+N+E (3)

[0023] in, represents the Gaussian noise in the hyperspectral image, represents the impulse noise in the hyperspectral image,

[0024] Therefore, the noise removal-spectral decomposition integrated observation model constructed is:

[0025] X=A L S L S L-1 …S2S1+N+E (4).

[0026] Furthermore, in step 3, the noise removal-spectral decomposition integrated observation model adopts a cost function based on Euclidean distance, which is divided into two stages: pre-training and fine-tuning. The cost functions can be expressed as:

[0027]

[0028]

[0029] Among them, formula (5) is the cost function of the model pre-training stage, X represents the data matrix to be decomposed in the current layer in the pre-training stage, A represents the endmember spectrum matrix obtained by the decomposition of the current layer, S represents the abundance matrix obtained by the decomposition of the current layer, and formula (6) is the cost function of the model fine-tuning stage; among them, ||·|| F represents the Frobenius norm, ||·|| 1 / 2 Represents L 1 / 2 Sparse constraint, the calculation method is shown in formula (7), λ is the control L 1 / 2 Regularization parameter for the sparsity constraint; ||·|| HTV represents the spatial total variation constraint of the hyperspectral image, and its calculation method is shown in Equation (8) and Equation (9), α is the regularization parameter controlling the spatial total variation constraint; ||·|| 1,2 Represents L 1,2 Sparse constraint, the calculation method is shown in formula (10), β is the control impulse noise L 1,2 Regularization parameter for sparsity constraints;

[0030]

[0031]

[0032]

[0033]

[0034] In formula (8) Represents the operation of reconstructing a vector containing N pixels into an m×n matrix, where m represents the number of rows of the original hyperspectral image and n represents the number of columns of the original hyperspectral image.

[0035] Furthermore, the alternating direction multiplier method is used for iterative optimization. In the pre-training stage, appropriate auxiliary variables are introduced and variable separation is performed, and all variables are updated alternately using the gradient descent method to decompose the complex objective function into several simple objective functions. In the fine-tuning stage, all variables obtained in the previous stage are fine-tuned to reduce the total reconstruction error.

[0036] In order to facilitate the derivation of subsequent iterative formulas, the auxiliary matrix K is introduced to transform the cost function (5) of the pre-training stage and the cost function (6) of the fine-tuning stage into:

[0037]

[0038]

[0039] The matrix S can be regarded as a combination of the matrix K and the noise. For the convenience of calculation, the constraint K = S is included in the cost function. Then the cost function of the pre-training stage and the cost function of the fine-tuning stage become:

[0040]

[0041]

[0042] Among them, μ is the regularization parameter that controls the similarity between the auxiliary matrix K and the abundance matrix S. The larger its value, the higher the similarity of the two matrices;

[0043] By optimizing and solving the cost function (13) in the pre-training stage, the iterative optimization formulas of the end member matrix A, abundance matrix S, auxiliary matrix K and impulse noise matrix E of each layer during the layer-by-layer optimization in the pre-training stage can be obtained as follows:

[0044] A←A.*((XE)S T ). / (ASS T ) (15)

[0045]

[0046]

[0047] E←soft β (X-AS) (18)

[0048] Among them, soft in formula (18) represents the soft threshold, and its calculation method is shown in the following formula:

[0049]

[0050] Among them, E i,: represents the i-th row of the impulse noise matrix E, Q = X-AS, Q i,: represents the i-th row of the matrix Q, 0 is the 1×N zero vector;

[0051] Similarly, the cost function (14) of the fine-tuning stage is optimized and solved. For the convenience of expression, the end member matrix of the l-th layer reconstruction is recorded as The reconstructed abundance matrix is Remember l-1 =S l-1 …S2S1;

[0052] The end-member matrix A of each layer in the fine-tuning stage can be obtained l , abundance matrix S l , auxiliary matrix K l And the impulse noise matrix E l The iterative optimization formulas are as follows:

[0053]

[0054]

[0055]

[0056] E l ←soft β (XA l S l ) (twenty three)

[0057] After the iterative optimization is completed, the end member matrix A and abundance matrix S obtained by the final mixed pixel decomposition are obtained through the following corresponding relationship:

[0058]

[0059] The endmember deep non-negative matrix factorization method under mixed noise proposed in this paper is the first to propose endmember deep non-negative matrix factorization by reverse modeling of the spectral deep linear mixing model. Compared with the abundance deep non-negative matrix factorization, the endmember deep non-negative matrix factorization takes into account the physical meaning of the intermediate process of deep non-negative matrix factorization, starting from the physical process of spectral deep linear mixing, making the deep non-negative matrix factorization more interpretable at the physical level.

[0060] At the same time, in order to solve the problem of mixed noise pollution in hyperspectral images, the method proposed in this paper not only considers the influence of Gaussian noise, but also takes into account the impulse noise, and incorporates noise removal into the end-member deep non-negative matrix decomposition framework, and constructs a noise removal-spectral decomposition integrated model to perform mixed pixel decomposition. In addition, the model also takes into account the sparsity of the abundance matrix and the smoothness of the spatial segmentation, and applies L 1 / 2 Sparse constraint and spatial total variation constraint. The method proposed in this invention provides an important basis and support for the subsequent quantitative applications of hyperspectral images such as object classification, image enhancement, and target detection. Therefore, the end-member deep non-negative matrix decomposition method under mixed noise not only has very important academic value, but also has important practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a technical framework diagram of the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention more clear, the endmember deep non-negative matrix decomposition method under mixed noise according to an embodiment of the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0063] The present invention provides a method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise, comprising the following steps:

[0064] Step 1: Based on the given hyperspectral image, the spectral depth linear mixing process is reversely modeled to construct a hyperspectral image endmember depth non-negative matrix decomposition model;

[0065] Step 2, coupling the hyperspectral image observation model and the endmember deep non-negative matrix decomposition model to construct a noise removal-spectral decomposition integrated observation model;

[0066] Step 3, based on the optimization theory of the alternating direction multiplier method, by introducing appropriate auxiliary variables and performing variable separation, the complex non-convex integrated model objective function is decomposed into several convex sub-objective functions to achieve mixed pixel decomposition and obtain the end member matrix and abundance matrix.

[0067] Moreover, in step 1, the given hyperspectral image is recorded as Among them, each represents a pixel point with B bands, N is the number of pixels, ignoring the influence of noise, is the pure and noise-free data matrix of the hyperspectral image, is the pure endmember spectrum matrix, p is the number of endmembers, The two are linearly mixed for the first time to obtain matrix A2, which is regarded as the data matrix formed after the electromagnetic wave is reflected once on the surface; then the electromagnetic wave is reflected for the second time on the surface, and the end member matrix A2 at this time is linearly mixed with its corresponding abundance matrix S2 again to generate a new data matrix A3; this is repeated, when the electromagnetic wave is reflected l times on the surface, the end member matrix A at this time is l The corresponding abundance matrix S l Perform linear mixing to obtain the matrix A l+1 Until the electromagnetic wave reaches the hyperspectral sensor after L times of linear mixing, the noise-free hyperspectral image data matrix Y after the final deep linear mixing will be obtained. Then the spectrum depth linear mixing model obtained after the electromagnetic wave undergoes multiple reflections and other effects and performs deep linear mixing is:

[0068] Y=A1S1S2…S L (1)

