Hyperspectral image classification method based on distance separable waveband optimization

Abstract

The invention relates to the technical field of optical remote sensing hyperspectral imaging, in particular to a hyperspectral image classification method based on distance separability waveband optimization. Firstly, an original image is preprocessed and then divided into a training set and a test set which are not overlapped in content; secondly, pairing ground feature categories needing to be identified in pairs to form a ground feature category combination, and calculating the wavebands in the image to obtain an optimal separable waveband set; then, according to the transformation wave band in the optimal separable wave band set, the training set and the test set are simplified, and only data consistent with the transformation wave band in the optimal separable wave band set are reserved; and finally, using the optimized training set to train a support vector machine, and using the trained support vector machine to complete classification labeling of ground feature categories on the optimized test set. According to the method, the Hughes phenomenon under limited samples can be effectively solved, the classification complexity and cost are greatly reduced, the classification speed is high, and the classification precision is high.

Description

Technical Field

[0001] The present invention relates to the technical field of optical remote sensing hyperspectral imaging, and in particular to a hyperspectral image classification method based on distance-separable band optimization. Background Art

[0002] Hyperspectral remote sensing images have both spatial and spectral characteristics, which makes them carry rich ground object information and have a wide range of applications in ground object classification, such as mineral distribution mapping, vegetation cover type determination, precision agriculture and other quantitative remote sensing fields.

[0003] With increasing spectral resolution, the spectral spatial dimension of hyperspectral images is increasing, and their ability to represent different land object categories is also improving. However, the high dimensionality, large data volume, and spectral variability of hyperspectral data also pose the Hughes problem for hyperspectral remote sensing image classification. Furthermore, the spectral and spatial dimensions of hyperspectral data are strongly correlated, resulting in significant data redundancy, which increases classification complexity and reduces efficiency. During the acquisition process, hyperspectral images are affected by factors such as the atmosphere, the sun, observation geometry, sensor noise, and spectral efficiency, resulting in the possibility of spectral aliasing between different land objects. Existing active learning methods require relatively high accuracy in the sample data for land object classification and require manual screening of large amounts of high-quality training datasets, making them unsuitable for widespread and rapid classification applications.

[0004] The existing technology proposes a hyperspectral remote sensing image classification based on minimum noise factorization (MNF) and support vector machine (SVM). The minimum noise factorization transform is used to remove noise and extract features, and the support vector machine classification method is used to classify the feature-reduced images. The classification results show that combining MNF and SVM can effectively improve the classification speed and overall classification accuracy. However, this method does not consider the correlation between the spatial and spectral dimensions of hyperspectral data, and its recognition ability for complex scenes is insufficient. In addition, there is human subjective judgment in the band selection after the MNF dimensionality reduction transform. Summary of the Invention

[0005] The purpose of the present invention is to provide a hyperspectral image classification method based on distance-separable band optimization, which mainly solves the problems existing in the above-mentioned prior art. On the basis of analyzing the spatial-spectral correlation of hyperspectral data, it designs an image feature classification algorithm based on distance-separable band optimization to ensure that high-precision classification results are obtained by using effectively selected bands.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is to provide a hyperspectral image classification method based on distance separable band optimization, which is characterized in that it is used to identify the category of ground objects in the original image, comprising the steps of:

[0007] Step S100, preprocessing the original image to obtain a preprocessed image;

[0008] Step S200, dividing the pre-processed images into two non-overlapping original training sets and original test sets;

[0009] Step S300, pairing the feature categories to be identified in pairs to form feature category combinations;

[0010] Step S400, based on the transformed bands in the preprocessed image, for each of the ground object category combinations, traverse all the transformed bands and select the optimal separable band corresponding to the ground object category combination;

[0011] Step S500, simplifying the original training set and the original test set according to the transformed bands in the optimal separable band set;

[0012] Step S600: Use the optimized training set to train a support vector machine, and use the support vector machine model obtained after training to complete the classification and labeling of the ground feature categories on the optimized test set.

[0013] In step S100 , the original image includes original data of multiple original bands; correspondingly, the preprocessed image also includes preprocessed data of multiple transformed bands.

