A method and system for rapid classification of sub-pixel objects in hyperspectral images

Abstract

The present invention proposes a method and system for rapid classification of sub-pixel level objects in hyperspectral images, comprising: processing the hyperspectral image, and constructing an endmember matrix and an abundance matrix for the object proportion corresponding to each endmember based on the object type in the hyperspectral image; taking the abundance distribution discreteness model constructed in advance based on the object type as a constraint condition, bringing the hyperspectral image into a pre-set objective function for iterative solution to obtain the endmember matrix and the abundance matrix; classifying the hyperspectral image based on the endmember matrix and the abundance matrix; the present invention uses the abundance distribution discreteness model as a constraint for an image classification method of non-negative matrix decomposition, which can effectively alleviate local extreme values ​​and improve the accuracy of object classification; at the same time, the present invention uses a quadratic function to characterize the abundance distribution discreteness, so that the algorithm iterative optimization process can support matrix multiplication iterative update, does not need to iteratively calculate pixel by pixel, and reduces the algorithm processing time.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing image processing, and in particular to a method and system for rapid classification of sub-pixel-level objects in hyperspectral images. Background Art

[0002] Pixel, also known as pixel point or pixel point. It is an image unit (picture element). It is the smallest unit that makes up a digital image. In remote sensing data acquisition, such as scanning imaging, it is the smallest unit for the sensor to scan and sample the ground scene; in digital image processing, it is the sampling point when scanning and digitizing analog images. It is the basic unit that constitutes remote sensing digital images and the sampling point in the remote sensing imaging process. Hyperspectral remote sensing is a kind of data that can simultaneously obtain the two-dimensional spatial information and spectral feature information of the ground object. The unification of the spectrum is an important feature of the hyperspectral image. The images of all bands are superimposed together to form a hyperspectral image data cube. Each pixel is a continuous spectral curve, and the points on the curve are the radiation values ​​on the images of each band. Therefore, in hyperspectral image processing, the pixel and ground object analysis of the hyperspectral image is mainly carried out from the two dimensions of image space and spectral space.

[0003] The ground reflection or emission spectral signals acquired by the remote sensor are recorded in pixels. A pixel that contains only one type is called a pure pixel. However, due to the complexity and diversity of imaging objects in nature, hyperspectral images are limited by the resolution of the sensor. The area corresponding to each pixel often contains different types of objects with different spectral response characteristics. A pixel that contains multiple types of surfaces is called a mixed pixel. Mixed pixels are commonly found in hyperspectral remote sensing images. When there are mixed pixels in a hyperspectral image, it is not rigorous to use conventional classification methods to forcibly classify the mixed pixels into a certain type of target. Conventional methods are difficult to meet the requirements of accurate quantitative analysis of hyperspectral images, which brings huge challenges to the fine classification and recognition of hyperspectral images.

[0004] The solution to the problem of mixed pixels for hyperspectral image classification is sub-pixel level classification technology, which decomposes mixed pixels to sub-pixel level. This process is also called mixed pixel analysis. Mixed pixel analysis decomposes the image to sub-pixel level, and uses geometric, statistical, modeling and other methods to decompose the different objects contained in the mixed pixels. Each type of object is called an end member, and the composition of each type of object, that is, the proportion it occupies, is called abundance. The mixed pixel analysis process mainly includes several links such as feature extraction (FeatureExtraction), end member extraction (Endmember Extraction), and abundance unmixing (unmixing) based on the mixed model. Using hyperspectral mixed pixel analysis technology, the proportion of each type of object in the mixed pixel can be obtained through the abundance unmixing of the end member, and the composition of each type of object in the pixel can be determined, which can reduce the possibility of image misclassification, thereby improving the accuracy of hyperspectral image object classification.

[0005] In recent years, a new non-negative matrix factorization blind signal separation method has been applied to the analysis of mixed pixels in hyperspectral remote sensing images, and has become an important branch of the research direction of hyperspectral image classification. The non-negative matrix factorization endmember extraction algorithm uses matrix operations to decompose the high-dimensional hyperspectral image data matrix into a low-dimensional endmember matrix and a corresponding abundance matrix. The algorithm retains the structure and characteristics of the endmember matrix itself, and compresses the data volume of the high-dimensional hyperspectral image original data, which can effectively improve the data processing speed. The unconstrained non-negative matrix factorization endmember solution process can support the matrix multiplication iterative rapid update of the objective function, and finally obtain the endmember matrix and abundance matrix at the same time. However, since the objective function is a non-convex function, this method has the problem of non-unique solution and the algorithm iteration process is prone to fall into local extreme values.

