A Hyperspectral Image Classification Method and System Based on Integrated Clustering Band Selection

Abstract

The present invention belongs to the technical field of image processing, and provides a hyperspectral image classification method and system based on hyperspectral band selection. The method includes: acquiring a hyperspectral image data set; based on the hyperspectral image data set, adopting a base clustering generation strategy, combining spectral clustering to obtain corresponding clustering partitions, and obtaining a base clustering set by setting different parameters; based on the base clustering set, using the LWEC method to calculate the entropy and ECI values of each cluster, and generating an LWCA matrix; based on the LWCA matrix, using a consensus function to find a segmentation point, and obtaining a clustering result according to the segmentation point; based on the clustering result, using a manifold ranking method to obtain representative bands; based on the representative bands, combining with the hyperspectral image data set, classifying the pixels of the hyperspectral image.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular, relates to a hyperspectral image classification method and system based on integrated clustering band selection. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] Hyperspectral images have rich spectral information and spatial information, making them widely used in fields such as ground object recognition and classification. However, with the increase in the number of bands, the correlation between different bands, especially adjacent bands, is very strong, resulting in a large amount of redundant information, forming the so-called "Hughes" phenomenon. This not only reduces the classification performance but also increases the computer processing time and storage space. Therefore, it is necessary to perform dimensionality reduction on hyperspectral data before ground object recognition and classification. Band selection is one of the important techniques for hyperspectral data dimensionality reduction. It selects representative bands from all bands as the optimal band subset to replace all bands for research, achieving data dimensionality reduction without changing the physical meaning of the original data.

[0004] Among band selection methods, clustering-based band selection methods have attracted much attention because they can effectively select rich information and representative bands. For example, Ward’s Linkage strategy Using Divergence (WaluDi), Enhanced Fast density-peak-based clustering (E-FDPC), and Normalized Cut based Optimal Clustering with ranking criteria using Information Entropy (NC-OC-IE), etc. These methods first divide all bands into multiple clusters and then select representative bands from each cluster. These clustering-based methods reduce the redundancy between bands and improve the classification accuracy.

[0005] However, these methods also have some deficiencies in the clustering process. For example, 1) Most of these methods use a single clustering algorithm. For high-dimensional data such as hyperspectral images, it is difficult to ensure the effectiveness and robustness of the clustering results. 2) Hyperspectral data contains a large amount of redundant and irrelevant information. The difference between bands may be caused by some features. Directly clustering the high-dimensional original data will reduce the effectiveness of the algorithm. 3) Most clustering-based methods ignore the utilization of information related to the band selection problem during the clustering process, thus limiting the performance of the algorithm.

[0006] Regarding the problem of high-dimensional data clustering, researchers at home and abroad mainly conduct research from aspects such as feature selection, subspace clustering, ensemble clustering, and deep learning. Among them, ensemble clustering can generate more effective clustering results through certain strategies from the results of multiple base clusterings, improving the stability of high-dimensional data clustering. In addition, ensemble clustering shows unique advantages in generating robust partitions, dealing with noisy features, and mining new structures. Generally speaking, ensemble clustering can be divided into two categories, namely the objective function-based method and the heuristic-based method. The objective function-based method takes the similarity measure between multiple partitions as an explicit global objective for designing an effective consensus function. Representative methods include combinatorial regularization and the class K-means algorithm. In contrast, heuristic-based methods, such as voting-based methods and consistency matrix-based methods, use some heuristics instead of an objective function to search for approximate solutions. For example, Huang et al. recently proposed a Locally Weighted Ensemble Clustering (LWEC), which uses entropy theory to estimate the uncertainty of each cluster in all base clusterings, and thus further improves the consensus clustering result using the local weighting strategy in the consensus function. Although existing ensemble clustering methods have greatly improved in clustering performance, they have not been applied to the band selection task. In addition, the method of locally weighted ensemble clustering was proposed without considering the characteristics of hyperspectral images. Therefore, how to introduce information related to the hyperspectral band selection problem into the clustering strategy and design an effective consensus function to generate better clustering results remains a challenging problem.

[0007] In recent years, representation learning has unique advantages in feature representation, which also provides a new solution idea for high-dimensional data clustering. Summary of the Invention

[0008] To solve the technical problems existing in the above background technology, the present invention provides a hyperspectral image classification method and system based on ensemble clustering band selection, which takes the methods of ensemble clustering and representation learning as the framework, fully utilizes the intrinsic characteristics of hyperspectral images, improves the clustering effect to facilitate the selection of more representative bands, and thus improves the accuracy of hyperspectral image classification.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] The first aspect of the present invention provides a hyperspectral image classification method based on ensemble clustering band selection.