[0014] In step S400, each of the ground feature category combinations corresponds to one of the optimal separable bands; all of the optimal separable bands corresponding to all of the ground feature category combinations are merged to obtain an optimal separable band set;

[0015] In step S500, among the preprocessed images in the original training set and the original test set, only the preprocessed data that is consistent with the transformed bands in the optimal separable band set is retained, and the preprocessed data corresponding to the transformed bands that are not in the optimal separable band set are deleted, thereby obtaining an optimized training set and an optimized test set.

[0016] Furthermore, in step S100, the pre-processing includes sub-steps:

[0017] Step S101, performing image processing on the original image to remove noise;

[0018] Step S102, performing standardization processing on the original image after noise removal to obtain a standardized image and an image data covariance matrix; the standardized image contains the same original band as the original image;

[0019] Step S103, performing decorrelation noise estimation on the standardized image based on space and spectrum to obtain the noise covariance matrix;

[0020] Step S104: combining the image data covariance matrix and the noise covariance matrix, further processing the normalized image after noise processing using a minimum noise separation transform to obtain the preprocessed image. The number of transformed bands contained in the preprocessed image is equal to the number of original bands.

[0021] Furthermore, in step S101, a median filter is used to remove noise from the original image.

[0022] Furthermore, in step S102,

[0023] First, each pixel value in the original image after noise removal and the original band are reduced from a three-dimensional matrix to a two-dimensional intermediate matrix of size [(m×n), P]; where m is the width of the original image, n is the height of the original image, and P is the number of the original bands;

[0024] Secondly, based on the two-dimensional intermediate matrix, the mean μ of each original band is calculated k and standard deviation σ k , and arranged into a mean matrix μ=[μ1,μ2,...,μ p ] and the standard deviation matrix σ=[σ1,σ2,…,σ p ];

[0025]

[0026] Among them, X i,k is the value of the i-th pixel in the k-th band in the two-dimensional intermediate matrix, and 1≤k≤p;

[0027] Then, the two-dimensional intermediate matrix X1 is transformed using the mean matrix μ and the standard deviation matrix σ to obtain the standardized image in a two-dimensional matrix form. The specific conversion formula is:

[0028] The standardized image = (the two-dimensional intermediate matrix - the mean matrix) / the standard deviation matrix.

[0029] Finally, the image data covariance matrix of the standardized image is calculated using the mean of the original band, satisfying the formula:

[0030]

[0031]

[0032] x i，kis the i-th pixel value of the k-th band of the two-dimensional matrix, where 1≤k≤p;

[0033] x l，k is the i-th pixel value of the l-th band of the two-dimensional matrix, where 1≤l≤p.

[0034] Furthermore, in step S103, decorrelation noise estimation is performed on the standardized image to obtain the noise covariance matrix, which includes:

[0035] First, the normalized image is divided into smaller w×h plot matrices;

[0036] Next, perform a multiple linear regression on each of the plot matrices and calculate the regression coefficients:

[0037]

[0038] Make the fitting condition Get the minimum value; among them,

[0039] is the residual value of each pixel value in the standardized image;

[0040] X j,i,k represents the pixel value at position (i, j) in the kth original band, and satisfies 1≤i≤w, 1≤j≤h, and 1≤k≤p; where X p,k satisfy:

[0041]

[0042] That is, for the images of the first original band and the last original band, only the adjacent original band is used for fitting;

[0043] Again, calculate the noise variance σ between the different transform bands k and l kl for:

[0044]

[0045] Wherein, M=w×h-1, i.e., the number of pixels in each plot matrix participating in the multiple linear regression, and k and l are both integers ranging from 1 to the total number of the original bands contained in the standardized image;

[0046] Then, the noise covariance matrix ∑ is obtained by using the noise variance value. N :

[0047]

[0048] Furthermore, in step S104, the minimum noise separation transformation includes:

[0049] First, the image data covariance matrix ∑ and the noise covariance matrix ∑ are used to calculate the image data covariance matrix ∑. N Get the transformation matrix so that the transformation matrix is the matrix ∑ -1 ∑ N The eigenvector matrix of

[0050] Then, the standardized image is transformed using the transformation matrix to obtain the preprocessed image: the preprocessed image = the transformation matrix ˙ the standardized image.

