Hyperspectral band selection method, system, computer storage medium and device for optimal combination strategy

Through subspace division and the combination of information entropy and local density, representative bands of hyperspectral remote sensing images are selected and combined evaluation is performed, which solves the problem of band selection being susceptible to noise and ignoring spectral context information in the prior art, and improves classification Accuracy and reduce data redundancy.

CN114758232BActive Publication Date: 2024-11-08XIAN UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210362647.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-07
Publication Date
2024-11-08
Estimated Expiration
2042-04-07

AI Technical Summary

Technical Problem

Existing hyperspectral remote sensing image band selection methods are susceptible to noise and ignore spectral context information, resulting in a reduced classification accuracy.

Method used

Using the optimal combination strategy, through the combination of subspace division and information entropy and local density, representative bands with a large amount of information and are not easily affected by noise are selected, and combined evaluation is carried out from a global and local perspective to select the optimal band. combination.

Benefits of technology

The band selection accuracy of hyperspectral remote sensing images is improved, the impact of noise is reduced, the amount of original information is retained, and data redundancy is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114758232B_ABST
    Figure CN114758232B_ABST
Patent Text Reader

Abstract

The present invention relates to a hyperspectral band selection method, system, computer storage medium and device of an optimal combination strategy to solve the technical problem that the existing clustering algorithm is easily affected by noise and ignores the spectral context information, thereby reducing the classification accuracy. The method includes: 1) dividing the first subspace; 2) forming the second subspace; 3) selecting representative bands; 4) using a search strategy to obtain the optimal band combination. The system includes a space division module, a representative band selection module and an optimal band combination selection module. The present invention uses subspace division, representative band selection and optimal band combination selection to achieve the band selection task, and the selected bands have a higher amount of information and less redundancy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a hyperspectral band selection method, and in particular to a hyperspectral band selection method, system, computer storage medium and device of an optimal combination strategy. Background Art

[0002] Hyperspectral remote sensing images are based on traditional remote sensing images, adding one-dimensional spectral information and combining it with two-dimensional image information, thus achieving "spectrum combination". Since different substances have different reflection intensities, the ground object information of the hyperspectral image can be easily identified based on the spectral response curve of the hyperspectral image. Because of this, hyperspectral remote sensing technology has been widely used in many fields, such as marine remote sensing, vegetation research, geological exploration, atmospheric environment and military applications.

[0003] However, as the number of bands of hyperspectral remote sensing images increases, the number of statistical parameters of the samples also increases; under limited conditions, the change in the number of bands used for classification accuracy shows the "Hughes" phenomenon, the so-called "curse of dimensionality". At the same time, the band widths of these high-dimensional remote sensing data are very narrow and overlap with each other, resulting in high data redundancy and high correlation between bands, which affects the subsequent classification accuracy.

[0004] Data dimensionality reduction is currently the best way to solve the "dimensionality curse" problem of hyperspectral remote sensing image data. The existing hyperspectral data dimensionality reduction methods mainly include band (feature) selection and feature extraction. Compared with feature extraction, band (feature) selection can retain the original characteristics and physical meaning of remote sensing data, and is widely used in the process of data dimensionality reduction.

[0005] Traditional band selection methods are divided into clustering-based and sorting-based methods: Clustering-based methods, such as fast density peak clustering (E-FDPC) and affinity propagation (AP), regard each band as a point and construct a similarity matrix through Euclidean distance or other criteria. Then, using this matrix, some clustering methods are used to cluster these bands so that the bands with high similarity are grouped together. Finally, the bands that are most related to other bands in each cluster are selected as representative bands. Unfortunately, this type of algorithm makes the similarity of the obtained subsets very small, but the selected bands may not have enough information. In addition, when performing clustering algorithms, hyperspectral bands are usually considered to be unordered. It leads to ignoring the contextual information of the entire spectral band. For sorting-based methods, typical methods are maximum variance principal component analysis (MVPCA) and linear constrained minimum variance based constrained band selection (LCMV-CBS). This method calculates the score of each hyperspectral band based on certain criteria. The scores are then sorted in ascending or descending order, and the bands corresponding to the maximum or minimum score values ​​are selected as the target subset according to the meaning of the score values. Although these methods have achieved satisfactory results, they also have two shortcomings. First, these algorithms are very sensitive to hyperspectral bands containing noise, which makes the selected bands likely to have strong similarities. Second, most algorithms change the original information in the hyperspectral image cube through spatial transformation, resulting in the loss of some key information. Summary of the invention

