Hyperspectral image band selection method based on superpixel segmentation and clustering
Through superpixel segmentation and clustering methods, the potential features of hyperspectral images are extracted and optimized, and the key bands are selected in combination with dynamic programming algorithms, which solves the problems of information loss and noise sensitivity in the band selection process in the prior art, and achieves high precision, Stable and efficient hyperspectral image dimensionality reduction.
Patent Information
- Application Number
- CN202411090470.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-09
AI Technical Summary
The existing hyperspectral image dimensionality reduction methods are prone to destroying the original physical structure during the band selection process, are sensitive to noise, and have unstable clustering performance, which fails to effectively utilize the spatial information and band order of the hyperspectral image.
Using a method based on superpixel segmentation and clustering, the first principal component of the image is divided into homogeneous regions, the k-nearest neighbor map is constructed, the potential feature matrix is extracted and optimized, and the key bands are selected in combination with the dynamic programming algorithm to form a representative one. subset of bands, considering the orderliness and spatial structure of bands.
It improves the classification accuracy and band selection stability of hyperspectral images, reduces information redundancy, preserves the global spatial structure of the image, and enhances robustness and clustering performance.
Smart Images

Figure CN119027751A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image processing, and in particular relates to a hyperspectral image band selection method based on superpixel segmentation and clustering. Background Art
[0002] Hyperspectral imaging technology can simultaneously obtain geometric, radiation and spectral information of objects, integrating the capabilities of cameras, radiometers and spectrometers. Compared with optical spatial two-dimensional imaging, it can perform spatial and spectral three-dimensional imaging of objects. Under a certain spatial resolution, it can obtain the unique continuous characteristic spectrum of objects in a wide spectral range, and has outstanding advantages in the fine classification and identification of objects. It has become an important frontier technology for remote sensing of the earth, and plays an increasingly important role in natural resource surveys, ecological environment monitoring, agriculture, forestry, animal husbandry and fishery, and marine and coastal monitoring. However, hyperspectral images contain hundreds or thousands of bands, which makes a lot of information hidden in hyperspectral images, and also brings redundancy, increases the complexity of calculations, and has a certain impact on the classification accuracy of objects. Therefore, it is necessary to perform dimensionality reduction processing on hyperspectral images.
[0003] Existing hyperspectral image dimensionality reduction methods are mainly divided into feature extraction and band selection. Feature extraction is to map high-dimensional data into low-dimensional space. The more commonly used methods include principal component analysis, independent component analysis and linear discriminant analysis. During feature extraction, the mapping operation destroys the physical structure of the original hyperspectral image, resulting in the loss of some key information. Band selection is to select a representative subset from all bands, which does not change the physical structure of the hyperspectral image. According to the availability of labels, band selection can be divided into unsupervised methods and supervised methods. Since the labels of hyperspectral images need to be manually marked by experts in various fields, which is time-consuming, labor-intensive and difficult to obtain, unsupervised band selection methods are of great interest to researchers.