[0051] Furthermore, in step S400, the calculation of the optimal separability band for each combination of the ground feature categories includes the following sub-steps:

[0052] Step S401, calculating the Bhattacharyya distance between two ground object categories on each of the transformed bands of the preprocessed image according to the selected ground object category combination;

[0053] Step S402, using the transformed band corresponding to the maximum Bhattacharyya distance calculated in step S401 as the optimal separability band corresponding to the selected feature category;

[0054] Step S403: Select the next ground feature category combination and proceed to step S401 until all ground feature category combinations are processed.

[0055] Furthermore, in step S401, the calculation formula of the Bhattacharyya distance is:

[0056]

[0057] in, and are respectively the mean values of the two land object categories in the land object category combination on the transformation band p, and and They are respectively the variances of the two ground object categories in the ground object category combination on the transformation band p, and p is an integer ranging from 1 to the total number of the transformation bands contained in the preprocessed image.

[0058] Furthermore, in step S600, for each combination of the ground feature categories, a Gaussian radial basis kernel function is used to train a support vector machine model.

[0059] Furthermore, completing the classification and labeling of the ground object categories on the test set in step S600 includes the following sub-steps:

[0060] Step S601, selecting a pixel in the pre-processed image test set;

[0061] Step S602, using the support vector machine model to classify the pixel through the support vector machine model of all the combinations of ground object categories;

[0062] Step S603, counting the number of votes obtained for each of the ground object categories, and taking the ground object category with the highest number of votes as the final category of the pixel;

[0063] Step S604 , selecting the next pixel in the pre-processed image, and jumping to step S602 for processing until all the pixels in the pre-processed image are processed.

[0064] In view of the above technical features, the present invention proposes a hyperspectral image classification method based on distance-separable band selection. Based on noise removal and standardization, this method performs a minimum noise separation transformation based on image noise assessment, calculates the maximum distance between categories, and selects the optimal classification band. Finally, a support vector machine classifier and voting decision-making are used to achieve fine-grained image classification. Compared with existing technologies, this method has significant advantages: it can effectively address the Hughes phenomenon under limited sample conditions, significantly simplifies classification complexity and cost, and achieves rapid classification and high classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a method flow chart of a preferred embodiment of the hyperspectral image classification method based on distance separable band optimization of the present invention;

[0066] Figure 2 This is a method flow chart of the preprocessing process in a preferred embodiment of the hyperspectral image classification method based on distance separable band optimization of the present invention;

[0067] Figure 3 This is a flow chart of a method for selecting the optimal separable band in a preferred embodiment of the hyperspectral image classification method based on distance separable band optimization of the present invention;

[0068] Figure 4 This is a flow chart of a method for training a support vector machine and using the support vector machine for classification and labeling in a preferred embodiment of the hyperspectral image classification method based on distance-separable band optimization of the present invention;

[0069] Figure 5 In an example of a preferred embodiment of the hyperspectral image classification method based on distance-separable band optimization of the present invention, an RGB pseudo-color image of Indian Pines, a visible-near-infrared hyperspectral image taken by an AVIRS sensor in Indiana, is used;

[0070] Figure 6 for Figure 5 The corresponding real feature classification map and corresponding feature category description;

[0071] Figure 7 for Figure 5 The classification results. DETAILED DESCRIPTION

[0072] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0073] Here, we take a hyperspectral image as an example to illustrate the specific implementation steps of a preferred embodiment of the present invention to complete image classification. The example image is the visible near-infrared hyperspectral image IndianPines taken by the AVIRS sensor in Indiana. Figure 5 Its RGB pseudo-color image, Figure 6 The following is a real-world object classification map and corresponding object category descriptions. The image is 145×145 pixels in length and width, with a wavelength range of 0.4μm to 2.5μm and 224 original bands. In this example, 200 valid bands are used for calculation, and the image coverage area includes 16 object categories.

