CN103150580B - A kind of high spectrum image semisupervised classification method and device - Google Patents

Abstract

The present invention relates to sorting technique and the device of remote sensing hyperspectral image, described method, comprises the following steps: step 1: that carries out spectral modeling weighting to high spectrum image obtains cluster indicative character based on kernel fuzzy C mean cluster; Step 2: carry out support vector machine (SVM) semisupervised classification to high spectrum image and obtain the first classified image Image1, carries out support vector machine (SVM) semisupervised classification to cluster indicative character and obtains the second classified image Image2; Step 3: build cluster and SVM collaborative framework, carries out Cooperative Analysis by Image1 and Image2 classification results embedding cluster and SVM collaborative framework and obtains final hyperspectral classification image.Described device, comprises cluster module, sort module, Cooperative Analysis module; This high spectrum image semisupervised classification method and device are a kind of collaborative high spectrum image semisupervised classification method of feasible, high-precision cluster and SVM and device.

Description

Hyperspectral image semi-supervised classification method and device

Technical Field

The invention relates to a method and a device for classifying remote sensing hyperspectral images, in particular to a method and a device for classifying hyperspectral images in a clustering and Support Vector Machine (SVM) cooperation mode in a semi-supervised mode.

Background

Currently, the hyperspectral image classification algorithms commonly used can be classified into supervised and unsupervised algorithms. The traditional supervision and classification method comprises a spectrum angle filling method, a parallelepiped method, a maximum likelihood method, a minimum distance method and a mahalanobis distance method; conventional unsupervised classification methods include the IsoData method, the K-Means method, and the like. In addition to the above conventional methods, there are new classification methods such as neural networks, decision trees, SVMs, expert systems, and the like.

However, the hyperspectral image has many wave bands and large data volume, the acquisition cost of the class label samples is high, and the spatial distribution of the remote sensing ground object class is difficult to accurately estimate by a small number of class label samples, so that the traditional supervised classification method is difficult to obtain a good classification effect.

The semi-supervised method can combine a small amount of class label samples with a large amount of class label-free samples to improve the generalization capability of learning. In a hyperspectral image, if sample points are located in the same cluster, their class label information consistency may be greater. Clustering reflects the internal data structure of the hyperspectral images to a great extent, and a hyperspectral image semi-supervised classification method capable of effectively combining clustering information does not exist at present.

Disclosure of Invention

The invention aims to solve the technical problem of providing a feasible and high-precision clustering and SVM cooperative hyperspectral image semi-supervised classification method and device aiming at the defects of the existing hyperspectral image classification technology.

The technical scheme adopted by the invention for solving the technical problems is as follows: a semi-supervised classification method for hyperspectral images comprises the following steps:

step 1: performing spectral angle weighting on the hyperspectral image, and clustering based on kernel function fuzzy C mean to obtain clustering indication characteristics;

step 2: carrying out Support Vector Machine (SVM) semi-supervised classification on the hyperspectral images to obtain a first classified Image1, and carrying out Support Vector Machine (SVM) semi-supervised classification on the clustering indication features to obtain a second classified Image 2;

and step 3: and constructing a clustering and SVM cooperative framework, and embedding the classification results of Image1 and Image2 into the clustering and SVM cooperative framework for cooperative analysis to obtain a final hyperspectral classified Image.

Preferably, the step 1 further comprises the following steps:

step 1.1: initializing a clustering center, and setting spectral angle weights of a sample and the clustering center to obtain a spectral angle weight matrix;

step 1.2: suppose the hyperspectral sample X ═ { X ═ X_1,x₂,…，x_N},x₁={x₁₁,x₁₂，…，x_1pP is the number of wave bands; class label is Y ═ Y₁,y₂,…,y_NFor class label y_i∈Y，y_i∈ {1,2, …, C }, wherein C is the number of classes, K is the number of clusters, and the clustering center of the K-th class is v_kThe matrix V two { V }₁,v₂,…，v_KWangwu contains all cluster centers; hyperspectral image a certain sample x_iBelong to a certain class j, i =1,2, …, n, j ∈ [1,2, …, k]Each high-dimensional feature space sample isObtaining a sample according to the spectral angle weight matrixClustering centers for clustering class j kernel

Step 1.3: lagrange function L for defining spectral angle weighting based on kernel function fuzzy C mean clustering_KSFCMTo obtain the minimum formula L_KSFCMIs a membership function u_ij；

