Hyperspectral image waveband selection method based on improved dung beetle algorithm
By improving the population initialization and convergence factor of the dung beetle algorithm, combining adaptive spiral search and triangle walking strategy, optimizing the band selection of hyperspectral images, the existing algorithms' slow convergence speed and easy to fall into local optimal problems, and efficient band subset selection and classification accuracy are achieved.
Patent Information
- Application Number
- CN202510151309.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-11
- Publication Date
- 2025-07-18
AI Technical Summary
The existing hyperspectral image band selection algorithm has problems such as slow convergence speed, easy to fall into local optimality, and poor global search performance, resulting in a decrease in classification accuracy.
The improved dung beetle algorithm is used to initialize the population through Tent chaotic mapping, and the convergence factor of the improved dung beetle algorithm is adaptive nonlinear decreasing, and the band subset is optimized by combining the hybrid spiral search and the triangle walk strategy of hybrid walk strategy, and SVM is used for geographic classification.
It accelerates the convergence speed of the dung beetle algorithm, improves the global search performance of band selection, and improves the classification accuracy and consistency of hyperspectral images.
Smart Images
Figure CN120339822A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hyperspectral image band selection, and particularly to a hyperspectral image band selection method based on an improved dung beetle algorithm. Background Art
[0002] Hyperspectral images solve problems that traditional images cannot by integrating spatial and spectral information. However, due to limitations in communication bandwidth and processing speed, there is a phenomenon where not all data can be transmitted and processed in real time. In addition, there is extremely high correlation and redundancy between bands, increasing the difficulty of subsequent processing.
[0003] Due to limited training samples, as the data dimension increases, the classification accuracy first increases and then decreases. Dimensionality reduction is a common method to reduce the computational complexity of hyperspectral images and improve classification performance, and it is also the best method to solve the "curse of dimensionality" problem of hyperspectral images at present. Band selection is an important technique for hyperspectral image dimensionality reduction.
[0004] In recent years, many scholars have used swarm intelligence optimization algorithms for band selection, such as genetic algorithms, grey wolf algorithms, gravitational search algorithms, etc. Among them, genetic algorithms have many parameters, resulting in complex algorithm execution and being prone to falling into local optima, with poor global search performance. The convergence factor of the grey wolf algorithm decreases linearly, resulting in slow algorithm convergence speed and being prone to falling into local optima. The gravitational search algorithm has a slow convergence speed and unsatisfactory global search effect. Summary of the Invention
[0005] According to the problems existing in the prior art, the present invention discloses a hyperspectral image band selection method based on an improved dung beetle algorithm, which specifically includes the following steps:
[0006] Define the hyperspectral image band sequence as a dung beetle population, assume that each dung beetle individual represents a subset of wavebands, and use Tent chaotic mapping to initialize the dung beetle population, so as to select an initial hyperspectral waveband subset;
[0007] Use the improved dung beetle optimization algorithm to optimize the hyperspectral waveband subset. In the reproduction stage of the dung beetle population, improve the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor to accelerate the convergence speed, and obtain a waveband subset;
[0008] In the reproduction stage and foraging stage of the dung beetle population, adopt a hybrid walking strategy that combines adaptive spiral search and triangular walking strategy to find a better waveband subset in the waveband subset;
[0009] Use SVM to perform ground object classification on the spectral reflectance of the hyperspectral image in the waveband subset, and use the overall accuracy OA value as the classification evaluation index, so as to select the individual corresponding to the optimal value as the optimal hyperspectral image band combination.
[0010] Furthermore, let the hyperspectral image data be HIM = {x1, x2,..., x n} ∈ R l×n , where l is the number of bands, n is the total number of pixels in each band image, k represents the sample category. Initialize the positions of the dung beetles using the Tent mapping, project the resulting values into the chaotic variable space using the Tent mapping relationship initially, and then map the obtained chaotic values back to the initial space of the algorithm through a linear transformation. The Tent mapping expression is:
[0011]
[0012] where t represents the iteration number of the dung beetle population, and x(t) represents the position information of the dung beetle at the t-th iteration.
