A hyperspectral band selection method based on improved iterative greedy algorithm

Abstract

This invention relates to the field of hyperspectral remote sensing image classification, and specifically to a hyperspectral band selection method based on an improved iterative greedy algorithm. The method includes: obtaining clustering results of hyperspectral image data using a k-means clustering algorithm; generating a nearest neighbor graph between bands by calculating the Euclidean distance between bands within each cluster; extracting the bands in each cluster with the closest Euclidean distance to the cluster center, and using the extracted band combinations as initial solutions; using the information entropy and mutual information of the band combinations as the objective function of an improved iterative greedy algorithm based on simulated annealing; iteratively optimizing the initial solution through destructive reconstruction and neighborhood search operations to find the maximum value of the objective function; the band combination corresponding to the maximum value of the objective function is the optimal solution. This method effectively improves the efficiency and accuracy of hyperspectral image band selection and classification.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral remote sensing image classification, and in particular to a hyperspectral band selection method based on an improved iterative greedy algorithm. Background Technology

[0002] Hyperspectral remote sensing is a pioneering Earth observation technology that uses hyperspectral imaging spectrometers to simultaneously image ground features across dozens or hundreds of spectral bands, creating complex hyperspectral images. These images combine spatial and spectral perspectives, enabling detailed object identification and material differentiation within a three-dimensional image cube. Therefore, hyperspectral remote sensing technology is widely used in numerous fields, highlighting its significant importance.

[0003] However, the vast amount of spectral information also presents challenges to related image processing techniques. The high correlation among numerous continuous spectral bands leads to significant redundancy in hyperspectral images. Furthermore, the detailed band information results in high dimensionality in hyperspectral images, greatly increasing the computational burden. Therefore, reducing the dimensionality of hyperspectral image data while preserving useful spectral information is crucial.

[0004] Feature selection, also known as band selection, has been identified as an effective dimensionality reduction method. It selects a subset of bands from a complex set of high-dimensional bands as a key determinant because it does not destroy the original information and has little impact on the subsequent classification process, and is therefore widely used.

[0005] Existing clustering-based methods treat each band as a data sample, grouping similar bands into the same cluster, and finally selecting corresponding bands from each cluster to form a band subset using certain evaluation criteria. Since the band subsets are composed of representative bands selected from each cluster, this method can reduce the correlation between bands and avoid high redundancy. However, precisely because these bands come from different clusters, while highlighting the importance of individual bands, it often neglects the overall performance of the band combination. Search-based methods, on the other hand, treat the band selection problem as a discrete combinatorial optimization problem. Many heuristic optimization algorithms have been used as solutions for hyperspectral band selection, such as genetic algorithms, particle swarm optimization, and the firefly algorithm. These algorithms iteratively optimize through different search strategies. Although they cannot guarantee finding the optimal solution, they can obtain suboptimal solutions within certain iteration limits. However, they still suffer from problems such as high computational cost, difficulty in balancing global exploration and local development, and a tendency to get trapped in local optima. Summary of the Invention

[0006] To address the shortcomings of existing technologies in computational efficiency and accuracy for hyperspectral image band selection and classification, this invention proposes a hyperspectral band selection method based on an improved iterative greedy algorithm, aiming to effectively improve the efficiency and accuracy of hyperspectral image band selection and classification.

[0007] The present invention provides a hyperspectral band selection method based on an improved iterative greedy algorithm, characterized in that:

[0008] Step 1: Obtain the clustering results of the hyperspectral image data using the k-means clustering algorithm;

[0009] Step 2: In the clustering results, a nearest neighbor graph between bands is generated by calculating the Euclidean distance between bands within each cluster.

