Hyperspectral image band selection method based on superpixel segmentation and clustering

Abstract

A hyperspectral image band selection method based on superpixel segmentation and clustering, comprising: superpixel segmentation; spatial feature extraction; key band identification and band cluster selection; verification result; the application selects key bands by optimizing the latent feature matrix and using a dynamic programming algorithm, improves the classification accuracy of the hyperspectral image, enhances the stability of the band selection by fusing the optimized latent feature matrix and maintaining the order of the bands in the frequency domain, reduces the dimension of the hyperspectral image by selecting a representative band subset, while avoiding the destruction of the physical structure of the original image, achieving the purpose of effective dimension reduction; by band selection rather than feature extraction, the information redundancy is reduced, and the data utilization efficiency is improved; the spatial structure information between superpixels is considered in the band selection process, and the global spatial structure of the hyperspectral image is more comprehensively retained.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for selecting hyperspectral image bands based on superpixel segmentation and clustering. Background Technology

[0002] Hyperspectral imaging technology can simultaneously acquire geometric, radiometric, and spectral information of ground features, integrating the capabilities of a camera, radiometer, and spectrometer. Compared to two-dimensional optical spatial imaging, it can perform three-dimensional spatial and spectral imaging of ground features. At a certain spatial resolution, it can acquire unique continuous spectral characteristics of ground features across a wide spectral range, offering significant advantages for fine classification and identification of ground features. It has become an important cutting-edge technology for Earth remote sensing, playing an increasingly important role in natural resource surveys, ecological environment monitoring, agriculture, forestry, animal husbandry, fisheries, and marine and coastal zone monitoring. However, hyperspectral images contain hundreds or even thousands of bands, which hides a large amount of information within them, introducing redundancy, increasing computational complexity, and impacting the accuracy of ground feature classification. Therefore, dimensionality reduction processing of hyperspectral images is necessary.

[0003] Existing hyperspectral image dimensionality reduction methods mainly fall into two categories: feature extraction and band selection. Feature extraction maps high-dimensional data to a low-dimensional space. Commonly used methods include principal component analysis (PCA), independent component analysis (ICA), and linear discriminant analysis (LDA). However, the mapping operation during feature extraction disrupts the physical structure of the original hyperspectral image, leading to the loss of some crucial information. Band selection, on the other hand, chooses a representative subset from all bands without altering the physical structure of the hyperspectral image. Based on the availability of labels, band selection can be further divided into unsupervised and supervised methods. Since labeling hyperspectral images requires manual work by experts in various fields, which is time-consuming, labor-intensive, and difficult to obtain, unsupervised band selection methods are of greater interest to researchers.

[0004] Existing unsupervised band selection methods can be roughly divided into two categories: ranking-based and clustering-based. The basic idea of ​​ranking-based band selection methods is to sort all hyperspectral bands according to a certain ranking rule, and then select the highest-ranking or lowest-ranking bands to form a new band subset. While these methods have achieved satisfactory results, they also have two drawbacks. First, these algorithms are very sensitive to noisy hyperspectral bands, which can lead to highly similar selected bands. Second, most algorithms alter the original information in the hyperspectral image cube through spatial transformation, resulting in the loss of some crucial information. The basic idea of ​​clustering-based band selection methods is to divide the original bands into multiple clusters using a clustering algorithm, and then select a representative band from each cluster to form the final band subset. Unfortunately, this algorithm results in a subset with low similarity, but the selected bands may not necessarily contain sufficient information. A common drawback of both methods is that most existing algorithms do not consider the spatial information of the hyperspectral image; they stretch each band into a single feature vector and then select hyperspectral bands according to a corresponding objective function. In practice, for a given band, different regions often correspond to different objects or land cover, resulting in different spectral characteristics. The high-dimensional pixel features of each band contain a large amount of redundant information, limiting the clustering performance of hyperspectral band selection. Therefore, it is not suitable to directly treat each band as a single feature vector. Furthermore, most existing methods, after extracting spatial features, tend to use the k-means method to cluster bands that have already undergone spatial dimensionality reduction. This leads to instability in the selected band subsets each time, introducing uncertainty into subsequent tasks. In addition, hyperspectral bands are often assumed to be unordered during clustering algorithms, leading to the neglect of the overall contextual information of the spectral band. However, in reality, bands that are closer together display more relevant image information, and experiments have demonstrated the orderliness of bands. Simultaneously, many methods using superpixel segmentation to extract spatial features from hyperspectral images do not consider the structural relationships between superpixels, but instead directly fuse the low-dimensional latent features of each small region, failing to comprehensively preserve the global spatial structure of the original hyperspectral image in the low-dimensional space. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a hyperspectral image band selection method based on superpixel segmentation and clustering with high classification accuracy, strong stability and reduced information redundancy.

