A hyperspectral image classification method, system and storage medium

Abstract

The present invention provides a hyperspectral image classification method, system and storage medium, and relates to the technical field of remote sensing image classification. The present invention uses superpixel segmentation and PCA to process hyperspectral images, obtains a reduced-dimensional hyperspectral image as spectral information, and performs filtering processing on the reduced-dimensional hyperspectral image through multiple edge-preserving filters to obtain multiple filtered images as spatial information, and then classifies the hyperspectral image in combination with the spatial information and the spectral information, making full use of the spectral information and spatial information of the hyperspectral image, with high classification accuracy, and the technical solution is different from the traditional global PCA processing of the hyperspectral image, but uses superpixel segmentation to divide the hyperspectral image into multiple uniform areas, and then performs PCA processing separately, avoiding the loss of image edge information and other independent components, and ensuring classification accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image classification, and more specifically, to a hyperspectral image classification method, system and storage medium. Background Art

[0002] Hyperspectral images usually contain information of dozens or even hundreds of bands. Due to the high dimensionality, similarity between spectra and mixed pixel characteristics of hyperspectral images, how to improve the classification accuracy has become a major difficulty. The classification accuracy of existing unsupervised learning classification methods represented by k-means and supervised learning classification methods represented by SVM, naive Bayes, decision trees, etc. is not ideal. For example, due to the small number of samples and high dimensionality of hyperspectral data, the naive Bayes estimation may produce the Hughes phenomenon, which affects the performance of the classification method. SVM, principal component analysis (PCA), and independent component analysis (ICA) only use spectral information and ignore spatial information. They cannot directly use the spatial information of HSI, which has a certain impact on the classification effect to a certain extent.

[0003] Publication number: CN103150580A, publication date: 2016-03-30, a method and device for semi-supervised classification of hyperspectral images, the invention provides a method and device for semi-supervised classification of hyperspectral images, performs support vector machine (SVM) semi-supervised classification on hyperspectral images to obtain a first classification image Image1, performs support vector machine (SVM) semi-supervised classification on clustering indicator features to obtain a second classification image Image2; constructs a clustering and SVM collaborative framework, embeds the classification results of Image1 and Image2 into the clustering and SVM collaborative framework for collaborative analysis to obtain the final hyperspectral classification image. However, this method only uses support vector machine (SVM) to classify hyperspectral images, does not fully utilize the spatial information in hyperspectral images, and the classification accuracy still has room for further improvement. Summary of the invention

[0004] In order to overcome the above technical problems, the present invention provides a hyperspectral image classification method, system and storage medium that can simultaneously utilize spatial information and spectral information in hyperspectral images, avoid the loss of image edge information and other independent components, and ensure classification accuracy.

[0005] The technical solution of the present invention is as follows:

[0006] A hyperspectral image classification method comprises the steps of:

[0007] S1, performing superpixel PCA dimensionality reduction on the original hyperspectral image to obtain a reduced-dimensional hyperspectral image;

[0008] S2, setting a plurality of edge-preserving filters to filter the reduced-dimensional hyperspectral images respectively to obtain a plurality of filtered images;

[0009] S3, performing PCA processing on the multiple filtered images to obtain multiple filter image principal components;

[0010] S4, constructing the dimension-reduced hyperspectral image obtained in step S1 into a spectral kernel based on a radial basis kernel function; constructing the plurality of filtered image principal components obtained in step S3 into a plurality of spatial kernels based on a radial basis kernel function; and using the spectral kernel to fuse with the plurality of spatial kernels to construct a plurality of composite kernels;

[0011] S5, inputting the multiple composite kernels into the SVM classifier to obtain multiple classification graphs;

[0012] S6. Perform decision fusion on multiple classification graphs to obtain the final classification result.

[0013] The present technical scheme proposes a hyperspectral image classification method, system and storage medium, which use superpixel segmentation and PCA to process the hyperspectral image to obtain a reduced-dimensional hyperspectral image as spectral information, and filter the reduced-dimensional hyperspectral image through multiple edge-preserving filters to obtain multiple filtered images as spatial information, and then classify the hyperspectral image in combination with the spatial information and the spectral information, making full use of the spectral information and spatial information of the hyperspectral image, with high classification accuracy. In addition, the present technical scheme is different from the traditional global PCA processing of the hyperspectral image. Instead, the hyperspectral image is segmented into multiple uniform areas by superpixel segmentation, and then PCA processing is performed separately, thereby avoiding the loss of image edge information and other independent components and ensuring classification accuracy.