[0011] A hyperspectral image classification method based on ensemble clustering band selection, comprising:

[0012] Obtain a hyperspectral image dataset;

[0013] Based on the hyperspectral image dataset, adopt a base clustering generation strategy, combine spectral clustering to obtain corresponding clustering partitions, and obtain a base clustering set by setting different parameters;

[0014] Based on the base clustering set, use the LWEC method to calculate the entropy and ECI values of each cluster, and generate an LWCA matrix;

[0015] Based on the LWCA matrix, adopt a consensus function to find the segmentation point, and obtain the clustering result according to the segmentation point;

[0016] Based on the clustering result, adopt a manifold ranking method to obtain representative bands;

[0017] Based on the representative bands, combine with the hyperspectral image dataset to classify the pixels of the hyperspectral image.

[0018] The second aspect of the present invention provides a hyperspectral image classification system based on integrated clustering band selection.

[0019] A hyperspectral image classification system based on integrated clustering band selection, comprising:

[0020] A data acquisition module, which is configured to: obtain a hyperspectral image dataset;

[0021] A base clustering module, which is configured to: based on the hyperspectral image dataset, adopt a base clustering generation strategy, combine spectral clustering to obtain corresponding clustering partitions, and obtain a base clustering set by setting different parameters;

[0022] A matrix generation module, which is configured to: based on the base clustering set, use the LWEC method to calculate the entropy and ECI values of each cluster, and generate an LWCA matrix;

[0023] A segmentation module, which is configured to: based on the LWCA matrix, adopt a consensus function to find the segmentation point, and obtain the clustering result according to the segmentation point;

[0024] A selection module, which is configured to: based on the clustering result, adopt a manifold ranking method to obtain representative bands;

[0025] A classification module, which is configured to: based on the representative bands, combine with the hyperspectral image dataset to classify the pixels of the hyperspectral image.

[0026] The third aspect of the present invention provides a computer-readable storage medium.

[0027] A computer-readable storage medium has a computer program stored thereon. When the program is executed by a processor, it implements the steps in the hyperspectral image classification method based on integrated clustering band selection described in the first aspect above.

[0028] The fourth aspect of the present invention provides a computer device.

[0029] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the hyperspectral image classification method based on integrated clustering band selection described in the first aspect above.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] The present invention selects representative bands and combines them with hyperspectral images to improve the classification performance of hyperspectral images.

[0032] The present invention introduces local weighted integrated clustering into the hyperspectral band selection method and makes improvements from the following four aspects: First, it combines a deep convolutional autoencoder with subspace clustering, utilizes spatial information and nonlinear feature transformation, and fully extracts the interactions between spectral bands, improving the quality of the base clustering in integrated clustering. Second, by combining with a large number of random diversity metrics, it enhances the diversity of the base clustering, which is beneficial to improving the consensus performance. Third, it proposes a new consensus function, namely the Continuous Similar Band Division Strategy (CSBDS), which effectively fuses the base clustering using the correlation between bands, making the final clustering result more in line with the needs of band selection. Fourth, it improves the manifold ranking and selects representative bands from each cluster. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0034] Figure 1 is a flowchart of hyperspectral band selection shown in an embodiment of the present invention;

[0035] FIG. 2(a) is an LWCA matrix in the CSBDS division process shown in an embodiment of the present invention;

[0036] FIG. 2(b) is the CSBDS division process shown in an embodiment of the present invention;

[0037] Figure 3(a) is a schematic diagram of the experimental results of using the SVM classifier on the Pavia University dataset shown in the embodiments of the present invention;

[0038] Figure 3(b) is a schematic diagram of the experimental results of using the KNN classifier on the Pavia University dataset shown in the embodiments of the present invention;

[0039] Figure 4(a) is a schematic diagram of the experimental results of using the SVM classifier on the Botswana dataset shown in the embodiments of the present invention;

[0040] Figure 4(b) is a schematic diagram of the experimental results of using the KNN classifier on the Botswana dataset shown in the embodiments of the present invention;

[0041] Figure 5(a) is a schematic diagram of the experimental results of using the SVM classifier on the Pavia Centre dataset shown in the embodiments of the present invention;

[0042] Figure 5(b) is a schematic diagram of the experimental results of using the KNN classifier on the Pavia Centre dataset shown in the embodiments of the present invention;