[0013] Furthermore, the process of improving the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor is as follows:
[0014] The calculation process of the linearly decreasing convergence factor R of the dung beetle algorithm is as follows:
[0015]
[0016] where t max represents the maximum number of iterations;
[0017] Improve the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor:
[0018]
[0019] Furthermore, the hybrid walking strategy is obtained in the following way:
[0020] In the reproduction stage, use the boundary selection strategy to simulate the egg-laying area of female dung beetles, defined as follows:
[0021] Lb * = max(X * × (1 - R), Lb)
[0022] Ub * = min(X * × (1 + R), Ub)
[0023] where X * represents the current local optimal value, and Lb * and Ub * represent the upper and lower limits of the region;
[0024] Introduce adaptive spiral search in the reproduction stage of the dung beetle population:
[0025] B i (t + 1) = X* +e dl ×cos(2πl)×b1×(B i (t)-Lb * )+
[0026] e dl ×cos(2πl)×b2×(B i (t)-Ub * )
[0027] where q is the coefficient of variation, q = 5, where b1 and b2 represent two independent random vectors of size 1×D, and D represents the dimension of the optimization problem;
[0028] Introduce the triangular walking strategy in the foraging stage of the dung beetle population to form a hybrid walking strategy:
[0029] x i (t + 1) = |X * -r×cosβ×L2|
[0030] L1 = X * -x i (t), L2 = rand()×L1
[0031] β = 2π×rand()
[0032] where β represents the walking direction, and r represents the ratio of the difference between the upper and lower boundaries to the total number of wavebands.
[0033] Furthermore, use the hybrid walking strategy to find a better waveband subset in the waveband subset:
[0034] Take the OA value as the objective function to improve the dung beetle algorithm;
[0035] Use the Tent chaotic map to replace the random initialization to generate the initial dung beetle population, and divide the population into 4 small populations of rolling dung beetles, breeding dung beetles, foraging dung beetles, and stealing dung beetles according to the ratio of 4:4:5:7. Each small population follows different position update rules;
[0036] Calculate the objective function of each population after the update, and repeat the above steps until the maximum number of iterations is reached and exit the loop to obtain the optimal solution. The population corresponding to the optimal solution is the selected optimal waveband subset sequence.
[0037] Due to the adoption of the above technical solution, the present invention provides a hyperspectral image band selection method based on an improved dung beetle algorithm. This method can effectively select a subset of bands suitable for classification. Considering that the basic dung beetle algorithm has a slow convergence speed and is prone to falling into local extrema, the population initialization method is changed, and at the same time, the convergence factor and the position update method of the population are optimized, which not only speeds up the algorithm convergence speed but also improves the search performance of the dung beetle algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. The drawings in the following description are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0039] Figure 1 is the overall flowchart of the band selection method described in the present invention;
[0040] Figure 2 is the false color map of Indian Pine data, the true ground object information map, and the classification result map of the present invention;
[0041] Figure 3 is the false color map of Pavia data, the true ground object information map, and the classification result map of the present invention;
[0042] Figure 4 is the false color map of Salinas data, the true ground object information map, and the classification result map of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0043] To make the technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention:
[0044] As Figure 1 shown, a hyperspectral image band selection method based on an improved dung beetle algorithm specifically includes the following steps:
[0045] S1: Define the hyperspectral image band sequence as the dung beetle population. Assume that each dung beetle individual represents a subset of bands. Use Tent chaotic mapping to initialize the dung beetle population instead of randomly initializing the population, so as to select the initial hyperspectral band subset, make the population distribution more uniform, and increase the diversity of the population;
[0046] S2: Optimize the hyperspectral waveband subset using the improved dung beetle optimization algorithm. During the reproduction stage of the dung beetle population, improve the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor. Corresponding to the larger search space in the early stage, the slower convergence speed avoids premature convergence, and the accelerated convergence speed in the later stage is conducive to accelerating exploration in a smaller solution space, flexibly balancing the global and local search capabilities of the algorithm;
[0047] S3: The process of finding a better hyperspectral waveband subset is the position update process of the dung beetle population. During the position update process in the reproduction stage and foraging stage of the dung beetle population, use a hybrid walking strategy that combines the adaptive spiral search and triangular walking strategies to find a better waveband subset;
[0048] S4: Use SVM to classify the spectral reflectance of hyperspectral images in the waveband subset, and use the overall accuracy OA value as the classification evaluation index, so as to select the individual corresponding to the optimal value as the selected optimal hyperspectral image band combination;
[0049] Furthermore: The solution process of using Tent chaotic mapping for the initialization of the dung beetle population is as follows:
[0050] Suppose the hyperspectral image data is represented by HIM = {x1, x2,..., x n} ∈ R l×n where l is the number of wavebands, n is the total number of pixels in each band image, and k represents the sample category. Use the Tent mapping to initialize the position of the dung beetles, project the result value onto the chaotic variable space using the Tent mapping relationship, and then map the obtained chaotic values back to the initial space of the algorithm through a linear transformation. The Tent mapping expression is:
[0051]
[0052] where t represents the iteration number of the dung beetle population, and x(t) represents the position information of the dung beetles at the t-th iteration.
[0053] Furthermore: The process of improving the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor is as follows:
[0054] S2-1: The calculation formula for the linearly decreasing convergence factor R of the dung beetle algorithm is as follows:
[0055]
[0056] where t max represents the maximum number of iterations.