[0074] See also Figure 1 The present invention discloses a hyperspectral image classification method based on distance separable band optimization, which is used to identify the category of ground objects in the original image. As shown in the figure, a preferred embodiment thereof includes the steps of:

[0075] Step S1: image preprocessing.

[0076] The original image contains all the information, but some information is not necessary for object recognition, and it also contains noise introduced during the acquisition process. Therefore, before actual object recognition, the original image is processed using a preprocessing algorithm to produce a preprocessed image. In this preprocessed image, information that affects object recognition is removed. The preprocessing process generally includes noise removal and image normalization.

[0077] In particular, raw images typically contain raw data collected from multiple original bands, such as red light and infrared bands. Accordingly, the preprocessing process also involves preprocessing the band information within the original image. For example, the minimum noise factorization transform (MNF) is applied to the original bands based on decorrelated noise estimation, resulting in transformed bands corresponding to the original bands. The preprocessed data then stores information within each transformed band. In this example image, the band range is 0.4μm to 2.5μm, and the number of original bands is 224.

[0078] Step S2: Divide the training set and the test set.

[0079] The preprocessed images are divided into a training set and a test set. In this embodiment, 30% of the preprocessed images are used as the training set, and the remaining 70% are used as the test set. The contents of the training set and the test set do not overlap, but their union covers all the preprocessed images.

[0080] Step S3: Pairing ground feature category combinations.

[0081] The object categories to be identified are known in advance and have nothing to do with the collected images. In the present invention, it is necessary to find the optimal band to distinguish two different object categories. Therefore, all the object categories to be identified are first paired to form an object category combination. That is, if there are T object categories C{C1, C2, ..., C T}, then we finally get T×(T-1) / 2 ground object category combinations. In this example, the image coverage area contains 16 ground object categories, so finally 16×(16-1) / 2=120 ground object category combinations are formed.

[0082] Step S4: Select the optimal separability band.

[0083] Different object categories have distinct spectral characteristics. Therefore, within each object category combination (pair of two object categories), the bands that best distinguish the two object categories will also vary. To optimize the subsequent support vector machine learning and classification process, it is necessary to select the bands that show the greatest difference between the two object categories from the perspective of the support vector machine to facilitate classification. Furthermore, because the number of transformed bands contained in the preprocessed image is limited, it is sufficient to select from the transformed bands contained in the preprocessed image, simplifying the selection process. In this example, 200 valid bands are used in the calculation. After denoising the original image (145×145×200), the resulting preprocessed image also contains 145×145 spatial dimensions, with 200 transformed bands.

[0084] Therefore, for each feature class combination, all transformed bands in the preprocessed image are traversed to select the optimal separable band. Ultimately, each feature class combination has a corresponding optimal separable band. Finally, the optimal separable bands corresponding to each feature class combination are merged to obtain the optimal separable band set. In this example, the optimal separable band set contains only 32 transformed bands.

[0085] Step S5: Screen the optimized training set and the optimized test set.

[0086] Based on the optimal separability bands, the spectra not in the optimal separability bands are removed from the original training and test sets to obtain the optimized training and test sets. In this example, the images in the original training and test sets both contain 200 original bands, while the optimized training and test sets only contain 32 transformed bands.

[0087] Step S6: train the support vector machine and complete classification labeling.

[0088] After selecting the optimal band set, the support vector machine model is trained using the optimized training set and the Gaussian radial basis kernel function. A single support vector machine model is trained for each feature class combination. After training, all trained support vector machine models are used to classify and label the features on the optimized test set.

[0089] In this example, there are 120 feature class combinations, so 120 support vector machines are trained. The optimal separability band set contains only 32 transform bands. Therefore, the support vector machine does not need to learn from 200 transform bands, but only needs to learn and classify them on these 32 transform bands. Furthermore, because the optimal separability bands are selected based on the maximum difference between the two feature classes within these selected transform bands, the support vector machine can easily classify any two feature classes, both during learning and classification, achieving optimal learning and classification results.

