CN104239900A - Polarized SAR image classification method based on K mean value and depth SVM - Google Patents

Abstract

The invention discloses a polarization SAR image classification method based on K-means and depth SVM, which mainly solves the problems of low classification accuracy and low classification efficiency existing in the existing polarization synthetic aperture radar SAR classification method. The steps that the present invention realizes are: (1) input image; (2) filter; (3) feature extraction; (4) set up wrong diversity set; (5) set up nearest neighbor sample set; (6) set up final training set; (7) ) to establish a deep support vector machine classifier; (8) classification; (9) calculation accuracy. The invention can realize accurate classification of polarization synthetic aperture radar SAR images, can effectively shorten the classification time of polarization synthetic aperture radar SAR images, and realize target recognition and tracking of polarization synthetic aperture radar SAR images.

Description

Polarization SAR image classification method based on K-means and depth SVM

technical field

The invention belongs to the technical field of image processing, and further relates to a polarization synthetic aperture radar (Synthetic Aperture Radar SAR) image classification based on K-means and deep support vector machine (Support Vector Machine SVM) in the technical field of machine learning and image classification method. The invention can be applied to the ground object classification of the polarimetric SAR image to realize target recognition and tracking. the

Background technique

Polarization SAR radar can obtain richer ground object information, and has extensive research in agriculture, forestry, ocean, military and other fields. There are many methods for polarimetric SAR image classification, which can be divided into supervised and unsupervised according to whether there is prior knowledge; according to the different classifiers used, they can be divided into statistics, neural networks, support vectors, decision trees, etc.; according to Whether to use spatial information can be divided into pixel-based and region-based. the

The patent "Polarization SAR image classification method based on SDIT and SVM" (patent application number: 201410089692.1, publication number: CN 103824084A) filed by Xidian University discloses a polarization SAR image classification method based on SDIT and SVM. In this method, the scattering features, polarization features, and texture features of polarimetric SAR data are combined into the feature combination SDIT of polarimetric SAR images, and then the polarimetric SAR images are classified by a support vector machine classifier. This method can not only avoid the interference between the polarization channels, but also maintain the polarization information and statistical correlation between the polarization channels, so that the edge of the image can be kept better. However, there are still shortcomings in that the extraction process of the combined feature SDIT of this method is complex, and the high-dimensional features will greatly increase the time complexity of training the support vector machine, and there are many misclassified points, and the classification accuracy is low. . the

The patent "Polarization SAR data classification method and system based on hybrid classifier" (patent application number: 201310310179, publication number: CN103366184A) filed by Wuhan University discloses a polarization SAR data classification method based on hybrid classifier. This method first obtains the initial polarization features by decomposing the polarization scattering matrix, and then uses a decision tree classifier to select the polarization features for classification from the initial polarization features. Finally, the selected polarization features , using a support vector machine classifier to classify polarimetric SAR data. Although this method combines the advantages of the decision tree classifier and the support vector machine classifier, the shortcomings of this method are that the classification accuracy is not much improved compared with the support vector machine classifier, the operation is complicated, and Only the scattering features are used, which is not enough to represent the actual ground objects. Therefore, there are many misclassified points in the classification of polarimetric SAR ground objects. the

Contents of the invention

The purpose of the present invention is to overcome the problem that the prior art can not achieve high classification accuracy and high classification efficiency at the same time for the classification of polarization SAR data, and propose a polarization SAR image classification method based on K-means clustering and deep SVM . Use the K-means clustering method to select the effective information in the initial training set as the final training set to train the SVM classifier, which can greatly reduce the training set and effectively save the time of training and prediction. Compared with other polarization SAR classification methods in the prior art, the present invention has high accuracy rate, strong anti-noise ability and low classification time complexity. the

The concrete steps that the present invention realizes comprise as follows:

(1) Input image:

Input an optional polarization synthetic aperture radar SAR image to be classified;

(2) Filtering:

The polarized refined Lee filtering method with a filter window size of 7*7 is used to filter the polarized synthetic aperture radar SAR image to be classified, remove the coherent speckle noise, and obtain the filtered polarized synthetic aperture radar SAR image;

(3) Feature extraction:

(3a) Extract the coherence matrix of the filtered polarization SAR image, wherein the coherence matrix is a matrix of 3*3*N, N represents the total number of pixels of the polarization SAR SAR, and each pixel is a 3 *3 matrix, the coherence matrix is constructed into a set of eigenvectors;

(3b) Within the range of [-1,1], normalize the value of the feature vector set to obtain the normalized feature vector set;

(4) Establish wrong diversity:

(4a) Randomly select five percent of the feature vectors from the normalized feature vector set to form an initial training set, and use the rest of the feature vector sets to form a test set;

(4b) Use the K-means clustering method to cluster the initial training set to obtain cluster labels, compare the real label and cluster label of each sample in the initial training set, and select the real label and cluster label of each sample in the initial training set For training samples with different class labels, these training samples are composed into a wrong diversity set;

(5) Establish the nearest neighbor sample set:

(5a) Using the Euclidean distance formula, calculate the Euclidean distance between each training sample in the misclassified set and each training sample in the initial training set;

(5b) Sort all Euclidean distance values from small to large;

(5c) sequentially select the training samples in the initial training set corresponding to the first 20 Euclidean distances, and add the selected training samples to the nearest neighbor sample set to obtain the nearest neighbor sample set;

(6) Establish the final training set:

(6a) Compare the real label and cluster label of each sample in the nearest neighbor sample set, and select the nearest neighbor sample whose real label and cluster label are different to form the final training set;

(6b) Compare the real label and the cluster label of each sample in the nearest neighbor sample set, select the first 20 nearest neighbor samples from each training sample in the wrong division set, and obtain the first 20 nearest neighbor samples;

(6c) Among the first 20 nearest neighbor samples selected, the 10 nearest neighbor samples whose cluster labels are equal to the real labels are added to the training samples in the misclassification set corresponding to the 10 nearest neighbor samples to the final training set, and the final Training set;

(7) Establish a deep support vector machine classifier:

(7a) Input the final training set into the support vector machine classifier for training, and obtain the support vectors of the training samples, the Lagrangian multipliers and labels corresponding to the support vectors;

(7b) Using the activation kernel function formula, calculate the activation value corresponding to each support vector;

(7c) input the activation value into the support vector machine classifier for training, and obtain the depth support vector machine classifier;

(8) Classification:

(8a) Use the deep support vector machine classifier to mark the polarization synthetic aperture radar SAR image to be classified, complete the classification, and obtain the classification result;

(8b) Statistical depth support vector machine classifier to classify the polarization synthetic aperture radar SAR image to be classified from the time used from the beginning of marking to the completion of classification, to obtain the classification time;

(9) Calculation accuracy:

Count the number of pixels in the polarimetric SAR SAR image to be classified that have the same category label as in the classification result, and calculate the percentage of the number of pixels with the same category label in the total number of pixels in the polarimetric SAR SAR image to be classified, and get classification accuracy. the

Compared with the prior art, the present invention has the following advantages:

First, because the present invention has adopted the K-means clustering method, effective information is selected from the initial training set as the final training set, thus the number of training sets can be greatly reduced, effectively saving the classification time of polarization synthetic aperture radar SAR , which overcomes the problem of high time complexity existing in the prior art, making the present invention more adaptable. the

Second, because the present invention adopts a deep support vector machine classifier, it overcomes the problem of more misclassification points caused by noise in the prior art, so that the classification accuracy of the present invention is better for polarimetric synthetic aperture radar SAR, and the Noise is more adaptable. the

Third, the present invention effectively combines the K-means clustering method and the deep support vector machine classifier, which improves the classification accuracy of the polarimetric synthetic aperture radar SAR and reduces the classification time of the polarimetric synthetic aperture radar SAR, making the present invention Inventions have a wider scope of application. the

Description of drawings

Fig. 1 is a flow chart of the present invention;

Figure 2 is the original image of the L-band multi-view polarization SAR data composite map obtained in Flevoland, Netherlands in 1989;

Figure 3 is the actual ground object marker map of the L-band multi-view polarization SAR image in the Flevoland, Netherlands area obtained in 1989;

Fig. 4 is a schematic diagram of the result of the present invention classifying the L-band multi-view polarization SAR data obtained in 1989 in the Flevoland, Netherlands region. the

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings. the

With reference to accompanying drawing 1, concrete implementation steps of the present invention are described in further detail:

Step 1, input image. the

Input an optional polarimetric SAR image to be classified. the

Step 2, filtering. the

The polarization refined Lee filtering method with a filter window size of 7*7 is used to filter the polarization synthetic aperture radar SAR image to be classified to remove the coherent speckle noise, and obtain the filtered polarization synthetic aperture radar SAR image. the

Step 3, feature extraction. the

Extract the coherence matrix of the filtered polarimetric synthetic aperture radar SAR image, where the coherent matrix is a 3*3*N matrix, and N represents the total number of pixels of the polarimetric synthetic aperture radar SAR, and each pixel is a 3*3 matrix, which constructs the coherence matrix as a set of eigenvectors. the

The coherence matrix is constructed as a matrix of N*9 eigenvector sets, where N represents the total number of pixels of the polarization synthetic aperture radar SAR, and each eigenvector set sample includes 9 elements, which are 3*3 eigenvector set samples The 3 elements on the diagonal of the coherence matrix of , the real part of the 3 elements of the upper triangular matrix of the eigenvector set sample coherence matrix, and the imaginary part of the 3 elements of the upper triangular matrix of the eigenvector set sample coherence matrix, a total of 9 elements. the

Step 4, establishing error diversity. the

From all the eigenvector sets, five percent of the eigenvectors are randomly selected to form the initial training set, and the rest of the eigenvector sets are formed into the test set. the

Use the K-means clustering method to cluster the initial training set to obtain cluster labels, compare the real label and cluster label of each sample in the initial training set, and select the real label and cluster label of each sample in the initial training set to be different. The same training samples, these training samples form a wrong diversity set. the

Step 5, establish the nearest neighbor sample set. the

Using the Euclidean distance formula, calculate the Euclidean distance between each training sample in the misclassified set and each training sample in the initial training set, sort the values of the Euclidean distance from small to large, and select the initial training set corresponding to the first 20 Euclidean distances in turn For training samples, add the selected training samples to the nearest neighbor sample set to obtain the nearest neighbor sample set. The Euclidean distance formula is as follows:

d(x,y)=||xy|| ₂

Among them, d(x, y) represents the Euclidean distance between two different training samples x and y in the nearest neighbor sample set, x and y respectively represent two different training samples in the nearest neighbor sample set, and ||·|| ₂ represents Two-norm operation.

Step 6, establish the final training set. the

Compare the real label of each sample in the nearest neighbor sample set with the cluster label, select the nearest neighbor samples whose real label and cluster label are different, and add them to the final sample set to form the final training set. the

Compare the real label and the cluster label of each sample in the nearest neighbor sample set, select the first 20 nearest neighbor samples from each training sample in the error division set, and obtain the top 20 nearest neighbor samples. Select 10 nearest neighbor samples whose cluster labels and real labels are equal among the first 20 nearest neighbor samples, and add the training samples in the misclassified set corresponding to the 10 nearest neighbor samples to the final training set to update the final training set. the

Step 7, establish a deep support vector machine classifier. the

Input the final training set into the support vector machine classifier for training, and obtain the support vectors of the training samples, the Lagrangian multipliers and labels corresponding to the support vectors. the

Among them, the support vector machine classifier uses the radial basis kernel function, and the formula of the radial basis kernel function is as follows:

$\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span>$

Among them, K(x, y) represents the radial basis kernel function of different training samples x and y in the final training set, ||xy|| ₂ represents the Euclidean distance of different training samples x and y in the final training set, and σ represents the radius To the width of the base kernel function, ||·|| ₂ means the two-norm operation.

Use the activation kernel function formula to calculate the activation value corresponding to each support vector. The activation kernel function formula is as follows:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span>$

Among them, h represents the support vector of the training sample and the activation value of the training sample, a represents the Lagrangian multiplier of the support vector of the training sample, t represents the label of the support vector of the training sample, and s represents the support vector of the training sample, x represents the training sample, and ||·|| ₂ represents the two-norm operation.

Input the obtained activation value into the support vector machine classifier for training, and obtain the depth support vector machine classifier;

Step 8, classification. the

Using the deep support vector machine classifier, the polarimetric synthetic aperture radar SAR image to be classified is marked, and the classification is completed, and the classification result of the polarimetric synthetic aperture radar SAR image is obtained. the

The time taken by the deep support vector machine classifier to be classified from the start of marking to the completion of classification is calculated to obtain the classification time of the polarimetric SAR image. the

Step 9, classification. the

Count the number of pixels in the polarimetric SAR SAR image to be classified that have the same category label as in the classification result, and calculate the percentage of the number of pixels with the same category label in the total number of pixels in the polarimetric SAR SAR image to be classified, and get Classification Accuracy of Polarimetric Synthetic Aperture Radar SAR Images. the

The present invention can be verified through the following simulation experiments. the

1. Simulation conditions:

In the simulation experiment of the present invention, a multi-view polarization SAR image obtained in 1989 in the Flevoland, Netherlands region as shown in Figure 2 was selected for the simulation experiment. The image size of the L-band multi-view polarimetric SAR image in the Flevoland, Netherlands area is 420 pixels × 380 pixels. the

The simulation experiment hardware platform of the present invention is: Intel Core2 Duo CPU i33.2GHZ, 3GB RAM, software platform: MATLAB R2012a. the

2. Simulation experiment results and analysis:

Figure 2 is the original image of the L-band multi-view polarimetric SAR data composite map obtained by the AIRSAR platform in Flevoland, Netherlands in 1989. Fig. 3 is the Flevoland obtained by the AIRSAR platform in 1989, the actual feature marker map in the Netherlands region, and Fig. 4 is the result of the present invention classifying the multi-view polarization SAR image of the L-band in the Flevoland obtained by the AIRSAR platform in 1989, the Netherlands region schematic diagram. the

After using the method of the present invention to classify the experimental image in Fig. 2, the schematic diagram of the result is shown in Fig. 4. It can be seen from Fig. 4 that the classification result obtained by the present invention is better, and the edges are relatively smooth and clearly identifiable. It can be seen that the method of the present invention is suitable for classifying ground objects on polarimetric synthetic aperture radar (SAR) images, and can obtain clear classification effects. the

The present invention and prior art depth support vector machine method are as shown in table 1 to the used classification time of Fig. 2 classification, as can be seen from table 1, the present invention uses the shortest time when polarimetric synthetic aperture radar SAR image classification, The effect is extremely obvious. the

Table 1 Classification time comparison table of two algorithms

the Classification time (s) Deep SVM 4398 this invention 1401

Method of the present invention and classic prior art support vector machine classification method and depth support vector machine method are as shown in table 2 to the accuracy rate of classification of Fig. 2, and in table 2, SVM represents support vector machine, depth SVM represents depth support vector Machine, categories 1 to 9 represent different ground object categories of the L-band multi-view polarization SAR image in the Flevoland, Netherlands area. the

Table 2 Comparison table of classification accuracy of three algorithms

the this invention SVM Deep SVM Category 1 0.9463 0.9515 0.9425 Category 2 0.9783 0.9802 0.9751 Category 3 0.9525 0.9713 0.9710 Category 4 0.9622 0.9358 0.9726 Category 5 0.9393 0.9382 0.9359 Category 6 0.9430 0.9322 0.9335 Category 7 0.9519 0.9327 0.9572 Category 8 0.8973 0.9115 0.8965 Category 9 0.8541 0.7878 0.8571 Average precision 0.9454 0.9361 0.9456

As can be seen from Table 2, the average classification accuracy of the present invention is higher than the classification accuracy of the classical support vector machine classification method, so, adopt the present invention, when classifying polarimetric synthetic aperture radar SAR images, classification efficiency and classification accuracy are both Improve to some extent, further verified the effect of the present invention. the

Claims (4)

1. The polarization SAR image classification method based on K mean value and depth SVM, comprises the following steps: (1) Input image: Input an optional polarization synthetic aperture radar SAR image to be classified; (2) Filtering: The polarization refined Lee filtering method with a filter window size of 7*7 is used to filter the polarization synthetic aperture radar SAR image to be classified, and the coherent speckle noise is removed to obtain the filtered polarization synthetic aperture radar SAR image; (3) Feature extraction: (3a) Extract the coherence matrix of the filtered polarization SAR image, wherein the coherence matrix is a matrix of 3*3*N, N represents the total number of pixels of the polarization SAR SAR, and each pixel is a 3 *3 matrix, the coherence matrix is constructed into a set of eigenvectors; (3b) Within the range of [-1,1], normalize the value of the feature vector set to obtain the normalized feature vector set; (4) Establish wrong diversity: (4a) Randomly select five percent of the feature vectors from the normalized feature vector set to form an initial training set; (4b) Use the K-means clustering method to cluster the initial training set to obtain cluster labels, compare the real label and cluster label of each sample in the initial training set, and select the real label and cluster label of each sample in the initial training set For training samples with different class labels, the training samples with different real labels and cluster labels are formed into a wrong diversity set; (5) Establish the nearest neighbor sample set: (5a) Using the Euclidean distance formula, calculate the Euclidean distance between each training sample in the misclassified set and each training sample in the initial training set; (5b) sort all the Euclidean distance values from small to large; (5c) sequentially select the training samples in the initial training set corresponding to the first 20 Euclidean distances, and add the selected training samples to the nearest neighbor sample set to obtain the nearest neighbor sample set; (6) Establish the final training set: (6a) Compare the real label and cluster label of each sample in the nearest neighbor sample set, and select the nearest neighbor sample whose real label and cluster label of the nearest neighbor sample are different to form the final training set; (6b) Compare the real label and the cluster label of each sample in the nearest neighbor sample set, select the first 20 nearest neighbor samples from each training sample in the wrong division set, and obtain the first 20 nearest neighbor samples; (6c) Among the first 20 nearest neighbor samples selected, the 10 nearest neighbor samples whose cluster labels are equal to the real labels are added to the training samples in the misclassification set corresponding to the 10 nearest neighbor samples to the final training set, and the final Training set; (7) Establish a deep support vector machine classifier: (7a) Input the final training set into the support vector machine classifier for training, and obtain the support vectors of the training samples, the Lagrangian multipliers and labels corresponding to the support vectors; (7b) Using the activation kernel function formula, calculate the activation value corresponding to each support vector; (7c) input the activation value into the support vector machine classifier for training, and obtain the deep support vector machine classifier; (8) Classification: (8a) Using a deep support vector machine classifier to mark the polarization synthetic aperture radar SAR image to be classified, complete the classification, and obtain the classification result; (8b) Statistical depth support vector machine classifier to be classified polarized synthetic aperture radar SAR image from the time used to complete the classification from the start of marking, to obtain the classification time; (9) Calculation accuracy: Count the number of pixels in the polarimetric SAR SAR image to be classified that have the same category label as in the classification result, and calculate the percentage of the number of pixels with the same category label in the total number of pixels in the polarimetric SAR SAR image to be classified, and get classification accuracy. 2. the polarization SAR image classification method based on K mean value and depth SVM according to claim 1, is characterized in that, step (3a) described coherent matrix is constructed into the matrix that feature vector set is N*9, wherein N represents the total number of pixels of polarimetric synthetic aperture radar SAR. Each eigenvector set sample includes 9 elements, which are the 3 elements on the diagonal of the coherence matrix of the eigenvector set sample 3*3, and the eigenvector set sample The real part of the 3 elements of the upper triangular matrix of the coherence matrix, the imaginary part of the 3 elements of the upper triangular matrix of the eigenvector set sample coherence matrix, a total of 9 elements. 3. the polarization SAR image classification method based on K mean value and depth SVM according to claim 1, is characterized in that, the Euclidean distance formula described in step (5) is as follows: d(x,y)=||xy|| ₂ Among them, d(x, y) represents the Euclidean distance between two different training samples x and y in the nearest neighbor sample set, x and y respectively represent two different training samples in the nearest neighbor sample set, and ||·|| ₂ represents Two-norm operation. 4. the polarization SAR image classification method based on K mean value and depth SVM according to claim 1, is characterized in that, the activation kernel function formula described in step (7b) is as follows: $\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span>$ Among them, h represents the support vector of the training sample and the activation value of the training sample, a represents the Lagrangian multiplier of the support vector of the training sample, t represents the label of the support vector of the training sample, and s represents the support vector of the training sample, x represents the training sample, and ||·|| ₂ represents the two-norm operation.