[0043] Figure 6(a) is a ground truth classification information diagram of the Pavia University dataset shown in the embodiments of the present invention;

[0044] Figure 6(b) is a classification diagram obtained by using the SVM classifier with 30 selected bands on the Pavia University dataset shown in the embodiments of the present invention;

[0045] Figure 6(c) is a classification diagram obtained by using the KNN classifier with 30 selected bands on the Pavia University dataset shown in the embodiments of the present invention;

[0046] Figure 7(a) is a ground truth classification information diagram of the Botswana dataset shown in the embodiments of the present invention;

[0047] Figure 7(b) is a classification diagram obtained by using the SVM classifier with 30 selected bands on the Botswana dataset shown in the embodiments of the present invention;

[0048] Figure 7(c) is a classification diagram obtained by using the KNN classifier with 30 selected bands on the Botswana dataset shown in the embodiments of the present invention;

[0049] Figure 8(a) is a ground truth classification information diagram of the Pavia Centre dataset shown in the embodiments of the present invention;

[0050] Figure 8(b) is the classification map obtained by using the SVM classifier with 30 selected bands on the Pavia Centre dataset as shown in the embodiments of the present invention;

[0051] Figure 8(c) is the classification map obtained by using the KNN classifier with 30 selected bands on the Pavia Centre dataset as shown in the embodiments of the present invention. Detailed implementation manners

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which the present invention belongs.

[0054] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of the described features, steps, operations, devices, components, and / or their combinations.

[0055] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of the methods and systems according to various embodiments of the present disclosure. It should be noted that each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the module, program segment, or part of code may include one or more executable instructions for implementing the logical functions specified in each embodiment. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. Similarly, it should be noted that each block in the flowchart and / or block diagram, as well as the combinations of blocks in the flowchart and / or block diagram, can be implemented using a dedicated hardware-based system for performing the specified functions or operations, or can be implemented using a combination of dedicated hardware and computer instructions.

[0056] Term explanation:

[0057] Integrated clustering: Integrated clustering is a method for effectively fusing the results of multiple clustering algorithms. It can give multiple data partitions based on the base clustering, providing more accurate and more robust clustering results.

[0058] Representation learning: Representation learning is to obtain an effective representation of data, which is to extract effective low-dimensional features from the complex features of high-dimensional data. Representation learning includes sparse representation, manifold learning, deep learning, reinforcement learning, autoencoders, etc. The deep convolutional autoencoder technology is used in the present invention.

[0059] Band selection: Band selection is one of the important techniques for hyperspectral data dimensionality reduction. It selects representative bands from all bands as the optimal band subset to replace all bands for research, and realizes data dimensionality reduction without changing the physical meaning of the original data.

[0060] Embodiment 1

[0061] This embodiment provides a hyperspectral image classification method based on integrated clustering band selection. This embodiment takes the application of this method to a server as an example. It can be understood that this method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is realized through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, web servers, cloud communications, middleware services, domain name services, security services CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication methods, and this application does not make any restrictions here. In this embodiment, the method includes the following steps:

[0062] Obtain a hyperspectral image data set;

[0063] Based on the hyperspectral image data set, adopt a base clustering generation strategy, combine spectral clustering to obtain corresponding clustering partitions, and obtain a base clustering set by setting different parameters;

[0064] Based on the base clustering set, use the LWEC method to calculate the entropy and ECI values of each cluster, and generate an LWCA matrix;

[0065] Based on the LWCA matrix, use a consensus function to find the segmentation point, and obtain a clustering result according to the segmentation point;

[0066] Based on the clustering result, use a manifold ranking method to obtain representative bands;

[0067] Based on the representative bands, combine with the hyperspectral image data set to classify the pixels of the hyperspectral image.

[0068] This embodiment includes two parts: integrated clustering and representative band selection. The overall process is asFigure 1 As shown, integrated clustering first needs to generate base clustering members, and then obtains the final clustering partition through a consensus function. After obtaining the integrated clustering result, representative bands are selected from each cluster through popularity ranking. In the generation stage of the base clustering, this method adopts two base clustering generation strategies. One is to use deep subspace clustering (DSC) based on a convolutional autoencoder, and the other is to use a method for measuring the diversity of high-dimensional data.