[0057] S2-2: Improve the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor:
[0058]
[0059] Furthermore, the process of determining the hybrid walking strategy for improving the dung beetle algorithm is as follows:
[0060] S3-1: During the reproduction stage, use the boundary selection strategy to simulate the egg-laying area of female dung beetles, defined as follows:
[0061] Lb * = max(X * × (1 - R), Lb)
[0062] Ub * = min(X * × (1 + R), Ub)
[0063] where X * represents the current local optimal value, and Lb * and Ub * represent the upper and lower limits of the region.
[0064] S3-2: Introduce adaptive spiral search during the reproduction stage of the dung beetle population:
[0065] B i (t + 1)= X * + e dl × cos(2πl)× b1× (B i (t)- Lb * )+
[0066] e dl × cos(2πl)× b2× (B i (t)- Ub * )
[0067] where q is the mutation coefficient, q = 5; where b1 and b2 represent two independent random vectors of size 1×D, and D represents the dimension of the optimization problem.
[0068] S3-3: Introduce the triangular walking strategy during the foraging stage of the dung beetle population:
[0069] x i (t + 1)= |X * - r× cosβ× L2|
[0070] L1 = X * - x i (t), L2 = rand()× L1
[0071] β = 2π× rand()
[0072] where β represents the walking direction, and r represents the ratio of the difference between the upper and lower boundaries to the total number of wavebands.
[0073] Furthermore: Search for the optimal waveband subset according to the improved dung beetle algorithm:
[0074] S4-1: Use the OA value as the objective function of the improved dung beetle algorithm;
[0075] S4-2: Use Tent chaotic mapping to replace random initialization to generate the initial dung beetle population, and divide the population into four small populations of rolling dung beetles, breeding dung beetles, foraging dung beetles, and stealing dung beetles according to the ratio of 4:4:5:7. Each small population follows different position update rules.
[0076] S4-3: Calculate the objective function of each population after update, repeat the above steps until the maximum iteration number is reached and exit the loop to obtain the optimal solution, and the population corresponding to the optimal solution is the selected optimal waveband subset sequence.
[0077] The following is an example to illustrate a hyperspectral image band selection method based on the improved dung beetle algorithm proposed by the present invention. The sample data comes from three publicly available hyperspectral data sets: Indian Pine data, Salinas data, and Pavia data. The Indian Pine data is a farmland image obtained by the AVIRIS sensor in the northwest of Indiana. Its spectral range is 0.4-2.5 μm, the spatial resolution is 20 m, the image size is 145×145, and there are 200 bands remaining after removing noise and water absorption bands, including 16 types of ground objects. Its false color map and real ground object information are as Figure 2 (a)(b) shown.
[0078] The Pavia data is an urban area image obtained by the ROSIS-03 sensor over the University of Pavia. The data image size is 610×340, the spatial resolution is 1.3 m, it contains 103 bands, and a total of 9 types of ground objects. Its false color map and real ground object information are as Figure 3 (a)(b) shown.
[0079] The Salinas data is obtained by the AVIRIS sensor in a certain area. The size of this image is 512×217, the spatial resolution is 3.7 m, and there are 204 bands remaining after removing noise and water absorption bands, including 16 types of ground objects. Its false color map and real ground object information are as Figure 4 (a)(b) shown.
[0080] Such as Figures 2 - 4As shown in (c), the initial population of the improved dung beetle algorithm is set to 20, and the maximum number of iterations is set to 50. SVM is used for land cover classification, where the Gaussian radial basis kernel function is selected for SVM. 10% of the data points are randomly selected as the training set from each class, and the rest are used as the test set. The overall accuracy (OA), average classification accuracy (AA), and Kappa coefficient of the consistency test are used to analyze the classification results.
[0081] The definition of OA is as follows:
[0082]
[0083] where S i is the number of samples of the i-th land cover type in the classification result, k is the number of land cover types, and N i is the number of samples of the i-th land cover type in the true land cover.
[0084] The classification accuracy of each land cover target class is calculated as follows:
[0085]
[0086] The definition of AA is as follows:
[0087]
[0088] The calculation formula of the Kappa coefficient is:
[0089]
[0090] where:
[0091]
[0092] N 1,i represents the number of pixels that are misclassified as other land cover types for the i-th land cover type in the classification result; N 2,i represents the number of pixels that are misclassified as the i-th land cover type for other land cover types.
[0093] Tables 1, 2, and 3 show the specific values of the classification accuracy and Kappa coefficient for each class of three groups of hyperspectral images. It can be seen that the selected wavelength subsets of the present invention have achieved good classification accuracy for different datasets, and at the same time, the Kappa coefficient is relatively high, indicating that the method can select more suitable wavelength subsets for classification.