[0006] In response to the local extremum problem of non-negative matrix decomposition methods, scholars at home and abroad usually adopt the solution of adding constraints to make the non-negative matrix decomposition end-member iteration process converge to the real end-member feature space direction to alleviate the local extremum problem of the objective function. However, this often destroys the multiplicative iterative optimization rule of the non-negative matrix decomposition method, and the second best method is to use pixel-by-pixel iterative update to optimize the objective function, such as gradient descent method or projected gradient method. For large-scale hyperspectral images, the pixel-by-pixel iterative optimization process has slow calculation speed and low calculation efficiency, which is difficult to meet the actual use requirements.

[0007] Therefore, how to effectively extract the end members of the ground objects through effective constraints and maintain the multiplicative iterative update rules of the iterative process of end member generation is the focus of the present invention. Summary of the invention

[0008] The purpose of the present invention is to provide a method for rapid classification of sub-pixel-level objects in hyperspectral images, which solves the problem that the sub-pixel hierarchical classification method based on pixel-by-pixel iterative optimization has slow calculation speed and low calculation efficiency in processing large-size hyperspectral images, and is difficult to meet actual use requirements.

[0009] The present invention provides a technical solution, including:

[0010] Process the hyperspectral image and construct the endmember matrix and abundance matrix based on the types of objects in the hyperspectral image;

[0011] Taking the pre-constructed abundance distribution dispersion model as a constraint condition, the hyperspectral image is brought into a pre-set objective function for iterative solution to obtain an end member matrix and an abundance matrix;

[0012] Classifying the hyperspectral image based on the endmember matrix and the abundance matrix;

[0013] Among them, the endmember matrix is ​​constructed based on the type of land objects, and the abundance matrix is ​​the proportion of land objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution values ​​of all endmembers in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix.

[0014] Preferably, the processing of the hyperspectral image and constructing the endmember matrix and the abundance matrix based on the types of land objects in the hyperspectral image include:

[0015] Set the number of end members p based on the number of feature types;

[0016] The hyperspectral image is subjected to noise whitening processing using MNF forward and inverse transformation;

[0017] Take the first p-1 bands of the hyperspectral image after noise whitening and perform inverse transformation to obtain the hyperspectral image R;

[0018] Randomly select p points in the hyperspectral image R as the initial endmember matrix E0, and set the initial abundance matrix C0=0.

[0019] Preferably, the pre-constructed abundance distribution dispersion model is used as a constraint condition, and the hyperspectral image is brought into a pre-set objective function for iterative solution to obtain an end member matrix and an abundance matrix, including:

[0020] Calculate the partial differential value of the objective function;

[0021] The objective function is iteratively solved based on the multiplicative iterative optimization iterative rule to obtain the end member matrix and abundance matrix;

[0022] The final end member matrix E and abundance matrix C are obtained through iterative operations, and the ground object composition of each pixel is obtained;

[0023] The iteration condition is as follows: the maximum number of iterations does not exceed 200.

[0024] Preferably, it is characterized in that: the calculation formula of the objective function is as follows:

[0025]

[0026] In the formula, J C (C) is the penalty function for the dispersion of abundance distribution, is the error, E is the endmember matrix, C is the abundance matrix, and α is the regularization parameter used to balance the model error value and abundance distribution constraints.

[0027] Preferably: the constraints of the objective function are as follows:

[0028] E≥0, C≥0.

[0029] Preferably: the penalty function of the abundance distribution dispersion is:

[0030] J C (C)=trace(C T C)

[0031] In the formula, J C (C) is a quadratic function, representing the abundance matrix C T The sum of all non-negative eigenvalues ​​of C.

[0032] Preferably: the calculation formula of the multiplication iterative optimization iterative rule is:

[0033]

[0034] Where α is the iteration step size of the matrix multiplication iterative rule.