[0004] Among the existing unsupervised band selection methods, they can be roughly divided into two categories: sorting-based and clustering-based. The basic idea of the sorting-based band selection method is to sort all hyperspectral bands according to certain sorting rules, and then select the top or bottom ranked bands to form a new band subset. Although these methods have achieved satisfactory results, they also have two disadvantages. First, these algorithms are very sensitive to noisy hyperspectral bands, 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. The basic idea of the clustering-based band selection method is to divide the original bands into multiple clusters through a clustering algorithm, and then select a representative band in each cluster to form the final band subset. Unfortunately, this algorithm makes the similarity of the obtained subsets very small, but the selected bands may not have enough information. The common defect of these two methods is that most of the 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 the corresponding objective function. In practice, for a certain band, different regions often correspond to different objects or land covers, and the spectral features are different. The high-dimensional pixel features of each band contain a lot of redundant information, which limits the clustering performance of hyperspectral band selection. Therefore, it is not suitable to directly regard each band as a single feature vector. In addition, most existing methods tend to use the k-means method to cluster on the bands that have been reduced in space after extracting spatial features, which makes the selected band subsets unstable each time, bringing uncertainty to subsequent tasks. In addition, when performing clustering algorithms, hyperspectral bands are usually considered to be disordered. It leads to the neglect of the contextual information of the entire spectral band. In fact, the closer the distance, the closer the image information displayed by the bands is, and experiments have also proved the orderliness of the bands. At the same time, there are a large number of methods that use superpixel segmentation technology to extract spatial features of hyperspectral images. They do not consider the structural relationship between superpixels, but directly fuse the low-dimensional potential features of each small area, and cannot more comprehensively retain 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 technical problems is: a hyperspectral image band selection method based on superpixel segmentation and clustering, comprising 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 areas. Each small area 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. Use the k nearest neighbor graph to construct a similarity matrix in each original superpixel, and the similarity between bands is measured by the Euclidean distance;
[0011] Step 2.2. Obtain the degree matrix through the similarity matrix, and then obtain the normalized Laplace matrix;
[0012] Step 2.3. The eigenvectors corresponding to the first d smallest eigenvalues of the Laplacian matrix form a latent feature matrix Y, with a total of N eigenvectors, where the size of the latent feature matrix Y is B×d;
[0013] Step 2.4. Optimize the latent feature matrix Y through the 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 cluster selection
[0016] A dynamic programming algorithm is used to find a series of key bands on the shared potential feature matrix F. The remaining bands are divided into different clusters according to the key bands. A band with the highest ranking value is selected in each cluster, and only one band can be selected in each cluster to form a band subset.
[0017] Step 4. Verify the results
[0018] The obtained band subset is verified using the classifier and the classification results are observed.
[0019] As a preferred technical solution, in step 1, the first principal component of the image is segmented into N homogeneous regions according to an entropy rate segmentation method.
[0020] As a preferred technical solution, in step 2.2, the degree matrix is obtained by the similarity matrix according to the following formula:
[0021]
[0022] Where D is the degree matrix, D mm is the mth element of the inter-band similarity matrix, Wlm is an element in the similarity matrix, indicating 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] Where D -1 / 2 is the semi-inverse matrix of the degree matrix, I is the identity matrix, and W is the inter-band similarity matrix.
[0026] As a preferred technical solution, the method for optimizing the potential feature matrix Y by the optimization function in step 2.4 is:
[0027] Step A1. Determine the average vector of each original superpixel according to the following formula;
[0028]
[0029] In the formula, μ i is the average vector of the i-th original superpixel, Q (i) is the i-th original superpixel, x is the original superpixel Q (i) Pixels in │Q (i) │ is the total number of pixels in the i-th superpixel;
[0030] Step A2. Process the average vector of the original superpixels using a similarity measurement method to obtain a spatial relationship matrix S;
[0031] Step A3. Use the latent feature matrix Y obtained by spectral embedding as the initial value of 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, is the mean vector of the ith latent feature matrix, Y (i) is the potential feature matrix corresponding to the i-th superpixel, and y is the i-th potential feature matrix Y (i) The elements in │Y (i) │ is the total number of elements in the i-th potential feature matrix;
[0035] Step A5. Use the same similarity measurement method as step A2 to obtain the weight matrix S';
[0036] Step A6. Update the latent feature matrix Y using the gradient descent method;
[0037]
[0038] Where η is the learning rate, S' ij is the similarity relationship between the i-th superpixel and the j-th superpixel after spatial feature extraction, S ij is 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 is exited until the objective function is satisfied or the number of iterations is reached. The objective function is as follows:
[0040]
[0041] The beneficial effects of the present invention are as follows:
[0042] The present invention improves the classification accuracy of hyperspectral images by optimizing the potential feature matrix and using a dynamic programming algorithm to select key bands. The stability of band selection is enhanced by fusing the optimized potential feature matrix and maintaining the orderliness of the bands in the frequency domain. The dimension of the hyperspectral image is reduced by selecting a representative band subset, while avoiding the destruction of the physical structure of the original image, thereby achieving the purpose of effective dimensionality reduction. By selecting bands rather than extracting features, information redundancy is reduced and data utilization efficiency is improved. In the band selection process, the spatial structure information between superpixels is considered, and the global spatial structure of the hyperspectral image is more comprehensively retained. Since the orderliness and spatial information of the bands are considered, the clustering performance of the band selection is improved. By optimizing the potential feature matrix, the sensitivity to noise is reduced, thereby enhancing the robustness of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a schematic diagram of the process of the present invention.