[0014] Furthermore, the hyperspectral image is a three-dimensional cube, and the three-dimensional cube is reshaped into a low-dimensional matrix X. The reshaped low-dimensional matrix X is a two-dimensional matrix, as shown in the following formula:

[0015] X=[x1,x2,…,x i , …, x P ]∈R L*P , P = MN

[0016] Where X represents the hyperspectral data after dimensionality reduction X∈R M×N×L , M, N and L represent the number of rows, columns and sampling wavelengths of the image respectively; each column represents a vector of a pixel; x i Represents the i-th pixel vector in the original image;

[0017] x i =[x i1 , x i2 , …, x il ] T .

[0018] Furthermore, the PCA processing in step S3 is to calculate the linear mapping of the pixels of the original image from the initial high-dimensional space to the low-dimensional space, and the calculation formula is:

[0019] y i =w T x i

[0020] Where w represents the transformation matrix, x i Represents the i-th pixel vector in the original image, y i A low-dimensional space matrix representing the i-th pixel vector.

[0021] Furthermore, in step S2, parameters of each edge preserving filter in the plurality of edge preserving filters are different, and the parameters include: the number of edge preserving filters EPF and parameters δs and δr controlling the smoothness of the edge preserving filters EPF.

[0022] Furthermore, step S4 constructs the spatial kernel and the spectral kernel using the following formulas respectively:

[0023]

[0024]

[0025] Among them, k spa (x i ,x j ) represents the spatial kernel, K spe (x i ,x j ) represents the spectral kernel, σ represents the width of the radial basis kernel function, which is a free parameter, and σ s is the width of the empty spectral kernel, σ w is the spectral kernel width; x i 、x j Normalized spectral data; For domain transformation transfer filtering from x i 、x j The obtained spatial-spectral characteristics, is the spectral information; the two radial basis kernel functions obtained above are constructed into a composite kernel.

[0026] Furthermore, the decision fusion in step S6 is achieved through a majority voting algorithm.

[0027] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the hyperspectral image classification method is implemented.

[0028] A hyperspectral image classification system, comprising: a superpixel segmentation unit, a PCA processing unit, an edge-preserving filter, a base kernel function building module, a SVM classifier, and a decision fusion module;

[0029] The superpixel segmentation unit performs superpixel segmentation on the input hyperspectral image to obtain a uniform area, the PCA processing unit performs PCA processing on the uniform area to obtain a low-dimensional matrix, and rearranges the low-dimensional matrix to obtain a reduced-dimensional hyperspectral image, the edge-preserving filter filters the reduced-dimensional hyperspectral image with different parameters to obtain a plurality of filtered images, the PCA processing unit performs PCA processing on the plurality of filtered images to obtain a plurality of filtered image principal components, the base kernel function construction module constructs the reduced-dimensional hyperspectral image into a spectral kernel based on a radial basis kernel function, the base kernel function construction module constructs the plurality of filtered image principal components into a plurality of spatial kernels based on a radial basis kernel function, and the spectral kernel and the plurality of spatial kernels are respectively fused to obtain a plurality of composite kernels, the plurality of composite kernels are input into an SVM classifier to obtain a plurality of classification graphs, and the decision fusion module performs decision fusion on the plurality of classification graphs to obtain a final classification result.

[0030] Furthermore, the hyperspectral image is a three-dimensional cube, and the three-dimensional cube is reshaped into a low-dimensional matrix X. The reshaped low-dimensional matrix X is a two-dimensional matrix, as shown in the following formula:

[0031] X=[x1,x2,…,x i , …, x P ]∈R L*P , P = MN

[0032] Where X represents the hyperspectral data after dimensionality reduction X∈R M×N×L , M, N and L represent the number of rows, columns and sampling wavelengths of the image respectively; each column represents a vector of a pixel; x i Represents the i-th pixel vector in the original image;

[0033] x i =[x i1 , x i2 , …, x il ] T .

