A hyperspectral image classification method based on spatial-spectral collaborative sparse unmixing

Abstract

The application provides a hyperspectral image classification method based on spatial spectrum collaborative sparse unmixing, comprising: obtaining a hyperspectral image Y and a spectrum library A; pruning the spectrum library A according to the hyperspectral image Y to obtain a spectrum library subset A1, then constructing a spatial spectrum collaborative sparse unmixing model, solving a sub-pixel representation {A1, X} thereof, and extracting non-zero rows of an abundance matrix X from the sub-pixel representation to form a matrix X1, and then combining category label information to train a preset SVM classifier model to realize classification of the hyperspectral image Y. The application can significantly improve accuracy by classifying the pixels by using the abundance matrix, and can reduce feature dimension, simplify the input of the classifier and improve the classification efficiency. In addition, the application can induce the similarity of the pixels in the local neighborhood by using a total variation term, can suppress different degrees of noise and spectral difference, can promote the merging of the categories to which similar pixels belong, and can achieve better classification effect.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral image classification technology, specifically relating to a hyperspectral image classification method based on spatial spectral collaborative sparse unmixing. Background Technology

[0002] Image classification is an important aspect of remote sensing image processing and application, aiming to uniquely classify each pixel in an image. Hyperspectral remote sensing images contain richer spectral information about ground features compared to other remote sensing images, making their classification more challenging. Current hyperspectral image classification methods mainly fall into three categories: pixel-level, object-level, and sub-pixel-level. Pixel-level classification utilizes pixel-level spectral information, employing methods such as feature extraction and difference measurement of spectral vectors to classify pixels. Machine learning algorithms can also be used to learn pixel label data, understand potential spectral differences, and construct supervised classifiers. Object-level classification considers the spatial information of the image, using image segmentation techniques, such as multi-scale segmentation algorithms, to identify different objects or regions, then extracting object-level features for classification. Sub-pixel-level classification considers the subtle material composition and structural information within pixels, exploring the similarities and differences in material composition features among pixels to achieve more refined classification of ground features. With the development of hyperspectral imaging technology and advancements in imaging equipment, the spatial and spectral resolution of hyperspectral images has been continuously improving, providing practical conditions for refined sub-pixel classification of hyperspectral images.

[0003] Existing hyperspectral image classification methods based on sub-pixel information mainly fall into two categories: ① Classification methods combining soft classification techniques and spectral unmixing (DOI of the paper on this method: 10.1109 / JSTSP.2010.2096798). This method addresses hyperspectral images with mixed pixels and proposes a spatial regularization method based on unmixing abundance information. First, a support vector machine (SVM) is used to classify pixels with relatively strong class determinism. For pixels with class uncertainty, the abundance coefficients of different classes are mined at the sub-pixel scale, and spatial regularization is performed using simulated annealing. This spatially locates the land cover class within each pixel, achieving sub-pixel-level classification of mixed pixels. ② Hyperspectral image sub-pixel classification methods based on SVM (Support Vector Machine) (DOI of the paper on this method: 10.13203 / j.whugis20150443). This method first trains an SVM classifier based on radial basis function kernels using simplified training samples. Then, a fully variational regularized denoising model is employed to first remove any random noise that may exist in the hyperspectral image. A fully constrained mixed pixel decomposition is then used to obtain the unmixed pixel results, and clean pixels are removed. Finally, the mixed pixels are fed into an SVM classifier for regression analysis to obtain the sub-pixel classification results. This method comprehensively considers the influence of the abundance values ​​of the central pixel and its neighboring pixels on the sub-pixel category assignment, and reduces the number of samples by removing clean pixels, thus improving algorithm efficiency while maintaining accuracy.

[0004] Classification methods combining soft classification techniques and spectral unmixing have the following limitations: First, it requires selecting an appropriate threshold to determine pixels with strong class certainty, and setting a scaling factor to determine the granularity of sub-pixel classification. Second, spatial regularization methods based on simulated degradation algorithms are sensitive to scaling factors; a larger scaling factor, while increasing the granularity of sub-pixel classification, leads to a significant increase in computational complexity. Furthermore, SVM-based hyperspectral image sub-pixel classification methods also have the following limitations: First, while removing pure pixels can improve classification efficiency, it may lose the differential features between different categories, blurring the classification boundaries between pixels of different categories. Second, using only a total variation regularization model for image denoising does not utilize the smoothing effect of the total variation term to achieve class merging and noise removal of hyperspectral similar pixels.