[0006] The technical solution adopted to solve the above-mentioned technical problems is: a hyperspectral image band selection method based on superpixel segmentation and clustering, including the following steps:

[0007] Step 1. Superpixel segmentation

[0008] On a hyperspectral dataset of size W×H×B, where W is the width of the image, H is the height of the image, and B is the number of channels of the image, the first principal component of the image is extracted, and the first principal component of the image is divided into N homogeneous small regions. Each small region is used as an original superpixel, and the structural relationship between each original superpixel is constructed using a k-nearest neighbor graph.

[0009] Step 2. Spatial Feature Extraction

[0010] Step 2.1. Construct a similarity matrix within each original superpixel using a k-nearest neighbor graph, where the similarity between bands is measured by Euclidean distance;

[0011] Step 2.2. Obtain the degree matrix from the similarity matrix, and then obtain the normalized Laplace matrix;

[0012] Step 2.3. The latent feature matrix Y is formed by the eigenvectors corresponding to the first d smallest eigenvalues ​​of the Laplacian matrix. There are N latent feature matrices Y, and the size of the latent feature matrix Y is B×d.

[0013] Step 2.4. Optimize the latent feature matrix Y using an optimization function to extract the spatial structure between superpixels;

[0014] Step 2.5. Fuse the N optimized latent feature matrices Y to obtain a shared latent feature matrix F of size B×d;

[0015] Step 3. Key Band Identification and Band Clustering Selection

[0016] A series of key bands are found using a dynamic programming algorithm on a shared latent feature matrix F. The remaining bands are divided into different clusters based on the key bands. Within each cluster, the band with the highest ranking value is selected, and only one band can be selected within each cluster to form a band subset.

[0017] Step 4. Verify the results

[0018] The obtained band subset is validated using a classifier, and the classification results are observed.

[0019] As a preferred technical solution, in step 1, the first principal component of the image is divided into N homogeneous regions according to the entropy rate segmentation method.

[0020] As a preferred technical solution, in step 2.2, the degree matrix is ​​obtained from the similarity matrix according to the following formula;

[0021]

[0022] In the formula, D is the degree matrix, D mm W is the m-th element of the inter-band similarity matrix.lm These are elements in the similarity matrix, representing the similarity between band l and band m;

[0023] The normalized Laplace matrix is ​​obtained according to the following formula;

[0024] L = ID -1 / 2 ·W·D -1 / 2

[0025] In the formula, D -1 / 2 It is the semi-inverse of the degree matrix, I is the identity matrix, and W is the similarity matrix between bands.

[0026] As a preferred technical solution, the method for optimizing the latent feature matrix Y using an optimization function in step 2.4 is as follows:

[0027] Step A1. Determine the average vector for each original superpixel according to the following formula;

[0028]

[0029] In the formula, μ i Q is the average vector of the i-th original superpixel. (i) x is the i-th primitive superpixel, and x is the Q-value of the primitive superpixel. (i) The pixels in |Q (i) | represents the total number of pixels within the i-th superpixel;

[0030] Step A2. Process the average vector of the original superpixels using a similarity metric to obtain the spatial relation matrix S;

[0031] Step A3. Use the latent feature matrix Y obtained from spectral embedding as the initial value for iterative optimization;

[0032] Step A4. Determine the average vector of each superpixel in the latent feature matrix Y according to the following formula;

[0033]

[0034] In the formula, Y is the average vector of the i-th latent feature matrix. (i) Let y be the latent feature matrix corresponding to the i-th superpixel, and let y be the latent feature matrix Y of the i-th superpixel. (i) The elements in, |Y (i) | represents the total number of elements in the i-th latent feature matrix;

[0035] Step A5. Obtain the weight matrix S' using the same similarity measurement method as in step A2;

[0036] Step A6. Update the latent feature matrix Y using gradient descent;

[0037]

[0038] In the formula, η is the learning rate, and S' ij S is the similarity relationship between the i-th superpixel and the j-th superpixel after spatial feature extraction. ij It represents the similarity relationship between the i-th original superpixel and the j-th original superpixel;

[0039] Step A7. After each iteration, recalculate the average vector of each superpixel after extracting spatial features. The iteration continues until the objective function is satisfied or the number of iterations is reached, at which point the iteration ends. The objective function is as follows:

[0040]

[0041] The beneficial effects of this invention are as follows:

[0042] This invention improves the classification accuracy of hyperspectral images by optimizing the latent feature matrix and using a dynamic programming algorithm to select key bands. It enhances the stability of band selection by fusing the optimized latent feature matrix and maintaining band order in the frequency domain. By selecting a representative subset of bands, it reduces the dimensionality of the hyperspectral image while avoiding damage to the original image's physical structure, achieving effective dimensionality reduction. By using band selection instead of feature extraction, it reduces information redundancy and improves data utilization efficiency. The spatial structure information between superpixels is considered during band selection, more comprehensively preserving the global spatial structure of the hyperspectral image. The consideration of band order and spatial information improves the clustering performance of band selection. Furthermore, by optimizing the latent feature matrix, it reduces sensitivity to noise, enhancing the robustness of the invention. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the present invention.