[0090] See also Figure 2 In the above step S1, the pre-processing includes multiple processing flows and therefore also includes multiple sub-steps of sequential processing, as follows:

[0091] Step S11: image denoising.

[0092] After obtaining the original image, the median filter algorithm is used to process all pixel values in the original image to remove noise.

[0093] Step S12: image normalization processing.

[0094] Different images may have systematic deviations due to variations in acquisition conditions when they are collected. In order to compare different images, the original images need to be normalized to remove the systematic deviations.

[0095] Specifically, a standardization process is performed on the original image after noise removal to obtain a standardized image; the standardized image includes the same original band as the original image;

[0096] Step S121: Generate a two-dimensional intermediate matrix.

[0097] Arrange the pixel values in the original image after noise removal in order, and then reduce the dimensionality of the 3D matrix together with the original bands to form a 2D intermediate matrix of size [(m × n), P], where m is the width of the original image, n is the height of the original image, and P is the number of original bands.

[0098] Step S122: Calculate the mean matrix and the standard deviation matrix.

[0099] Read the elements in the two-dimensional intermediate matrix and calculate the mean μ corresponding to each original band using the following formula k and standard deviation σ k .

[0100]

[0101] Among them, X i,k is the value of the i-th pixel in the k-th band in the two-dimensional intermediate matrix, and 1≤k≤p;

[0102] The calculated values are arranged into a mean matrix μ = [μ1, μ2, ..., μ p ] and the standard deviation matrix σ=[σ1,σ2,…,σ p ].

[0103] Step S123: obtaining a standardized image.

[0104] Using the mean matrix μ and the standard deviation matrix σ, the two-dimensional intermediate matrix X1 is transformed into a standardized image. The specific conversion formula is:

[0105] Normalized image = (two-dimensional intermediate matrix - mean matrix) / standard deviation matrix.

[0106] The normalized image is also in the form of a two-dimensional matrix with the specification of [(m*n),P].

[0107] Specifically, different bands are processed separately. For a given band, the mean of all pixel values within it is subtracted, and then divided by the standard deviation of the band. The result is used as the new pixel value corresponding to the original pixel position in the standardized image. Therefore, the standardized image has the same pixel specifications as the two-dimensional intermediate matrix.

[0108] Step S124: Calculate the image covariance matrix.

[0109] The standardized image data after normalization has the specification of [(m*n),p], which can be expressed as where x i is a column vector of [m*n,1], then its corresponding covariance matrix specification is [p,p], and the calculation formula is as follows:

[0110]

[0111] x i，k is the pixel value of the i-th pixel in the k-th band in the two-dimensional matrix of the standardized image, and satisfies 1≤k≤p. l，k is the value of the i-th pixel in the l-th band in the two-dimensional matrix of the standardized image, and satisfies 1≤l≤p.

[0112] Step S13: decorrelation noise estimation.

[0113] Decorrelated noise estimation is performed on the normalized image based on spatial and spectral bands.

[0114] Step S131, dividing the plot matrix.

[0115] The matrix of the normalized image is further divided into smaller w×h plot matrices.

[0116] Step S132: Perform multiple linear regression on the plot matrix.

[0117] Perform multiple linear regression on each plot matrix and calculate the regression coefficient:

[0118]

[0119] Make the fitting condition Get the minimum value.

[0120] in, is the residual value of each pixel value in the standardized image. j,i,k represents the pixel value at position (i, j) in the kth original band, and satisfies 1≤i≤w, 1≤j≤h, and 1≤k≤p. p,k Satisfies the following formula:

[0121]

[0122] In particular, for the images of the first original band and the last original band, only one original band adjacent to them is used for fitting.

[0123] Step S133: Calculate the noise variance value.

[0124] Calculate the noise variance σ between different transform bands k and l kl , satisfying the following formula:

[0125]

[0126] Where M = w × h - 1, i.e., the number of pixels in each plot matrix participating in the multivariate linear regression. k and l are both integers ranging from 1 to the total number of original bands contained in the standard image, i.e., traversing all transformed bands.