[0010] Step 3: Extract the band in each cluster that is closest to the cluster center by Euclidean distance, and combine the extracted bands as the initial solution;

[0011] Step 4: Use the information entropy and mutual information of the band combination as the objective function of the improved iterative greedy algorithm based on simulated annealing, and calculate the objective function value of the initial solution;

[0012] Step 5: Iteratively optimize the initial solution through destruction reconstruction and neighborhood search operations to find the maximum value of the objective function. The band combination corresponding to the maximum value of the objective function is the optimal solution.

[0013] Furthermore, hyperspectral image data ,in, Represents the number of pixels. The method for determining the nearest neighbor graph among bands in each cluster, representing the number of bands, is as follows.

[0014] For each band ,turn up ( ) bands , making

[0015]

[0016] in, It is a band and band The Euclidean distance between them.

[0017] Furthermore, the information entropy of the corresponding band in the objective function is determined by the following formula:

[0018]

[0019]

[0020] in, For band Information entropy This represents the overall information entropy of a band combination. Indicates the first The first band The probability distribution of pixel values;

[0021] Mutual information of corresponding bands in the objective function Represented as,

[0022]

[0023] in, Indicates band and The joint entropy;

[0024] The objective function of the improved iterative greedy algorithm based on simulated annealing The calculation formula is expressed as follows:

[0025]

[0026] in, The number of bands to be selected. These are weight parameters. Indicates the band number.

[0027] Furthermore, the improved destructive reconstruction strategy of the simulated annealing-based iterative greedy algorithm is as follows:

[0028] When destroying the reconstructed solution, Bands at each position are removed, and new bands are selected from the cluster to which the removed bands belong to update the solution using the Lévy flight strategy. The Lévy flight step size is... The calculation formula is as follows:

[0029]

[0030] Among them, the stability index Control the shape of the step size distribution. and It is a random variable drawn from a normal distribution with a mean of zero, specifically represented as,

[0031]

[0032]

[0033] in, Equals 1, The calculation method is as follows:

[0034]

[0035] After determining the Lévy flight step size, the process of updating the solution is represented as follows:

[0036]

[0037] in, This represents the band combination in the current solution. This refers to the band combination in the updated solution.

[0038] Furthermore, the improved neighborhood search strategy of the simulated annealing-based iterative greedy algorithm is as follows: when performing the neighborhood search, the... For each position, the bands are removed. The nearest neighbor bands of the current band are traversed from the nearest neighbor graph to which the removed band belongs. The objective function of the improved iterative greedy algorithm is used to determine the band with the largest objective function as the component band in the new solution.

[0039] Furthermore, the improved iterative optimization method of the simulated annealing-based iterative greedy algorithm is as follows: calculate the objective function value of the band combination corresponding to the initial solution and update the optimal solution; after destructing and reconstructing the initial solution, calculate the objective function value of the corresponding band combination and update the optimal solution; after performing a neighborhood search, calculate the objective function value of the corresponding band combination, update the optimal solution, and use the simulated annealing strategy to probabilistically accept the difference solution.

[0040] The probability of a simulated annealing strategy accepting a different solution. The calculation formula is as follows:

[0041]

[0042] in, For the current solution The objective function value, A new solution is generated within the neighborhood of the current solution. The objective function value, For the temperature parameter at this time, if the objective function value of the new solution is higher than that of the current solution, the new solution is accepted; otherwise, the difference solution is accepted according to the probability calculated by formula (8).

[0043] Repeat the above process until the maximum number of iterations is reached. The band combination corresponding to the optimal solution at this point is the hyperspectral image band to be selected.