[0044] Figure 2 This is a detailed architectural diagram of the present invention.

[0045] Figure 3 This is the result obtained from the embodiments of the present invention.

[0046] Figure 4 This is the result obtained using the RLFFC method. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0048] Example

[0049] exist Figure 1 ,2 The hyperspectral image band selection method based on superpixel segmentation and clustering in this embodiment includes the following steps:

[0050] Step 1. Superpixel segmentation

[0051] On a hyperspectral dataset of size W×H×B, where W is the image width, H is the image height, and B is the number of image channels, the first principal component of the image is extracted. This first principal component is then segmented into N homogeneous small regions Q = [Q...]. (1) Q (2) ,…,Q (N) ], Q (i) The size is i×B. Each small region is treated as an original superpixel, and the structural relationships between the original superpixels are constructed using a k-nearest neighbor graph.

[0052] Step 2. Spatial Feature Extraction

[0053] Step 2.1. Construct a similarity matrix within each original superpixel using a k-nearest neighbor graph, where the similarity between bands is measured by Euclidean distance;

[0054] Step 2.2. Obtain the degree matrix from the similarity matrix using the following formula:

[0055]

[0056] In the formula, D mm It is a degree matrix, W lm These are elements in the band similarity matrix, representing the similarity between band l and band m;

[0057] The normalized Laplace matrix is ​​obtained according to the following formula;

[0058] L = ID -1 / 2 ·W·D -1 / 2

[0059] In the formula, D -1 / 2 It is the semi-inverse of the degree matrix, I is the identity matrix, and W is the similarity matrix between bands;

[0060] Step 2.3. The latent feature matrix Y is formed by the eigenvectors corresponding to the first d smallest eigenvalues ​​of the Laplacian matrix. There are N latent feature matrices Y, and the size of the latent feature matrix Y is B×d.

[0061] Step 2.4. Optimize the latent feature matrix Y using an optimization function to extract the spatial structure between superpixels;

[0062] The method for optimizing the latent feature matrix Y using the optimization function is as follows:

[0063] Step A1. Determine the average vector for each original superpixel according to the following formula;

[0064]

[0065] In the formula, μ i Q is the average vector of the i-th original superpixel. (i) x is the i-th primitive superpixel, and x is the Q-value of the primitive superpixel. (i) The pixels in |Q (i) | represents the total number of pixels within the i-th superpixel;

[0066] Step A2. Process the average vector of the original superpixels using a similarity metric to obtain the spatial relation matrix S;

[0067] Step A3. Use the latent feature matrix Y obtained from spectral embedding as the initial value for iterative optimization;

[0068] Step A4. Determine the average vector of each superpixel in the latent feature matrix Y according to the following formula;

[0069]

[0070] In the formula, Y is the average vector of the i-th latent feature matrix. (i) Let y be the latent feature matrix corresponding to the i-th original superpixel, and y be the latent feature matrix Y of the i-th element. (i) The elements in, |Y (i) | represents the total number of elements in the i-th latent feature matrix;

[0071] Step A5. Obtain the weight matrix S' using the same similarity measurement method as in step A2;

[0072] Step A6. Update the latent feature matrix Y using the gradient descent method as follows:

[0073]

[0074] In the formula, η is the learning rate, and S' ij S is the similarity relationship between the i-th superpixel and the j-th superpixel after spatial feature extraction. ij It represents the similarity relationship between the i-th original superpixel and the j-th original superpixel;

[0075] Step A7. After each iteration, recalculate the average vector of each superpixel after extracting spatial features. The iteration continues until the objective function is satisfied or the number of iterations is reached, at which point the iteration ends. The objective function is as follows:

[0076]

[0077] Step 2.5. Fuse the N optimized latent feature matrices Y to obtain a shared latent feature matrix F of size B×d;

[0078] Step 3. Key Band Identification and Band Clustering Selection

[0079] A series of key bands are found using a dynamic programming algorithm on a shared latent feature matrix F. The remaining bands are divided into different clusters based on the key bands. Within each cluster, the band with the highest ranking value is selected, and only one band can be selected within each cluster to form a band subset.

[0080] Step 4. Verify the results

[0081] The obtained band subset is validated using an SVM classifier to obtain the classification results.