[0006] The purpose of the present invention is to solve the technical problem that the existing clustering algorithm is easily affected by noise and ignores the spectral context information, thereby reducing the classification accuracy. A hyperspectral band selection method, system, computer storage medium and device with an optimal combination strategy are proposed. By introducing a subspace partitioning method, the spatial information of the spectrum is used for clustering and the influence of noise bands is reduced. On this basis, two representative bands are selected in each subspace using the information volume and correlation of the bands. Starting from the global strategy optimality and the local strategy optimality, these representative bands are combined and the performance of the corresponding combination is evaluated to select the optimal band combination.

[0007] The technical solution of the present invention is:

[0008] A hyperspectral band selection method of an optimal combination strategy is characterized in that it includes the following steps:

[0009] A1. According to the number of bands of the hyperspectral image, the bands are divided into m first subspaces in order, where m ≥ 1 and m is an integer;

[0010] A2. Move the boundary points of adjacent first subspaces based on local density, divide the bands with high similarity into the same subspace, and divide the bands with large differences into different subspaces, forming m second subspaces;

[0011] A3, calculating the information entropy of all bands in the hyperspectral image, and selecting two bands with large information entropy and high similarity in each second subspace as representative bands, so as to ensure that the representative bands have a large amount of information and are not noise bands;

[0012] A4. Global strategy to select the optimal band combination: select a representative band from each second subspace according to the order of the second subspace to form multiple band combinations; solve the average value E of the information entropy of each representative band in the band combination and the average value D of the distance between adjacent representative bands, calculate the result of E / D as the score of the band combination, and select the band combination with the highest score as the optimal band combination.

[0013] Furthermore, in step A2, the boundary points of the adjacent first subspaces are moved based on the local density, specifically:

[0014] A2.1. Calculate the similarity matrix of all bands in the hyperspectral image and normalize it;

[0015] A2.2. Calculate the local density of all bands according to the similarity matrix;

[0016] A2.3. Select m-1 partition regions [F L ,F R ], where F L is the center band of the previous first subspace, F R For the center band of the next first subspace, the band with the lowest local density is selected as the boundary point of the new adjacent subspace in the divided area, so that the bands with high similarity are divided into the same subspace, and the bands with large differences are divided into different subspaces, completing a boundary point division;

[0017] A2.4. Set the maximum number of boundary point divisions. When the number of divisions reaches the maximum number of boundary point divisions or the boundary points of the subspace do not change, stop the subspace division to obtain m second subspaces.

[0018] Further, in step A4, the multiple band combinations are 5m band combinations.

[0019] The present invention also provides a hyperspectral band selection method of an optimal combination strategy, which is special in that it includes the following steps:

[0020] B1. According to the number of bands of the hyperspectral image, the bands are divided into m first subspaces in order, where m ≥ 1 and m is an integer;

[0021] B2. Move the boundary points of adjacent first subspaces based on local density, divide the bands with high similarity into the same subspace, and divide the bands with large differences into different subspaces, forming m second subspaces;

[0022] B3, calculating the information entropy of all bands in the hyperspectral image, and selecting two bands with large information entropy and high similarity in each second subspace as representative bands to ensure that the representative bands have a large amount of information and are not noise bands;

[0023] B4. Local strategy to select the optimal band combination:

[0024] B4.1. Select a representative band with the largest information entropy from the two representative bands in the first second subspace, and form a sub-band combination with two representative bands in the adjacent second subspace respectively. Calculate the average value E' of the information entropy of the representative bands in the sub-band combination and the average value D' of the distance between adjacent representative bands in the sub-band combination. Calculate the result of E' / D' as the score of the sub-band combination, and select the sub-band combination with the highest score as the optimal sub-band combination.