[0044] This invention provides a hyperspectral band selection method based on an improved iterative greedy algorithm. The method first uses k-means clustering to cluster hyperspectral image data, simplifying spectral information while preserving spatial proximity relationships. Within each cluster, Euclidean distances between bands are calculated to generate a nearest neighbor graph, facilitating subsequent neighborhood searches and reducing the size of the algorithm's solution space. Next, the band with the closest Euclidean distance to the cluster center is selected as the initial solution in each cluster, ensuring its quality. Finally, information entropy and mutual information are combined as the objective function of the improved iterative greedy algorithm. The initial solution is iteratively optimized through destructive reconstruction and neighborhood search operations to find the band combination corresponding to the maximum value of the objective function. Ultimately, the optimized band combination is the optimal solution. In summary, this invention effectively improves the efficiency and accuracy of hyperspectral image band selection and classification. Attached Figure Description

[0045] Figure 1 This is a schematic diagram illustrating the working principle of the present invention;

[0046] Figure 2 This is a real-world map of Indian Pine data;

[0047] Figure 3 This is a true ground feature map of Salinas data;

[0048] Figure 4 This is a real-world map of Botswana data;

[0049] Figure 5 This is a diagram showing the data classification results of the Indian Pine of this invention;

[0050] Figure 6 This is a diagram showing the Salinas data classification results of the present invention;

[0051] Figure 7 This is a diagram showing the Botswana data classification results of the present invention. Detailed Implementation

[0052] like Figure 1 As shown, the hyperspectral band selection method based on an improved iterative greedy algorithm provided by this invention is mainly implemented through the following process.

[0053] Step 1: Obtain the clustering results of the hyperspectral image data using the k-means clustering algorithm.

[0054] Step 2: In each cluster contained in the clustering results, a nearest neighbor graph between bands is generated by calculating the Euclidean distance between each band. Specifically, the hyperspectral image data uses... It means that, among them, Represents the number of pixels. This represents the number of bands. The specific method for determining the nearest neighbor graph among bands in each cluster is to, for each band... ,turn up ( ) bands , making

[0055]

[0056] in, It is a band and band The Euclidean distance between them.

[0057] Step 3: Extract the bands with the closest Euclidean distance to the cluster center, and use the extracted band combination as the initial solution. Each band comes from a cluster, and only one band is selected from each cluster.

[0058] Step 4: Using the information entropy and mutual information of the band combination as the objective function of the improved iterative greedy algorithm based on simulated annealing, calculate the objective function value of the initial solution.

[0059] Specifically, the information entropy of the corresponding band in the objective function is determined by the following formula.

[0060]

[0061]

[0062] in, For the information entropy of a single band, This represents the overall information entropy of a band combination. Indicates the first The first band The probability distribution of pixel values.

[0063] Mutual information of corresponding bands in the objective function It can be represented as,

[0064]

[0065] in, Representing variables and The joint entropy.

[0066] The objective function of the improved iterative greedy algorithm based on simulated annealing The calculation formula can be expressed as follows:

[0067]

[0068] in, The number of bands to be selected. These are weight parameters. Indicates the band number.

[0069] Step 5: Iteratively optimize the initial solution through destruction reconstruction and neighborhood search operations to find the maximum value of the objective function. The band combination corresponding to the maximum value of the objective function is the optimal solution.

[0070] Specifically, the destructive reconstruction strategy for the improved simulated annealing-based iterative greedy algorithm is determined as follows. ,Will Bands at each position are removed, and new bands are selected from the cluster to which the removed bands belong to update the solution using the Lévy flight strategy. The Lévy flight step size is... The calculation method is as follows

[0071] Among them, the stability index Control the shape of the step size distribution. and It is a random variable drawn from a normal distribution with a mean of zero, specifically represented as:

[0072]

[0073]

[0074] in, Equals 1, The calculation method is as follows:

[0075]

[0076] Determine Levi's flight stride Then, the process of updating the solution can be represented as follows:

[0077]

[0078] in, This represents the band combination in the current solution. This refers to the band combination in the updated solution.

[0079] The improved neighborhood search strategy for the iterative greedy algorithm based on simulated annealing is determined as follows. When performing the neighborhood search, the strategy is... For each position, the bands are removed. The nearest neighbor bands of the current band are traversed from the nearest neighbor graph to which the removed band belongs. The objective function of the improved iterative greedy algorithm is used to determine the band with the largest objective function as the component band in the new solution.