[0005] Although other hyperspectral image classification methods utilize sub-pixel information, most use them as auxiliary tools, and image classification methods relying solely on hyperspectral unmixing information remain scarce. Furthermore, existing sub-pixel-level image classification methods do not fully explore and utilize the ground feature composition information of images, nor do they effectively combine the spatial distribution features of images. Therefore, it is necessary to explore effective ways to utilize the sub-pixel composition information of hyperspectral images, combining it with prior knowledge of the spatial structure of hyperspectral images to achieve low-dimensional feature representation and difference measurement at the sub-pixel level, and to establish efficient hyperspectral image sub-pixel classification methods. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a hyperspectral image classification method based on spatial spectral collaborative sparse unmixing. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] A hyperspectral image classification method based on spatial spectral cooperative sparse unmixing includes:

[0008] S100, acquire the hyperspectral image Y to be classified, the applicable spectral library A, and the category labels of some pixels in the spectral library A, wherein the spectral library A includes the spectral features of many basic substances, and the pixel spectrum of the hyperspectral image is composed of the spectral features of the basic substances in the spectral library A, and the composition coefficient is represented by the abundance matrix X.

[0009] S200, the spectral library A is pruned according to the hyperspectral image Y to obtain a matching subset A1 of the spectral library;

[0010] S300, using a subset of the spectral library A1 to construct a spatial spectral cooperative sparse unmixing model, and solve its sub-pixel representation {A1, X};

[0011] S400, extract the non-zero rows of the abundance matrix X in the sub-pixel representation {A1, X}, and reorganize them to obtain the sub-pixel representation information X1;

[0012] S500, using any column of the sub-pixel representation information X1 and the corresponding category label, train a preset SVM classifier model;

[0013] S600, using the trained SVM classifier model, classify the remaining columns in the sub-pixel representation information X1 to obtain category labels, and use them as the classification results of the corresponding pixels in the hyperspectral image Y.

[0014] Beneficial effects:

[0015] This invention provides a hyperspectral image classification method based on spatial spectral cooperative sparse unmixing, comprising: acquiring a hyperspectral image Y to be classified and an applicable spectral library A; trimming the spectral library A according to the hyperspectral image Y to obtain a matching spectral library subset A1; constructing a spatial spectral cooperative sparse unmixing model using the spectral library subset A1 and solving its sub-pixel representation {A1, X}; extracting the non-zero rows of the abundance matrix X in the sub-pixel representation {A1, X} and recombining them to obtain sub-pixel representation information X1; training a preset SVM classifier model using the sub-pixel representation information X1 and the category label information in S100; and using the trained SVM classifier model to classify the columns of matrix X1 to obtain the classification result of the hyperspectral image Y. This invention obtains the abundance matrix X from the hyperspectral image Y by unmixing it, and then classifies pixels based on this matrix, significantly improving accuracy. Furthermore, the abundance matrix-based image classification method reduces feature dimensions, simplifies the classifier input, and thus improves classification efficiency, especially for processing large-scale hyperspectral image data. Moreover, this invention utilizes total variational terms to induce similarity in the composition of pixels in local neighborhoods, which can suppress noise and occlusion to varying degrees. In addition, by inducing similarity in the abundance vectors of adjacent pixels, it can promote the merging of similar pixel categories, achieving better classification results.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a hyperspectral image classification method based on spatial spectral collaborative sparse unmixing provided by the present invention;

[0018] Figure 2 This is a comparison chart of the classification results of the proposed method and the pixel-based spectral classification method on the Pavia dataset;

[0019] Figure 3 This is a heatmap showing the classification accuracy of the classification method of this invention under different regularization parameters. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0021] Before introducing the present invention, the technical concept of the present invention will be briefly described first.

[0022] Hyperspectral images, due to their excellent spectral resolution, contain a wealth of pixel spectral information, which can be used for detailed interpretation of ground targets. However, existing hyperspectral image classification methods do not adequately utilize sub-pixel-level information in images. Therefore, this invention aims to explore effective ways to utilize sub-pixel information in hyperspectral images. By mining the low-dimensional feature representations of images at the sub-pixel level, it further analyzes the similarity and differences of image pixels in sub-pixel composition, thereby improving the accuracy and efficiency of hyperspectral classification. Therefore, this invention proposes an image classification method based on the synergistic sparse unmixing information of hyperspectral neighborhood spatial similarity. This method has advantages such as low computational complexity, strong noise robustness, strong algorithm applicability, and strong generalization ability.

[0023] The details of the present invention are described below.

[0024] like Figure 1 As shown, this invention provides a hyperspectral image classification method based on spatial spectral cooperative sparse unmixing, comprising:

[0025] S100, acquire the hyperspectral image Y to be classified, the applicable spectral library A, and the category labels of some pixels in the spectral library A, wherein the spectral library A includes the spectral features of many basic substances, and the pixel spectrum of the hyperspectral image is composed of the spectral features of the basic substances in the spectral library A, and the composition coefficient is represented by the abundance matrix X.