Step 1.4: according to membership function u_ijObtaining each sample x_iIs characteristic of cluster indication r_i。

Preferably, the step 2 further comprises the following steps:

step 2.1: carrying out Support Vector Machine (SVM) semi-supervised classification on the hyperspectral Image, obtaining two classifiers C1 and C2 by utilizing SVM training class label samples, respectively predicting the class-label-free samples by C1 and C2, adding the obtained class-label-free samples with high confidence coefficient and prediction labels thereof into a class-label sample training set until the sample classification is finished, and obtaining a first classification Image 1;

step 2.2: selecting a clustering center as a class label sample, and establishing two classifiers by utilizing SVM (support vector machine) to indicate a characteristic r of the clustering_iPerforming semi-supervised classification to obtain a second classified Image 2;

4. the hyperspectral image semi-supervised classification method according to claim 2 is characterized in that: the spectrum angle weight matrix determines the weight by using the size of the spectrum angle; the spectral response of n wave bands of each pixel on the hyperspectral image is used as a vector of an n-dimensional space, and the spectral angle can be expressed by an inverse cosine:

$\mathrm{\θ}={\cos }^{-1}\left\{\frac{{\mathrm{\Σ}}_{i=1}^n\mathrm{tr}}{{\left[{\mathrm{\Σ}}_{i=1}^n{t}^2\right]}^{1/2}{\left[{\mathrm{\Σ}}_{i=1}^n{r}^2\right]}^{1/2}}\right\}$

wherein n is the number of wave bands, t and r are respectively a clustering center spectrum and a certain sample spectrum,the spectral angle weight matrix is:

where p is the number of bands, then the kernel clustering center

The Lagrange function of the fuzzy C mean value clustering based on the kernel function weighted by the spectrum angle is as follows:

its minimum formula L_KSFCMThe membership function of (a) is:

$u_{\mathrm{ij}}=\frac{1}{{\mathrm{\Σ}}_{k=1}^c{\left(\frac{K(x_i,x_i)-2K(x_i,v_{\mathrm{ij}})+K(v_{\mathrm{ij}},v_{\mathrm{ij}})}{K(x_i,x_i)-2K(x_i,v_{\mathrm{ik}})+K(v_{\mathrm{ik}},v_{\mathrm{ik}})}\right)}^{\frac{1}{(m-1)}}}$

wherein,

$K(x_i,v_{\mathrm{ij}})={\mathrm{\Σ}}_{k=1}^n{\mathrm{SAW}}_{\mathrm{ik}}{u}_{\mathrm{kj}}^{m}K(x_i,x_k)/{\mathrm{\Σ}}_{k=1}^n{\mathrm{SAW}}_{\mathrm{ik}}{u}_{\mathrm{kj}}^m$

$K(v_{\mathrm{ij}},v_{\mathrm{ij}})={\mathrm{\Σ}}_{k=1}^n{\mathrm{\Σ}}_{t=1}^n\mathrm{SA}{W}_{\mathrm{ik}}{u}_{\mathrm{kj}}^m\mathrm{SA}{W}_{\mathrm{it}}{u}_{\mathrm{tj}}^{m}K(x_k,x_t)/{\left({\mathrm{\Σ}}_{k=1}^n\mathrm{SA}{W}_{\mathrm{ik}}{u}_{\mathrm{kj}}^m\right)}^2$

sample x_iIs characteristic of cluster indication r_iI.e. representing a sample x_iThe membership to each cluster center is:

r_i={u_1i，u_2i，…，u_ki}

wherein k is the number of clusters, r_iIs a sample x_iMembership vector to each cluster, thus, satisfying e^Tr_i=1, i ∈ {1, …, N }, e being the unit column vector, r_ik≥0,k∈{1,…,K}。

Preferably, the Support Vector Machine (SVM) semi-supervised classification further comprises the following steps:

step 2.1.1: is provided with a sample set X = { X₁,X_uIn which X is₁For class label sample sets, X_uInputting class label sample set X for class label-free sample set₁Class label free sample set X_u；

Step 2.1.2: SVM pair X₁Training to obtain classifiers C1 and C2, wherein the parameters of C1 are default values, and the parameters of C2 are the preferred parameters of the genetic algorithm;

step 2.1.3: x pair by classifier C1_uPrediction is carried out, a marking result p1 is obtained, and a classifier C2 is used for X_uPredicting and obtaining a marking result p 2;

step 2.1.4: comparing p1 and p2, selecting unlabeled sample with high confidence and its prediction label to be added into training set, that is, adding sample with consistent labeling result into training set X₁In, and update X₁And exiting the loop until the iteration termination condition is met.