[0025] B4.2. According to the scoring calculation method of step B4.1, the optimal sub-band combination is combined with the representative bands of the subsequent second subspaces in the order of the second subspace, and the combination ends at the last second subspace to obtain the optimal band combination.

[0026] Furthermore, in step B2, the boundary points of the adjacent first subspaces are moved based on the local density, specifically:

[0027] B2.1. Calculate the similarity matrix of all bands of the hyperspectral image and normalize it;

[0028] B2.2. Calculate the local density of all bands according to the similarity matrix;

[0029] B2.3, select m-1 partition regions [F L ,F R ], where F L is the center band of the previous first subspace, F R For the center band of the next first subspace, select the band with the lowest local density in the divided area as the boundary point of the new adjacent subspace, so that the bands with high similarity are divided into the same subspace, and the bands with large differences are divided into different spaces, completing a boundary point division;

[0030] B2.4. Set the maximum number of boundary point divisions. When the number of divisions reaches the maximum number of boundary point divisions or the boundary points of the subspace do not change, stop the subspace division to obtain m second subspaces.

[0031] The present invention also provides a hyperspectral band selection system of an optimal combination strategy, which is special in that it includes a subspace division module, a representative band selection module and an optimal band combination selection module;

[0032] The output of the subspace division module is connected to the input of the representative band selection module, and the subspace division module is used to divide the hyperspectral band into multiple subspaces;

[0033] The output of the representative band selection module is connected to the input of the optimal band combination selection module, and the representative band selection module is used to select two representative bands with large information entropy and high similarity from each subspace;

[0034] The optimal band combination selection module is used to select the optimal band combination from multiple band combinations formed by representative band combinations of each subspace.

[0035] Furthermore, the optimal band combination selection module is an optimal band combination global selection module or an optimal band combination local selection module.

[0036] The present invention also provides a computer storage medium on which a computer program is stored, and the special feature of the computer storage medium is that when the computer program is executed by a processor, the steps of the hyperspectral band selection method of the optimal combination strategy are implemented.

[0037] At the same time, the present invention also provides a computer device, including a processor, a memory connected to the processor, and a computer program that can be run on the processor, the special feature of which is that when the processor executes the computer program, the steps of the hyperspectral band selection method of the above-mentioned optimal combination strategy are implemented.

[0038] Beneficial effects of the present invention:

[0039] 1. The present invention adopts local density to divide the second subspace, and quickly and effectively divides the hyperspectral image into multiple second subspaces with extremely high similarity inside and large differences outside.

[0040] 2. The present invention uses information entropy and local density to select two representative bands from each second subspace. These two representative bands can fully represent each second subspace.

[0041] 3. The present invention proposes two search strategies, namely, a global strategy (OCS-G) and a local strategy (OCS-L), from the global and local perspectives respectively, and combines and screens representative bands to obtain the optimal band combination.

[0042] 4. The present invention utilizes subspace partitioning, representative band selection and optimal band combination selection to achieve the band selection task, and the selected bands have a higher amount of information and less redundancy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the process flow of the hyperspectral band selection method of the optimal combination strategy of the present invention. DETAILED DESCRIPTION

[0044] This embodiment provides a hyperspectral band selection method of an optimal combination strategy, which includes the following specific implementation methods according to different optimal band combination selection strategies.

[0045] Example 1

[0046] See also Figure 1 ,The optimal band combination selection of hyperspectral bands using the global ,strategy includes the following steps:

[0047] A1. According to the number of bands of the hyperspectral image, the bands are divided into m first subspaces in order, where m ≥ 1 and m is an integer; the first subspace is a coarse subspace obtained by coarse division according to the number and order of the bands.

[0048] A2. Move the boundary points of adjacent first subspaces based on local density to form m second subspaces.

[0049] Specifically, the boundary points of the adjacent first subspace are moved based on the local density:

[0050] A2.1. Calculate the similarity matrix of all bands in the hyperspectral image and normalize them.