[0035] Based on the same inventive concept, the present invention also provides a hyperspectral image sub-pixel level rapid classification system for ground objects, comprising:

[0036] An initialization module is used to process the hyperspectral image and construct an end member matrix and an abundance matrix based on the types of objects in the hyperspectral image;

[0037] A matrix calculation module is used to use a pre-constructed abundance distribution dispersion model as a constraint condition, bring the hyperspectral image into a pre-set objective function for iterative solution, and obtain an end member matrix and an abundance matrix;

[0038] A ground object classification module, used for classifying the ground objects in the hyperspectral image based on the endmember matrix and the abundance matrix;

[0039] Among them, the endmember matrix is ​​constructed based on the type of land objects, and the abundance matrix is ​​the proportion of land objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution values ​​of all endmembers in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix.

[0040] Preferably, the initialization module includes:

[0041] The dimensionality reduction processing submodule is used to set the number of end members p based on the number of ground object types; it is also used to perform noise whitening processing on the hyperspectral image using MNF forward and inverse transformation; it is also used to take the first p-1 bands of the hyperspectral image after noise whitening processing and perform inverse transformation to obtain the hyperspectral image R;

[0042] The endmember matrix initialization submodule is used to randomly select p points in the hyperspectral image R as the initial endmember matrix E0;

[0043] The abundance matrix initialization module is used to set the initial abundance matrix C0=0.

[0044] Preferably, the matrix calculation module includes:

[0045] The objective function calculation submodule is used to calculate the partial differential value of the objective function;

[0046] The matrix calculation submodule is used to iteratively solve the objective function based on the multiplication iterative optimization iterative rule to obtain the end member matrix and the abundance matrix; wherein the iterative condition is: the maximum number of iterations does not exceed 200

[0047] The feature classification submodule is used to obtain the final end member matrix E and abundance matrix C through iterative operations to obtain the feature composition of each pixel.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention proposes a method and system for rapid classification of sub-pixel-level objects in hyperspectral images, comprising: processing a hyperspectral image, and constructing an endmember matrix and an abundance matrix based on the object type in the hyperspectral image; taking a pre-constructed abundance distribution dispersion model as a constraint condition, bringing the hyperspectral image into a pre-set objective function for iterative solution to obtain an endmember matrix and an abundance matrix; classifying objects in the hyperspectral image based on the endmember matrix and the abundance matrix; wherein the endmember matrix is ​​constructed based on the object type, and the abundance matrix is ​​the proportion of objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution value distribution of all endmembers existing in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix, wherein the abundance distribution dispersion model is used as a constraint in the image classification method of non-negative matrix decomposition, which can effectively alleviate local extreme values, improve the accuracy of object classification, and improve computational efficiency;

[0050] The technical solution provided by the present invention uses a quadratic function to characterize the discreteness of the abundance distribution, so that the algorithm iterative optimization process can support matrix multiplication iterative updates, without the need for pixel-by-pixel iterative calculations, thereby reducing the algorithm processing time. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flow chart of the method for rapid sub-pixel classification of hyperspectral images provided by the present invention;

[0052] Figure 2 A schematic diagram of a process of a sub-pixel-level rapid classification method of a hyperspectral image based on abundance distribution discreteness constraint of the present invention;

[0053] Figure 3 This is a structural diagram of the hyperspectral image sub-pixel level rapid classification system provided by the present invention. DETAILED DESCRIPTION

[0054] The present invention discloses a method for sub-pixel-level classification of hyperspectral images. The key points of the technical solution include the following steps: ①. Preprocessing the hyperspectral image and initializing the endmember matrix and the abundance matrix; ②. Optimizing the construction method of the abundance distribution dispersion model of the hyperspectral image, simplifying the characterization of the distribution dispersion, and being used for fast optimization and iteration of hyperspectral image endmember extraction; ③. Further proposing an objective function based on the abundance distribution dispersion constraint, and using the abundance distribution dispersion model as a constraint for the image classification method of non-negative matrix decomposition; ④. Further optimizing the iterative method of image classification based on the abundance distribution constraint non-negative matrix decomposition, so that the method can converge to the real endmember through the distribution dispersion constraint, and can also maintain the matrix multiplication iterative update rule to improve the convergence speed of the algorithm. The present invention obtains a fast and effective hyperspectral image sub-pixel-level classification solution for large-size hyperspectral images.

[0055] The abundance distribution discreteness constraint adopted by the present invention is a feature only related to the abundance distribution situation, the feature characterization formula has low calculation complexity, and the abundance distribution discreteness is used as a constraint to classify land objects.