[0044] Figure 2 It is a specific architecture diagram of the present invention.
[0045] Figure 3 This is the result obtained by the embodiment of the present invention.
[0046] Figure 4 is the result obtained by the RLFFC method. DETAILED DESCRIPTION
[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples, but the present invention is not limited to the following embodiments.
[0048] Example
[0049] exist Figure 1 ,2 In the embodiment, the hyperspectral image band selection method based on superpixel segmentation and clustering includes the following steps:
[0050] Step 1. Superpixel segmentation
[0051] 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 using the entropy rate Q=[Q (1) ,Q (2) ,…,Q (N) ], Q (i) The size of is i×B, Each small area is taken as an original superpixel, and the k-nearest neighbor graph is used to construct the structural relationship between the original superpixels;
[0052] Step 2. Spatial feature extraction
[0053] Step 2.1. Use the k nearest neighbor graph to construct a similarity matrix in each original superpixel, and the similarity between bands is measured by the Euclidean distance;
[0054] Step 2.2. Obtain the degree matrix through the similarity matrix according to the following formula:
[0055]
[0056] Where D mm is the degree matrix, W lm is an element in the inter-band similarity matrix, indicating 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] Where D -1 / 2 is the semi-inverse matrix of the degree matrix, I is the identity matrix, and W is the inter-band similarity matrix;
[0060] Step 2.3. The eigenvectors corresponding to the first d smallest eigenvalues of the Laplacian matrix form a latent feature matrix Y, with a total of N eigenvectors, where the size of the latent feature matrix Y is B×d;
[0061] Step 2.4. Optimize the latent feature matrix Y through the optimization function to extract the spatial structure between superpixels;
[0062] Among them, the method for optimizing the potential feature matrix Y by the optimization function is:
[0063] Step A1. Determine the average vector of each original superpixel according to the following formula;
[0064]
[0065] In the formula, μ i is the average vector of the i-th original superpixel, Q (i) is the i-th original superpixel, x is the original superpixel Q (i) Pixels in │Q (i) │ is the total number of pixels in the i-th superpixel;
[0066] Step A2. Process the average vector of the original superpixels using a similarity measurement method to obtain a spatial relationship matrix S;
[0067] Step A3. Use the latent feature matrix Y obtained by spectral embedding as the initial value of 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, is the mean vector of the ith latent feature matrix, Y (i) is the potential feature matrix corresponding to the i-th original superpixel, and y is the i-th potential feature matrix Y (i) The elements in │Y (i) │ is the total number of elements in the i-th potential feature matrix;
[0071] Step A5. Use the same similarity measurement method as step A2 to obtain the weight matrix S';
[0072] Step A6. Update the latent feature matrix Y using the following gradient descent method:
[0073]
[0074] Where η is the learning rate, S' ij is the similarity relationship between the i-th superpixel and the j-th superpixel after spatial feature extraction, S ij is 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 is exited until the objective function is satisfied or the number of iterations is reached. 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 cluster selection
[0079] A dynamic programming algorithm is used to find a series of key bands on the shared potential feature matrix F. The remaining bands are divided into different clusters according to the key bands. A band with the highest ranking value is selected in each cluster, and only one band can be selected in each cluster to form a band subset.
[0080] Step 4. Verify the results
[0081] The obtained band subset is verified using the SVM classifier to obtain the classification result.