[0069] 1) Subspace clustering based on a deep autoencoder

[0070] Given a hyperspectral dataset where N represents the number of bands, P represents the number of pixels, and B is the input image, is the reconstructed image. The encoder has two modules. Each module includes two convolutional layers with batch normalization (BN), RELU activation, and a max pooling layer. The encoder is expressed as a function of the input B and the parameter θ e and the output μ, μ = E(B; θ e ). Generally, μ is called the latent representation, which reveals the inherent spatial information of the input band image. Similarly, the decoder also consists of two modules. Instead, deconvolution operations are applied to reconstruct the input. The decoder can be defined as where the latent representation μ is used as the input to the decoder. Accordingly, DSC adopts the mean squared error as the loss function:

[0071]

[0072] For the self-expression layer, first, the latent representation μ is unfolded into a d-dimensional vector. Assume that N band images are processed in n different subspaces. That is, S = S1 ∪ S2 ∪... ∪ S n , where S1, S2, …… S n represent n subspaces with dimensions d1, d2, ……, d n respectively. And it satisfies Mathematically, this assumption is expressed as:

[0073] Min ‖C‖ p s.t. U = UC, diag(C) = 0, (2)

[0074] where U = [μ1, μ2,..., μ N T is the latent representation matrix. is the coefficient matrix, and the i-th column represents μ i ​Linear representation. diag(C)=0 constrains the diagonal elements of C to be equal to 0 to avoid trivial solutions. To avoid the coefficient solution being "too sparse" due to the high correlation of hyperspectral image bands and enable it to be trained with CAE, DSC uses the L2-norm in the model. Equation (2) is rewritten as follows

[0075]

[0076] where λ is the balance coefficient used to weigh these two terms. Equation (3) uses C as the connection weight matrix and ignores the bias, which can actually be regarded as a fully connected layer. Therefore, this layer is called the self-expression layer. Its goal is to learn the subspace representation of the high-level convolutional features in CAE. The complete loss function is expressed as:

[0077]

[0078] where B is the input band image, represents the reconstructed band image, and α is the coefficient of the self-expression term. After network training, the similarity matrix C can be obtained, and then spectral clustering is used to obtain the corresponding clustering partition. By setting different parameters, multiple base clustering members can be obtained.

[0079] 2) Generation of base clusters for diversity measurement

[0080] A. Random kernel similarity measurement. For the hyperspectral dataset B = {b1,..., b N} containing N bands, where is the i-th band and D is the number of pixels (features). The kernel similarity between b i and b j is defined as:

[0081]

[0082] where u is the hyperparameter, ε ij is the scale parameter, d(b i , b j ) is the Euclidean distance between b i and b j , and N k (b i ) represents the set of the smallest neighborhoods of b i . The average distance between b i and its k nearest neighbors can be calculated as:

[0083]

[0084] To simultaneously consider the neighborhood of b i and the neighborhood of b jThe neighborhood thereof and the distance, and define the scale parameter ε ij as ρ k (b i ), ρ k (b j ), and the average value of d(b i , b j ). It is expressed by the formula as follows:

[0085]

[0086] The above kernel is a variant of the Gaussian kernel. It has two parameters, namely the hyperparameter u and the number k of the nearest neighbors. Randomly select these two parameters u ∈ [u min , u max and k ∈ [k min , k max , which are expressed respectively as follows:

[0087] u = u min + σ1(u max - u min ), (8)

[0088]

[0089] wherein, σ1 ∈ [0, 1] and σ2 ∈ [0, 1] are two uniformly distributed random variables, representing the lower limit of the output real number. By randomizing the parameters u and k, a large number of diverse metrics are generated to enhance the diversity. Assume that it is randomly selected M times, then a set of M pairs of u and k can be obtained, which correspond to the random kernel similarity metrics of the dataset B and are expressed as:

[0090]

[0091] wherein, u i and k i are a pair of their random parameters.

[0092] B. Random subspace sampling. Assume that F = {f1,..., f d} is the feature set in the dataset B, where f i represents the i-th feature. The random subspace is a set of a certain number of features randomly sampled from the original feature set. Randomly sample M times, and M random subspaces can be obtained, which are expressed as F1,..., Fm. Combine the generated random subspaces with the random kernel similarity metrics to obtain M random metric-subspace pairs, which are expressed as:

[0093]

[0094] If it is necessary to calculate b i and b in the m-th subspacej The similarity between them is first calculated according to the m-th metric-subspace pair Map b i and b j into the subspace associated with the component data set B m Then, use the randomly selected parameters u m and k m to calculate their kernel similarity. One subspace data set can obtain one similarity matrix. Therefore, M similarity matrices regarding the metric-subspace pairs can be obtained as follows:

[0095] S = {S (1) , S (2) , S (m) , …, S (M)}, (12)

[0096] where

[0097]

[0098]

[0099] After constructing M similarity matrices using different metric-subspace pairs, spectral clustering is used to generate the base clustering members. Specifically, for the m-th similarity matrix S (m) , each data sample is regarded as a graph node to construct a similarity graph as follows:

[0100] G (m) = (V, E (m) ), (15)

[0101] where V = B is the node set and E (m) is the edge set. The weight of the edge is determined by the similarity matrix S (m) . Assume K m represents the number of clusters among them. The goal of spectral clustering is to partition the graph G (m) into K m disjoint subsets. To this end, a normalized Laplacian matrix is constructed as follows:

[0102]

[0103] where the degree matrix is a diagonal matrix, and its (i, i)-th term is defined as the sum of the i-th row of S (m) . Calculate the first K m eigenvalues and the corresponding eigenvectors of L m , and arrange the eigenvectors to form a new matrix Then, by performing L1 normalization on U m , the matrix Take each row of T m as a data point and perform K-means clustering to obtain the m-th base cluster based on the similarity matrix S (m) . The m-th base cluster is represented as:

[0104]

[0105] Finally, based on different similarity matrices, M base clusters can be generated. Merge the generated base clusters with the base clusters generated by DSC into a matrix. Then, use the method of locally weighted ensemble clustering to obtain the Locally Weighted Co-Association (LWCA).

[0106] Specifically, given a cluster G i ∈π z , the entropy of G i with respect to the base cluster π z (z = 1, 2,..., Z) is defined as

[0107]

[0108]

[0109] where n z represents the number of clusters in π z ; denotes the j-th cluster in π z ; the symbol ∩ refers to the intersection of two clusters; |G i | represents the number of data points in G i . According to formulas (18) and (19), the entropy of G i with respect to the set Π of the entire base cluster is defined as

[0110]

[0111] where Z represents the number of base clusters in Π.

[0112] To measure the effectiveness of each cluster in all base clusters, LWEC proposes the Ensemble-driven Cluster Index (ECI) based on the entropy of each cluster with respect to the set of the entire base cluster. Given a set Π containing Z base clusters, the ECI of cluster G i is defined as

[0113]

[0114] Among them, θ represents the parameter used to adjust the influence of entropy on ECI. According to formula (21), it can be seen that the smaller the entropy of the cluster, the larger the corresponding ECI value. Based on ECI, an LWCA matrix is constructed to reflect the probability that two objects in multiple base clusters are divided into the same cluster. The LWCA matrix W is defined as

[0115]

[0116]

[0117]

[0118]

[0119] Among them, Gls z (o i ) represents the cluster to which the data object o i belongs.

[0120] Based on the LWCA matrix, this embodiment proposes a new consensus function CSDBS. This consensus function is designed based on the fact that adjacent bands have a high probability of being in the same cluster. Because the characteristics of hyperspectral data are that the bands are arranged in sequence, and each band has a strong correlation with adjacent bands within a certain range, while having a lower correlation with farther bands. At the same time, CSDBS also considers the number of bands in each cluster, and each time a division is made, it is made in a larger cluster. This takes into account both the correlation between adjacent bands and the size of each cluster, and is more in line with the needs of hyperspectral band selection. Therefore, the problem of hyperspectral band clustering is actually transformed into the problem of finding a series of segmentation points.

[0121] Specifically, CSDBS first regards all bands as a cluster G0, defined as

[0122] G0 = {b1, b2, b n , …, b N}, (26)

[0123] Among them, b n represents the nth band of the target hyperspectral image, N represents the number of bands. In order to find the appropriate segmentation points, G0 is divided into two new clusters. CSDBS uses the LWCA matrix W as the similarity matrix, which represents the similarity between any pair of bands. The similarity between adjacent bands b i and b i+1 is represented by Γ(b i , b i+1 ), and is defined as follows

[0124] Γ(b i , b i+1 ) = Wi,i+1 , (27)

[0125] Secondly, based on G0 and formula (27), the first splitting point t1 is determined by finding a pair of adjacent bands with the lowest similarity. t1 is obtained by the following formula:

[0126]

[0127] After the splitting point t1 is determined, two new clusters G1 = {b1, b2, …, b i} and G2 = {b i+1 , b i+2 , …, b v} can be obtained. Then, in order to find the remaining splitting points, the following steps are performed iteratively:

[0128] 1) Find the cluster Gj with the largest number of bands from the obtained clusters G j = {b m , b m+1 , ···, b v , b v+1 , ···, bn};

[0129] 2) Determine the splitting point t v by finding a pair of adjacent bands b v+1 and b j with the lowest similarity from Gj, and its formula is expressed as:

[0130]

[0131] 3) When the number of obtained clusters is equal to the number of classes the user wants to cluster into, the partitioning ends, thus obtaining the final consensus clustering result classes

[0132] To further describe the process of finding the splitting point in each iteration, an example is given to illustrate, as shown in Figures 2(a) and 2(b). Given G0 = {b1, b2, b3, b4, b5} and the LWCA matrix W shown in Figure 2(a), it can be found that the similarity between bands b2 and b3 is the lowest among all adjacent bands, and the first splitting point t1 = 2 can be determined. Therefore, G0 is divided into two clusters G1 = {b1, b2} and G2 = {b3, b4, b5}. Next, since the number of bands in G2 is greater than that in G1, the cluster G2 is selected for partitioning. According to formula (29), the second splitting point t2 = 3 is calculated. Therefore, G2 is divided into two clusters {b3} and {b4, b5}.

[0133] After obtaining the final consensus clustering result, representative bands can be selected from each cluster. Most traditional clustering-based methods select representative bands from the obtained clusters through certain criteria, such as information divergence, band noise estimation, and the distance from the band to the cluster center, etc. However, through research, it is found that these algorithms have deficiencies in selecting representative bands because the most representative band in each cluster may not be the most representative for the entire band. To address this problem, in this embodiment, the manifold ranking method is improved to rank the entire band set, and only one representative band is selected from each cluster.

[0134] Specifically, given y = (y1, y2, …, y n , …, y N ) is an indicator vector. If y n = 1, it means the corresponding band b n is in the representative band set Φ. Additionally, define the vector a = (a1, a2, …, a n , …, a N ) to indicate whether a representative band has been selected for the cluster in the final integrated clustering. If a n = 1, it means the cluster where the band b n is located has selected a representative band. Otherwise, it means the cluster has not selected a band yet. First, select the band b i with the maximum variance from the entire band set and put it into the initial representative band set Φ. Then, set the value of through formula (30).

[0135]

[0136] Meanwhile, set the value of in , where n = 1, 2, …, N. If b n and b i are in the same cluster, the value of is set to 1, otherwise it is set to 0. Formally expressed as:

[0137]

[0138] where Gls * (b i ) ∈ π * represents the cluster to which the band b i belongs.

[0139] Next, to select the remaining representative bands, the following process is iteratively executed:

[0140] (1) Calculate the ranking score vector according to the ranking function. The calculation formula is expressed as:

[0141] q = (D - αW) -1 y, (32)

[0142] D = diag{d 11 , d 22 , …, d nn , …, d NN}, (33)

[0143] where W is the LWCA matrix, and α is a balancing parameter.

[0144] (2) Select a representative band b s = B s: , where the index position of the band is determined by formula (34)

[0145]

[0146] (3) Put the selected frequency band into the representative frequency band set Φ.

[0147] (4) Update y (k) and a (k) respectively by formulas (35) and (36).

[0148]

[0149]

[0150] Finally, when the number of bands contained in Φ is equal to the selected number of bands L, the algorithm stops. Correspondingly, the bands contained in Φ are the selected representative band subset. After obtaining the selected band subset, the pixels of the hyperspectral image can then be classified using a support vector machine (SVM) or a K-nearest neighbor (KNN) classifier. Generally speaking, SVM has a better classification effect than KNN and is faster.

[0151] Based on the above analysis, the algorithm process of the proposed embodiment is given in Table 1.

[0152]

[0153] To verify the effectiveness of this embodiment, this embodiment is simulated on the experimental datasets Pavia University, Botswana, and Pavia Centre, using a support vector machine and a K-nearest neighbor classifier, and six typical band selection methods MVPCA, E-FDPC, NC-OC-IE, LWEA, MDEC, and DSC are selected for comparison.

[0154] Table 2 Comparison of the average performance of two classifiers on three datasets (black bold and italics indicate the best and suboptimal values, respectively)

[0155]

[0156] Table 2 shows the average classification performance comparison of the seven band selection methods on the three data sets. The best and second-best results are indicated in black bold and italics, respectively. The band range is [5,50]. It can be seen from Table 1 that this embodiment achieves the best performance on all three data sets. From the average performance, this embodiment improves the classification performance of hyperspectral images by improving the base clustering and consensus function, which proves the effectiveness of the method.