[0034] Furthermore, the base kernel function construction module constructs the spatial kernel and the spectral kernel using the following formulas respectively:

[0035]

[0036]

[0037] Among them, k spa (x i , x j ) represents the spatial kernel, Kspe (x i , x j ) represents the spectral kernel, σ represents the width of the radial basis kernel function, which is a free parameter, and σ s is the width of the empty spectral kernel, σ w is the spectral kernel width; x i 、x j Normalized spectral data; For domain transformation transfer filtering from x i 、x j The obtained spatial-spectral characteristics, is the spectral information; the two radial basis kernel functions obtained above are constructed into a composite kernel.

[0038] The technical solution proposes a hyperspectral image classification method, system and storage medium. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: superpixel segmentation and PCA are used to process hyperspectral images to obtain reduced-dimensional hyperspectral images as spectral information, and the reduced-dimensional hyperspectral images are filtered by multiple edge-preserving filters to obtain multiple filtered images as spatial information, and then the hyperspectral images are classified in combination with the spatial information and the spectral information, which makes full use of the spectral information and spatial information of the hyperspectral images and has high classification accuracy. In addition, the technical solution is different from the traditional global PCA processing of hyperspectral images. Instead, the hyperspectral images are divided into multiple uniform areas by superpixel segmentation, and then PCA processing is performed separately, which avoids the loss of image edge information and other independent components and ensures classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the steps of the hyperspectral image classification method of the present invention;

[0040] Figure 2 Schematic diagram of the processing process of the hyperspectral image classification system;

[0041] Figure 3 Schematic diagram of the main projection directions of global PCA and class-specific PCA;

[0042] Figure 4 It is the dimensionality reduction process of superpixel PCA. DETAILED DESCRIPTION

[0043] In order to clearly illustrate a hyperspectral image classification method, system and storage medium of the present invention, the present invention is further described in conjunction with embodiments and drawings, but this should not limit the protection scope of the present invention.

[0044] Example 1

[0045] A hyperspectral image classification method, such as Figure 1As shown, the method comprises the following steps:

[0046] S1, performing superpixel PCA dimensionality reduction on the original hyperspectral image to obtain a reduced-dimensional hyperspectral image;

[0047] S2, setting a plurality of edge-preserving filters to filter the reduced-dimensional hyperspectral images respectively to obtain a plurality of filtered images;

[0048] S3, performing PCA processing on the multiple filtered images to obtain multiple filter image principal components;

[0049] S4, constructing the dimension-reduced hyperspectral image obtained in step S1 into a spectral kernel based on a radial basis kernel function; constructing the plurality of filtered image principal components obtained in step S3 into a plurality of spatial kernels based on a radial basis kernel function; and using the spectral kernel to fuse with the plurality of spatial kernels to construct a plurality of composite kernels;

[0050] S5, inputting the multiple composite kernels into the SVM classifier to obtain multiple classification graphs;

[0051] S6. Decision fusion is performed on multiple classification images to obtain the final classification result. This technical solution uses superpixel segmentation and PCA to process the hyperspectral image, obtains a reduced-dimensional hyperspectral image as spectral information, and filters the reduced-dimensional hyperspectral image through multiple edge-preserving filters to obtain multiple filtered images as spatial information, and then classifies the hyperspectral image in combination with the spatial information and the spectral information, making full use of the spectral information and spatial information of the hyperspectral image, with high classification accuracy. In addition, this technical solution is different from the traditional global PCA processing of the hyperspectral image. Instead, it uses superpixel segmentation to divide the hyperspectral image into multiple uniform areas, and then performs PCA processing separately, avoiding the loss of image edge information and other independent components, and ensuring classification accuracy.

