A hyperspectral anomaly detection method based on spectral-spatial feature fusion

Abstract

The application discloses a hyperspectral anomaly detection method based on space spectrum feature fusion, which comprises the following steps: step one, performing principal component analysis on a hyperspectral image to obtain a principal component image; step two, performing a morphological closing operation on the principal component image by using a computer and multiple structure elements to obtain a closing operation feature image; step three, performing a morphological opening operation on the principal component image by using a computer and multiple structure elements to obtain an opening operation feature image; step four, fusing the closing operation feature image, the principal component image and the opening operation feature image to obtain a feature matrix based on space spectrum feature fusion; and step five, performing anomaly score calculation on the feature matrix based on space spectrum feature fusion by using a computer and a Reed-Xiaoli algorithm, and then obtaining anomaly points and background points in the hyperspectral image according to the anomaly scores. The method is simple in steps and reasonable in design, and effectively improves anomaly detection precision by fusing spectral features and spatial features.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a hyperspectral anomaly detection method based on spatial-spectral feature fusion. Background Technology

[0002] Hyperspectral remote sensing imagery is a type of remote sensing data with integrated image and spectral characteristics. It provides dozens or even hundreds of continuous and finely detailed spectral information points for the same scene, exhibiting extremely high spectral resolution. Leveraging the spectral and spatial characteristics of hyperspectral images, various hyperspectral image processing techniques have emerged, such as hyperspectral image classification, spectral unmixing, and target detection. Hyperspectral image anomaly detection falls under unsupervised target detection, its main purpose being to identify target pixels that differ from background features without any prior spectral information, making it highly practical.

[0003] However, most current hyperspectral anomaly detection algorithms only utilize the spectral information of hyperspectral images and cannot effectively utilize their spatial information. Therefore, hyperspectral anomaly detection based on spatial-spectral feature fusion will help improve detection performance. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a hyperspectral anomaly detection method based on spatial-spectral feature fusion. The method has simple steps and a reasonable design. It obtains a feature matrix based on spatial-spectral feature fusion by fusing closed-operation feature maps, principal component images, and open-operation feature maps. Then, it detects anomaly points and background points in hyperspectral images by calculating anomaly scores based on the feature matrix based on spatial-spectral feature fusion. In this way, the anomaly detection accuracy is effectively improved by fusing spectral features and spatial features.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a hyperspectral anomaly detection method based on spatial-spectral feature fusion, characterized in that the method includes the following steps:

[0006] Step 1: Perform principal component analysis on the hyperspectral image to obtain the principal component image;

[0007] Step 2: Use a computer to perform morphological closing operations on the principal component image using multiple structuring elements to obtain the closing operation feature map;

[0008] Step 3: Use a computer to perform morphological opening operations on the principal component image using multiple structuring elements to obtain the opening operation feature map;

[0009] Step 4: Fuse the closing operation feature map, principal component image, and opening operation feature map to obtain a feature matrix based on spatial spectrum feature fusion;

[0010] Step 5: Use a computer to calculate the anomaly score of the feature matrix based on spatial-spectral feature fusion using the Reed-Xiaol i algorithm, and then obtain the anomaly points and background points in the hyperspectral image based on the anomaly score.

[0011] The above-mentioned hyperspectral anomaly detection method based on spatial-spectral feature fusion is characterized in that: in step one, principal component analysis is performed on the hyperspectral image to obtain the principal component image, and the specific process is as follows:

[0012] Step 101: Using a computer hyperspectral image denoted as I, the hyperspectral image matrix X is represented as... Where, x j,i Let d represent the data corresponding to the i-th band at the j-th pixel in hyperspectral image I, where j is a positive integer and its value ranges from 1 to n; i and d are both positive integers, and 1 ≤ i ≤ d, n represents the total number of pixels in hyperspectral image I, and d is the number of bands in hyperspectral image I.

[0013] Step 102: Use a computer to perform principal component analysis on the matrix X of the hyperspectral image to obtain A principal components;

[0014] Step 103: Use a computer to record the image corresponding to the a-th principal component as the a-th principal component image; where a and A are positive integers, and the value range of a is 1≤a≤A; and A is less than d.

[0015] The above-mentioned hyperspectral anomaly detection method based on spatial-spectral feature fusion is characterized in that: in step two, a computer is used to perform morphological closing operations on the principal component image using multiple structuring elements to obtain a closing operation feature map. The specific process is as follows:

[0016] Step 201: Using a computer, perform morphological closing operations on the image of the a-th principal component using the b-th structuring element to obtain the a-th principal component image. b A closed operation feature map; where b is a positive integer, and 1≤b≤β, β is the total number of structuring elements, and the value of β ranges from 3 to 6;

[0017] Step 202: Use a computer to process the a-th... b The matrix corresponding to the feature map of the closing operation is denoted as the a-th... b Matrix B a,b ;and Where T represents transpose. Indicates the a-th b Matrix B a,b The eigenvalue corresponding to the j-th pixel;

[0018] Step 203: Following the methods described in steps 201 to 202, obtain the a-th... β The matrix corresponding to the feature map of the closing operation is denoted as the a-th...β Matrix B a,β ,and in, Indicates the a-th β Matrix B a,β The eigenvalue corresponding to the j-th pixel.

[0019] The above-mentioned hyperspectral anomaly detection method based on spatial-spectral feature fusion is characterized in that: in step three, a computer is used to perform morphological opening operations on the principal component image using multiple structuring elements to obtain an opening operation feature map. The specific process is as follows:

[0020] Step 301: Using a computer, perform a morphological opening operation on the a-th principal component image using the b-th structuring element to obtain the a-th principal component image. b One open operation feature map;

[0021] Step 302: Use a computer to process the a-th... b The matrix corresponding to the a-th opening operation feature map is denoted as the a-th matrix. b Matrix B′ a,b ;and Where T represents transpose. Indicates the a-th b Matrix B′ a,b The eigenvalue corresponding to the j-th pixel;

[0022] Step 303: Following the methods described in steps 301 to 302, obtain the a-th... β The matrix corresponding to the a-th opening operation feature map is denoted as the a-th matrix. β Matrix B′ a,β ,and in, Indicates the a-th β Matrix B′ a,β The eigenvalue corresponding to the j-th pixel.

[0023] The above-mentioned hyperspectral anomaly detection method based on spatial-spectral feature fusion is characterized in that: in step four, the closing operation feature map, the principal component image, and the opening operation feature map are fused to obtain a feature map based on spatial-spectral feature fusion. The specific process is as follows:

[0024] Step 401: Using a computer, denote the a-th principal component matrix corresponding to the a-th principal component image as I(a), and I(a) = [I1(a) … I j (a) … I n (a)] T Among them, I j (a) represents the eigenvalue corresponding to the j-th pixel in the a-th principal component matrix I(a);

[0025] Step 402: Use a computer to process the a-th... b Matrix B a,b The a-th principal component matrix I(a), the a-th principal component matrix b Matrix B a ′ ,b The eigenvalues ​​corresponding to the j-th pixel are fused to obtain the eigenvalue matrix corresponding to the j-th pixel, denoted as C. (j) ,and Among them, C (j) The size is A(2β+1)×1;

[0026] Step 403: Use a computer to generate the eigenvalue matrix C corresponding to the first pixel. (1) The eigenvalue matrix C corresponding to the j-th pixel is... (j) ... to the eigenvalue matrix C corresponding to the nth pixel (n) The feature matrix C = [C] is obtained by fusing spatial spectral features. (1) …C (j) …C (n) ] T .

[0027] The above-mentioned hyperspectral anomaly detection method based on spatial-spectral feature fusion is characterized in that: in step five, a computer is used to calculate anomaly scores on the feature matrix based on spatial-spectral feature fusion using the Reed-Xiaoli algorithm, and then anomaly points and background points in the hyperspectral image are obtained based on the anomaly scores. The specific process is as follows:

[0028] Step 501: Use a computer to process the feature matrix based on spatial-spectral feature fusion according to the formula. The average matrix of the feature matrix obtained based on spatial spectral feature fusion is obtained.

[0029] Step 502: Use a computer to calculate the formula. The covariance matrix E of the feature matrix based on spatial spectral feature fusion is obtained;

[0030] Step 503: Use a computer to calculate the formula. Obtain the anomaly score for the j-th pixel. Among them, E -1 Denotes the inverse of the covariance matrix E;

[0031] Step 504: Use a computer to calculate the anomaly score of the j-th pixel in the feature matrix based on spatial-spectral feature fusion. Compare with the threshold τ, when If the value is greater than the threshold τ, then the j-th pixel in the hyperspectral image is an outlier; when... If the value is less than or equal to the threshold τ, then the j-th pixel in the hyperspectral image is a background point.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] 1. The method of the present invention has simple steps and reasonable design, which solves the problem that most current hyperspectral anomaly detection algorithms only utilize the spectral information of hyperspectral images and cannot effectively utilize the spatial information of hyperspectral images.

[0034] 2. This invention uses morphological opening and closing operations to extract spatial information based on morphology, which has low time and space computational complexity and can be effectively applied to the processing and analysis of remote sensing images.

[0035] 3. This invention extracts spatial structural information of different scales and contrasts in hyperspectral images by performing morphological opening and closing operations on multiple structural elements of known size and shape, thereby fully exploring the spatial information of hyperspectral images and laying a good foundation for subsequent anomaly detection.

[0036] 4. This invention utilizes the Reed-Xiaol i algorithm to calculate anomaly scores on feature matrices based on spatial-spectral feature fusion, and then obtains anomaly points and background points in hyperspectral images based on the anomaly scores, which can effectively improve the accuracy of background and anomaly detection.

[0037] In summary, the method of this invention is simple in steps and reasonable in design. It obtains a feature matrix based on spatial-spectral feature fusion by fusing closed operation feature map, principal component image and open operation feature map. Then, it realizes the detection of abnormal points and background points in hyperspectral image by calculating the anomaly score based on the feature matrix based on spatial-spectral feature fusion. In this way, the anomaly detection accuracy is effectively improved by fusing spectral features and spatial features.

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] like Figure 1 As shown, the hyperspectral anomaly detection method based on spatial-spectral feature fusion of the present invention includes the following steps:

[0041] Step 1: Perform principal component analysis on the hyperspectral image to obtain the principal component image;

[0042] Step 2: Use a computer to perform morphological closing operations on the principal component image using multiple structuring elements to obtain the closing operation feature map;

[0043] Step 3: Use a computer to perform morphological opening operations on the principal component image using multiple structuring elements to obtain the opening operation feature map;

[0044] Step 4: Fuse the closing operation feature map, principal component image, and opening operation feature map to obtain a feature matrix based on spatial spectrum feature fusion;

[0045] Step 5: Use a computer to calculate the anomaly score of the feature matrix based on spatial-spectral feature fusion using the Reed-Xiaoli algorithm, and then obtain the anomaly points and background points in the hyperspectral image based on the anomaly score.

[0046] In this embodiment, the principal component image is obtained by performing principal component analysis on the hyperspectral image in step one. The specific process is as follows:

[0047] Step 101: Using a computer hyperspectral image denoted as I, the hyperspectral image matrix X is represented as... Where, x j,i Let d represent the data corresponding to the i-th band at the j-th pixel in hyperspectral image I, where j is a positive integer and its value ranges from 1 to n; i and d are both positive integers, and 1 ≤ i ≤ d, n represents the total number of pixels in hyperspectral image I, and d is the number of bands in hyperspectral image I.

[0048] Step 102: Use a computer to perform principal component analysis on the matrix X of the hyperspectral image to obtain A principal components;

[0049] Step 103: Use a computer to record the image corresponding to the a-th principal component as the a-th principal component image; where a and A are positive integers, and the value range of a is 1≤a≤A; and A is less than d.

[0050] In this embodiment, step two involves using a computer to perform morphological closing operations on the principal component image using multiple structuring elements to obtain a closing operation feature map. The specific process is as follows:

[0051] Step 201: Using a computer, perform morphological closing operations on the image of the a-th principal component using the b-th structuring element to obtain the a-th principal component image. b A closed operation feature map; where b is a positive integer, and 1≤b≤β, β is the total number of structuring elements, and the value of β ranges from 3 to 6;

[0052] Step 202: Use a computer to process the a-th... b The matrix corresponding to the feature map of the closing operation is denoted as the a-th... b Matrix B a,b ;and Where T represents transpose. Indicates the a-th b Matrix B a,b The eigenvalue corresponding to the j-th pixel;

[0053] Step 203: Following the methods described in steps 201 to 202, obtain the a-th... β The matrix corresponding to the feature map of the closing operation is denoted as the a-th... β Matrix B a,β ,and in, Indicates the a-th β Matrix B a,β The eigenvalue corresponding to the j-th pixel.

[0054] In this embodiment, step three involves using a computer to perform morphological opening operations on the principal component image using multiple structuring elements to obtain an opening operation feature map. The specific process is as follows:

[0055] Step 301: Using a computer, perform a morphological opening operation on the a-th principal component image using the b-th structuring element to obtain the a-th principal component image. b One open operation feature map;

[0056] Step 302: Use a computer to process the a-th... b The matrix corresponding to the a-th opening operation feature map is denoted as the a-th matrix. b Matrix B′ a,b ;and Where T represents transpose. Indicates the a-th b Matrix B′ a,b The eigenvalue corresponding to the j-th pixel;

[0057] Step 303: Following the methods described in steps 301 to 302, obtain the a-th... β The matrix corresponding to the a-th opening operation feature map is denoted as the a-th matrix. β Matrix B′ a,β ,and in, Indicates the a-th β Matrix B′ a,β The eigenvalue corresponding to the j-th pixel.

[0058] In this embodiment, step four involves fusing the closing operation feature map, the principal component image, and the opening operation feature map to obtain a feature map based on spatial spectrum feature fusion. The specific process is as follows:

[0059] Step 401: Using a computer, denote the a-th principal component matrix corresponding to the a-th principal component image as I(a), and I(a) = [I1(a) … I j (a) … I n (a)] T Among them, I j (a) represents the eigenvalue corresponding to the j-th pixel in the a-th principal component matrix I(a);

[0060] Step 402: Use a computer to process the a-th... b Matrix B a,b The a-th principal component matrix I(a), the a-th principal component matrix b Matrix B′ a,b The eigenvalues ​​corresponding to the j-th pixel are fused to obtain the eigenvalue matrix corresponding to the j-th pixel, denoted as C. (j) ,and Among them, C (j) The size is A(2β+1)×1;

[0061] Step 403: Use a computer to generate the eigenvalue matrix C corresponding to the first pixel. (1) The eigenvalue matrix C corresponding to the j-th pixel is... (j) ... to the eigenvalue matrix C corresponding to the nth pixel (n) The feature matrix C = [C] is obtained by fusing spatial spectral features. (1) …C (j) …C (n) ] T .

[0062] In this embodiment, step five uses a computer to calculate anomaly scores on the feature matrix based on spatial-spectral feature fusion using the Reed-Xiaoli algorithm, and then obtains anomaly points and background points in the hyperspectral image based on the anomaly scores. The specific process is as follows:

[0063] Step 501: Use a computer to process the feature matrix based on spatial-spectral feature fusion according to the formula. The average matrix of the feature matrix obtained based on spatial spectral feature fusion is obtained.

[0064] Step 502: Use a computer to calculate the formula. The covariance matrix E of the feature matrix based on spatial spectral feature fusion is obtained;

[0065] Step 503: Use a computer to calculate the formula. Obtain the anomaly score for the j-th pixel. Among them, E -1 Denotes the inverse of the covariance matrix E;

[0066] Step 504: Use a computer to calculate the anomaly score of the j-th pixel in the feature matrix based on spatial-spectral feature fusion. Compare with the threshold τ, when If the value is greater than the threshold τ, then the j-th pixel in the hyperspectral image is an outlier; when... If the value is less than or equal to the threshold τ, then the j-th pixel in the hyperspectral image is a background point.

[0067] In this embodiment, C (j) In the expression, 'a' takes values ​​from 1 to A, and B represents the a1-th matrix. a,1 The eigenvalue corresponding to the j-th pixel in the image. B represents the a1-th matrix. a ′ ,1 The eigenvalue corresponding to the j-th pixel.

[0068] In this embodiment, the hyperspectral image is an actual hyperspectral image AVIRIS-I taken in San Diego, California, USA. A 120×120 pixel portion of the image was captured using the Airborne Visible / Infrared Imaging Spectroradiometer (AVIRIS) and contains data from 189 bands within the 370nm–2510nm spectral band. Therefore, the hyperspectral image has 100 rows and 100 columns, so n = 100×100, and the number of bands d is 189.

[0069] In this embodiment, the value of β is 4, so 1≤b≤4.

[0070] In this embodiment, the size of the first structuring element is 3*3, the size of the second structuring element is 5*5, the size of the third structuring element is 7*7, the size of the fourth structuring element is 9*9, and the element value of each structuring element is 1.

[0071] In this embodiment, the value of A is 5, and C (j) The size is 45×1.

[0072] In this embodiment, the threshold τ is determined as follows:

[0073] Step A: Following the method described in step 504, obtain the anomaly score s1 of the first pixel to the anomaly score s2 of the nth pixel. n And the anomaly score s~1 of the first pixel to the anomaly score s~n of the nth pixel. n Arrange them in ascending order to obtain the maximum anomaly score;

[0074] Step B: The threshold τ is 5% of the maximum anomaly score.

[0075] In summary, the method of this invention is simple in steps and reasonable in design. It obtains a feature matrix based on spatial-spectral feature fusion by fusing closed operation feature map, principal component image and open operation feature map. Then, it realizes the detection of abnormal points and background points in hyperspectral image by calculating the anomaly score based on the feature matrix based on spatial-spectral feature fusion. In this way, the anomaly detection accuracy is effectively improved by fusing spectral features and spatial features.

[0076] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A hyperspectral anomaly detection method based on spectral-spatial feature fusion, characterized in that, The method comprises the following steps: Step one, principal component analysis is performed on the hyperspectral image to obtain a principal component image; Step two, a computer is used to perform morphological closing operation on the principal component image by using multiple structure elements to obtain a closing operation feature map; Step three, a computer is used to perform morphological opening operation on the principal component image by using multiple structure elements to obtain an opening operation feature map; Step four, the closing operation feature map, the principal component image and the opening operation feature map are fused to obtain a feature matrix based on spatial-spectral feature fusion; Step five, a computer is used to perform abnormal score calculation on the feature matrix based on spatial-spectral feature fusion by using the Reed-Xiaoli algorithm, and then the abnormal points and background points in the hyperspectral image are obtained according to the abnormal scores; In step one, the principal component image is obtained by performing principal component analysis on the hyperspectral image, and the specific process is as follows: Step 101: Record using computer hyperspectral images Then the matrix of the hyperspectral image Represented as ;in, Represents hyperspectral images The Middle 1 pixel Data corresponding to each band It is a positive integer, and The value range is 1~ ; and All are positive integers, and , Represents hyperspectral images Total number of pixels For hyperspectral images Number of bands; Step 102, using a computer to perform principal component analysis on the matrix of the hyperspectral image principal components are obtained.​ Step 103: Use a computer to process the first... The image corresponding to the principal component is denoted as the i-th principal component. One principal component image; among which... and It is a positive integer, and The range of values ​​for is 1≤ ≤ ;and Less than ; In step two, the computer is used to perform morphological closing operation on the principal component image by using multiple structure elements to obtain a closing operation feature map, and the specific process is as follows: Step 201, using a computer to perform morphological closing operation on the first principal component image by using the first structural element to obtain a first closed operation feature image; wherein, n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. n is a positive integer, and m is an integer greater than 0. Step 202, using a computer to record the matrix corresponding to the first closed operation feature map as the first matrix ; and ; wherein, denotes transposition, denotes the eigenvalue corresponding to the first pixel in the first matrix ​ Step 203, obtaining the matrix corresponding to the first closed operation feature map according to the method in steps 201-202 , and denoting the first matrix , and ; wherein, denotes the feature value corresponding to the first pixel in the first matrix . Step 203, obtaining the matrix corresponding to the first closed operation feature map according to the method in steps 201-202 , and denoting the first matrix , and ; wherein, denotes the feature value corresponding to the first pixel in the first matrix . Step 203, obtaining the matrix corresponding to the first closed operation feature map according to the method in steps 201-202 , and 2.The hyperspectral anomaly detection method based on the fusion of spectral features according to claim 1, characterized in that: In step three, the computer is used to perform morphological opening operation on the principal component image by using multiple structure elements to obtain an opening operation feature map, and the specific process is as follows: Step 301: Using a computer to utilize the... The structure element pairs with the first Morphological opening operations are performed on the principal component images to obtain the first... One open operation feature map; Step 302, using a computer to record the matrix corresponding to the first feature map of the first matrix ; and ; wherein, denotes the transpose, denotes the eigenvalue corresponding to the first matrix at the first pixel Step 303: Following the methods described in steps 301 to 302, obtain the first... The matrix corresponding to the i-th opening operation feature map is denoted as the i-th... matrix ,and ;in, Indicates the first matrix The Middle The feature value corresponding to each pixel. 3.The hyperspectral anomaly detection method based on the fusion of spectral features according to claim 2, characterized in that: In step four, the closing operation feature map, the principal component image and the opening operation feature map are fused to obtain a feature map based on spatial-spectral feature fusion, and the specific process is as follows: Step 401: Use a computer to process the first... The first principal component image corresponds to the first... The principal component matrix is ​​denoted as ,and ;in, Indicates the first Principal component matrices The Middle The eigenvalues ​​corresponding to each pixel; Step 402, using a computer to fuse the feature values corresponding to the first pixel in the first matrix, the first principal component matrix, the second matrix and the first matrix to obtain a feature value matrix corresponding to the first pixel and recorded as ; wherein, the size of the first principal component matrix is ;​​​​​​​​​​ Step 403, using a computer to fuse the feature value matrix corresponding to the first pixel ,..., the feature value matrix corresponding to the nth pixel ,..., the feature value matrix corresponding to the nth pixel ,..., the feature value matrix corresponding to the nth pixel ,..., the feature value matrix corresponding to the nth pixel , to obtain a feature matrix based on the fusion of the spectral feature .

4. The hyperspectral anomaly detection method based on fusion of spectral features according to claim 3, characterized in that: In step five, the computer is used to perform abnormal score calculation on the feature matrix based on spatial-spectral feature fusion by using the Reed-Xiaoli algorithm, and then the abnormal points and background points in the hyperspectral image are obtained according to the abnormal scores, and the specific process is as follows: Step 501, using a computer to obtain an average matrix of the feature matrix based on the fusion of spatial and spectral features according to the formula . ; Step 502, obtaining a covariance matrix of the feature matrix based on the spectral feature fusion by using a computer according to a formula ;​ Step 503: Use a computer to calculate the formula. , obtained the Anomaly score of each pixel ;in, Represents the covariance matrix The reverse; Step 504, using a computer to compare the anomaly score of the i-th pixel in the feature matrix based on the spectral feature fusion with a threshold value, when the anomaly score is greater than the threshold value, the i-th pixel in the hyperspectral image is an anomaly point; when the anomaly score is less than or equal to the threshold value, the i-th pixel in the hyperspectral image is a background point. ​​​​​​​​​

Images