[0157] Table 3 Comparison of single-category OA in PaviaUniverisy dataset (black bold and italics indicate optimal and suboptimal values, respectively)

[0158]

[0159] Table 4 Comparison of single-category OA in Pavia Centre dataset (black bold and italics indicate optimal and suboptimal values, respectively)

[0160]

[0161] Table 5 Comparison of single-category classification accuracy of Botswana dataset (black bold and italics indicate the optimal and suboptimal values, respectively)

[0162]

[0163] Tables 3, 4 and 5 list the single-category classification performance on the three datasets using 30 selected bands. Among them, black bold and italics represent the best and second-best results, respectively. This embodiment obtains the best or second-best results in most cases, and is also better than other methods in terms of average classification accuracy.

[0164] Figures 3(a) - 3(b) , Figures 4(a) - 4(b) and Figures 5(a) - 5(b) The comparison of OA values ​​of seven methods using support vector machine and K nearest neighbor classifier on three data sets is shown. In general, the overall classification performance of this embodiment is better than that of other methods.

[0165] Figures 6(a) - 6(c) , Figures 7(a) - 7(c)Figs. 8(a) - 8(c) show the comparison between the classification maps and the ground truth classification information when 30 bands are selected using this embodiment. It can be seen from the figures that the differences between the classification maps of the three datasets and the ground truth classification maps are very subtle, and satisfactory classification results are achieved with fewer bands. In summary, this embodiment uses different classifiers on different datasets, and good classification results are obtained in both overall classification and single - category classification, verifying the effectiveness and stability of this embodiment.

[0166] Embodiment 2

[0167] This embodiment provides a hyperspectral image classification system based on integrated clustering band selection.

[0168] A hyperspectral image classification system based on integrated clustering band selection, comprising:

[0169] A data acquisition module, which is configured to: acquire a hyperspectral image dataset;

[0170] A base clustering module, which is configured to: based on the hyperspectral image dataset, adopt a base clustering generation strategy, combine spectral clustering to obtain corresponding clustering partitions, and obtain a base clustering set by setting different parameters;

[0171] A matrix generation module, which is configured to: based on the base clustering set, use the LWEC method to calculate the entropy and ECI values of each cluster, and generate an LWCA matrix;

[0172] A segmentation module, which is configured to: based on the LWCA matrix, adopt a consensus function to find the segmentation points, and obtain the clustering results according to the segmentation points;

[0173] A selection module, which is configured to: based on the clustering results, adopt a manifold ranking method to obtain representative bands;

[0174] A classification module, which is configured to: based on the representative bands, combine with the hyperspectral image dataset to classify the pixels of the hyperspectral image.

[0175] It should be noted here that the above - mentioned data acquisition module, base clustering module, matrix generation module, segmentation module, selection module, and classification module have the same examples and application scenarios as the steps in Embodiment 1, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above - mentioned modules, as part of the system, can be executed in a computer system such as a set of computer - executable instructions.

[0176] Embodiment 3

[0177] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the hyperspectral image classification method based on integrated clustering band selection as described in Embodiment 1 above.

[0178] Embodiment 4

[0179] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the hyperspectral image classification method based on hyperspectral band selection as described in Embodiment 1 above.

[0180] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0181] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0182] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0183] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide means for implementing the functions specified in one Figure 1Steps of functions specified in one or more processes and / or boxes Figure 1 Steps of functions specified in one or more boxes

[0184] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0185] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A hyperspectral image classification method based on integrated clustering band selection, characterized in that Including: Obtain a hyperspectral image dataset; Based on the hyperspectral image dataset, adopt a base clustering generation strategy, combine spectral clustering to obtain corresponding clustering partitions, and obtain a base clustering set by setting different parameters; Based on the base clustering set, use the LWEC method to calculate the entropy and ECI value of each cluster, and generate an LWCA matrix; Based on the LWCA matrix, adopt a consensus function to find the segmentation point, and obtain the clustering result according to the segmentation point; Based on the clustering result, adopt a manifold ranking method to obtain representative bands; Based on the representative bands, combine with the hyperspectral image dataset to classify the pixels of the hyperspectral image; The specific process of based on the LWCA matrix, adopting a consensus function to find the segmentation point and obtaining the clustering result according to the segmentation point includes: Regarding the hyperspectral image dataset as a cluster G0, where, ; The consensus function uses the LWCA matrix W as the similarity matrix, which represents the similarity between any pair of bands; for adjacent bands b i and b i+1 , the similarity is represented by Γ(b i , b i+1 ) and is defined as follows: Based on G0 and the similarity calculation formula of the above adjacent bands, the first segmentation point t1 is determined by finding a pair of adjacent bands with the lowest similarity: After the splitting point t1 is determined, two new clusters G1 = {b 1, b2, …, b i} and G2 = {b i+1, b i+2, …, b v} are obtained; To find the remaining segmentation points, iteratively execute the process of 1)-3): 1) Find the cluster G with the largest number of bands from the obtained clusters j ={b m ,b m+1 ,···,b v ,b v+1 ,···,bn}; 2) By finding a pair of adjacent bands b with the lowest similarity from Gj v and b v+1 , to determine the segmentation point t j , which is expressed by the formula: 3) The partitioning ends when the number of obtained clusters is equal to the number of classes to be clustered, and the final consensus clustering result classes are obtained. .