[0051] A2.2. Calculate the local density of all bands according to the similarity matrix.

[0052] A2.3. Select m-1 partition regions [F L ,F R ], where F L is the center band of the previous first subspace, F R The center band of the latter first subspace is selected, and the band with the lowest local density in the divided area is selected as the boundary point of the new adjacent subspace, so that the bands with high similarity according to the similarity matrix are divided into the same subspace, and the bands with large differences are divided into different subspaces, completing a boundary point division; wherein, the center band is obtained by dividing the total number of bands in the first subspace by 2 and taking an integer.

[0053] A2.4. According to experimental verification, considering the accuracy of band selection, the maximum number of boundary point divisions is set. When the number of divisions reaches the maximum number of boundary point divisions or the boundary points of the subspace do not change, the subspace division is stopped to obtain m second subspaces. It can be understood that the second subspace is a fine subspace obtained by subdividing the first subspace according to the local density calculated based on the similarity matrix of each band.

[0054] A3. Calculate the information entropy of all bands in the hyperspectral image, and select two bands with large information entropy and high similarity in each second subspace as representative bands to ensure that the representative bands have a large amount of information and are not noise bands. It can be understood that when selecting representative bands, firstly, the representative band of each second subspace should be greater than 1, and secondly, when the representative band of each second subspace is greater than or equal to 3, the amount of analysis and calculation will increase exponentially, and the algorithm for evaluating representative bands will become complicated. After comprehensive consideration, it is most appropriate to select two representative bands.

[0055] A4. Select a representative band from each second subspace in the order of the second subspace to form 5m band combinations. It can be understood that if the number of band combinations is too small, the performance of the algorithm cannot guarantee the optimal band combination. If the number is too large, the algorithm time will be affected. When the number of second subspaces is m, 5m band combinations are more appropriate. In addition, when the value of m is 1, 2 or 3, there are repeated band combinations in the band combination.

[0056] The average value E of the information entropy of each representative band in the band combination and the average value D of the distance between adjacent representative bands are solved, and the result of calculating E / D is used as the score of the band combination, and the band combination with the highest score is selected as the optimal band combination.

[0057] The global strategy is to adopt a random search strategy for the band combinations composed of representative band combinations, and then screen these band combinations to select the band combination with the best performance; this global strategy can obtain the band combination composed of almost all representative bands from a global perspective, and can obtain the band combination that is infinitely close to the global optimal solution.

[0058] Example 2

[0059] See also Figure 1 ,The optimal band combination selection of hyperspectral bands using local ,strategy includes the following steps:

[0060] B1. According to the number of bands of the hyperspectral image, the bands are divided into m first subspaces in order, where m ≥ 1 and m is an integer; the first subspace is a coarse subspace obtained by coarse division according to the number and order of the bands.

[0061] B2. Move the boundary points of adjacent first subspaces based on local density, divide the bands with high similarity into the same subspace, and divide the bands with large differences into different subspaces, forming m second subspaces.

[0062] Specifically, the boundary points of the adjacent first subspace are moved based on the local density:

[0063] B2.1. Calculate the similarity matrix of all bands of the hyperspectral image and normalize it.

[0064] B2.2. Calculate the local density of all bands according to the similarity matrix.

[0065] B2.3, select m-1 partition regions [F L ,F R ], where F L is the center band of the previous first subspace, F R For the center band of the latter first subspace, the band with the lowest local density in the divided area is selected as the boundary point of the new adjacent subspace, so that the bands with high similarity are divided into the same subspace, and the bands with large differences are divided into different spaces, completing a boundary point division.

[0066] B2.4. Set the maximum number of boundary point divisions. When the number of divisions reaches the maximum number of boundary point divisions or the boundary points of the subspace do not change, stop the subspace division to obtain m second subspaces. It can be understood that the second subspace is a fine subspace obtained by subdividing the first subspace according to the local density calculated based on the similarity matrix of each band.