[0069] The above-mentioned spectral depth linear mixing process is reversely simulated. In the first layer, the noise-free hyperspectral image data matrix Y is subjected to a non-negative matrix decomposition to obtain the endmember spectral matrix A1 after the first decomposition and its corresponding abundance matrix S1; due to the spectral depth linear mixing, the endmember spectral matrix obtained after the first decomposition is not pure enough, so a second decomposition is performed to obtain the endmember spectral matrix A2 after the second decomposition and its corresponding abundance matrix S2; this is repeated. When the first decomposition is performed, the endmember spectral matrix A obtained by the previous decomposition is l-1 Decomposed into a purer endmember spectrum matrix A l and its corresponding abundance matrix S l ; After L decompositions, it is considered that a pure end member spectrum matrix A is obtained. L and its corresponding abundance matrix S L The endmember deep non-negative matrix decomposition model of hyperspectral image is obtained as follows:

[0070] Y=A L S L S L-1 …S2S1 (2)

[0071] Moreover, in step 2, the hyperspectral image observation model is:

[0072] X=Y+N+E (3)

[0073] in, represents the Gaussian noise in the hyperspectral image, Represents the impulse noise in the hyperspectral image. Therefore, the noise removal-spectral decomposition integrated observation model is constructed as follows:

[0074] X=A L S L S L-1 …S2S1+N+E (4)

[0075] Moreover, in step 3, the noise removal-spectral decomposition integrated model adopts a cost function based on Euclidean distance, which is divided into two stages: pre-training and fine-tuning. The cost functions can be expressed as:

[0076]

[0077]

[0078] Wherein, formula (5) is the cost function of the model pre-training stage, X represents the data matrix to be decomposed in the current layer in the pre-training stage, A represents the endmember spectrum matrix obtained by decomposition of the current layer, S represents the abundance matrix obtained by decomposition of the current layer, and formula (6) is the cost function of the model fine-tuning stage. Wherein, ||•|| Frepresents the Frobenius norm, ||·|| 1 / 2 Represents L 1 / 2 Sparse constraint, the calculation method is shown in formula (7), λ is the control L 1 / 2 Regularization parameter for the sparsity constraint; ||·|| HTV represents the spatial total variation constraint of the hyperspectral image, and its calculation method is shown in Equation (8) and Equation (9), α is the regularization parameter controlling the spatial total variation constraint; ||·|| 1,2 Represents L 1,2 Sparse constraint, the calculation method is shown in formula (10), β is the control impulse noise L 1,2 Regularization parameter for the sparsity constraint.

[0079]

[0080]

[0081]

[0082]

[0083] In formula (8) Represents the operation of reconstructing a vector containing N pixels into an m×n matrix, where m represents the number of rows of the original hyperspectral image and n represents the number of columns of the original hyperspectral image.

[0084] Hyperspectral remote sensing images have many bands and large data volumes, so the alternating direction multiplier method is used for iterative optimization. In the pre-training stage, by introducing appropriate auxiliary variables and performing variable separation, all variables are updated alternately using the gradient descent method to decompose the complex objective function into several simple objective functions; in the fine-tuning stage, all variables obtained in the previous stage are fine-tuned to reduce the total reconstruction error.

[0085] In order to facilitate the derivation of subsequent iterative formulas, the auxiliary matrix K is introduced to transform the cost function (5) of the pre-training stage and the cost function (6) of the fine-tuning stage into:

[0086]

[0087]

[0088] The matrix S can be regarded as a combination of the matrix K and the noise. For the convenience of calculation, the constraint K = S is included in the cost function. Then the cost function of the pre-training stage and the cost function of the fine-tuning stage become:

[0089]

[0090]

[0091] Among them, μ is the regularization parameter that controls the similarity between the auxiliary matrix K and the abundance matrix S. The larger its value, the higher the similarity of the two matrices.

[0092] By optimizing and solving the cost function (13) in the pre-training stage, the iterative optimization formulas of the end member matrix A, abundance matrix S, auxiliary matrix K and impulse noise matrix E of each layer during the layer-by-layer optimization in the pre-training stage can be obtained as follows:

[0093] A←A.*((XE)S T ). / (ASS T ) (15)

[0094]

[0095]

[0096] E←soft β (X-AS) (18)

[0097] Among them, soft in formula (18) represents the soft threshold, and its calculation method is shown in the following formula:

[0098]

[0099] Among them, E i,: represents the i-th row of the impulse noise matrix E, Q = X-AS, Q i,: Represents the i-th row of matrix Q, and 0 is a 1×N zero vector.