[0056] In order to better understand the present invention, the content of the present invention is further described below in conjunction with the accompanying drawings and examples.

[0057] Example 1

[0058] like Figure 1 As shown, the present invention provides a method for rapid classification of sub-pixel-level objects in hyperspectral images, comprising:

[0059] S1, processing the hyperspectral image, and constructing an endmember matrix and an abundance matrix based on the type of land objects in the hyperspectral image; wherein the endmember matrix is ​​composed of spectral vectors corresponding to all endmembers, each endmember corresponds to a land object type, and each spectral vector corresponds to a spectrum; each element in the abundance matrix corresponds to an endmember, and the value of each element in each abundance matrix is ​​the proportion of the land object;

[0060] S2, using a pre-constructed abundance distribution dispersion model as a constraint condition, bringing the hyperspectral image into a pre-set objective function for iterative solution to obtain an end member matrix and an abundance matrix;

[0061] S3, classifying the hyperspectral image based on the endmember matrix and the abundance matrix;

[0062] Among them, the endmember matrix is ​​constructed based on the type of land objects, and the abundance matrix is ​​the proportion of land objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution values ​​of all endmembers in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix.

[0063] The specific process is as follows Figure 2 As shown,

[0064] The hyperspectral image preprocessing in the step s1 is to preset the number of end members p, and the original hyperspectral image is subjected to noise whitening processing using MNF forward and inverse transformations. The first p-1 bands are taken for inverse transformation to obtain a noise whitened image, and the hyperspectral image is reduced to p-1 dimensions as the hyperspectral image R.

[0065] The method for initializing the endmember matrix and the abundance matrix in step s1 is to randomly select p points in the hyperspectral image R as initial endmembers E0, and set the initial abundance matrix C0=0.

[0066] The step s2 uses the pre-constructed abundance distribution dispersion model as a constraint condition, brings the hyperspectral image into a pre-set objective function for iterative solution, and obtains the end member matrix and abundance matrix including:

[0067] S2-1 optimized the method for constructing the abundance distribution dispersion model of hyperspectral images, simplified the representation of distribution dispersion, and can be used for fast optimization and iteration of hyperspectral image endmember extraction;

[0068] S2-2 further proposed an objective function based on the abundance distribution discreteness constraint, and used the abundance distribution discreteness model as a constraint for the image classification method of non-negative matrix decomposition; the representation method has low computational complexity and high algorithm computational efficiency;

[0069] S2-3 further optimizes the iterative method of image classification based on abundance distribution constrained non-negative matrix decomposition, so that the method can converge to the true end member through the distribution discreteness constraint, and maintain the matrix multiplication iterative update rule to improve the convergence speed of the algorithm, and obtains a fast and effective hyperspectral image sub-pixel level classification solution for large-size hyperspectral images.

[0070] The hyperspectral image abundance distribution dispersion model in step S2-1 is the divergence of the hyperspectral image abundance matrix C trace.

[0071] The hyperspectral image abundance distribution dispersion model J in step S2-2 is C The construction method of (C) is: C (C)=trace(C T C), J C (C) is a quadratic function, representing the matrix C T The sum of all non-negative eigenvalues ​​of C.

[0072] Non-negative matrix decomposition refers to the process of decomposing any given non-negative matrix into the product of two non-negative matrices. This method has good expression ability for the characteristics and structure of the original data, can reflect the most intuitive feeling of the whole composed of parts, and is conducive to discovering the potential local structure in the original data. Therefore, the algorithm has strong practicality and is widely used in pattern recognition and image processing. According to the matrix representation of the linear spectral mixture model, the non-negative matrix decomposition problem can be described as: given a non-negative matrix R, find a non-negative matrix E and a non-negative matrix C such that

[0073] R≈EC

[0074] Where: R is the sub-pixel level ground feature pixel set obtained above

[0075] The matrix of E represents the end member matrix, and C is the abundance distribution matrix corresponding to the pixel. The non-negative matrix decomposition problem is essentially an optimization problem.