Preferably, the clustering and SVM collaborative framework is constructed by a clustering loss function (cluster), a class consistency function (CaC), a Class Differentiation (CD), and a Sample Differentiation (SD);

the clustering loss function is:

$\mathrm{CuL}={\mathrm{\Σ}}_i^{1+u}{\mathrm{\Σ}}_k{r}_{\mathrm{ik}}^2{\left|\right|x_i-v_k\left|\right|}^2$

wherein l represents the number of class label samples, and u represents the number of non-class label samples;

the classification consistency function CaC includes CaCO and CaCC, where CaCO represents the classifier consistency for classifying the raw data, and CaCC represents the classifier consistency for classifying the clustering indication features:

CaC＝CaCO+CaCC

$\mathrm{CaCO}={\mathrm{\Σ}}_i^1\log p\left(y_i\right|x_i)$

$\mathrm{CaCC}={\mathrm{\Σ}}_i^1\log p\left(y_i\right|r_i)$

classification diversity function CD uses Jensen-Shannon divergence:

$\mathrm{CD}={\mathrm{\Σ}}_i^{1+u}(\frac{1}{2}\mathrm{KL}\left(p\left(c\right|x_i)\right|\left|p\left(c\right|r_i)\right)+\frac{1}{2}\mathrm{KL}\left(p\left(c\right|r_i)\right|\left|p\left(c\right|x_i)\right))$

wherein C = {1,2, …, C };

the sample difference function uses the euclidian distance:

$\mathrm{SD}={\mathrm{\Σ}}_{\mathrm{ij}}^{1+u}({\left|\right|\log p\left(c\right|x_i)-\log p\left(c\right|x_i)\left|\right|}^2+{\left|\right|\log p\left(c\right|r_i)-\log p\left(c\right|r_j)\left|\right|}^2)$

the clustering and SVM collaborative framework is as follows:

minS = CuL-lambda₁CaC ten lambda₂CD ten lambda₃SD

And solving the minimum value of the objective function S to ensure that the clustering loss is minimum, the classification consistency is highest, the classification difference is minimum and the sample difference is minimum, so that the optimal classification result is obtained.

The invention discloses a hyperspectral image semi-supervised classification device which comprises a clustering module, a classification module and a collaborative analysis module;

the clustering module is used for performing spectral angle weighted kernel function fuzzy C mean value based clustering on the hyperspectral image to obtain clustering indication characteristics;

the classification module is used for executing an SVM classifier twice on the hyperspectral Image to obtain a first classified Image1 and a second classified Image2, wherein Image1 is a classification result of the SVM classifier on the original hyperspectral Image, and Image2 is a classification result of clustering indication features of the SVM classifier;

the collaborative analysis module is used for collaboratively analyzing the first classified Image1 and the second classified Image2 obtained by the two SVM classifiers to construct a clustering and SVM collaborative framework so as to obtain a final hyperspectral classified Image;

the clustering module is connected with the classifying module in parallel and then connected with the collaborative analysis module in series.

The clustering and SVM cooperated hyperspectral image semi-supervised classification method and device have the following beneficial effects: the clustering indication features are generated by adopting semi-supervised learning and utilizing a clustering algorithm for classification, and the clustering and classification are cooperatively analyzed by combining respective advantages of clustering and classification, so that the problem of difficulty in selecting class label samples is avoided, the problem of wrong fraction caused by the fact that the class with the maximum membership degree is used as the final sample class in the traditional clustering algorithm and the problem that the number of support vectors is linearly increased along with the increase of training samples are solved, and a clustering loss function, a classification consistent function, classification difference and sample difference are provided, so that a target function is minimized, and an optimal classification result is obtained.

Drawings

FIG. 1: the invention relates to a flow chart of a hyperspectral image semi-supervised classification method.

FIG. 2: the invention discloses a structural schematic diagram of a hyperspectral image semi-supervised classification device.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

Referring to fig. 1, fig. 1 is a flowchart of a semi-supervised classification method for hyperspectral images, the method of the invention comprises the following steps:

step 1: performing spectral angle weighting on the hyperspectral image, and clustering based on kernel function fuzzy C mean to obtain clustering indication features, wherein the step comprises the following substeps:

step 1.1: initializing a clustering center, and setting spectral angle weights of the samples and the clustering center to obtain a spectral angle weight matrix. The spectrum angle weight matrix determines the weight value by using the size of the spectrum angle. The spectrum angle judges the approximation degree between the spectrums, the information of the spectrum dimension is fully utilized, the shape characteristic of the spectrums is emphasized, and the smaller the included angle is, the larger the similarity is, the larger the matrix weight is. The spectral response of n wave bands of each pixel on the hyperspectral image is used as a vector of an n-dimensional space, and the spectral angle can be expressed by an inverse cosine:

$\mathrm{\θ}={\cos }^{-1}\left\{\frac{{\mathrm{\Σ}}_{i=1}^n\mathrm{tr}}{{\left[{\mathrm{\Σ}}_{i=1}^n{t}^2\right]}^{1/2}{\left[{\mathrm{\Σ}}_{i=1}^n{r}^2\right]}^{1/2}}\right\}$

wherein n is the number of bands, and t and r are the clustering center spectrum and a certain sample spectrum respectively.The spectral angle quantity considers the spectral shape characteristics, and the influence of factors such as illumination, terrain and the like in the hyperspectral image classification process can be eliminated to a certain extent, so that the spectral angle weight matrix can fully utilize the spectral information of the sample. Two samples with small spectral angle theta belong to the same category, and such samples should have higher weight, i.e. the matrix spectral angle weight is large, and the similarity of samples with large spectral angle theta is small, so the matrix spectral angle weight is small.

Step 1.2: example hyperspectral sample X ═ X (X)₁，x₂，...，x_N}，x₁＝{x₁₁，x₁₂，...，x_1pP is the number of wave bands; class label Y ═ Y₁，y₂，...，y_NFor class label y_i∈Y，y_i∈ {1, 2.., C }, where C is the number of classes, K is the number of clusters, the K-th class of clustersCenter is v_kThe matrix V ═ V₁，v₂，...，v_KContains all cluster centers; hyperspectral image certain identity x_iBelong to a certain class j, i ═ 1,2]Each high-dimensional feature space sample isObtaining a sample according to the spectral angle weight matrixClustering centers for clustering class j kernel

Step 1.3: lagrange function L for defining spectral angle weighting based on kernel function fuzzy C mean clustering_KSFCMTo obtain the minimum formula L_KSFCMIs a membership function u_ij；

The Lagrange function of the fuzzy C mean value clustering based on the kernel function weighted by the spectrum angle is as follows:

its minimum formula L_KSFCMThe membership function of (a) is:

$u_{\mathrm{ij}}=\frac{1}{{\mathrm{\Σ}}_{k=1}^c{\left(\frac{K(x_i,x_i)-2K(x_i,v_{\mathrm{ij}})+K(v_{\mathrm{ij}},v_{\mathrm{ij}})}{K(x_i,x_i)-2K(x_i,v_{\mathrm{ik}})+K(v_{\mathrm{ik}},v_{\mathrm{ik}})}\right)}^{\frac{1}{(m-1)}}}$

wherein,

$K(x_i,v_{\mathrm{ij}})={\mathrm{\Σ}}_{k=1}^n{\mathrm{SAW}}_{\mathrm{ik}}{u}_{\mathrm{kj}}^{m}K(x_i,x_k)/{\mathrm{\Σ}}_{k=1}^n{\mathrm{SAW}}_{\mathrm{ik}}{u}_{\mathrm{kj}}^m$

$K(v_{\mathrm{ij}},v_{\mathrm{ij}})={\mathrm{\Σ}}_{k=1}^n{\mathrm{\Σ}}_{t=1}^n\mathrm{SA}{W}_{\mathrm{ik}}{u}_{\mathrm{kj}}^m\mathrm{SA}{W}_{\mathrm{it}}{u}_{\mathrm{tj}}^{m}K(x_k,x_t)/{\left({\mathrm{\Σ}}_{k=1}^n\mathrm{SA}{W}_{\mathrm{ik}}{u}_{\mathrm{kj}}^m\right)}^2$

step 1.4: according to membership function u_ijObtaining each sample x_iIs characteristic of cluster indication r_i;

Sample x_iIs characteristic of cluster indication r_iI.e. representing a sample x_iThe membership to each cluster center is:

r_i={u_1i,u_2i,…，u_ki}

wherein k is the number of clusters, r_iIs a sample x_iMembership vector to each cluster, thus, satisfying e^Tr_i=1, i ∈ {1, …, N }, e being the unit column vector, r_ikAnd the cluster indication feature describes the internal structural features of the data from the clustering point of view so as to establish connection between the clusters and the classifications, thereby realizing the advantage complementation of the clusters and the classifications.

Step 2: carrying out Support Vector Machine (SVM) semi-supervised classification on the hyperspectral images to obtain classified images Image1, carrying out SVM semi-supervised classification on the clustering indication features to obtain classified images Image2, and the steps comprise the following substeps:

step 2.1: carrying out vector machine SVM semi-supervised classification on the hyperspectral images, wherein a sample set X = { X in the embodiment₁，X_uIn which X is₁For class label sample sets, X_uInputting class label sample set X for class label-free sample set₁Class label free sample set X_uSVM pair X₁Training to obtain classifiers C1 and C2, wherein the parameters of C1 are default values, and the parameters of C2 are the preferred parameters of the genetic algorithm; x pair by classifier C1_uPredicting and obtaining a marking result p 1; x pair by classifier C2_uPredicting and obtaining a marking result p 2; comparing p1 and p2, selecting unlabeled sample with high confidence and its prediction label to be added into training set, that is, adding sample with consistent labeling result into training set X₁In, and update X₁If the iteration termination condition is met, exiting the loop to obtain a classified Image 1;

step 2.2: selecting a clustering center as a class label sample, and utilizing an SVM to indicate a characteristic r of the clustering_iPerforming semi-supervised classificationA classified Image2 is obtained.

And step 3: constructing a clustering and SVM collaborative framework, embedding the classification results of Image1 and Image2 into the clustering and SVM collaborative framework for collaborative analysis to obtain a final hyperspectral classified Image;

the clustering and SVM collaborative framework is constructed by a clustering loss function (CuL), a classification consistency function (CaC), a Classification Difference (CD) and a Sample Difference (SD);

the clustering loss function CuL is mainly used for judging clustering loss, and the smaller the CuL value is, the better the clustering result is:

$\mathrm{CuL}={\mathrm{\Σ}}_i^{1+u}{\mathrm{\Σ}}_k{r}_{\mathrm{ik}}^2{\left|\right|x_i-v_k\left|\right|}^2$

wherein l represents the number of class label samples, and u represents the number of non-class label samples; because the membership matrix is the representation of the membership of each sample to the clustering center, the clustering loss function can obtain the clustering center which minimizes the clustering loss according to the membership of each sample to each clustering center. In the traditional hyperspectral image classification by using a clustering algorithm, the clustering class with the maximum membership degree of each sample in a membership matrix is always assigned to the sample, so that false classification can be caused, and a clustering loss function can avoid the problem and reduce the false classification rate of the sample;

the classification consistency function CaC is mainly used for judging the classification loss of the classifier, CaCO represents the classifier consistency for classifying the raw data, and CaCC represents the classifier consistency for classifying the clustering indication features:

CaC=CaCO+CaCC

$\mathrm{CaCO}={\mathrm{\Σ}}_i^1\log p\left(y_i\right|x_i)$

$\mathrm{CaCC}={\mathrm{\Σ}}_i^1\log p\left(y_i\right|r_i)$

and the classification consistency function is used for judging the consistency of the classification result of the SVM classifier on the original hyperspectral data and the clustering indication feature and the original class label according to probability statistics and by utilizing class label information. The higher the CaC value is, the higher the consistency between the classification result and the class label is, and the better the classification effect is. A classification consistent function may be used to constrain the classification error samples so that the error rate is minimized;

the classification difference function CD is used for judging the difference of the classification results of the two classifiers, the classification effect is better as the class information is certain, the smaller the CD value between the two classifiers is, and the Jensen-Shannon divergence is adopted in the CD calculation:

$\mathrm{CD}={\mathrm{\Σ}}_{\mathrm{ij}}^{1+u}(\frac{1}{2}\mathrm{KL}\left(p\left(c\right|x_i)\right|\left|p\left(c\right|r_i)\right)+\frac{1}{2}\mathrm{KL}\left(p\left(c\right|r_i)\right|\left|p\left(c\right|x_i)\right))$

where C = {1,2, …, C }. Target classification results of the two classifiers are kept consistent, so that the error fraction can be reduced under the difference constraint condition, and the maximum accuracy of the classification results is ensured;

the sample difference function SD is used to determine the size of the sample difference within the category. The smaller the SD value of two samples in the same category is, the better the classification effect is. SD adopts euclidean distance:

$\mathrm{SD}={\mathrm{\Σ}}_{\mathrm{ij}}^{1+u}({\left|\right|\log p\left(c\right|x_i)-\log p\left(c\right|x_i)\left|\right|}^2+{\left|\right|\log p\left(c\right|r_i)-\log p\left(c\right|r_j)\left|\right|}^2)$

the classification criterion shows that the smaller the difference in the sample class is, the greater the similarity between the samples is, and the better the classification effect is;

the clustering and SVM collaborative framework is as follows:

monS=CuL-λ₁CaC+λ₂CD+λ₃SD

and solving the minimum value of the objective function S to ensure that the clustering loss is minimum, the classification consistency is highest, the classification difference is minimum and the sample difference is minimum, so that the optimal classification result is obtained. The clustering loss function is constraint of kernel function fuzzy C-means-based clustering for weighting the spectral angle, and ensures that clustering indication features obtained by a clustering algorithm can represent the internal structure of the hyperspectral data to the maximum extent; the classification consistency is that the classification result of the two classifiers is verified by utilizing class label samples; the classification difference is to limit the results of the two classifiers and reduce the error fraction of the samples; the sample difference function is an evaluation factor of the algorithm and is used as an evaluation index for judging the classification effect of the algorithm.

Referring to fig. 2, fig. 2 is a schematic structural diagram of the hyperspectral image semi-supervised classification device, which comprises a clustering module, a classification module and a collaborative analysis module. The clustering module is connected with the classifying module in parallel and then connected with the collaborative analysis module in series.

The clustering module is used for carrying out kernel function fuzzy C mean value clustering based on spectral angle weighting on the hyperspectral image, and introduces a spectral angle weight matrix on the basis of the traditional kernel function fuzzy C mean value clustering, so that the center of each kernel clustering is different along with different spectral information among samples. And obtaining the clustering indication characteristic of each sample by utilizing kernel function fuzzy C-means-based clustering weighted by spectral angles, namely the characteristic of the internal structure of the data described from the clustering angle. The classification module is used for establishing two semi-supervised SVM classifiers, one classifier is used for carrying out SVM semi-supervised classification on the hyperspectral Image to obtain a classification result Image1, and the other classifier is used for carrying out SVM semi-supervised classification on the clustering indication feature to obtain a classification result Image 2. And the collaborative analysis module is used for constructing a clustering and SVM collaborative framework according to the clustering loss function, the classification consistent function, the classification difference and the sample difference. The clustering loss function is constraint of kernel function fuzzy C-means-based clustering for weighting the spectral angle, and ensures that clustering indication features obtained by a clustering algorithm can represent the internal structure of the hyperspectral data to the maximum extent; the classification consistency is that the classification result of the two classifiers is verified by utilizing class label samples; the classification difference is to limit the results of the two classifiers, reduce the error fraction of the samples and take the prediction classification result with high confidence as the final classification result; the sample difference function is an evaluation factor of the algorithm and is used as an evaluation index for judging the classification effect of the algorithm.

In conclusion, the clustering and SVM cooperated hyperspectral image semi-supervised classification method and device provided by the invention have the advantages that a large number of unlabeled samples and a small number of class-labeled samples are utilized to better reflect the distribution characteristics of a sample space, so that a trained classifier has better popularization performance, aiming at the characteristic that hyperspectral data class-labeled samples are difficult to obtain. Meanwhile, the clustering and SVM collaborative hyperspectral image semi-supervised classification method and device carry out collaborative analysis by combining respective advantages of clustering and classification, and also avoid the problem of wrong fraction caused by the fact that the category with the maximum membership degree of a clustering algorithm is used as the category of the final sample and the problem that the support vector data linearly increases along with the increase of training samples.

The foregoing is a more detailed description of the invention, taken in conjunction with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments disclosed. It will be understood by those skilled in the art that various changes in detail may be effected therein without departing from the scope of the invention as defined by the appended claims.

Claims (4)

1. A semi-supervised classification method for hyperspectral images comprises the following steps:

step 1: performing spectral angle weighting on the hyperspectral image, and clustering based on kernel function fuzzy C mean to obtain clustering indication characteristics;

the step 1 further comprises the following steps:

step 1.1: initializing a clustering center, and setting spectral angle weights of a sample and the clustering center to obtain a spectral angle weight matrix;

step 1.2: suppose the hyperspectral sample X ═ { X ═ X₁,x₂,…,x_N}，x₁＝{x₁₁,x₁₂,…,x_1pP is the number of wave bands; class label is Y ═ Y₁，y₂,…,y_NFor class label y_i∈Y，y_i∈ {1,2, …, C }, wherein C is the number of classes, K is the number of clusters, and the clustering center of the K-th class is v_kThe matrix V ═ V₁,v₂,…,v_KContains all cluster centers; hyperspectral image a certain sample x_iBelong to a certain class of j, i ═ 1,2, …, n, j ∈ [1,2, …, k]Each high-dimensional feature space sample isObtaining a sample according to the spectral angle weight matrixClustering centers for clustering class j kernel

Step 1.3: lagrange function L for defining spectral angle weighting based on kernel function fuzzy C mean clustering_KSFCMTo obtain the minimum formula L_KSFCMIs a membership function u_ij；

Step 1.4: according to membership function u_ijObtaining each sample x_iIs characteristic of cluster indication r_i；

Step 2: carrying out Support Vector Machine (SVM) semi-supervised classification on the hyperspectral images to obtain a first classified Image1, and carrying out Support Vector Machine (SVM) semi-supervised classification on the clustering indication features to obtain a second classified Image 2;

the step 2 further comprises the following steps:

step 2.1: carrying out Support Vector Machine (SVM) semi-supervised classification on the hyperspectral Image, obtaining two classifiers C1 and C2 by utilizing SVM training class label samples, respectively predicting the class-label-free samples by C1 and C2, adding the obtained class-label-free samples with high confidence coefficient and prediction labels thereof into a class-label sample training set until the sample classification is finished, and obtaining a first classification Image 1;

step 2.2: selecting a clustering center as a class label sample, and establishing two classifiers by utilizing SVM (support vector machine) to indicate a characteristic r of the clustering_iPerforming semi-supervised classification to obtain a second classified Image 2;

and step 3: constructing a clustering and SVM collaborative framework, embedding the classification results of Image1 and Image2 into the clustering and SVM collaborative framework for collaborative analysis to obtain a final hyperspectral classified Image;

the clustering and SVM collaborative framework is constructed by a clustering loss function CuL (Cluster loss, CuL), a classification consistent function CaC (ClassConsistent, CaC), a classification difference function CD (CD) and a sample difference function SD (SD);

the cluster loss function CuL is:

$CuL=\underset{i}{\overset{l+u}{\Σ}}\underset{k}{\Σ}{r}_{ik}^2\left|\right|x_i-v_k|{|}^2$

wherein l represents the number of class label samples, and u represents the number of non-class label samples;

the classification consistency function CaC includes CaCO and CaCC, where CaCO represents the classifier consistency for classifying the raw data, and CaCC represents the classifier consistency for classifying the clustering indication features:

CaC＝CaCO+CaCC

$CaCO=\underset{i}{\overset{l}{\Σ}}\log p\left(y_i\right|x_i)$

$CaCC=\underset{i}{\overset{l}{\Σ}}\log p\left(y_i\right|r_i)$

classification diversity function CD uses Jensen-Shannon divergence:

$CD=\underset{i}{\overset{l+u}{\Σ}}(\frac{1}{2}KL\left(p\left(c|x_i\right)\left|\right|p\left(c|r_i\right)\right)+\frac{1}{2}KL(p\left(c|x_i\right)\left|\right|p\left(c|r_i\right)\left)\right)$

wherein C ═ {1,2, …, C };

the sample difference function SD uses the euclidian distance:

$SD={\mathrm{\Σ}}_{ij}^{l+u}\left(\right|\left|\log p\right(c|x_i)-\log p(c|x_j\left)\right|{|}^2+\left|\right|\log p\left(c|r_i\right)-\log p\left(c|r_j\right)\left|{|}^2\right)$

the clustering and SVM collaborative framework is as follows:

minS＝CuL-λ₁CaC+λ₂CD+λ₃SD

and solving the minimum value of the objective function S to ensure that the clustering loss is minimum, the classification consistency is highest, the classification difference is minimum and the sample difference is minimum, so that the optimal classification result is obtained.

2. The hyperspectral image semi-supervised classification method according to claim 1 is characterized in that: the spectrum angle weight matrix determines the weight by using the size of the spectrum angle; the spectral response of n wave bands of each pixel on the hyperspectral image is used as a vector of an n-dimensional space, and the spectral angle can be expressed by an inverse cosine:

$\mathrm{\θ}={\cos }^{-1}\left\{\frac{{\mathrm{\Σ}}_{i=1}^{n}tr}{{\[{\mathrm{\Σ}}_{i=1}^n{t}^2\]}^{1/2}{\[{\mathrm{\Σ}}_{i=1}^n{r}^2\]}^{1/2}}\right\}$

wherein n is the number of wave bands, t and r are respectively a clustering center spectrum and a certain sample spectrum,the spectral angle weight matrix is:

where p is the number of bands, then the kernel clustering centerComprises the following steps:

the Lagrange function of the fuzzy C mean value clustering based on the kernel function weighted by the spectrum angle is as follows:

its minimum formula L_KSFCMThe membership function of (a) is:

$u_{ij}=\frac{1}{{\mathrm{\Σ}}_{k=1}^c{\left(\frac{K(x_i,x_i)-2K(x_i,v_{ij})+K(v_{ij},v_{ij})}{K(x_i,x_i)-2K(x_i,v_{ik})+K(v_{ik},v_{ik})}\right)}^{\frac{1}{(m-1)}}}$

wherein,

$K(x_i,v_{ij})=\underset{k=1}{\overset{n}{\Σ}}{\mathrm{SAW}}_{ik}{u}_{kj}^{m}K(x_i,x_k)/\underset{k=1}{\overset{n}{\Σ}}{\mathrm{SAW}}_{ik}{u}_{kj}^m$

$K(v_{ij},v_{ij})=\underset{k=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{n}{\Σ}}{\mathrm{SAW}}_{ik}{u}_{kj}^m{\mathrm{SAW}}_{it}{u}_{tj}^{m}K(x_k,x_t)/{\left(\underset{k=1}{\overset{n}{\Σ}}{\mathrm{SAW}}_{ik}{u}_{kj}^m\right)}^2$

sample x_iIs characteristic of cluster indication r_iI.e. representing a sample x_iThe membership to each cluster center is:

r_i＝{u_1i,u_2i,…,u_ki}

wherein k is the number of clusters, r_iIs a sample x_iMembership vector to each cluster, thus, satisfying e^Tr_i1, i ∈ {1, …, N }, e being the unit column vector, r_ik≥0，k∈{1,…,K}。

3. The hyperspectral image semi-supervised classification method according to claim 1 is characterized in that: the Support Vector Machine (SVM) semi-supervised classification further comprises the following steps:

step 2.1.1: is provided with a sample set X ═ X₁,X_uIn which X is₁For class label sample sets, X_uInputting class label sample set X for class label-free sample set₁Class label free sample set X_u；

Step 2.1.2: SVM pair X₁Training to obtain classifiers C1 and C2, wherein the parameters of C1 are default values, and the parameters of C2 are the preferred parameters of the genetic algorithm;

step 2.1.3: x pair by classifier C1_uPrediction is carried out, a marking result p1 is obtained, and a classifier C2 is used for X_uPredicting and obtaining a marking result p 2;

step 2.1.4: comparing p1 and p2, selecting unlabeled sample with high confidence and its prediction label to be added into training set, that is, adding sample with consistent labeling result into training set X₁In, and update X₁And exiting the loop until the iteration termination condition is met.

4. A hyperspectral image semi-supervised classification device for realizing the hyperspectral image semi-supervised classification method of claim 1 comprises a clustering module, a classification module and a collaborative analysis module;

the clustering module is used for carrying out spectrum angle weighted kernel function fuzzy C mean value clustering on the hyperspectral image to obtain clustering indication characteristics;

the classification module is used for carrying out Support Vector Machine (SVM) semi-supervised classification on the hyperspectral images to obtain a first classified Image1, and carrying out Support Vector Machine (SVM) semi-supervised classification on the clustering indication features to obtain a second classified Image 2;

the collaborative analysis module is used for constructing a clustering and SVM collaborative framework, embedding the classification results of Image1 and Image2 into the clustering and SVM collaborative framework for collaborative analysis to obtain a final hyperspectral classified Image;

the clustering module is connected with the classifying module in parallel and then connected with the collaborative analysis module in series.