[0100] Similarly, the cost function (14) of the fine-tuning stage is optimized and solved. For the convenience of expression, the end member matrix of the l-th layer reconstruction is recorded as The reconstructed abundance matrix is Remember l-1 =S l-1 …S2S1.

[0101] The end-member matrix A of each layer in the fine-tuning stage can be obtained l , abundance matrix S l , auxiliary matrix K l And the impulse noise matrix E l The iterative optimization formulas are as follows:

[0102]

[0103]

[0104]

[0105] E l ←softβ (XA l S l ) (twenty three)

[0106] After the iterative optimization is completed, the end member matrix A and abundance matrix S obtained by the final mixed pixel decomposition are obtained through the following corresponding relationship:

[0107] A=A L (twenty four)

[0108] S=S L S L-1 ...S l ...S2S1

[0109] The present invention aims at the robust performance of the deep non-negative matrix decomposition model against noise, uses the powerful inclusiveness of non-negative matrix decomposition, couples the hyperspectral image observation model and the end-member deep non-negative matrix decomposition model, establishes a noise removal-spectral decomposition integrated observation model, and establishes a theoretical framework for integrated processing based on the Bayesian inverse problem solving theory, so as to achieve a common description of the two processing tasks. In addition, the present invention fully mines the hierarchical feature information of the image by reversely modeling the spectral deep linear mixed model, thereby enhancing the practicality and interpretability of the model.

[0110] The present invention can be implemented by computer software technology. Figure 1 The specific steps of the method for deep non-negative matrix decomposition of image end members in the embodiment are described in detail.

[0111] Step 1: Based on the given hyperspectral image, a more accurate deep modeling of the mixing process of the spectral signals of the ground objects is performed. By back-propagating the model layer by layer, the hierarchical feature information in the hyperspectral image is extracted, and a deep non-negative matrix decomposition model of the hyperspectral image endmembers is constructed.

[0112] The endmember deep non-negative matrix decomposition method under mixed noise proposed in the present invention reversely establishes an endmember deep non-negative matrix decomposition model by simulating a spectral deep linear mixed model. In the embodiment, a cost function based on Euclidean distance is adopted, and based on the Bayesian inverse problem solving theory, an alternating direction multiplier method is used for optimization.

[0113] Step 2: Construct the corresponding noise removal-spectral decomposition integrated observation model, input the hyperspectral image X into the model, and then set the model parameters, iteratively optimize according to the cost function until the model converges or reaches the maximum number of iterations.

[0114] The image size in the embodiment is 100×100, containing a total of 198 bands. There are 4 types of endmembers in the image. In the specific implementation, those skilled in the art can use the endmember number estimation algorithm to determine the number of endmembers in the image. Since the initialization of the matrix in the endmember depth non-negative matrix decomposition model must have a certain physical meaning, the vertex component analysis method is used in the embodiment to extract the endmember matrix, and the fully constrained least squares method is used to estimate the abundance matrix as the initialization of the non-negative matrix decomposition. In the pre-training stage of the model, this embodiment adopts the alternating direction multiplier method optimization theory, introduces appropriate auxiliary variables and performs variable separation, and uses the gradient descent method to alternately update all variables; in the fine-tuning stage of the model, all variables obtained in the pre-training stage are fine-tuned as a whole to reduce the total reconstruction error. The abundance sparse regularization parameter λ is set to 0.2, the spatial total variation regularization parameter α is set to 0.05, the impulse noise sparse regularization parameter β is set to 0.005, and the maximum number of iterations of the model is 500. In the specific implementation, those skilled in the art can select near the empirical value according to the specific image used.