[0076] The objective function based on the abundance distribution discreteness constraint in step S2-3 is:

[0077]

[0078] Among them J C (C)=trace(C T C) is the penalty function for the dispersion of abundance distribution, is the model error of the minimization model based on the matrix non-negative constraint, E is the end member matrix, C is the abundance matrix, and α is the regularization parameter used to balance the model error value and the abundance distribution constraint. The present invention uses the abundance distribution dispersion as a constraint to classify the land objects. The abundance distribution dispersion constraint is only related to the features of the abundance distribution, and the feature characterization formula has low calculation complexity.

[0079] Here, the endmember matrix is ​​a matrix of size M×p, and the abundance matrix is ​​a matrix of size pXN, where M is the spectral band of the hyperspectral image, N is the number of pixels of the hyperspectral image, and p is the number of endmembers.

[0080] The sub-pixel level rapid classification method of hyperspectral images based on abundance distribution discreteness constraint, wherein the multiplicative iterative method for image classification based on abundance distribution constraint non-negative matrix decomposition in step S2-3 is: a process of finding E and C by an iterative optimization method so that f(E,C) reaches a minimum value.

[0081] The multiplicative iterative method for image classification based on abundance distribution constrained non-negative matrix decomposition in step S2-3 comprises the following steps:

[0082] A) Calculate the partial differential value of the objective function

[0083] B) The matrix multiplication iteration method is used to perform the objective function optimization iteration, and the objective function multiplication iteration optimization iteration rule is generated as follows:

[0084]

[0085] Wherein α is the iterative step length of the matrix multiplication iterative rule. In the present invention, α is selected as 0.01, which can realize the multiplication iterative update under the conditions of the sum being unity and non-negative total constraints.

[0086] C) Set the iterative calculation termination condition to the maximum number of iterations condition, and the selected maximum number of iterations is 200.

[0087] D) Considering and as a constraint, replace the matrices R and E in the formulas in (A) and (B) in step d) with and Choose σ = 20;

[0088] E) When the iteration termination condition is met, the iteration is terminated, and the final end member matrix E and abundance matrix C are obtained through iterative operations, and the ground feature composition of each pixel is obtained. The multiplicative iterative update under the conditions of unity and non-negative full constraints can be realized, and the processing efficiency is higher.

[0089] Example 2

[0090] In order to implement the above method, the present invention also provides a hyperspectral image sub-pixel level rapid classification system. Figure 3 As shown, it contains the following modules:

[0091] 1. Initialization module, used to process the hyperspectral image and construct the end member matrix and abundance matrix based on the ground object type in the hyperspectral image;

[0092] 2. A matrix calculation module, which is used to use a pre-constructed abundance distribution dispersion model as a constraint condition, bring the hyperspectral image into a pre-set objective function for iterative solution, and obtain an end member matrix and an abundance matrix;

[0093] 3. A ground object classification module, used for classifying the ground objects in the hyperspectral image based on the end member matrix and the abundance matrix;

[0094] Among them, the endmember matrix is ​​constructed based on the type of land objects, and the abundance matrix is ​​the proportion of land objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution values ​​of all endmembers in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix.

[0095] In practical applications of sub-pixel classification methods for hyperspectral images, certain requirements are placed on the processing efficiency of large-size hyperspectral images. Based on this, the system of the present invention completes the design and compilation of the software and realizes the rapid and accurate classification of hyperspectral images.

[0096] The initialization module includes:

[0097] The dimensionality reduction processing submodule is used to set the number of end members p based on the number of ground object types; it is also used to perform noise whitening processing on the hyperspectral image using MNF forward and inverse transformation; it is also used to take the first p-1 bands of the hyperspectral image after noise whitening processing and perform inverse transformation to obtain the hyperspectral image R;

[0098] The endmember matrix initialization submodule is used to randomly select p points in the hyperspectral image R as the initial endmember matrix E0;

[0099] The abundance matrix initialization module is used to set the initial abundance matrix C0=0.

[0100] Matrix calculation module, including:

[0101] The objective function calculation submodule is used to calculate the partial differential value of the objective function;

[0102] The matrix calculation submodule is used to iteratively solve the objective function based on the multiplication iterative optimization iterative rule to obtain the end member matrix and the abundance matrix; wherein the iterative condition is: the maximum number of iterations does not exceed 200

[0103] The feature classification submodule is used to obtain the final end member matrix E and abundance matrix C through iterative operations to obtain the feature composition of each pixel.