[0067] B3. Calculate the information entropy of all bands in the hyperspectral image, and select two bands with large information entropy and high similarity in each second subspace as representative bands to ensure that the representative bands have a large amount of information and are not noise bands; because the information entropy and similarity must be met at the same time, the information entropy of the two representative bands may not be the largest.

[0068] B4. Select the optimal band combination:

[0069] B4.1. Select a representative band with the largest information entropy from the two representative bands in the first second subspace, and form a sub-band combination with the two representative bands in the adjacent second subspace. Calculate the average value E′ of the information entropy of the representative bands in the sub-band combination and the average value D′ of the distance between adjacent representative bands in the sub-band combination. Calculate the result of E′ / D′ as the score of the sub-band combination, and select the sub-band combination with the highest score as the optimal sub-band combination.

[0070] B4.2. According to the scoring calculation method of step B4.1, the optimal sub-band combination is combined with the selected representative bands of the subsequent second subspaces in the order of the second subspace, and the combination ends at the last second subspace to obtain the optimal band combination.

[0071] The local strategy adopts a forward search algorithm from a local perspective to screen the representative bands subspace by subspace. Not only is the performance not much different from the global strategy, but it also ensures the stability of the algorithm and reduces the running time of the algorithm. Although this strategy is called a local strategy, it is only a local search strategy, and the main body it searches is still all the representative bands. Therefore, the band combination it obtains is infinitely close to the global optimal band combination.

[0072] This embodiment also provides a corresponding system for implementing the hyperspectral band selection method of the above-mentioned optimal combination strategy, including a subspace division module, a representative band selection module and an optimal band combination selection module;

[0073] The output of the subspace partitioning module is connected to the input of the representative band selection module, and the subspace partitioning module is used to divide the hyperspectral band into multiple subspaces;

[0074] The output of the representative band selection module is connected to the input of the optimal band combination selection module. The representative band selection module is used to select two representative bands with large information entropy and high similarity from each subspace; the optimal band combination selection module is used to select the optimal band combination from multiple band combinations formed by representative band combinations of each subspace. The optimal band combination selection module is an optimal band combination global selection module or an optimal band combination local selection module; the optimal band combination global selection module is used to select the optimal band combination of hyperspectral bands through a global strategy; the optimal band combination local selection module is used to select the optimal band combination of hyperspectral bands through a local strategy.

[0075] An embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the steps of the hyperspectral band selection method of the optimal combination strategy described in the above embodiment are implemented; the storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM) or an optical disk, etc., which can store program codes.

[0076] An embodiment of the present invention also provides a computer device, which includes a processor, a memory connected to the processor, and a computer program that can be run on the processor, and the processor executes the computer program to implement the steps of the high-spectral band selection method of the optimal combination strategy described in the above embodiment; the computer device can be a portable computer such as a laptop computer and a tablet computer, or it can be a computer device such as a desktop computer, etc. Since the computer device has a computer program that can implement the engine system design method with parameter association in the above embodiment, it has the same beneficial effects and will not be repeated in this embodiment.

[0077] In order to better illustrate the effect of this embodiment, the following specific verification test was carried out.

[0078] (1) Test conditions: When the CPU is The experiments were conducted using MATLAB on an i5-10400F 3.4GHz CPU, 16G memory, AMD6600M GPU, and Win10 operating system.

[0079] (2) Comparison method: In order to demonstrate the superiority of the method of this embodiment, the sorting algorithm MVPCA and the clustering algorithm E-FDPC are selected as comparison algorithms. Specifically, MVPCA first constructs the covariance matrix of the bands and performs eigenvalue decomposition on the similarity matrix to construct the load factor matrix, and then sorts all the bands according to the matrix; E-FDPC first clusters all the bands, and then calculates and normalizes the local density and intra-class distance of each band to sort the bands.

[0080] (3) Datasets: The hyperspectral band datasets to be selected include the Indian Pines dataset, the Salinas dataset, and the Botswana dataset.