[0115] Step 3: After the model converges or reaches the maximum number of iterations, Figure 1 Integrate the obtained variables and finally get the end member matrix A = A L , the abundance matrix S = S L S L-1 ...S l ...S2S1, realize mixed pixel decomposition.

[0116] Those skilled in the art can understand that the present invention can not only break through the limitations of traditional mixed pixel decomposition methods by performing scientific mathematical modeling of singular noise points in hyperspectral images under the influence of mixed noise, but also fully explore the deep feature information and spatial information of the image in the process of deep decomposition, provide technical support for high-precision mixed pixel decomposition, and effectively enhance the application potential of hyperspectral images.

[0117] It should be noted and understood that the above is only a preferred embodiment of the present invention, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations can be made to the present invention described in detail above without departing from the spirit and principle of the present invention required by the attached claims, all of which should be within the scope of protection of the present invention. Therefore, the scope of the technical solution claimed is not limited by any specific exemplary teaching given.

Claims

1. A method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise, characterized in that: The following steps are involved: Step 1: Based on the given hyperspectral image, the spectral depth linear mixing process is reversely modeled to construct a hyperspectral image endmember depth non-negative matrix decomposition model; Step 2, coupling the hyperspectral image observation model and the endmember deep non-negative matrix decomposition model to construct a noise removal-spectral decomposition integrated observation model; Step 3, construct the cost function of the noise removal-spectral decomposition integrated observation model, and based on the optimization theory of the alternating direction multiplier method, introduce auxiliary variables and perform variable separation to decompose the complex non-convex integrated observation model cost function into several convex sub-objective functions, realize mixed pixel decomposition, and obtain the end member matrix and abundance matrix; In step 3, the noise removal-spectral decomposition integrated observation model adopts a cost function based on Euclidean distance, which is divided into two stages: pre-training and fine-tuning. The cost functions can be expressed as: Wherein, formula (5) is the cost function of the model pre-training stage, X represents the data matrix to be decomposed in the current layer in the pre-training stage, A represents the endmember spectrum matrix decomposed in the current layer, S represents the abundance matrix decomposed in the current layer, and formula (6) is the cost function of the model fine-tuning stage; where, ‖·‖ F represents the Frobenius norm, ||·|| 1 / 2 Represents L 1 / 2 Sparse constraint, the calculation method is shown in formula (7), λ is the control L 1 / 2 Regularization parameter for the sparsity constraint; ||·|| HTV represents the spatial total variation constraint of the hyperspectral image, and its calculation method is shown in Equation (8) and Equation (9), α is the regularization parameter controlling the spatial total variation constraint; ||·|| 1,2 Represents L 1,2 Sparse constraint, the calculation method is shown in formula (10), β is the control impulse noise L 1,2 Regularization parameter for sparsity constraints; In formula (8) Represents the operation of reconstructing a vector containing N pixels into an m×n matrix, where m represents the number of rows of the original hyperspectral image and n represents the number of columns of the original hyperspectral image.