[0127] Step S134: Calculate the noise covariance matrix.

[0128] The noise variance values are combined into the noise covariance matrix ∑ N :

[0129]

[0130] Step S14: performing minimum noise separation transformation processing on the standardized image.

[0131] The noise-processed normalized image is further processed using the minimum noise separation transform to obtain a preprocessed image. In the preprocessed image, the original bands are converted into transformed bands, so the number of transformed bands is equal to the number of original bands.

[0132] Step S141, obtaining a transformation matrix.

[0133] Using the image covariance matrix ∑ obtained in step S124 and the noise covariance matrix ∑ obtained in step S134 N , solve to get the transformation matrix. The transformation matrix is the matrix ∑ -1 ∑ N The eigenvector matrix of .

[0134] Step S142: Obtain pre-processed image

[0135] Multiply the transformation matrix and the matrix of the standardized image to obtain the preprocessed image, that is:

[0136] The preprocessed image = the transformation matrix ˙ the standardized image.

[0137] See also Figure 3In the above step S4, the optimal separability bands for training and classification are selected from all the transformed bands to ensure that the distance between each ground object category combination in the optimal separability band is the largest and easy to separate. Specifically, step S4 includes the following sub-steps:

[0138] Step S41 , calculating the distance of each ground feature category combination in each transformed band.

[0139] The same feature class combination exhibits different distances in different bands, so the distances for each feature class combination in different bands must be calculated. On the other hand, since we only need to filter the bands included in the preprocessed image, we do not need to calculate distances for bands not present in the image. Therefore, for each feature class combination, the distance between two feature classes is calculated for each transformed band included in the preprocessed image.

[0140] In the present invention, the distance is specifically selected as the Bhattacharyya distance, and its calculation formula is:

[0141]

[0142] Among them, μ c1 and μ c2 are the means of the two land object categories in the land object category combination on the transformed band p, and ∑ c1 and ∑ c2 are the variances of these two land feature categories in the transformed band p.

[0143] For each transformed band p, its Bhattacharyya distance is calculated, so p is an integer ranging from 1 to the total number of transformed bands included in the preprocessed image.

[0144] Step S42: Select the optimal separability band.

[0145] For each feature class combination, the Bhattacharyya distance for each transformed band can be obtained. The transformed band corresponding to the largest Bhattacharyya distance is selected as the optimal separability band for the feature class combination.

[0146] Step S43: traverse all combinations of ground feature categories.

[0147] If there is still a feature category combination for which the Bhattacharyya distance has not been calculated, the feature category combination is selected and the process proceeds to step S41. If the corresponding optimal separability bands are selected for all feature category combinations, then all the work in step S4 is completed.

[0148] See also Figure 4 In the above step S6, the trained support vector machine is used to complete the classification and labeling of the ground object categories on the test set, which includes the following sub-steps:

[0149] Step S61: Select a pixel.

[0150] Select a pixel in the preprocessed image.

[0151] Step S62: Vote for the feature category.

[0152] Using the pixel as input, the algorithm traverses all combinations of feature categories and classifies the pixel using the trained support vector machine corresponding to each feature category. After each support vector machine obtains the classification result, it votes for the feature category it selected.

[0153] Step S63: Determine the final category.

[0154] After all feature class combinations have voted, the number of votes received by each feature class is counted. The feature class with the highest number of votes is the optimal classification result for that pixel and is considered the final class for that pixel.

[0155] Step S64, traverse all pixels.

[0156] If there are still unclassified pixels in the pre-processed image, select this pixel as the next pixel and jump to step S62 for processing. If all pixels in the pre-processed image are processed, the object classification of the original image corresponding to the pre-processed image is completed and a result is obtained.

[0157] Next, the classification results of an embodiment of the hyperspectral image classification method based on distance separable band optimization of the present invention are compared with the classification results of the hyperspectral image classification method based on the same Figure 5 However, a comparison of the results using different algorithms is shown in Table 1. The algorithms included in the comparison are MLC (maximum likelihood classification) after preprocessing, MLC after band optimization, and SVM after preprocessing. The overall classification accuracy compared to the band optimization method proposed in this invention is shown below.

[0158] Table 1. Comparison of overall classification accuracy

[0159]

[0160]

[0161] As can be seen from Table 1, after band optimization, the classification accuracy of both MLC and SVM methods is improved. The overall classification accuracy proposed by the present invention is the highest, reaching 97.74%. The image classification results of the present invention are as follows: Figure 7 shown.

[0162] In summary, the present invention uses preprocessing methods such as noise removal, image normalization, and a minimum noise separation transform based on spatial / spectral decorrelation noise assessment. It then uses a band selection method based on class separability to select the optimal band combination after preprocessing. Finally, a support vector machine (SVM) and voting decision method are used to obtain the classification results. This effectively reduces data redundancy in the hyperspectral spatial spectral dimension, lowers classification complexity, and achieves high overall classification accuracy.

[0163] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A hyperspectral image classification method based on distance separable band optimization, characterized in that: Used to identify the object category in the original image, including the steps: Step S100, preprocessing the original image to obtain a preprocessed image; Step S200, dividing the pre-processed images into two non-overlapping original training sets and original test sets; Step S300, pairing the feature categories to be identified in pairs to form feature category combinations; Step S400, based on the transformed bands in the preprocessed image, for each of the ground object category combinations, traverse all the transformed bands and select the optimal separable band corresponding to the ground object category combination; Step S500, simplifying the original training set and the original test set according to the transformed bands in the optimal separable band set; Step S600: Use the optimized training set to train a support vector machine, and use the support vector machine model obtained after training to complete the classification and labeling of the ground feature categories on the optimized test set.

2. The hyperspectral image classification method based on distance separability band optimization according to claim 1 is characterized in that: In step S100 , the original image includes original data of multiple original bands; correspondingly, the preprocessed image also includes preprocessed data of multiple transformed bands.

3. The hyperspectral image classification method based on distance separability band optimization according to claim 1 is characterized in that: In step S100, the pre-processing includes sub-steps: Step S101, performing image processing on the original image to remove noise; Step S102, performing standardization processing on the original image after noise removal to obtain a standardized image and an image data covariance matrix; the standardized image contains the same original band as the original image; Step S103, performing decorrelation noise estimation on the standardized image based on space and spectrum to obtain the noise covariance matrix; Step S104, combining the image data covariance matrix and the noise covariance matrix, and further processing the noise-processed normalized image using a minimum noise separation transform to obtain the preprocessed image; The number of the transformed bands contained in the preprocessed image is equal to the number of the original bands.

4. The hyperspectral image classification method based on distance separability band optimization according to claim 3 is characterized in that: In step S101, median filtering is used to remove noise from the original image.

5. The hyperspectral image classification method based on distance separability band optimization according to claim 3 is characterized in that: In step S102, First, each pixel value in the original image after noise removal and the original band are reduced from a three-dimensional matrix to a two-dimensional intermediate matrix of size [(m×n), P]; where m is the width of the original image, n is the height of the original image, and P is the number of the original bands; Secondly, based on the two-dimensional intermediate matrix, the mean μ of each original band is calculated k and standard deviation σ k , and arranged into a mean matrix μ=[μ1,μ2,...,μ p ] and the standard deviation matrix σ=[σ1,σ2,…,σ p ]; Among them, X i,k is the value of the i-th pixel in the k-th band in the two-dimensional intermediate matrix, and 1≤k≤p; Then, the two-dimensional intermediate matrix X1 is transformed using the mean matrix μ and the standard deviation matrix σ to obtain the standardized image in a two-dimensional matrix form. The specific conversion formula is: The standardized image = (the two-dimensional intermediate matrix - the mean matrix) / the standard deviation matrix; Finally, the image data covariance matrix of the standardized image is calculated using the mean of the original band, satisfying the formula: x i，k is the i-th pixel value of the k-th band of the two-dimensional matrix, where 1≤k≤p; x l,k is the i-th pixel value of the l-th band of the two-dimensional matrix, where 1≤l≤p.