[0104] The hyperspectral image preprocessing in the dimensionality reduction processing submodule is a preset end member number p, the original hyperspectral image is subjected to noise whitening processing using MNF forward and inverse transformation, the first p-1 bands are taken for inverse transformation, a noise whitening image is obtained, and the hyperspectral image is reduced to p-1 dimensions as the hyperspectral image R. The abundance distribution discreteness constraint adopted by the present invention is a feature only related to the abundance distribution, the feature characterization formula has low computational complexity, and the abundance distribution discreteness is used as a constraint for object classification.

[0105] The method for initializing the endmember matrix and the abundance matrix in the endmember matrix initialization submodule is to randomly select p points in the hyperspectral image R as initial endmembers E0, and set the initial abundance matrix C0=0.

[0106] The hyperspectral image abundance distribution dispersion model in the matrix calculation module is the divergence of the hyperspectral image abundance matrix C trace.

[0107] The hyperspectral image abundance distribution dispersion model in the matrix calculation module is C The construction method of (C) is: C (C)=trace(C T C), J C (C) is a quadratic function, representing the matrix C T The sum of all non-negative eigenvalues ​​of C.

[0108] The objective function based on the abundance distribution discreteness constraint in the matrix calculation module is:

[0109]

[0110] Among them J C (C)=trace(C T C) is the penalty function for the dispersion of abundance distribution, is the model error of the minimization model based on the matrix non-negative constraint, E is the end member matrix, C is the abundance matrix, and α is the regularization parameter used to balance the model error value and the abundance distribution constraint. The present invention uses the abundance distribution dispersion as a constraint to classify the land objects. The abundance distribution dispersion constraint is only related to the features of the abundance distribution, and the feature characterization formula has low calculation complexity.

[0111] The multiplicative iterative method for image classification based on abundance distribution constrained non-negative matrix decomposition in the matrix calculation module is: a process of finding E and C through an iterative optimization method so that f(E, C) reaches a minimum value.

[0112] The multiplicative iterative method for image classification based on abundance distribution constrained non-negative matrix decomposition in the matrix calculation module comprises the following steps:

[0113] A) Calculate the partial differential value of the objective function

[0114] B) The matrix multiplication iteration method is used to perform the objective function optimization iteration, and the objective function multiplication iteration optimization iteration rule is generated as follows:

[0115]

[0116] Wherein α is the iterative step length of the matrix multiplication iterative rule. In the present invention, α is selected as 0.01, which can realize the multiplication iterative update under the conditions of the sum being unity and non-negative total constraints.

[0117] C) Set the iterative calculation termination condition to the maximum number of iterations condition, and the selected maximum number of iterations is 200.

[0118] D) Considering that and are a constraint, replace the matrices R and E in the formulas in (A) and (B) with and Choose σ = 20;

[0119] E) When the iteration termination condition is met, the iteration is terminated, and the final end member matrix E and abundance matrix C are obtained through iterative operations, and the ground feature composition of each pixel is obtained. The multiplicative iterative update under the conditions of unity and non-negative full constraints can be realized, and the processing efficiency is higher.

[0120] The present invention found that there is a positive proportional relationship between the discrete degree of end-member abundance distribution and the end-member extraction result in the study of hierarchical classification of hyperspectral images. Based on this positive proportional relationship, the present invention proposes a sub-pixel-level rapid classification method for hyperspectral images based on the constraint of abundance distribution discreteness, and introduces the abundance distribution discreteness as a constraint into the non-negative matrix decomposition method. The introduction of this constraint can effectively alleviate local extreme values ​​and improve the accuracy of ground object classification. In addition, the present invention uses a quadratic function to characterize the abundance distribution discreteness, so that the algorithm iterative optimization process can support matrix multiplication iterative updates, without the need for iterative calculation pixel by pixel, and reduces the algorithm processing time.