2. The hyperspectral image classification method based on integrated clustering band selection according to claim 1, wherein, The base clustering generation strategy includes: combining a deep convolutional autoencoder with subspace clustering, using spatial information and non-linear feature transformation to extract the interaction between spectral bands to obtain a first similarity matrix; based on the first similarity matrix, adopt spectral clustering to obtain corresponding clustering partitions, and obtain multiple first base clustering members by setting different parameters.

3. The hyperspectral image classification method based on integrated clustering band selection according to claim 1, characterized in that The base clustering generation strategy includes: using a high-dimensional data diversification measurement method, combining the hyperspectral image dataset with a random diversification measurement to obtain a random kernel similarity measurement; constructing a random subspace sampling of the hyperspectral image dataset; combining the random subspace sampling with the random kernel similarity measurement to generate a second similarity matrix; based on the second similarity matrix, adopt spectral clustering to obtain corresponding clustering partitions, and multiple second base clustering members can be obtained by setting different parameters.

4. The hyperspectral image classification method based on integrated clustering band selection according to claim 3, wherein Combine multiple first base clustering members and multiple second base clustering members to construct a base clustering set.

5. The hyperspectral image classification method based on integrated clustering band selection according to claim 1, characterized in that, The specific process of based on the clustering result, adopting a manifold ranking method to obtain representative bands includes: Given is an indication vector, if , it indicates that the corresponding band b n is in the representative band set ; Define vector Used to indicate whether the representative band has been selected for the cluster in the final integrated clustering. If , it means that the cluster where band b n is located has selected the representative band; otherwise, it means that the cluster has not selected the band yet.

6. The hyperspectral image classification method based on integrated clustering band selection according to claim 5, wherein The process of selecting representative bands is as follows: Select the band b with the maximum variance from the entire band set i and place it in the initial representative band set ; then set value Meanwhile, set the value in where ; if b n and b i are in the same cluster ， the value is set to 1, otherwise set to 0; To select the remaining representative bands, iteratively execute the following process: (1) Calculate the sorted score vector according to the sorting function , and the calculation formula is expressed as: where W is the LWCA matrix, , is a balance parameter; (2)Select a representative band , where the index position of the band is: (3) Put the selected bands into the representative band set ; (4) Update respectively by the formulas and ; When the number of bands included therein is equal to the selected number of bands L, the algorithm stops; correspondingly, the bands included therein are the selected representative subset of wavebands.

7. A system for implementing the hyperspectral image classification method based on integrated clustering band selection according to any one of claims 1-6, characterized in that, Including: A data acquisition module, which is configured to: obtain a hyperspectral image dataset; A base clustering module, which is configured to: based on the hyperspectral image dataset, adopt a base clustering generation strategy, combine spectral clustering to obtain corresponding clustering partitions, and obtain a base clustering set by setting different parameters; A matrix generation module, which is configured to: based on the base clustering set, use the LWEC method to calculate the entropy and ECI value of each cluster, and generate an LWCA matrix; A segmentation module, which is configured to: based on the LWCA matrix, adopt a consensus function to find the segmentation point, and obtain the clustering result according to the segmentation point; A selection module, which is configured to: based on the clustering result, adopt a manifold ranking method to obtain representative bands; A classification module, which is configured to: based on the representative bands, combine with the hyperspectral image dataset to classify the pixels of the hyperspectral image.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps in the hyperspectral image classification method based on integrated clustering band selection as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the hyperspectral image classification method based on integrated clustering band selection as described in any one of claims 1-6.

Images