[0082] experiment

[0083] To verify the beneficial effects of the present invention, the inventors conducted a comparative experiment using the method of the embodiments of the present invention and the existing RLFFC method.

[0084] Experiments were conducted using 10 pre-defined test and training sets. The average of the 10 results was taken as the final result. The evaluation metrics were overall accuracy (OA), average accuracy (AA), and Kappa coefficient. The classification results of the method in this embodiment are as follows: Figure 3 The classification results of the RLFFC method are as follows: Figure 4 .

[0085] The classification results of the two methods show that the bands selected by the RLFFC-OCF method proposed in this invention are more representative than those of the original RLFFC method in most cases. Specifically, when selecting 10 and 15 bands to form the final band subset, the original RLFFC method is better than the method proposed in this invention (RLFFC-OCF); while when the number of selected bands is 5, 20, 25, 30, 35, 40, 45, and 50, the RLFFC-OCF proposed in this invention is better than the original RLFFC.

Claims

1. A method for selecting bands in hyperspectral images based on superpixel segmentation and clustering, characterized in that, Includes the following steps: Step 1. Superpixel segmentation In size W H On the hyperspectral dataset B, where W is the width of the image, H is the height of the image, and B is the number of channels in the image, the first principal component of the image is extracted, and the first principal component of the image is divided into N homogeneous small regions. Each small region is used as an original superpixel, and the structural relationship between each original superpixel is constructed using a k-nearest neighbor graph. Step 2. Spatial Feature Extraction Step 2.

1. Construct a similarity matrix within each original superpixel using a k-nearest neighbor graph, where the similarity between bands is measured by Euclidean distance; Step 2.

2. Obtain the degree matrix from the similarity matrix, and then obtain the normalized Laplace matrix; Step 2.

3. The latent feature matrix Y is formed by the eigenvectors corresponding to the first d smallest eigenvalues ​​of the Laplacian matrix. There are N latent feature matrices Y, and the size of the latent feature matrix Y is B×d. Step 2.

4. Optimize the latent feature matrix Y using an optimization function to extract the spatial structure between superpixels: Step A1. Determine the average vector for each original superpixel according to the following formula; : = In the formula, It is the average vector of the i-th original superpixel. It is the i-th original superpixel. It is the original superpixel Pixels in It is the total number of pixels within the i-th superpixel; Step A2. Process the average vector of the original superpixels using a similarity metric to obtain the spatial relation matrix S; Step A3. Use the latent feature matrix Y obtained from spectral embedding as the initial value for iterative optimization; Step A4. Determine the average vector of each superpixel in the latent feature matrix Y according to the following formula; ： = In the formula, It is the average vector of the i-th latent feature matrix. It is the latent feature matrix corresponding to the i-th superpixel. It is the i-th latent feature matrix The elements in It is the total number of elements in the i-th latent feature matrix; Step A5. Obtain the weight matrix using the same similarity measurement method as in Step A2. ; Step A6. Update the latent feature matrix Y using gradient descent; ： ← - In the formula, It's the learning rate. It represents the similarity relationship between the i-th superpixel and the j-th superpixel after spatial feature extraction. It represents the similarity relationship between the i-th original superpixel and the j-th original superpixel; Step A7. After each iteration, recalculate the average vector of each superpixel after extracting spatial features. The iteration continues until the objective function is satisfied or the number of iterations is reached, at which point the iteration ends. The objective function is as follows: ； Step 2.

5. Fuse the N optimized latent feature matrices Y to obtain a shared latent feature matrix F of size B×d; Step 3. Key Band Identification and Band Clustering Selection A series of key bands are found using a dynamic programming algorithm on a shared latent feature matrix F. The remaining bands are divided into different clusters based on the key bands. Within each cluster, the band with the highest ranking value is selected, and only one band can be selected within each cluster to form a band subset. Step 4. Verify the results The obtained band subset is validated using a classifier, and the classification results are observed.

2. The hyperspectral image band selection method based on superpixel segmentation and clustering according to claim 1, characterized in that, In step 1, the first principal component of the image is segmented into N homogeneous regions according to the entropy rate segmentation method.

3. The hyperspectral image band selection method based on superpixel segmentation and clustering according to claim 1, characterized in that, In step 2.2, the degree matrix is ​​obtained from the similarity matrix according to the following formula; = In the formula, It is a degree matrix. It is the m-th element of the inter-band similarity matrix. These are elements in the similarity matrix, representing the similarity between band l and band m; The normalized Laplace matrix is ​​obtained according to the following formula; In the formula, It is the semi-inverse of the degree matrix. It is the identity matrix. It is a similarity matrix between bands.

Images