6. The hyperspectral image classification method based on distance separability band optimization according to claim 3 is characterized in that: In step S103, decorrelation noise estimation is performed on the standardized image to obtain the noise covariance matrix, which includes: First, the normalized image is divided into smaller w×h plot matrices; Next, perform a multiple linear regression on each of the plot matrices and calculate the regression coefficients: Make the fitting condition Get the minimum value; among them, is the residual value of each pixel value in the standardized image; X j,i,k represents the pixel value at position (i, j) in the kth original band, and satisfies 1≤i≤w, 1≤j≤h, and 1≤k≤p; where X p,k satisfy: That is, for the images of the first original band and the last original band, only the adjacent original band is used for fitting; Again, calculate the noise variance σ between the different transform bands k and l kl for: Wherein, M=w×h-1, i.e., the number of pixels in each plot matrix participating in the multiple linear regression, and k and l are both integers ranging from 1 to the total number of the original bands contained in the standardized image; Then, the noise covariance matrix ∑ is obtained by using the noise variance value. N :

7. The hyperspectral image classification method based on distance separability band optimization according to claim 3 is characterized in that: In step S104, the minimum noise separation transformation includes: First, the image data covariance matrix ∑ and the noise covariance matrix ∑ are used to calculate the image data covariance matrix ∑. N Get the transformation matrix so that the transformation matrix is the matrix ∑ -1 ∑ N The eigenvector matrix of Then, the standardized image is transformed using the transformation matrix to obtain the preprocessed image: the preprocessed image = the transformation matrix ˙ the standardized image.

8. The hyperspectral image classification method based on distance separability band optimization according to claim 1 is characterized in that: In step S400 , each of the land feature category combinations corresponds to one of the optimal separable bands; all of the optimal separable bands corresponding to all of the land feature category combinations are merged to obtain an optimal separable band set.

9. The hyperspectral image classification method based on distance separability band optimization according to claim 1, characterized in that: In step S400, the optimal separability band is calculated for each combination of the ground feature categories, including the following sub-steps: Step S401, calculating the Bhattacharyya distance between two ground object categories on each of the transformed bands of the preprocessed image according to the selected ground object category combination; Step S402, using the transformed band corresponding to the maximum Bhattacharyya distance calculated in step S401 as the optimal separability band corresponding to the selected feature category; Step S403: Select the next ground feature category combination and proceed to step S401 until all ground feature category combinations are processed.

10. The hyperspectral image classification method based on distance separability band optimization according to claim 9, characterized in that: In step S401, the calculation formula of the Bhattacharyya distance is: in, and are respectively the mean values of the two land object categories in the land object category combination on the transformation band p, and and They are respectively the variances of the two ground object categories in the ground object category combination on the transformation band p, and p is an integer ranging from 1 to the total number of the transformation bands contained in the preprocessed image.

11. The hyperspectral image classification method based on distance separability band optimization according to claim 1, characterized in that: In step S500, among the preprocessed images in the original training set and the original test set, only the preprocessed data that is consistent with the transformed bands in the optimal separable band set is retained, and the preprocessed data corresponding to the transformed bands that are not in the optimal separable band set are deleted, thereby obtaining an optimized training set and an optimized test set.

12. The hyperspectral image classification method based on distance separability band optimization according to claim 1, characterized in that: In step S600, for each combination of the ground feature categories, a support vector machine model is trained using a Gaussian radial basis kernel function.

13. The hyperspectral image classification method based on distance separability band optimization according to claim 1, characterized in that: The step S600 of completing the classification and labeling of the feature categories on the test set includes the following sub-steps: Step S601, selecting a pixel in the pre-processed image test set; Step S602, using the support vector machine model to classify the pixel through the support vector machine model of all the combinations of ground object categories; Step S603, counting the number of votes obtained for each of the ground object categories, and taking the ground object category with the highest number of votes as the final category of the pixel; Step S604 , selecting the next pixel in the pre-processed image, and jumping to step S602 for processing until all the pixels in the pre-processed image are processed.

Images