2. The method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise according to claim 1, characterized in that: In step 1, the spectral depth linear mixing process is as follows; The given hyperspectral image is denoted as Among them, each represents a pixel point with B bands, N is the number of pixels, ignoring the influence of noise, is the pure and noise-free data matrix of the hyperspectral image, is the pure endmember spectrum matrix, p is the number of endmembers, The two are linearly mixed for the first time to obtain matrix A2, which is regarded as the data matrix formed after the electromagnetic wave is reflected once on the surface; then the electromagnetic wave is reflected for the second time on the surface, and the end member matrix A2 at this time is linearly mixed with its corresponding abundance matrix S2 again to generate a new data matrix A3; this is repeated, when the electromagnetic wave is reflected l times on the surface, the end member matrix A at this time is l The corresponding abundance matrix S l Perform linear mixing to obtain the matrix A l+1 ; Until the electromagnetic wave reaches the hyperspectral sensor after L times of linear mixing, the noise-free hyperspectral image data matrix Y after the final deep linear mixing will be obtained. Then the spectrum depth linear mixing model obtained after the electromagnetic wave is deeply linearly mixed after multiple reflections is: Y=A1S1S2…S L (1)。 3. The method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise according to claim 2, characterized in that: In step 1, the spectral depth linear mixing process is reversely simulated. In the first layer, the noise-free hyperspectral image data matrix Y is subjected to a non-negative matrix decomposition to obtain the endmember spectral matrix A1 after the first decomposition and its corresponding abundance matrix S1; due to the spectral depth linear mixing, the endmember spectral matrix obtained after the first decomposition is not pure enough, so a second decomposition is performed to obtain the endmember spectral matrix A2 after the second decomposition and its corresponding abundance matrix S2; this is repeated. When the first decomposition is performed, the endmember spectral matrix A obtained by the previous decomposition is l-1 Decomposed into a purer endmember spectrum matrix A l and its corresponding abundance matrix S l ; After L decompositions, it is considered that a pure end member spectrum matrix A is obtained. L and its corresponding abundance matrix S L , then the endmember deep non-negative matrix decomposition model of hyperspectral image is obtained as: Y=A L S L S L-1 …S2S1 (2)。 4. The method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise according to claim 3, characterized in that: In step 2, the hyperspectral image observation model is: X=Y+N+E (3) in, represents the Gaussian noise in the hyperspectral image, represents the impulse noise in the hyperspectral image, Therefore, the noise removal-spectral decomposition integrated observation model constructed is: X=A L S L S L-1 …S2S1+N+E (4)。 5. The method for endmember deep non-negative matrix decomposition of hyperspectral images under mixed noise according to claim 4, characterized in that: The alternating direction multiplier method is used for iterative optimization. In the pre-training stage, appropriate auxiliary variables are introduced and variable separation is performed. All variables are updated alternately using the gradient descent method to decompose the complex objective function into several simple objective functions. In the fine-tuning stage, all variables obtained in the previous stage are fine-tuned to reduce the total reconstruction error. In order to facilitate the derivation of subsequent iterative formulas, the auxiliary matrix K is introduced to transform the cost function (5) of the pre-training stage and the cost function (6) of the fine-tuning stage into: The matrix S can be regarded as a combination of the matrix K and the noise. For the convenience of calculation, the constraint K = S is included in the cost function. Then the cost function of the pre-training stage and the cost function of the fine-tuning stage become: Among them, μ is the regularization parameter that controls the similarity between the auxiliary matrix K and the abundance matrix S. The larger its value, the higher the similarity of the two matrices; By optimizing and solving the cost function (13) in the pre-training stage, the iterative optimization formulas of the end member matrix A, abundance matrix S, auxiliary matrix K and impulse noise matrix E of each layer during the layer-by-layer optimization in the pre-training stage can be obtained as follows: A←A.*((X-E)S T ). / (ASS T )(15) E←soft β (X-AS)(18) Among them, soft in formula (18) represents the soft threshold, and its calculation method is shown in the following formula: Among them, E i,: represents the i-th row of the impulse noise matrix E, Q = X-AS, Q i,: represents the i-th row of the matrix Q, 0 is the 1×N zero vector; Similarly, the cost function (14) of the fine-tuning stage is optimized and solved. For the convenience of expression, the end member matrix of the l-th layer reconstruction is recorded as The reconstructed abundance matrix is Remember l-1 =S l-1 …S2S1; The end-member matrix A of each layer in the fine-tuning stage can be obtained l , abundance matrix S l , auxiliary matrix K l And the impulse noise matrix E l The iterative optimization formulas are as follows: E l ←soft β (X-A l S l )(23) After the iterative optimization is completed, the end member matrix A and abundance matrix S obtained by the final mixed pixel decomposition are obtained through the following corresponding relationship: