CN107292336A - A kind of Classification of Polarimetric SAR Image method based on DCGAN - Google Patents

Abstract

The invention discloses a polarimetric SAR image classification method based on DCGAN, comprising the following steps: 1) Obtain odd-order scattering coefficients, even-order scattering coefficients, and volume scattering coefficients, and then construct a pixel-based feature matrix F; The value of each element in the pixel-based feature matrix F is normalized to [0, 1], and the normalized result is recorded as the feature matrix F1; 3) Pass each element in the feature matrix F1 through its surrounding 64×64 image blocks are replaced to obtain the feature matrix F2 based on the image block; 4) Construct the feature matrix W1 of the unlabeled training data set D1 and the feature matrix W2 of the labeled training data set D2; 5) Construct the test data set T The feature matrix W3 of the superpixel clustering center; 6) Obtain the training network model DCGAN after training; 7) Construct the discriminant classification network model, and then classify the feature matrix W3 through the discriminant classification network model, this method can realize polarimetric SAR Image classification, and the classification accuracy is high.

Description

A Polarization SAR Image Classification Method Based on DCGAN

technical field

The invention belongs to the technical field of image processing, and relates to a polarimetric SAR image classification method based on DCGAN.

Background technique

Polarization SAR is a high-resolution active microwave remote sensing imaging radar, which has the advantages of all-weather, all-time, high resolution, side-view imaging, etc., and can obtain richer information on targets. The purpose of polarimetric SAR image classification is to use the polarization measurement data obtained by airborne or spaceborne polarimetric SAR sensors to determine the category to which each pixel belongs. It has a wide range of applications in agriculture, forestry, military, geology, hydrology, and ocean research and application value. The classic polarimetric SAR image classification methods are:

In 1992, Lee et al. researched that the multi-view polarization SAR image can be expressed in the form of polarization covariance matrix, and the matrix approximately obeys the complex Wishart distribution. On this basis, he proposed a simple and effective Wishart classification algorithm and used To classify forests, urban areas, oceans, sea ice and other types of land features.

In 1998, Lee et al. used the features extracted by the H/Alpha decomposition method to initially cluster the images and obtained 8 cluster centers; then they used the Wishart iterative classifier describing the multi-view covariance matrix to classify the images (referred to as H/Alpha -Wishart classifier).

In 2000, Pottier et al. proposed the H/Alpha/A-Wishart classifier. On the basis of the H/Alpha decomposition, the A feature was added to group the images into 16 categories, and then the Wishart iterative classification was performed on the images.

Due to its late start, the development of polarimetric SAR is still immature. Many core technologies, such as filtering technology, polarimetric target decomposition technology, and classification technology, need to be improved urgently. In particular, polarimetric SAR image classification still lacks efficient and reliable algorithms. The machine learning theories and methods of have not been applied in polarimetric SAR image classification. The classic polarization SAR image classification method is difficult to adapt to more and more polarization SAR data, so it is difficult to fully learn and utilize the distribution characteristics of polarization SAR data, it is difficult to extract good features, and it cannot achieve high classification accuracy .

Contents of the invention

The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art, and to provide a polarimetric SAR image classification method based on DCGAN, which can realize the classification of the polarimetric SAR image, and has high classification accuracy.

In order to achieve the above object, the polarimetric SAR image classification method based on DCGAN of the present invention comprises the following steps:

1) Obtain the polarized scattering matrix S, perform Pauli decomposition on the polarized scattering matrix S, and obtain odd-order scattering coefficients, even-order scattering coefficients, and volume scattering coefficients, and then use odd-order scattering coefficients, even-order scattering coefficients, and volume scattering coefficients as The three-dimensional image features of the polarimetric SAR image to be classified construct a pixel-based feature matrix F;

2) Normalize the values of each element in the pixel-based feature matrix F to [0, 1], and record the normalized result as the feature matrix F1;

3) Each element in the feature matrix F1 is replaced by a 64×64 image block around it to obtain a feature matrix F2 based on the image block;

4) Utilize the feature matrix F2 based on the image block to construct the feature matrix W1 of the unlabeled training data set D1 and the feature matrix W2 of the labeled training data set D2;

5) Use the feature matrix F2 based on the image block to construct the test data set T, and then use the SLIC superpixel algorithm to divide the superpixel block in the feature matrix F2 based on the pixel point, and obtain the clustering center of the superpixel block, and then in the image block based Construct the feature matrix W3 of the superpixel clustering center of the test data set T in the feature matrix F2 of ;

6) Train the training network model DCGAN through the unlabeled training data set D1 to obtain the trained training network model DCGAN;

7) Replace the binary classifier in the discriminator D in the trained training network model DCGAN with a softmax classifier, and then use the replaced discriminator D as the classification network model;

8) Input the feature matrix W2 of the labeled training data set D2 into the classification network model, and update the parameters of the softmax classifier, then update the parameters of the entire classification network model through the feature matrix W2 of the labeled training data set D2, and then pass The discriminative classification network model classifies the feature matrix W3 of the superpixel cluster center of the test data set T, and then marks the class labels of the test data set T to realize the polarimetric SAR image classification based on DCGAN.

In step 1), the Pauli decomposition is performed on the polarized scattering matrix S to obtain the odd-order scattering coefficient, even-order scattering coefficient and volume scattering coefficient as follows:

1a) Set the Pauli basis {S ₁ ,S ₂ ,S ₃ }, where,

Among them, S ₁ is odd scattering, S ₂ is even scattering, S ₃ is volume scattering;

1b) Defined by Pauli decomposition:

Among them, a is the odd-order scattering coefficient, b is the even-order scattering coefficient, and c is the volume scattering coefficient;

1c) Solve the formula (2) to get the odd-order scattering coefficient a, the even-order scattering coefficient b, and the volume scattering coefficient c, where,

The specific operation of constructing the feature matrix F based on pixels in step 1) is:

Set the characteristic matrix of size M1×M2×3, then assign the odd scattering coefficient a, the even scattering coefficient b and the volume scattering coefficient c to the characteristic matrix of size M1×M2×3, and obtain the pixel-based characteristic matrix F, where M1 is the length of the polarimetric SAR image to be classified, and M2 is the width of the polarimetric SAR image to be classified.

The specific operation of step 2) is: solving the maximum value max(F) of the feature matrix F based on the pixel point, and then dividing each element in the feature matrix F based on the pixel point by the maximum value max(F), Get the feature matrix F1.

In step 6), the specific operation of training the training network model DCGAN through the unlabeled training data set D1 is as follows:

6a) Let the generator G in the training network model DCGAN include the input layer a, the first deconvolution layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer and the output layer, where the input of the input layer a is a 100-dimensional noise vector; the number of feature maps of the first deconvolution layer is 512, and the filter size of the first deconvolution layer is 5; the feature of the second deconvolution layer The number of maps is 256, the filter size of the second deconvolution layer is 5; the number of feature maps of the third deconvolution layer is 128, the filter size of the third deconvolution layer is 5; the fourth deconvolution layer The number of feature maps in the product layer is 64, and the filter size of the fourth deconvolution layer is 5; the output layer outputs a pseudo-color map with a size of 64×64×3;

6b) Let the discriminator D in the training network model DCGAN include the input layer b, the first convolutional layer a, the second convolutional layer a, the third convolutional layer a, the fourth convolutional layer a, and the second convolutional layer connected in sequence. Classifier, wherein the number of feature maps of the input layer b is 3; the number of feature maps of the first convolutional layer a is 64, and the filter size of the first convolutional layer a is 5; the second layer of convolution The number of feature maps of layer a is 128, and the filter size of the second convolutional layer a is 5; the number of feature maps of the third convolutional layer a is 256, and the filter size of the third convolutional layer a is is 5; the number of feature maps of the fourth convolutional layer a is 512, and the filter size of the fourth convolutional layer a is 5; the binary classifier outputs a scalar;

6c) Input 100-dimensional uniform noise into the generator G of the training network model DCGAN, input the feature matrix W1 of the unlabeled training data set D1 and the output of the generator G into the discriminator D in the training network model DCGAN, and pass the training The generator G and the discriminator D in the network model DCGAN compete against each other for learning and training, and complete the training of the training network model DCGAN.

The discriminative classification network model includes an input layer c, a first convolutional layer b, a second convolutional layer b, a third convolutional layer b, a fourth convolutional layer b and a softmax classifier connected in sequence, wherein the softmax classifier The number of feature maps is 5.

The present invention has the following beneficial effects:

The DCGAN-based polarization SAR image classification method of the present invention is in specific operation, by training the training network model DCGAN, then reusing the discriminator D in the training network model DCGAN, and replacing the binary classifier with a softmax classifier Construct a discriminant classification network model, and realize the classification of polarimetric SAR images through the discriminant classification network model. Compared with other deep learning feature extraction methods, the present invention can also realize the classification of polarimetric SAR images without heuristic loss function, thus effectively Improving the classification accuracy of polarimetric SAR images. In addition, it should be noted that, by training the training network model DCGAN in the present invention, the training network model DCGAN can learn the distribution characteristics of the data from a large number of unlabeled data samples, making it possible to still It can achieve high classification accuracy and overcome the problems of fewer labeled samples and poor classification accuracy in the traditional polarization SAR image classification process.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is the manual marking diagram of the image to be classified in the present invention;

Fig. 3 is a diagram of classification results of images to be classified using the present invention.

detailed description

The present invention is described in further detail below in conjunction with accompanying drawing:

With reference to Fig. 1, the polarization SAR image classification method based on DCGAN of the present invention comprises the following steps:

1) Obtain the polarized scattering matrix S, perform Pauli decomposition on the polarized scattering matrix S, and obtain odd-order scattering coefficients, even-order scattering coefficients, and volume scattering coefficients, and then use odd-order scattering coefficients, even-order scattering coefficients, and volume scattering coefficients as The three-dimensional image features of the polarimetric SAR image to be classified construct a pixel-based feature matrix F;

Among them, the polarization SAR image to be classified in this embodiment is selected from April 2008, the San Francisco Bay full polarization data image of San Francisco Bay, the resolution is 50m, the image is L-band, the image size is 1800*1380, and all pixels have a total of Divided into 5 categories.

In step 1), the Pauli decomposition is performed on the polarized scattering matrix S to obtain the odd-order scattering coefficient, even-order scattering coefficient and volume scattering coefficient as follows:

1a) Set the Pauli basis {S ₁ ,S ₂ ,S ₃ }, where,

Among them, S ₁ is odd scattering, S ₂ is even scattering, S ₃ is volume scattering;

1b) Defined by Pauli decomposition:

Among them, a is the odd-order scattering coefficient, b is the even-order scattering coefficient, and c is the volume scattering coefficient;

1c) Solve the formula (2) to get the odd-order scattering coefficient a, the even-order scattering coefficient b, and the volume scattering coefficient c, where,

The specific operation of constructing the feature matrix F based on pixels in step 1) is:

Set the characteristic matrix of size M1×M2×3, then assign the odd scattering coefficient a, the even scattering coefficient b and the volume scattering coefficient c to the characteristic matrix of size M1×M2×3, and obtain the pixel-based characteristic matrix F, where M1 is the length of the polarimetric SAR image to be classified, and M2 is the width of the polarimetric SAR image to be classified.

2) Normalize the values of each element in the pixel-based feature matrix F to [0, 1], and record the normalized result as the feature matrix F1;

Among them, the commonly used normalization methods include feature linear scaling, feature standardization, and feature whitening. The specific operation of step 2) in this embodiment is: solve the maximum value max(F) of the feature matrix F based on the pixel point, and then divide each element in the feature matrix F based on the pixel point by the maximum value max (F), get the feature matrix F1.

3) Each element in the feature matrix F1 is replaced by a 64×64 image block around it to obtain a feature matrix F2 based on the image block;

4) Utilize the feature matrix F2 based on the image block to construct the feature matrix W1 of the unlabeled training data set D1 and the feature matrix W2 of the labeled training data set D2;

In the present embodiment, the concrete operation of step 4) is:

4a) Divide the polarimetric SAR image into 5 categories, record the position of the pixels corresponding to each category in the image to be classified, and generate 5 positions A1, A2, A3, A4, A5, where A1 corresponds to the position of the pixel of the first type of feature in the image to be classified, A2 corresponds to the position of the pixel of the second type of feature in the image to be classified, and A3 corresponds to the pixel of the third type of feature in the image to be classified The position in the image, A4 corresponds to the position of the pixel point of the fourth type of ground object in the image to be classified, and A5 corresponds to the position of the pixel point of the fifth type of ground object in the image to be classified;

4b) Randomly select 0.5% of the elements from the A1, A2, A3, A4, A5, and generate 5 positions B1, B2, B3, B4, B5 corresponding to the pixel points of different types of ground objects selected as the training data set , where B1 is the position of the pixels corresponding to the first type of ground objects selected as the training data set in the image to be classified, and B2 is the position of the pixels corresponding to the second type of ground objects selected as the training data set in the image to be classified The position in the image, B3 is the position of the pixel selected as the training data set corresponding to the third type of ground object in the image to be classified, and B4 is the position of the pixel selected as the training data set corresponding to the fourth type of ground object in The position in the image to be classified, B5 is the position in the image to be classified of the pixel selected as the training data set corresponding to the fifth type of ground object, and the elements in B1, B2, B3, B4, and B5 are combined to form the training The position L1 of all pixels in the data set in the image to be classified;

4c) Define the feature matrix W1 of the unlabeled training data set D1, randomly select 3.5% of the elements of the total pixels in the feature matrix F2 based on image blocks to form the feature matrix W1 of the unlabeled training data set D1;

4d) Define the feature matrix W2 of the labeled training data set D2, take the value at the corresponding position according to L1 in the feature matrix F2 based on the image block, and assign it to the feature matrix W2 of the training data set D2.

5) Use the feature matrix F2 based on the image block to construct the test data set T, and then use the SLIC superpixel algorithm to divide the superpixel block in the feature matrix F2 based on the pixel point, and obtain the clustering center of the superpixel block, and then in the image block based Construct the feature matrix W3 of the superpixel clustering center of the test data set T in the feature matrix F2 of ;

The concrete operation of step 5) is:

5a) Use the remaining 99.5% of the elements in A1, A2, A3, A4, and A5 described in step 4 to generate 5 positions C1, C2, C3, C4, and C5 corresponding to pixels of different types of ground objects selected as test data sets , where C1 is the position of the pixel selected as the test data set corresponding to the first type of ground object in the image to be classified, and C2 is the position of the pixel selected as the test data set corresponding to the second type of ground object in the image to be classified The position in the image, C3 is the position of the pixel selected as the test data set corresponding to the third type of ground object in the image to be classified, and C4 is the pixel point selected as the test data set corresponding to the fourth type of ground object in The position in the image to be classified, C5 is the position of the pixel selected as the test data set corresponding to the fifth type of ground object in the image to be classified, and the elements in C1, C2, C3, C4, and C5 are combined to form a test The position L2 of all pixels in the data set in the image to be classified;

5b) According to L2, the pixel point set at the corresponding position is taken as the test data set T;

5c) Initialize the seed points, according to the set number of superpixels K=80000, evenly distribute the seed points in the image, assuming that the picture has a total of N pixels, pre-divided into K superpixels of the same size, each The size of the superpixel is N/K, then the distance (step size) between adjacent seed points is approximately:

Calculate the size of each pixel block as (1800×1380)÷80000≈32, and the side length of the pixel block is about 6;

5d) Reselect the seed point in the n*n neighborhood of the seed point, the specific method is: calculate the gradient value of all the pixel points in the neighborhood, and move the seed point to the position with the smallest gradient in the neighborhood;

5e) Assign a class label (that is, which cluster center it belongs to) to each pixel point in the neighborhood around each seed point, and the search range is limited to 2S*2S;

5f) The distance measure includes color distance and space distance. For each searched pixel point, the distance between it and the seed point is calculated respectively. The distance calculation method is as follows:

Among them, d _c represents the color distance, d _c represents the spatial distance, N _s is the maximum spatial distance within the class, defined as N _s = S, applicable to each cluster, the maximum color distance N _c varies with different pictures, It also varies with different clusters. Since each pixel point will be searched by multiple seed points, each pixel point will have a distance from the surrounding seed points, and the seed point corresponding to the minimum value will be taken as the clustering value of the pixel point. class center;

5g) Iterate the above steps until the error converges, and re-divide the original image with a side length of 6, then there will be (1800×1380)÷36=69000 superpixel blocks, which is the final number of superpixel blocks, record these superpixels The position L3 of the center pixel of the block;

5h) Define the feature matrix W3 of the superpixel clustering center of the test data set T, take the value at the corresponding position according to L3 in the feature matrix F2 based on the image block, and assign it to the superpixel clustering center of the test data set T feature matrix W3;

6) Train the training network model DCGAN through the unlabeled training data set D1 to obtain the trained training network model DCGAN;

7) Replace the binary classifier in the discriminator D in the trained training network model DCGAN with a softmax classifier, and then use the replaced discriminator D as the classification network model;

8) Input the feature matrix W2 of the labeled training data set D2 into the classification network model, and update the parameters of the softmax classifier, then update the parameters of the entire classification network model through the feature matrix W2 of the labeled training data set D2, and then pass The discriminative classification network model classifies the feature matrix W3 of the superpixel cluster center of the test data set T, and then marks the class labels of the test data set T to realize the polarimetric SAR image classification based on DCGAN.

In step 6), the specific operation of training the training network model DCGAN through the unlabeled training data set D1 is as follows:

6a) Let the generator G in the training network model DCGAN include the input layer a, the first deconvolution layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer and the output layer, where the input of the input layer a is a 100-dimensional noise vector; the number of feature maps of the first deconvolution layer is 512, and the filter size of the first deconvolution layer is 5; the feature of the second deconvolution layer The number of maps is 256, the filter size of the second deconvolution layer is 5; the number of feature maps of the third deconvolution layer is 128, the filter size of the third deconvolution layer is 5; the fourth deconvolution layer The number of feature maps in the product layer is 64, and the filter size of the fourth deconvolution layer is 5; the output layer outputs a pseudo-color map with a size of 64×64×3;

6b) Let the discriminator D in the training network model DCGAN include the input layer b, the first convolutional layer a, the second convolutional layer a, the third convolutional layer a, the fourth convolutional layer a, and the second convolutional layer connected in sequence. Classifier, wherein the number of feature maps of the input layer b is 3; the number of feature maps of the first convolutional layer a is 64, and the filter size of the first convolutional layer a is 5; the second layer of convolution The number of feature maps of layer a is 128, and the filter size of the second convolutional layer a is 5; the number of feature maps of the third convolutional layer a is 256, and the filter size of the third convolutional layer a is is 5; the number of feature maps of the fourth convolutional layer a is 512, and the filter size of the fourth convolutional layer a is 5; the binary classifier outputs a scalar;

6c) Input 100-dimensional uniform noise into the generator G of the training network model DCGAN, input the feature matrix W1 of the unlabeled training data set D1 and the output of the generator G into the discriminator D in the training network model DCGAN, and pass the training The generator G and the discriminator D in the network model DCGAN compete against each other for learning and training, and complete the training of the training network model DCGAN.