[0121] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A method for rapid classification of sub-pixel objects in hyperspectral images, characterized in that: include: Process the hyperspectral image and construct the endmember matrix and abundance matrix based on the types of objects in the hyperspectral image; Taking the pre-constructed abundance distribution dispersion model as a constraint condition, the hyperspectral image is brought into a pre-set objective function for iterative solution to obtain an end member matrix and an abundance matrix; Classifying the hyperspectral image based on the endmember matrix and the abundance matrix; The endmember matrix is ​​constructed based on the type of land objects, and the abundance matrix is ​​the proportion of land objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution values ​​of all endmembers in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix; The pre-constructed abundance distribution dispersion model is used as a constraint condition, and the hyperspectral image is brought into a pre-set objective function for iterative solution to obtain an end member matrix and an abundance matrix, including: Calculate the partial differential value of the objective function; The objective function is iteratively solved based on the multiplicative iterative optimization iterative rule to obtain the end member matrix and abundance matrix; The final end member matrix E and abundance matrix C are obtained through iterative operations, and the ground object composition of each pixel is obtained; The iteration condition is as follows: the maximum number of iterations does not exceed 200; The objective function is calculated as follows: In the formula, is the penalty function for the dispersion of abundance distribution, is the error, E is the end member matrix, C is the abundance matrix, is the regularization parameter used to balance the model error value and abundance distribution constraint; The calculation formula of the multiplication iterative optimization iteration rule is: In the formula, is the iteration step size of the matrix multiplication iteration rule.

2. The method for rapid classification of sub-pixel objects in hyperspectral images according to claim 1 is characterized by: The processing of the hyperspectral image and the construction of the endmember matrix and the abundance matrix based on the ground object types in the hyperspectral image include: Set the number of end members p based on the number of feature types; The hyperspectral image is subjected to noise whitening processing using MNF forward and inverse transformation; Take the first p-1 bands of the hyperspectral image after noise whitening and perform inverse transformation to obtain the hyperspectral image R; Randomly select p points in the hyperspectral image R as the initial endmember matrix E0, and set the initial abundance matrix C0=0.

3. The method for rapid classification of sub-pixel objects in hyperspectral images according to claim 1 is characterized by: The constraints of the objective function are as follows: E≥0, C≥0.

4. The method for rapid classification of sub-pixel objects in hyperspectral images according to claim 3 is characterized by: The penalty function for the abundance distribution discreteness is: In the formula, is a quadratic function, representing the abundance matrix The sum of all nonnegative eigenvalues.

5. A hyperspectral image sub-pixel level rapid classification system, characterized by: include: An initialization module is used to process the hyperspectral image and construct an end member matrix and an abundance matrix based on the types of objects in the hyperspectral image; A matrix calculation module is used to use a pre-constructed abundance distribution dispersion model as a constraint condition, bring the hyperspectral image into a pre-set objective function for iterative solution, and obtain an end member matrix and an abundance matrix; A ground object classification module, used for classifying the ground objects in the hyperspectral image based on the end member matrix and the abundance matrix; The endmember matrix is ​​constructed based on the type of land objects, and the abundance matrix is ​​the proportion of land objects corresponding to each endmember; the abundance distribution dispersion of the hyperspectral image is a measure of the degree of dispersion of the contribution values ​​of all endmembers in the hyperspectral image to each mixed pixel; the abundance distribution dispersion model is constrained by the non-negativity of the endmember matrix and the abundance matrix; The matrix calculation module comprises: The objective function calculation submodule is used to calculate the partial differential value of the objective function; A matrix calculation submodule, for iteratively solving the objective function based on a multiplication iterative optimization iterative rule to obtain an end member matrix and an abundance matrix; wherein the condition for the iteration is: the maximum number of iterations does not exceed 200; The ground object classification submodule is used to obtain the final end member matrix E and abundance matrix C through iterative operations to obtain the ground object composition of each pixel; The objective function is calculated as follows: In the formula, is the penalty function for the dispersion of abundance distribution, is the error, E is the end member matrix, C is the abundance matrix, is the regularization parameter used to balance the model error value and abundance distribution constraint; The calculation formula of the multiplication iterative optimization iteration rule is: In the formula, is the iteration step size of the matrix multiplication iteration rule.

6. The hyperspectral image sub-pixel level rapid classification system according to claim 5, characterized in that: The initialization module includes: The dimensionality reduction processing submodule is used to set the number of end members p based on the number of ground object types; it is also used to perform noise whitening processing on the hyperspectral image using MNF forward and inverse transformation; it is also used to take the first p-1 bands of the hyperspectral image after noise whitening processing and perform inverse transformation to obtain the hyperspectral image R; The endmember matrix initialization submodule is used to randomly select p points in the hyperspectral image R as the initial endmember matrix E0; The abundance matrix initialization module is used to set the initial abundance matrix C0=0.

Images