[0080] The improved iterative optimization method of the simulated annealing-based iterative greedy algorithm is as follows: Calculate the objective function value of the band combination corresponding to the initial solution and update the optimal solution; after destroying and reconstructing the initial solution, calculate the objective function value of the corresponding band combination and update the optimal solution; after performing a neighborhood search, calculate the objective function value of the corresponding band combination, update the optimal solution, and use the simulated annealing strategy to probabilistically accept the difference solution;

[0081] The probability of a simulated annealing strategy accepting a different solution. The calculation formula is as follows:

[0082]

[0083] in, For the current solution The objective function value, A new solution is generated within the neighborhood of the current solution. The objective function value, For the temperature parameter at this time, if the objective function value of the new solution is higher than that of the current solution, the new solution is accepted; otherwise, the difference solution is accepted according to the probability calculated by formula (8).

[0084] Repeat the above process until the maximum number of iterations is reached. The band combination corresponding to the optimal solution at this point is the hyperspectral image band to be selected.

[0085] The following specific embodiment of the present invention uses a real dataset to illustrate the application of the hyperspectral band selection method based on an improved iterative greedy algorithm provided by the present invention, and further describes the present invention.

[0086] The real datasets are derived from three publicly available hyperspectral datasets: Indian Pine data, Botswana data, and Salinas data. The Indian Pine data contains hyperspectral imagery collected in the Indian Pine Forest region of Indiana, USA. After removing bands in water absorption areas, 200 bands remain, each with 145 bands. 145 pixels, wavelength range of 0.4 2.5 The available real-world features are divided into 16 categories. Their real-world feature maps are shown below. Figure 2 As shown.

[0087] Botswana data comes from satellite remote sensing images of Africa. After removing noise and water-absorbing bands, 145 bands remain, each with 1476 bands. 256 pixels, wavelength range 400 2500 nanometers. Contains 14 identified real-world feature categories. Its real-world feature map is shown below. Figure 3 As shown.

[0088] Salinas data comes from the California Valley, USA. After discarding 20 water absorption bands, it has a total of 204 bands, with 512 bands per band. 217 pixels, containing 16 real-world feature categories. Its real-world feature map is shown below. Figure 4 As shown.

[0089] like Figures 5-7 As shown, the improved iterative greedy algorithm based on simulated annealing is configured with 30 selected bands, a maximum number of iterations of 1000, and a simulated annealing temperature of... 1000 data points were classified using SVM with a Gaussian radial basis function kernel. 20% of the data points in each class were randomly selected as the training set, and the remainder were used as the test set. The classification results were analyzed using overall accuracy (OA), average accuracy (AA), and the Kappa coefficient for consistency testing. The results are shown in Tables 1, 2, and 3.

[0090] Table 1. Classification accuracy of Indian Pine data

[0091]

[0092] Table 2 Classification accuracy of Botswana data

[0093]

[0094] Table 3 Classification accuracy of Salinas data

[0095]

[0096] As can be seen from the specific values ​​of the classification accuracy and kappa coefficient of each land cover in the three sets of hyperspectral image data in Tables 1, 2 and 3, the band subset selected in this invention has achieved good classification accuracy and kappa coefficient for multiple datasets, indicating that this method can select a band subset suitable for hyperspectral image classification.

[0097] This invention combines clustering with a global optimization algorithm. Through clustering, it simplifies spectral information while preserving the spatial proximity relationships of hyperspectral data, thus reducing the size of the algorithm's solution space. Furthermore, it generates a nearest neighbor graph between bands within each cluster, improving the efficiency and accuracy of neighborhood search. Using an objective function combining information entropy and mutual information, it selects a subset of bands more suitable for hyperspectral image classification. In summary, the application of this invention can effectively improve the efficiency and accuracy of hyperspectral image band selection and classification.