[0026] It is worth noting that the parameters of the spectral library A-endmembers, the number of random unit vectors, the population size, the number of neighborhood subproblems, and the maximum number of iterations in this invention can all be preset and adjusted according to the actual situation.

[0027] S200, the spectral library A is pruned according to the hyperspectral image Y to obtain a matching subset A1 of the spectral library;

[0028] In addition to pruning by calculating the projection operator, this step can also prune the spectral library A using the subspace matching method.

[0029] S300, using a subset of the spectral library A1 to construct a spatial spectral cooperative sparse unmixing model, and solve its sub-pixel representation {A1, X};

[0030] S400, extract the non-zero rows of the abundance matrix X in the sub-pixel representation {A1, X}, and reorganize them to obtain the sub-pixel representation information X1;

[0031] S500, using any column of the sub-pixel representation information X1 and the corresponding category label, train a preset SVM classifier model;

[0032] S600, using the trained SVM classifier model, classify the remaining columns in the sub-pixel representation information X1 to obtain category labels, and use them as the classification results of the corresponding pixels in the hyperspectral image Y.

[0033] Traditional methods classify hyperspectral images Y using spectral vectors. This invention uses an abundance matrix X to classify the hyperspectral image Y. The abundance matrix X serves as the representation information of Y, possessing low dimensionality, which is beneficial for classification. Furthermore, since each column of the hyperspectral image Y corresponds to each column of X, this invention uses the classification result of the abundance matrix X as the classification result of Y. The classification result is a matrix composed of the category labels of each pixel in the hyperspectral image Y.

[0034] In one specific embodiment of the present invention, S200 includes:

[0035] S210, Calculate a set of orthogonal bases E in the subspace where the hyperspectral image Y is located;

[0036] S220, using the aforementioned set of orthogonal bases E, calculate the projection operator on the orthogonal complement space of the subspace containing the hyperspectral image Y; the projection operator is expressed as:

[0037] ;

[0038] in, Represents the identity matrix;

[0039] S230, calculate the projection length of each endmember in the spectral library A on the projection operator P, and normalize it according to the normalization formula to obtain the normalized projection length.

[0040] The normalization formula is expressed as:

[0041] ;

[0042] in, Represents the first in spectral library A List, This represents the normalized projection length.

[0043] S240, sort the normalized projection lengths in ascending order, and extract the basic substance spectra corresponding to the first k values ​​to form a spectral library subset A1.

[0044] In an optional embodiment of the present invention, S300 includes:

[0045] S310, introduce row sparsity regularization and neighborhood space regularization terms for the abundance matrix X, and construct an initial sparse and unmixed model using the spectral library subset A1; the initial sparse and unmixed model is expressed as:

[0046]

[0047] in, and Both represent regularization parameters. This represents the number of endmembers in the spectral library subset A1. Let X represent the total variation term of matrix X, where Let i be the set of neighborhood cell indices, with superscripts... T Indicates transpose;

[0048] The regularization parameters of the row sparse regularization term and the neighborhood space regularization term in this invention can be adjusted.

[0049] S320, auxiliary variables Z1 and Z2 are introduced into the initial spalling unmixing model to obtain a spatial spectral cooperative sparse unmixing model; the spatial spectral cooperative sparse unmixing model is expressed as:

[0050] ;

[0051] S330, the spatial spectral cooperative sparse unmixing model is transformed using the Lagrange multiplier method to obtain the transformed model; the transformed model is expressed as:

[0052] ;

[0053] in, They represent the Lagrange multipliers, >0 indicates the regular expression parameter, subscript F This represents the Frobenius norm.

[0054] S340, The ADMM algorithm is used to iteratively solve the transformed model to obtain the sub-pixel representation {A1, X} of the hyperspectral image.

[0055] In an optional embodiment of the present invention, S400 includes:

[0056] S410, Extract the abundance matrix X from the sub-pixel representation {A1, X};

[0057] S420, calculate the squares of the elements in the abundance matrix X, and sum them row by row to obtain the summed column vector S;

[0058] S430, determine the row containing the non-zero element in column vector S, and denote its index vector as IdX;

[0059] S440, extract the row vectors corresponding to the index vector IdX in the abundance matrix X, and reorganize them to form sub-pixel representation information X1.

[0060] In an optional embodiment of the present invention, S500 includes:

[0061] S510: Select the abundance vector corresponding to the training sample from the sub-pixel representation information X1, and pair the corresponding category label with the abundance vector to form training data;

[0062] S520, Establish an SVM classifier model;

[0063] In addition to using an SVM classifier model with a linear kernel function, this invention can also use an SVM model with a radial basis function kernel.

[0064] S530, The SVM classifier model is trained using the training data to obtain the trained SVM classifier model.

[0065] The classification effect of this invention will be illustrated below through simulation.