[0081] Indian Pines dataset: This scene was acquired by the AVIRIS sensor at the Indian Pines test site in northwestern Indiana. It has a total of 220 bands, but because the 104th-108th, 150th-163rd, and 220th bands cannot be reflected by water, we generally use the remaining 200 bands after removing these 20 bands as the research object; the data has a total of 21,025 pixels.

[0082] Salinas dataset: This scene is captured by the 224-band AVIRIS sensor over the Salinas Valley in California. It has a total of 224 bands. Generally, images of the remaining 204 bands are used after removing the 108th-112th, 154th-167th, and 224th bands that cannot be reflected by water. The size of the image is 512×217, so it contains a total of 111,104 pixels.

[0083] Botswana dataset: This scene was acquired by NASA EO-1 satellite in the Okavango Delta of Botswana from 2001 to 2004. It has a total of 242 bands, and generally uses 145 bands after removing 97 bands affected by noise. The image size is 1476×256, so it contains 377,856 pixels.

[0084] (4) Classifier: In order to test the performance of the band selection method, two classifiers were introduced: KNN classifier and SVM classifier, which were used to classify the Indian Pines dataset, Salinas dataset, and Botswana dataset respectively. The same parameter settings were set in the experiments to ensure the accuracy of the experiments. The parameter of KNN was set to 5, and the penalty factor of SVM was set to 1×10 4 , the gamma coefficient is 0.5. For both classifiers, 10% of the training samples are used, and the other samples are used as test samples. The test is run 5 times for the results of each algorithm, and the average value is taken as the final calculation result. The number of representative bands in the band combination is taken in the range of [5,50], and the overall accuracy (OA) and average overall accuracy (AOA) are used as metrics to test the performance of the method provided in this embodiment, the selected sorting algorithm MVPCA and the clustering algorithm E-FDPC.

[0085] (5) Experimental results analysis:

[0086] Table 1 OA (%) of KNN classifier against four algorithms on Indian Pines dataset

[0087] algorithm 5 8 10 12 15 20 25 30 40 50 OCS-G 62.23 66.33 67.37 66.35 66.92 66.40 66.44 65.85 67.17 66.29 OCS-L 61.85 65.72 67.24 66.27 66.68 67.45 66.64 65.58 66.99 67.39 MVPCA 45.06 46.70 54.24 56.89 56.33 58.56 59.07 59.49 61.04 63.55 E-FDPC 58.48 59.86 60.52 60.05 60.53 61.61 62.38 63.15 63.78 65.12

[0088] Table 2 OA (%) of SVM classifier for four algorithms on Indian Pines dataset

[0089] algorithm 5 8 10 12 15 20 25 30 40 50 OCS-G 68.03 71.19 73.80 77.23 77.04 77.20 78.01 78.75 79.51 79.32 OCS-L 67.90 70.65 73.83 76.99 77.87 78.07 78.67 79.18 78.97 79.71 MVPCA 50.49 51.35 58.53 60.34 63.93 66.36 69.44 69.78 71.62 73.12 E-FDPC 59.16 62.65 64.10 65.22 65.89 67.58 68.26 70.92 72.87 74.46

[0090] It can be seen from Table 1 and Table 2 that on the Indian Pines dataset, the global strategy (OCS-G) algorithm and the local strategy (OCS-L) algorithm have strong performance advantages: whether in the low frequency band or in the medium and high frequency band, the accuracy of the OCS-G algorithm and the OCS-L algorithm is higher than that of the MVPCA algorithm and the E-FDPC algorithm, and even reaches a 20% accuracy gap under extreme conditions; and by comparing the OCS-G algorithm and the OCS-L algorithm, it can be found that the OCS-G algorithm has a greater advantage in the low frequency band, and the OCS-L algorithm has a greater advantage in the medium and high frequency bands.