The discriminative classification network model includes an input layer c, a first convolutional layer b, a second convolutional layer b, a third convolutional layer b, a fourth convolutional layer b and a softmax classifier connected in sequence, wherein the softmax classifier The number of feature maps is 5;

The specific operation of the softmax classifier for training is:

The feature matrix W2 of the labeled training data set D2 is used as the input of the discriminant classification network model, and the category of each pixel in the labeled training data set D2 is used as the output of the discriminative classification network model, and the softmax classifier is trained. By solving the above categories and The errors between the manually marked correct categories are propagated backwards, and only the parameters of the softmax classifier are updated to obtain a trained softmax classifier. The manually marked correct class labels are shown in Figure 2.

In step 8), the specific operation of classifying the feature matrix W3 of the superpixel cluster center of the test data set T through the discriminant classification network model is as follows:

The feature matrix W3 of the superpixel clustering center of the test data set T is used as the input of the trained discriminant classification network model, and the output of the trained discriminant classification network model is the classification category of the superpixel cluster center of the test data set T, Then, each pixel in the test set T is marked as the category of the cluster center of the superpixel block where the pixel is located, so as to complete the classification of the test set.

Simulation

Simulation condition: hardware platform is: HP Z840; Software platform is: TensorFlow; Simulation content and result: carry out experiment under above-mentioned simulation condition with the inventive method, at first select the unlabeled sample of 3.5% (accounting for all samples total number 1800*1380 3.5%, a block of about 8w pixels) for unsupervised learning, and then randomly select 0.5% of labeled samples from each category of polarized SAR data (accounting for 0.5% of the total number of labeled samples), and then train softmax The classification network is fine-tuned, and the remaining marked pixels are used as test samples, and the classification results shown in Figure 3 are obtained, which are divided into 5 types of ground objects.

It can be seen from Figure 3 that the regional consistency of the classification results is good, the edges of different regions are clearly identifiable, and the detailed information is maintained, and the noise in the classification result map is relatively small.

The test data set classification accuracy of the present invention and convolutional neural network CNN is compared, and the result is as shown in table 1:

Table 1

Classification convolutional neural network this invention Category 1 (%) 99.9849 99.9913 Category 2 (%) 96.9208 98.5980 Category 3 (%) 89.4749 98.5554 Category 4 (%) 98.6352 99.6494 Category 5(%) 98.4097 99.0883 total accuracy 97.254 99.4346

Reduce the labeled samples to 0.2%, and compare it with the classification accuracy of the test data set of the convolutional neural network CNN. The classification accuracy is shown in Table 2 below:

Table 2

Classification convolutional neural network this invention Category 1 (%) 99.0036 99.9317 Category 2 (%) 92.6015 97.6998 Category 3 (%) 92.8231 98.8479 Category 4 (%) 93.2422 97.8111 Category 5(%) 96.0703 97.3789 total accuracy 95.9239 98.9805

It can be seen from Table 1 and Table 2 that under the conditions of 0.5% and 0.2% of labeled samples, the classification accuracy of each category of the test data set of the present invention is higher than that of the convolutional neural network, which can improve the classification accuracy.

In summary, the present invention uses DCGAN for feature extraction, which can learn data distribution characteristics from a large amount of unlabeled data, and has a good feature representation ability, even if a small number of labeled samples are used to classify polarimetric SAR data. High classification accuracy.

Claims (6)

1. a kind of Classification of Polarimetric SAR Image method based on DCGAN, it is characterised in that comprise the following steps：

1) polarization scattering matrix S is obtained, Pauli decomposition is carried out to polarization scattering matrix S, odd scattering coefficient, even scattering is obtained Coefficient and volume scattering coefficient, then it regard odd scattering coefficient, even scattering coefficient and volume scattering coefficient as polarization SAR figure to be sorted Eigenmatrix F of the 3-D view feature construction of picture based on pixel；

2) each element value in the eigenmatrix F based on pixel is normalized in [0,1], and normalized result is denoted as Eigenmatrix F1；

3) each element in eigenmatrix F1 is replaced by around it 64 × 64 image block, obtained based on image block Eigenmatrix F2；

4) construct the eigenmatrix W1 without label training dataset D1 using the eigenmatrix F2 based on image block and have label instruction Practice data set D2 eigenmatrix W2；

5) the eigenmatrix F2 construction test data set T based on image block, then the profit in the eigenmatrix F based on pixel are utilized With SLIC super-pixel algorithm partition super-pixel block, the cluster centre of super-pixel block is obtained, then in the eigenmatrix based on image block The eigenmatrix W3 of test data set T super-pixel cluster centre is constructed in F2；

6) by being trained without label training dataset D1 to training network model DCGAN, the training network after being trained Model DCGAN；

7) two graders in arbiter D in the training network model DCGAN after training are replaced by softmax graders, then It regard the arbiter D after replacing as sorter network model；

8) the eigenmatrix W2 for having label training dataset D2 is input in sorter network model, and updates softmax classification The parameter of device, then by there is label training dataset D2 eigenmatrix W2 to update the parameter of whole sorter network model, then The eigenmatrix W3 of test data set T super-pixel cluster centre is classified by identification and classification network model, then marked Test data set T category, realizes the Classification of Polarimetric SAR Image based on DCGAN.

2. the Classification of Polarimetric SAR Image method according to claim 1 based on DCGAN, it is characterised in that step 1) in it is right Polarization scattering matrix S carries out Pauli decomposition, and the operation for obtaining odd scattering coefficient, even scattering coefficient and volume scattering coefficient is：

Pauli bases { S 1a) is set₁,S₂,S₃, wherein,

<mrow> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>S</mi> <mn>1</mn> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <msqrt> <mn>2</mn> </msqrt> </mfrac> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mn>0</mn> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>S</mi> <mn>2</mn> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <msqrt> <mn>2</mn> </msqrt> </mfrac> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mn>0</mn> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>S</mi> <mn>3</mn> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <msqrt> <mn>2</mn> </msqrt> </mfrac> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mn>0</mn> </mtd> </mtr> </mtable> </mfenced> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow>

Wherein, S₁Scattered for odd, S₂Scattered for even, S₃For volume scattering；

1b) decomposed and defined by Pauli：

<mrow> <mi>S</mi> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <msub> <mi>S</mi> <mrow> <mi>H</mi> <mi>H</mi> </mrow> </msub> </mtd> <mtd> <msub> <mi>S</mi> <mrow> <mi>H</mi> <mi>V</mi> </mrow> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>S</mi> <mrow> <mi>H</mi> <mi>V</mi> </mrow> </msub> </mtd> <mtd> <msub> <mi>S</mi> <mrow> <mi>V</mi> <mi>V</mi> </mrow> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>=</mo> <msub> <mi>aS</mi> <mn>1</mn> </msub> <mo>+</mo> <msub> <mi>bS</mi> <mn>2</mn> </msub> <mo>+</mo> <msub> <mi>cS</mi> <mn>3</mn> </msub> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2</mn> <mo>)</mo> </mrow> </mrow>

Wherein, a is odd scattering coefficient, and b is even scattering coefficient, and c is volume scattering coefficient；

Formula (2) 1c) is solved, odd scattering coefficient a, even scattering coefficient b and volume scattering coefficient c is obtained, wherein,

<mrow> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <mi>a</mi> <mo>=</mo> <mfrac> <mn>1</mn> <msqrt> <mn>2</mn> </msqrt> </mfrac> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mrow> <mi>H</mi> <mi>H</mi> </mrow> </msub> <mo>+</mo> <msub> <mi>S</mi> <mrow> <mi>V</mi> <mi>V</mi> </mrow> </msub> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mi>b</mi> <mo>=</mo> <mfrac> <mn>1</mn> <msqrt> <mn>2</mn> </msqrt> </mfrac> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mrow> <mi>H</mi> <mi>H</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>S</mi> <mrow> <mi>V</mi> <mi>V</mi> </mrow> </msub> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mi>c</mi> <mo>=</mo> <msqrt> <mn>2</mn> </msqrt> <msub> <mi>S</mi> <mrow> <mi>H</mi> <mi>V</mi> </mrow> </msub> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>3</mn> <mo>)</mo> </mrow> <mo>.</mo> </mrow>

3. the Classification of Polarimetric SAR Image method according to claim 1 based on DCGAN, it is characterised in that step 1) in structure The concrete operations for building the eigenmatrix F based on pixel are：

If size is the eigenmatrix of M1 × M2 × 3, then by odd scattering coefficient a, even scattering coefficient b and volume scattering coefficient c The eigenmatrix that size is M1 × M2 × 3 is assigned to, the eigenmatrix F based on pixel is obtained, wherein, M1 is polarization SAR to be sorted The length of image, M2 is the width of Polarimetric SAR Image to be sorted.

4. the Classification of Polarimetric SAR Image method according to claim 1 based on DCGAN, it is characterised in that step 2) tool Gymnastics conduct：The maximum max (F) of the eigenmatrix F based on pixel is solved, then by the eigenmatrix F based on pixel Each element is equal divided by the maximum max (F), obtain eigenmatrix F1.

5. the Classification of Polarimetric SAR Image method according to claim 1 based on DCGAN, it is characterised in that step 6) in lead to Cross no label training dataset D1 is to the training network model DCGAN concrete operations being trained：

6a) setting maker G in training network model DCGAN includes being sequentially connected the input layer a connect, the first warp lamination, the Two warp laminations, the 3rd warp lamination, the 4th warp lamination and output layer, wherein, input layer a input for 100 dimension noises to Amount；The Feature Mapping map number of first warp lamination is 512, and the filter size of the first warp lamination is 5；Second warp lamination Feature Mapping map number be 256, the filter size of the second warp lamination is 5；The Feature Mapping figure number of 3rd warp lamination Mesh is 128, and the filter size of the 3rd warp lamination is 5；The Feature Mapping map number of 4th warp lamination is 64, the 4th warp The filter size of lamination is 5；Output layer exports the pcolor of 64 × 64 × 3 sizes；

6b) setting the arbiter D in training network model DCGAN includes being sequentially connected the input layer b connect, the first convolutional layer a, second Convolutional layer a, the 3rd convolutional layer a, Volume Four lamination a and two graders, wherein, input layer b Feature Mapping map number is 3；The One layer of convolutional layer a Feature Mapping map number is 64, and first layer convolutional layer a filter size is 5；Second layer convolutional layer a's Feature Mapping map number is 128, and second layer convolutional layer a filter size is 5；Third layer convolutional layer a Feature Mapping figure number Mesh is 256, and third layer convolutional layer a filter size is 5；4th layer of convolutional layer a Feature Mapping map number is the 512, the 4th Layer convolutional layer a filter size is 5；Two graders export a scalar；

6c) the Uniform noise that input 100 is tieed up into training network model DCGAN maker G, will be without label training dataset D1 Eigenmatrix W1 and maker G output be input in training network model DCGAN in arbiter D, pass through training network mould Maker G and arbiter D in type DCGAN compete with one another for resisting learning training, complete training network model DCGAN training.

6. the Classification of Polarimetric SAR Image method according to claim 1 based on DCGAN, it is characterised in that step 7) in sentence Other sorter network model includes being sequentially connected the input layer c connect, the first convolutional layer b, the second convolutional layer b, the 3rd convolutional layer b, the Four convolutional layer b and softmax graders, wherein the Feature Mapping map number of softmax graders are 5.