[0066] refer to Figure 2 and Figure 3 , Figure 2 A comparison of the classification results of this algorithm (left), the pixel-based spectral classification algorithm (middle), and the reference image (right) when the training samples account for 5% of the pixels. Figure 3 A heatmap showing the classification accuracy as a function of regularization parameters when training samples comprise 5% of the pixels. Figure 2 It can be concluded that the classification result obtained by the present invention can group neighboring pixels into the same class compared with the classification result based on pixel spectrum, greatly suppressing the generation of noise class and being closer to the reference image. Figure 3 The results show that the classification accuracy of the algorithm varies with the model parameters. Figure 3 The results in the lower left corner indicate that the appropriate parameter range is relatively wide, and parameter selection is relatively easy.

[0067] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A hyperspectral image classification method based on spatial spectral collaborative sparse unmixing, characterized in that, include: S100, acquire the hyperspectral image Y to be classified, the applicable spectral library A, and the category labels of some pixels in the spectral library A, wherein the spectral library A includes the spectral features of many basic substances, and the pixel spectrum of the hyperspectral image is composed of the spectral features of the basic substances in the spectral library A, and the composition coefficient is represented by the abundance matrix X. S200, the spectral library A is pruned according to the hyperspectral image Y to obtain a matching subset A1 of the spectral library; S300, using a subset of the spectral library A1 to construct a spatial spectral cooperative sparse unmixing model, and solve its sub-pixel representation {A1, X}; S400, extract the non-zero rows of the abundance matrix X in the sub-pixel representation {A1, X}, and reorganize them to obtain the sub-pixel representation information X1; S500, using any column of the sub-pixel representation information X1 and the corresponding category label, train a preset SVM classifier model; S600, using the trained SVM classifier model, classify the remaining columns in the sub-pixel representation information X1 to obtain category labels, and use them as the classification results of the corresponding pixels in the hyperspectral image Y; S200 includes: S210, Calculate a set of orthogonal bases E in the subspace where the hyperspectral image Y is located; S220, using the set of orthogonal bases E, calculate the projection operator P on the orthogonal complement space of the subspace where the hyperspectral image Y is located; S230, calculate the projection length of each endmember in the spectral library A on the projection operator P, and normalize it according to the normalization formula to obtain the normalized projection length. S240, sort the normalized projection lengths in ascending order, and extract the basic substance spectra corresponding to the first k values ​​to form a spectral library subset A1; The S300 includes: S310, introduce row sparsity regularization term and neighborhood space regularization term for abundance matrix X, and construct initial sparse mixed model using the spectral library subset A1; S320, Introduce auxiliary variables Z1 and Z2 into the initial spalling and mixing model to obtain a spatial spectral synergistic spalling and unmixing model; S330, The spatial spectral cooperative sparse unmixing model is transformed using the Lagrange multiplier method to obtain the transformed model; S340, The ADMM algorithm is used to iteratively solve the transformed model to obtain the sub-pixel representation {A1, X} of the hyperspectral image; The initial spalling hybrid model in S310 is expressed as follows: ； in, and Both represent regularization parameters. This represents the number of endmembers in the spectral library subset A1. Let X represent the total variation term of matrix X, where Let i be the set of neighborhood cell indices, with superscripts... T Indicates transpose; The spatial spectral cooperative sparse unmixing model in S320 is expressed as follows: The transformed model in S330 is represented as follows: ； in, They represent the Lagrange multipliers, >0 indicates the regular expression parameter, subscript F This represents the Frobenius norm.

2. The hyperspectral image classification method based on spatial spectral cooperative sparse unmixing according to claim 1, characterized in that, The projection operator in S220 is expressed as: Where I represents the identity matrix; The normalization formula in S230 is expressed as follows: in, Represents the first in spectral library A List, This represents the normalized projection length.

3. The hyperspectral image classification method based on spatial spectral cooperative sparse unmixing according to claim 1, characterized in that, The S400 includes: S410, Extract the abundance matrix X from the sub-pixel representation {A1, X}; S420, calculate the squares of the elements in the abundance matrix X, and sum them row by row to obtain the summed column vector S; S430, determine the row containing the non-zero element in column vector S, and denote its index vector as IdX; S440, extract the row vectors corresponding to the index vector Idx in the abundance matrix X, and reorganize them to form the sub-pixel representation information X1.

4. The hyperspectral image classification method based on spatial spectral cooperative sparse unmixing according to claim 1, characterized in that, The S500 includes: S510: Select the abundance vector corresponding to the training sample from the sub-pixel representation information X1, and pair the corresponding category label with the abundance vector to form training data; S520, Establish an SVM classifier model; S530, The SVM classifier model is trained using the training data to obtain the trained SVM classifier model.

Images