[0091] Table 3 OA (%) of KNN classifier for four algorithms on Salinas dataset

[0092] algorithm 5 8 10 12 15 20 25 30 40 50 OCS-G 86.51 88.06 88.94 89.05 89.14 89.12 89.02 88.93 88.86 88.89 OCS-L 86.16 87.85 89.12 88.90 89.10 89.05 89.01 88.94 89.08 88.96 MVPCA 73.52 76.39 78.96 79.77 80.14 84.96 85.12 85.70 87.22 87.71 E-FDPC 86.01 88.00 88.20 88.47 89.03 89.01 89.06 89.27 89.48 89.49

[0093] Table 4 OA (%) of SVM classifier for four algorithms on Salinas dataset

[0094]

[0095]

[0096] It can be seen from Tables 3 and 4 that in some low-frequency bands of the Salinas dataset, the classification accuracy of the E-FDPC algorithm achieves better results than the algorithm provided in this embodiment. This is because E-FDPC, as a clustering algorithm, has its own advantages in datasets with relatively simple scenarios such as the Salinas dataset. However, in most frequency bands, the classification accuracy of E-FDPC is lower than that of the OCS-G algorithm and the OCS-L algorithm, but the gap is not as large as that on the Indian Pines dataset; at the same time, it also reflects the robustness of the OCS-G algorithm and the OCS-L algorithm.

[0097] Table 5 OA (%) of KNN classifier for four algorithms on Botswana dataset

[0098] algorithm 5 8 10 12 15 20 25 30 40 50 OCS-G 80.40 81.93 83.26 84.58 83.22 85.29 85.34 83.46 84.35 83.18 OCS-L 80.42 81.35 83.41 84.07 82.24 85.66 85.47 85.35 84.46 82.55 MVPCA 60.25 59.81 59.56 58.01 77.38 80.14 81.53 80.14 80.77 81.62 E-FDPC 75.19 80.27 81.53 82.27 83.58 83.48 83.78 84.53 83.98 84.65

[0099] Table 6 OA (%) of SVM classifier for four algorithms on Botswana dataset

[0100] algorithm 5 8 10 12 15 20 25 30 40 50 OCS-G 84.09 85.93 86.95 87.93 88.22 90.35 90.26 89.63 90.84 91.07 OCS-L 83.70 85.35 87.19 87.64 87.10 90.60 90.28 90.75 90.87 90.16 MVPCA 68.23 70.10 70.33 70.35 84.97 86.33 88.51 88.62 88.36 89.48 E-FDPC 75.42 84.90 85.65 86.64 87.26 87.86 87.80 88.08 88.66 88.78

[0101] It can be seen from Table 5 and Table 6 that the OCS-G algorithm and the OCS-L algorithm show great advantages on the Botswana data set. Whether in the medium and high frequency bands or the low frequency bands, the OCS-G algorithm and the OCS-L algorithm in this embodiment show great performance advantages.

[0102] Table 7 AOA of KNN classifier for four algorithms on three data sets (%)

[0103] algorithm Indian pins dataset Salinas Dataset Botswana Dataset OCS-G 66.13 88.65 83.50 OCS-L 66.18 88.62 83.50 MVPCA 56.09 87.71 71.92 E-FDPC 61.55 89.49 82.33

[0104] Table 8 AOA of SVM classifier for four algorithms on three data sets (%)

[0105] algorithm Indian pins dataset Salinas Dataset Botswana Dataset OCS-G 66.52 88.87 83.81 OCS-L 66.61 88.86 83.81 MVPCA 57.20 82.79 73.09 E-FDPC 61.86 88.86 83.04

[0106] It can be seen from Tables 7 and 8 that, whether it is the KNN classifier or the SVM classifier, the average classification accuracy of the OCS-G algorithm and the OCS-L algorithm on the three data sets is better than that of the MVPCA algorithm and the E-FDPC algorithm, which fully demonstrates the performance advantages and robustness of the OCS-G algorithm and the OCS-L algorithm, and further proves the effectiveness of the global strategy and the local strategy to obtain the optimal band combination.

Claims

1. A hyperspectral band selection method based on an optimal combination strategy, characterized in that: The following steps are involved: A1. According to the number of bands of the hyperspectral image, the bands are divided into m first subspaces in order, where m ≥ 1 and m is an integer; A2, based on the local density, move the boundary points of the adjacent first subspaces to form m second subspaces; The boundary points of the adjacent first subspace are moved based on the local density, specifically: A2.

1. Calculate the similarity matrix of all bands in the hyperspectral image and normalize it; A2.

2. Calculate the local density of all bands according to the similarity matrix; A2.

3. Select m-1 partition regions [F L ,F R ], where F L is the center band of the previous first subspace, F R For the center band of the next first subspace, select the band with the lowest local density in each divided area as the boundary point of the new adjacent subspace to complete a boundary point division; A2.

4. Set the maximum number of boundary point divisions. When the number of divisions reaches the maximum number of boundary point divisions or the boundary points of the subspace do not change, stop the subspace division to obtain m second subspaces. A3. Calculate the information entropy of all bands in the hyperspectral image, and select two bands with large information entropy and high similarity in each second subspace as representative bands; A4. Global strategy to select the optimal band combination: select a representative band from each second subspace according to the order of the second subspace to form multiple band combinations; solve the average value E of the information entropy of each representative band in the band combination and the average value D of the distance between adjacent representative bands, calculate the result of E / D as the score of the band combination, and select the band combination with the highest score as the optimal band combination.

2. The hyperspectral band selection method of the optimal combination strategy according to claim 1, characterized in that: In step A4, the plurality of band combinations are 5m band combinations.

3. A hyperspectral band selection method with an optimal combination strategy, characterized in that: The following steps are involved: B1. According to the number of bands of the hyperspectral image, the bands are divided into m first subspaces in order, where m ≥ 1 and m is an integer; B2, moving the boundary points of adjacent first subspaces based on local density to form m second subspaces; The boundary points of the adjacent first subspace are moved based on the local density, specifically: B2.

1. Calculate the similarity matrix of all bands in the hyperspectral image and normalize it; B2.

2. Calculate the local density of all bands according to the similarity matrix; B2.3, select m-1 partition regions [F L ,F R ], where F L is the center band of the previous first subspace, F R For the center band of the next first subspace, select the band with the lowest local density in each divided area as the boundary point of the new adjacent subspace to complete a boundary point division; B2.

4. Set the maximum number of boundary point divisions. When the number of divisions reaches the maximum number of boundary point divisions or the boundary points of the subspace do not change, stop the subspace division to obtain m second subspaces. B3, calculating the information entropy of all bands in the hyperspectral image, and selecting two bands with large information entropy and high similarity in each second subspace as representative bands; B4. Local strategy to select the optimal band combination: B4.

1. Select a representative band with the largest information entropy from the two representative bands in the first second subspace, and form a sub-band combination with two representative bands in the adjacent second subspace respectively. Calculate the average value E' of the information entropy of the representative bands in the sub-band combination and the average value D' of the distance between adjacent representative bands in the sub-band combination. Calculate the result of E' / D' as the score of the sub-band combination, and select the sub-band combination with the highest score as the optimal sub-band combination. B4.

2. According to the scoring calculation method of step B4.1, the optimal sub-band combination is combined with the representative bands of the subsequent second subspaces in the order of the second subspace, and the combination ends at the last second subspace to obtain the optimal band combination.

4. A hyperspectral band selection system with an optimal combination strategy, characterized in that: It includes a subspace division module, a representative band selection module and an optimal band combination selection module, which are used to implement the hyperspectral band selection method of the optimal combination strategy described in claim 1 or 3; The output of the subspace division module is connected to the input of the representative band selection module, and the subspace division module is used to divide the hyperspectral band into multiple subspaces; The output of the representative band selection module is connected to the input of the optimal band combination selection module, and the representative band selection module is used to select two representative bands with large information entropy and high similarity from each subspace; The optimal band combination selection module is used to select the optimal band combination from multiple band combinations formed by representative band combinations of each subspace.

5. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the hyperspectral band selection method of the optimal combination strategy according to any one of claims 1 to 3 are implemented.

6. A computer device comprising a processor, a memory connected to the processor, and a computer program executable on the processor, characterized in that: When the processor executes the computer program, the steps of the hyperspectral band selection method of the optimal combination strategy according to any one of claims 1 to 3 are implemented.