[0082] experiment
[0083] In order to verify the beneficial effects of the present invention, the inventors conducted a comparative experiment using the method of the embodiment of the present invention and the existing RLFFC method.
[0084] The experiment was conducted using 10 different test sets and training sets that were divided in advance. The average of the 10 results was taken as the final result. The evaluation indicators were the overall accuracy OA, the average accuracy AA and the Kappa coefficient. The classification results of the method of the embodiment of the present invention 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 the present invention are more representative than those of the original RLFFC method in most cases. Specifically, when 10 bands and 15 bands are selected to form the final band subset, the original RLFFC method is better than the method proposed in the present invention (RLFFC-OCF); and when the number of selected bands is 5, 20, 25, 30, 35, 40, 45 and 50, the RLFFC-OCF proposed in the present invention is better than the original RLFFC.
Claims
1. A hyperspectral image band selection method based on superpixel segmentation and clustering, characterized in that: The following steps are involved: Step 1. Superpixel segmentation 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 areas. Each small area 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. Use the k nearest neighbor graph to construct a similarity matrix in each original superpixel, and the similarity between bands is measured by the Euclidean distance; Step 2.
2. Obtain the degree matrix through the similarity matrix, and then obtain the normalized Laplace matrix; Step 2.
3. The eigenvectors corresponding to the first d smallest eigenvalues of the Laplacian matrix form a latent feature matrix Y, with a total of N eigenvectors, where the size of the latent feature matrix Y is B×d; Step 2.
4. Optimize the latent feature matrix Y through the optimization function to extract the spatial structure between superpixels; 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 cluster selection A dynamic programming algorithm is used to find a series of key bands on the shared potential feature matrix F. The remaining bands are divided into different clusters according to the key bands. A band with the highest ranking value is selected in each cluster, and only one band can be selected in each cluster to form a band subset. Step 4. Verify the results The obtained band subset is verified using the 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 an 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 by the similarity matrix according to the following formula; Where D is the degree matrix, D mm is the mth element of the inter-band similarity matrix, W Im is an element in the similarity matrix, indicating the similarity between band l and band m; The normalized Laplace matrix is obtained according to the following formula; L=I-D -1 / 2 ·W·D -1 / 2 Where D -1 / 2 is the semi-inverse matrix of the degree matrix, I is the identity matrix, and W is the inter-band similarity matrix.
4. The hyperspectral image band selection method based on superpixel segmentation and clustering according to claim 1, characterized in that: The method for optimizing the potential feature matrix Y by optimizing the function in step 2.4 is: Step A1. Determine the average vector of each original superpixel according to the following formula; In the formula, μ i is the average vector of the i-th original superpixel, Q (i) is the i-th original superpixel, x is the original superpixel Q (i) Pixels in │Q (i) │ is the total number of pixels in the i-th superpixel; Step A2. Process the average vector of the original superpixels using a similarity measurement method to obtain a spatial relationship matrix S; Step A3. Use the latent feature matrix Y obtained by spectral embedding as the initial value of 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, is the mean vector of the ith latent feature matrix, Y (i) is the potential feature matrix corresponding to the i-th superpixel, and y is the i-th potential feature matrix Y (i) The elements in │Y (i) │ is the total number of elements in the i-th potential feature matrix; Step A5. Use the same similarity measurement method as step A2 to obtain the weight matrix S'; Step A6. Update the latent feature matrix Y using the gradient descent method; Where η is the learning rate, S' ij is the similarity relationship between the i-th superpixel and the j-th superpixel after spatial feature extraction, S ij is 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 is exited until the objective function is satisfied or the number of iterations is reached. The objective function is as follows:
Citation Information
Patent Citations
wrapper-type hyperspectral waveband selection method based on pixel clustering
CN105184314A
Hyperspectral classification identification method based on superpixel segmentation
CN115937685A
Global weighted least square filtering method based on RoEWA operator
CN116309145A
Local view auxiliary discrimination hyperspectral band selection method and system
CN116778339A
Deep sea hyperspectral image target identification method based on multi-modal fusion
CN117437533A