Claims

1. A hyperspectral band selection method based on an improved iterative greedy algorithm, characterized in that, Includes the following steps, Step 1: Obtain the clustering results of the hyperspectral image data using the k-means clustering algorithm; Step 2: In the clustering results, a nearest neighbor graph between bands is generated by calculating the Euclidean distance between bands within each cluster. Step 3: Extract the band in each cluster that is closest to the cluster center by Euclidean distance, and combine the extracted bands as the initial solution; Step 4: Use the information entropy and mutual information of the band combination as the objective function of the improved iterative greedy algorithm based on simulated annealing, and calculate the objective function value of the initial solution; Step 5: Iteratively optimize the initial solution through destructive reconstruction and neighborhood search operations. In the destructive reconstruction... Bands at certain positions are removed, and new bands are selected from the clusters to which the removed bands belong to update the solution; during the neighborhood search, For each position, bands are removed. The nearest neighbor bands of the current band are traversed from the nearest neighbor graph to which the removed band belongs. The objective function of the improved iterative greedy algorithm is used to determine the band with the largest objective function as the component bands in the new solution. The maximum value of the objective function is searched, and the combination of bands corresponding to the maximum value of the objective function is the optimal solution.

2. The hyperspectral band selection method based on an improved iterative greedy algorithm according to claim 1, further characterized in that, Hyperspectral image data ,in, Represents the number of pixels. The method for determining the nearest neighbor graph among bands in each cluster, representing the number of bands, is as follows. For each band ,turn up ( ) bands , making ； in, It is a band and band The Euclidean distance between them.

3. The hyperspectral band selection method based on an improved iterative greedy algorithm according to claim 2, further characterized in that, The information entropy of the corresponding band in the objective function is determined by the following formula. ； ； in, For band Information entropy This represents the overall information entropy of a band combination. Indicates the first The first band The probability distribution of pixel values; the mutual information of corresponding bands in the objective function. Represented as, ； in, Indicates band and The joint entropy; The objective function of the improved iterative greedy algorithm based on simulated annealing The calculation formula is expressed as follows: ； in, The number of bands to be selected. These are weight parameters. Indicates the band number.

4. The hyperspectral band selection method based on an improved iterative greedy algorithm according to claim 3, further characterized in that, The improved destructive reconstruction strategy of the simulated annealing-based iterative greedy algorithm is as follows: When destroying the reconstructed solution, Bands at each position are removed, and new bands are selected from the cluster to which the removed bands belong to update the solution using the Lévy flight strategy. The Lévy flight step size is... The calculation formula is as follows: ； Among them, the stability index Control the shape of the step size distribution. and It is a random variable drawn from a normal distribution with a mean of zero, specifically represented as, ； ； in, Equals 1, The calculation method is as follows: ； After determining the Lévy flight step size, the process of updating the solution is represented as follows: ； in, This represents the band combination in the current solution. This refers to the band combination in the updated solution.

5. The hyperspectral band selection method based on an improved iterative greedy algorithm according to claim 4, further characterized in that, The improved iterative optimization method for the simulated annealing-based iterative greedy algorithm is as follows: Calculate the objective function value of the band combination corresponding to the initial solution and update the optimal solution; after destructive reconstruction of the initial solution, calculate the objective function value of the corresponding band combination and update the optimal solution; after performing a neighborhood search, calculate the objective function value of the corresponding band combination, update the optimal solution, and use simulated annealing strategy to probabilistically accept the difference solution; The probability of a simulated annealing strategy accepting a different solution. The calculation formula is as follows: ； in, For the current solution The objective function value, A new solution is generated within the neighborhood of the current solution. The objective function value, For the temperature parameter at this time, if the objective function value of the new solution is higher than that of the current solution, the new solution is accepted; otherwise, the difference solution is accepted according to the probability calculated by formula (8). Repeat the above process until the maximum number of iterations is reached. The band combination corresponding to the optimal solution at this point is the hyperspectral image band to be selected.

Images