[0094] Table 1
[0095]
[0096] Table 2
[0097]
[0098] Table 3
[0099]
[0100] The present invention uses an improved swarm intelligence optimization algorithm for band selection to select a subset of wavebands suitable for classification. The chaotic mapping technique is applied to population initialization to increase the diversity of the population. The convergence factor of the dung beetle algorithm is improved from linear decrease to adaptive non-linear decrease, corresponding to a larger search space and slower convergence speed in the early stage to avoid premature convergence, and an accelerated convergence speed in the later stage is beneficial to accelerate exploration in a smaller solution space, flexibly balancing the global and local search capabilities of the algorithm. A hybrid walking strategy is designed to optimize the position update process of the dung beetle population, enabling the algorithm to jump out of the local optimum while ensuring accuracy. The subset of wavebands selected by the present invention has good classification accuracy for classification purposes.
[0101] As described above, only the specific preferred embodiments of the present invention are given, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered by the protection scope of the present invention.
Claims
1. A hyperspectral image band selection method based on an improved dung beetle algorithm, characterized in that Including: Define the hyperspectral image band sequence as the dung beetle population. Assume that each dung beetle individual represents a subset of wavebands, and use the Tent chaotic map to initialize the dung beetle population, so as to select the initial hyperspectral waveband subset. Use the improved dung beetle optimization algorithm to optimize the hyperspectral waveband subset. In the reproduction stage of the dung beetle population, improve the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor to accelerate the convergence speed and obtain the waveband subset. In the reproduction stage and foraging stage of the dung beetle population, adopt a hybrid walking strategy that combines the adaptive spiral search and the triangular walking strategy to find a better waveband subset in the waveband subset. Use SVM to classify the ground objects based on the spectral reflectance of the hyperspectral image in the waveband subset, and use the overall accuracy OA value as the classification evaluation index, so as to select the individual corresponding to the optimal value as the optimal hyperspectral image band combination.
2. The hyperspectral image band selection method based on the improved dung beetle algorithm according to claim 1, characterized in that: Let the hyperspectral image data be HIM = {x1, x2,..., x n} ∈ R l×n , where l is the number of bands, n is the total number of pixels in each band image, k represents the sample category. Initialize the positions of the dung beetles using the Tent map, project the resulting values onto the chaotic variable space using the Tent map relationship, and then map the obtained chaotic values back to the initial space of the algorithm through a linear transformation. The Tent map expression is: Where t represents the iteration number of the dung beetle population, and x(t) represents the position information of the dung beetle at the t-th iteration.
3. A hyperspectral image band selection method based on an improved dung beetle algorithm according to claim 1, characterized in that: The process of improving the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor is as follows: The calculation process of the linearly decreasing convergence factor R of the dung beetle algorithm is as follows: where t max represents the maximum number of iterations; Improve the linearly decreasing convergence factor to an adaptive non-linearly decreasing convergence factor:
4. A hyperspectral image band selection method based on an improved dung beetle algorithm according to claim 1, characterized in that: The above-mentioned hybrid walking strategy is obtained in the following way: In the reproduction stage, use the boundary selection strategy to simulate the spawning area of female dung beetles, which is defined as follows: Lb * = max(X * ×(1 - R), Lb) Ub * = min(X * ×(1 + R), Ub) where X * represents the current local optimum value, Lb * and Ub * represent the upper and lower limits of the region; Introduce the adaptive spiral search in the reproduction stage of the dung beetle population: B i (t + 1)= X * + e dl × cos(2πl)× b1× (B i (t)- Lb * ) + e dl × cos(2πl) × b2 × (B i (t) - Ub * ) wherein q is the coefficient of variation, q = 5, where b1 and b2 represent two independent random vectors of size 1×D, and D represents the dimension of the optimization problem; Introduce the triangular walking strategy in the foraging stage of the dung beetle population to form a hybrid walking strategy: x i (t + 1) = |X * -r × cosβ × L2 L1 = X * -x i (t), L2 = rand() × L1 β = 2π × rand() Where β represents the walking direction, and r represents the ratio of the difference between the upper and lower boundaries to the total number of wavebands.
5. A hyperspectral image band selection method based on an improved dung beetle algorithm according to claim 1, characterized in that: When using the hybrid walking strategy to find a better waveband subset in the waveband subset: Use the OA value as the objective function of the improved dung beetle algorithm; Use the Tent chaotic map to replace the random initialization to generate the initial dung beetle population, and divide the population into four small populations of tumble dung beetles, reproductive dung beetles, foraging dung beetles, and kleptoparasitic dung beetles according to the ratio of 4:4:5:
7. Each small population follows different position update rules; Calculate the objective function of each population after the update, and repeat the above steps until the maximum iteration number is reached and exit the loop to obtain the optimal solution. The population corresponding to the optimal solution is the selected optimal waveband subset sequence.