[0052] Example 2

[0053] A hyperspectral image classification method, such as Figure 1 As shown, the method comprises the following steps:

[0054] S1, performing superpixel PCA dimensionality reduction on the original hyperspectral image to obtain a reduced-dimensional hyperspectral image;

[0055] The superpixel PCA dimensionality reduction comprises the steps of:

[0056] S11, performing super-pixel segmentation on the hyperspectral image to obtain a uniform region, wherein the uniform region is represented by a matrix, and the columns of the matrix are spectral vectors of pixels in the uniform region;

[0057] S12, performing PCA processing on the uniform area to obtain a low-dimensional matrix, and rearrange and combine the low-dimensional matrix to obtain a hyperspectral image with reduced dimensionality;

[0058] S2, setting a plurality of edge-preserving filters to filter the reduced-dimensional hyperspectral images respectively to obtain a plurality of filtered images;

[0059] S3, performing PCA processing on the multiple filtered images to obtain multiple filter image principal components;

[0060] S4, constructing the dimension-reduced hyperspectral image obtained in step S1 into a spectral kernel based on a radial basis kernel function; constructing the plurality of filtered image principal components obtained in step S3 into a plurality of spatial kernels based on a radial basis kernel function; and using the spectral kernel to fuse with the plurality of spatial kernels to construct a plurality of composite kernels;

[0061] S5, inputting the multiple composite kernels into the SVM classifier to obtain multiple classification graphs;

[0062] S6. Perform decision fusion on multiple classification graphs to obtain the final classification result.

[0063] Among them, the superpixel segmentation in step S11 adopts entropy rate superpixel segmentation. The edge-preserving filter in step S2 is a domain transformation recursive filter. The parameters of each edge-preserving filter in the multiple edge-preserving filters in step S2 are different, and the parameters include: the number of edge-preserving filters EPF and parameters δs and δr that control the smoothness of the edge-preserving filter EPF. The PCA processing in step S3 calculates the linear mapping of the pixels of the original image from the initial high-dimensional space to the low-dimensional space, and the calculation formula is:

[0064] y i =w T x i

[0065] Where w represents the transformation matrix, x i Represents the i-th pixel vector in the original image, y i A low-dimensional space matrix representing the i-th pixel vector.

[0066] The decision fusion in step S6 is achieved through a majority voting algorithm.

[0067] Example 3

[0068] A hyperspectral image classification system, the system processes the hyperspectral image as follows Figure 2 As shown, the hyperspectral image classification system of this embodiment includes: a superpixel segmentation unit, a PCA processing unit, an edge-preserving filter, a base kernel function construction module, a SVM classifier, and a decision fusion module;

[0069] The super-pixel segmentation unit performs super-pixel segmentation on the input hyperspectral image to obtain a uniform area, the PCA processing unit performs PCA processing on the uniform area to obtain a low-dimensional matrix, and rearranges and combines the low-dimensional matrix to obtain a reduced-dimensional hyperspectral image;

[0070] The edge-preserving filter filters the reduced-dimensional hyperspectral image with different parameters to obtain multiple filtered images. In this embodiment, three filtered images are obtained by edge-preserving filters with three different parameters: EPF1, EPF2, and EPF3.

[0071] The PCA processing unit performs PCA processing on multiple filtered images to obtain multiple filtered image principal components. The base kernel function construction module constructs the reduced-dimensional hyperspectral image into a spectral kernel based on the radial basis kernel function. The base kernel function construction module constructs multiple filtered image principal components into multiple spatial kernels based on the radial basis kernel function, and fuses the spectral kernel and the multiple spatial kernels respectively to obtain multiple composite kernels. The multiple composite kernels are input into the SVM classifier to obtain multiple classification graphs. The decision fusion module performs decision fusion on the multiple classification graphs to obtain the final classification result.

[0072] The segmentation method of the super pixel segmentation unit is entropy rate super pixel segmentation. The edge preserving filter is a domain transform recursive filter (Domain Transform Recursive Filtering, DTRF).

[0073] Example 4

[0074] This embodiment discloses a hyperspectral image classification method based on superpixel PCA and edge-preserving filtering. The method for implementing this classification includes the following parts. Principal component analysis PCA, as an unsupervised dimensionality reduction method, is widely considered to be an efficient image preprocessing method for hyperspectral image processing and analysis. Conventional PCA performs dimensionality reduction on the entire image, i.e., global PCA. Since different uniform regions of a hyperspectral image represent different objects, the classification performance will be reduced to some extent.

[0075] Therefore, this embodiment adopts the superpixel PCA method as image preprocessing. Specifically: a hyperspectral image (HSI cube (x∈R MXNXL )) consists of hundreds of spectral bands with a spectral band resolution of 5-10nm. The three-dimensional HSI cube is reshaped into a two-dimensional matrix X. The reshaped two-dimensional matrix X is shown as follows:

[0076] X=[x1,x2,…,x i , …, x P ]∈R L*P ,P=MN (1)

[0077] Where X represents the hyperspectral data after dimensionality reduction X∈R M×N×L , M, N and L represent the number of rows, columns and sampling wavelengths of the image respectively; each column represents a vector of a pixel; x i Represents the i-th pixel vector in the original image;

[0078] x i =[x i1 , x i2 , …, x il ] T (2)

[0079] PCA achieves dimensionality reduction by calculating a low-dimensional representation that maximizes the variance of the data in the dimensionality-reduced space. Specifically, it finds a linear mapping that can reduce the original L-dimensional space to the D-dimensional space, i.e., formula (3).

[0080] y i =w T x i (3)

[0081] Where w represents the transformation matrix, x i represents the i-th pixel vector in the observed cube. i A low-dimensional space matrix representing the i-th pixel vector.

[0082] In general mathematics, the linear transformation matrix w is obtained by solving formula (4):

[0083] W * = argmaxTr(w T Cov(x)w)

[0084] w T w=I (4)

[0085] Where Cov(X) represents the covariance matrix of the data set X, Tr(X) represents the trajectory of an nXn square matrix X, and w represents the transformation matrix to be obtained.

[0086] Global PCA considers the entire data space and finds the best transformation vector for that space, which ignores the differences in homogeneous regions (different homogeneous regions represent different objects). Even the best transformation vector is different from the transformation vectors of each region. Figure 3 As shown:

[0087] The uniform region mentioned above is specifically: The dimension reduction method used in this embodiment is a strategy based on PCA to extract dimension reduction for each uniform region feature. The same number of main components are extracted from each homogeneous region, and then the components after dimension reduction are combined. Since entropy rate superpixel segmentation has good performance in terms of efficiency and effectiveness, the homogeneous region generated by the hyperspectral image in this embodiment adopts entropy rate superpixel segmentation technology. First, the original hyperspectral image is subjected to dimension reduction extraction before the main component I f , in I f The entropy rate superpixel segmentation is performed on the image, as shown in formula (5):

[0088]

[0089] Where S represents the number of superpixels, χ k represents the kth superpixel, χ g represents the g-th superpixel. f It represents the main component before dimensionality reduction extraction of the original hyperspectral image, I f Captures the main information of hyperspectral images.

[0090] After the above steps, the hyperspectral image I after dimensionality reduction is obtained, and then the image I is subjected to edge-preserving filtering. In order to obtain the multi-scale structural information retained in the EPFs obtained under different parameter settings, different parameters are set for the domain transform recursive filter. The filtered image is very different in preserving edge information. When the degree of blur is large, the spectral separability of different objects can be reduced. In addition, highly smooth filtering can effectively reduce noise and improve the spectral purity of the area. This means that EPFs with different degrees of blur have specific advantages in representing objects or features of different scales. The number of EPFs X and δs and δr control the smoothness of the EPF, and the parameters of the EPF can be constructed, namely Get EPF with complementary information,

[0091] In this embodiment, three sets of δs and δr values ​​are set, and different parameters are set for edge preserving filtering to obtain EPF with complementary information. The values ​​of the three sets of parameters are:

[0092] These three sets of parameters are used together to extract useful spatial information features from the image after dimensionality reduction. EPF1, EPF2, and EPF3 are obtained using formula (6).

[0093] As shown below:

[0094]

[0095] Where DTRF is the domain transformation filter, I is the hyperspectral image after dimensionality reduction, The xth parameter setting used by the domain transform recursive filter controls the smoothness of the EPF. It is the image after image filtering using domain transform filtering (DTRF) according to equation (6).

[0096] The above-obtained EPF1, EPF2, and EPF3 are subjected to PCA dimensionality reduction operations as shown in formula (9) to obtain EPF1-PCA (P1), EPF2-PCA (P2), and EPF3-PCA (P3). In order to simultaneously utilize spatial spectral information, EPF1-PCA, EPF2-PCA, and EPF3-PCA are fused with the spectral image after dimensionality reduction in step 1 as shown in formula (10), and the EPF1, EPF2, and EPF3 composite kernels are used as inputs of SVM to obtain their respective classification maps. Next, the decision fusion method is used to obtain the final classification result for each classification map.

[0097] P1=PCA(F1,L) (9)

[0098] The same is true for P2 and P3.

[0099]

[0100]

[0101] in is the weight coefficient, k spectral represents the spectral information, that is, the spectral information obtained after the original hyperspectral image is reduced by superpixel PCA. spatial It is the spatial information obtained after domain transformation filtering and PCA. It is the composite information after the fusion of spectral information and spatial information.

[0102] Will get The composite information is used as the input of the SVM classifier to obtain a classification map. Since the edge-preserving filter image in this embodiment is obtained by setting different parameters, each EPF has its own unique characteristics, so the classification map (i.e. Figure 2 Categories in Figure 1 ,Classification Figure 2 ,Classification Figure 3 ) have complementary information, so in this stage, decision fusion is further used to fuse the three classification maps to obtain the final classification result. Specifically, majority voting (MV) is used pixel by pixel in the three classification maps obtained. The number of occurrences of each class is calculated and represented as Count1, Count2, Count3, ..., CountM (where M is the class number). The class label of the test pixel can be obtained by the following formula (11):

[0103] L=max(Count1,Count2,…,Count M ) (11)

[0104] In rare cases, that is, for some test samples, the labels of the three classification maps are different, and it is difficult to select the correct label as the final classification result. To solve this problem, this embodiment defines each classification map (i.e., classification Figure 1 ,Classification Figure 2 ,Classification Figure 3 ) has an initial value of 0. When a majority vote is performed (i.e., the vote belongs to a majority vote), the confidence of the corresponding graph increases by 1. Finally, the classification graph label with the highest confidence is selected as the label of those uncertain test samples, thereby obtaining the classification result.

[0105] Example 5

[0106] This embodiment discloses a hyperspectral image classification method based on superpixel PCA and edge-preserving filtering, comprising the following steps:

[0107] 1. Image Dimensionality Reduction

[0108] The dimensionality reduction method adopted in this embodiment is the superpixel PCA dimensionality reduction method, which is an improvement on the traditional PCA dimensionality reduction method combined with the superpixel characteristics. First, the hyperspectral image is divided into some uniform areas through superpixel segmentation. The columns of the matrix in the hyperspectral image data are the spectral vectors of the pixels. Secondly, PCA is applied to these high-dimensional matrices to obtain the dimensionality reduction matrix. Finally, the obtained dimensionality reduction matrix is ​​rearranged and combined to obtain the dimensionality reduction hyperspectral image. The dimensionality reduction process of superpixel PCA is as follows: Figure 4 shown.

[0109] The uniform region refers to relatively uniform and regular superpixels. Specifically, segmentation can be regarded as the detailed division of the observed image into many different regions, each of which is considered uniform. These regions form a segmentation map with similar color, texture, brightness and other characteristics. Through entropy rate superpixel segmentation, the superpixels it produces are relatively uniform and regular.

[0110] 2. Feature extraction of edge-preserving filtering

[0111] For the reduced-dimensional hyperspectral image obtained in step 1, the edge-preserving filters with different parameter settings are used to filter the hyperspectral image after the dimension reduction in step 1, and filter images with different parameters are obtained. Different parameter settings will obtain different filtering results, representing different characteristics of hyperspectral images and having their special advantages in classifying objects or features of different scales. Therefore, the combination of edge features with different parameters can extract additional spatial information from the image and improve the classification accuracy. The parameters include the size of the local window that controls the filtered pixels, and are used to define how much the weight of the pixel is reduced due to the intensity difference between the reference pixels. The motivation of this step is to obtain the multi-scale structural information retained in the EPF obtained with different parameter settings. Different filtered images are formed by adjusting the parameters of the domain transform filter. The filtered images will look very different in terms of retaining edges and boundaries. When the degree of blur is relatively large, the spectral separability of pixels belonging to different objects will be reduced. In addition, for smooth areas with noise, filtering with high smoothness can effectively reduce noise and improve the spectral purity of pixels in the area.

[0112] 3. Dimensionality reduction of filtered images

[0113] The generated EPF contains not only supplementary information but also a lot of redundant information. Although edge-preserving filtering can smooth the image while preserving the boundary, the smoothing operation can still reduce the absolute difference between pixels belonging to different categories. To solve this problem, PCA can not only reduce the dimension of the data, but also improve the separability of the spectral pixels. Therefore, PCA technology is used in this step to reduce the dimension of the edge-preserving filtering images with different parameters obtained in step 2.

[0114] P=PCA(F,L) (12)

[0115] Where L is the number of main components retained in P, and F is the image after edge-preserving filtering. In other words, PCA is performed on the filtered image, so the XxK-dimensional EPF will be converted into L main components. It can represent the main information after filtering.

[0116] 4. EPF Fusion

[0117] In this step, the hyperspectral image after dimension reduction obtained in step 1 is used as the spectral information feature and fused with the spatial information feature obtained by applying the edge-preserving filter in step 3. The fused information is used as the input of SVM to obtain the classification map.

[0118] The filtered images obtained by EPF with different parameter settings, i.e., the object-level features (spatial features), are used to construct two different kernels: spatial kernel and spectral kernel.

[0119]

[0120]

[0121] Among them, k spa (x i ,x j ) represents the spatial kernel, K spe (x i ,x j ) represents the spectral kernel, σ represents the width of the radial basis kernel function, which is a free parameter, and σ s is the width of the empty spectral kernel, σ w is the spectral kernel width; x i , xj is the spectral data after normalization; For domain transformation transfer filtering from x i 、x j The obtained spatial-spectral characteristics, is the spectral information; the two radial basis kernel functions obtained above are constructed into a composite kernel.

[0122] This embodiment does not directly use the spatial features after dimensionality reduction in step 3 as SVM input for classification, but instead forms a composite kernel with the spectral features obtained in step 1 and the spatial features obtained in step 3, thereby making full use of the spectral features and the spatial features.

[0123] The kernel strategy is usually adopted in SVM, where the kernel function is defined as:

[0124] K(x i ,x j )=<φ(x i ),φ(x j )> (15)

[0125] Substituting equation (16) into equation (15) and solving the dual Lagrangian problem, we can obtain the final decision function of test x:

[0126]

[0127] where a i is the Lagrange multiplier. In order to further utilize the spatial information of HSI to construct two different types of kernels, namely K spa (x i ,x j ),K spe (x i ,x j ), and then the composite kernel is calculated by the weighted average method as shown in (17):

[0128]

[0129] in Represents the weight, k spe represents the spectral information, k spa Represents spatial information.

[0130] 5. Generate classification map

[0131] The composite kernels with different parameter settings obtained in step 4 are used as the input of SVM to obtain the classification graph.

[0132] 6. Decision Fusion

[0133] The support vector machine classifier predicts the fusion of spatial information and spectral information obtained by EPF with different parameter settings, and generates the final classification result through the majority voting decision fusion strategy.

Claims

1. A hyperspectral image classification method, characterized in that: Includes steps: S1. Performing superpixel PCA dimensionality reduction on the original hyperspectral image, wherein the input hyperspectral image is segmented by superpixels to obtain uniform regions, the uniform regions are processed by PCA to obtain low-dimensional matrices, and the low-dimensional matrices are rearranged to obtain a hyperspectral image with reduced dimensionality; S2, setting a plurality of edge-preserving filters to filter the reduced-dimensional hyperspectral images respectively to obtain a plurality of filtered images; S3, performing PCA processing on the multiple filtered images to obtain multiple filter image principal components; S4, constructing the dimension-reduced hyperspectral image obtained in step S1 into a spectral kernel based on a radial basis kernel function; constructing the plurality of filtered image principal components obtained in step S3 into a plurality of spatial kernels based on a radial basis kernel function; and using the spectral kernel to fuse with the plurality of spatial kernels to construct a plurality of composite kernels; S5, inputting the multiple composite kernels into the SVM classifier to obtain multiple classification graphs; S6. Perform decision fusion on multiple classification graphs to obtain the final classification result.

2. A hyperspectral image classification method according to claim 1, characterized in that: The hyperspectral image is a three-dimensional cube, and the three-dimensional cube is reshaped into a low-dimensional matrix X. The reshaped low-dimensional matrix X is a two-dimensional matrix, as shown in the following formula: X=[x1,x2,...,x i ,...,x P ]∈R L*P P=MN Where X represents the hyperspectral data after dimensionality reduction X∈R M×N×L , M, N and L represent the number of rows, columns and sampling wavelengths of the image respectively; each column represents a vector of a pixel; x i Represents the i-th pixel vector in the original image; x i =[x i1 ,x i2 ,…,x il ] T 。 3. A hyperspectral image classification method according to claim 1, characterized in that: The PCA processing in step S3 is to calculate the linear mapping of the pixels of the original image from the initial high-dimensional space to the low-dimensional space, and the calculation formula is: y i =w T x i Where w represents the transformation matrix, x i Represents the i-th pixel vector in the original image, y i A low-dimensional space matrix representing the i-th pixel vector.

4. A hyperspectral image classification method according to claim 1, characterized in that: In step S2, the parameters of each edge preserving filter in the plurality of edge preserving filters are different, and the parameters include: the number of edge preserving filters EPF and parameters δs and δr controlling the smoothness of the edge preserving filters EPF.

5. The hyperspectral image classification method according to claim 1, characterized in that: Step S4 constructs the spatial kernel and the spectral kernel using the following formulas: Among them, k spa (x i ,x j ) represents the spatial kernel, K spe (x i ,x j ) represents the spectral kernel, σ represents the width of the radial basis kernel function, which is a free parameter, σ s is the width of the empty spectral kernel, σ w is the spectral kernel width; x i 、x j Normalized spectral data; For domain transformation transfer filtering from x i 、x j The obtained spatial-spectral characteristics, is the spectral information; the two radial basis kernel functions obtained above are constructed into a composite kernel.

6. A hyperspectral image classification method according to claim 1, characterized in that: The decision fusion in step S6 is achieved through a majority voting algorithm.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the hyperspectral image classification method according to any one of claims 1 to 6 is implemented.

8. A hyperspectral image classification system, characterized in that: include: Superpixel segmentation unit, PCA processing unit, edge-preserving filter, base kernel function building module, SVM classifier, decision fusion module; The superpixel segmentation unit performs superpixel segmentation on the input hyperspectral image to obtain a uniform area, the PCA processing unit performs PCA processing on the uniform area to obtain a low-dimensional matrix, and rearranges the low-dimensional matrix to obtain a reduced-dimensional hyperspectral image, the edge-preserving filter filters the reduced-dimensional hyperspectral image with different parameters to obtain a plurality of filtered images, the PCA processing unit performs PCA processing on the plurality of filtered images to obtain a plurality of filtered image principal components, the base kernel function construction module constructs the reduced-dimensional hyperspectral image into a spectral kernel based on a radial basis kernel function, the base kernel function construction module constructs the plurality of filtered image principal components into a plurality of spatial kernels based on a radial basis kernel function, and the spectral kernel and the plurality of spatial kernels are respectively fused to obtain a plurality of composite kernels, the plurality of composite kernels are input into an SVM classifier to obtain a plurality of classification graphs, and the decision fusion module performs decision fusion on the plurality of classification graphs to obtain a final classification result.

9. A hyperspectral image classification system according to claim 8, characterized in that: The hyperspectral image is a three-dimensional cube, and the three-dimensional cube is reshaped into a low-dimensional matrix X. The reshaped low-dimensional matrix X is a two-dimensional matrix, as shown in the following formula: X=[x1,x2,...,x i ,...,x P ]∈R L*P P=MN Where X represents the hyperspectral data after dimensionality reduction X∈R M×N×L , M, N and L represent the number of rows, columns and sampling wavelengths of the image respectively; each column represents a vector of a pixel; x i Represents the i-th pixel vector in the original image; x i =[x i1 ,x i2 ,…,x il ] T 。 10. A hyperspectral image classification system according to claim 8, characterized in that: The base kernel function building module constructs the spatial kernel and the spectral kernel using the following formulas: Among them, k spa (x i , x j ) represents the spatial kernel, K spe (x i , x j ) represents the spectral kernel, σ represents the width of the radial basis kernel function, which is a free parameter, σ s is the width of the empty spectral kernel, σ w is the spectral kernel width; x i 、x j Normalized spectral data; For domain transformation transfer filtering from x i 、x j The obtained spatial-spectral characteristics, is the spectral information; the two radial basis kernel functions obtained above are constructed into a composite kernel.

Images