CN106067042A - Polarization SAR sorting technique based on semi-supervised degree of depth sparseness filtering network - Google Patents

Abstract

The invention discloses a kind of semi-supervised degree of depth learning method based on semi-supervised sparseness filtering.Solving conventional depth learning method parameter regulation complexity, the technical problem that nicety of grading is not high when having relatively low label data, its step includes: input polarimetric SAR image data to be sorted；Extract training sample and test sample；Seek Wishart neighbour's sample of training sample；Initialize the parameter of degree of depth sparseness filtering network；To degree of depth sparseness filtering network pre-training；Degree of depth sparseness filtering network is finely tuned；Class prediction is carried out to test sample；Export classification chart picture and the nicety of grading of Polarimetric SAR Image to be sorted.The present invention is by building novel degree of depth sparseness filtering network model, and adds the method for semi-supervised regular terms during pre-training, reduces the complexity of degree of depth learning network parameter regulation, improves the precision of Polarimetric SAR Image terrain classification.Can be used for the technical fields such as environmental monitoring, earth resources survey and military system.

Description

Polarization SAR classification method based on semi-supervised deep sparse filtering network

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a polarized SAR image ground feature classification method, in particular to a polarized SAR classification method based on a semi-supervised deep sparse filter network. The method can be used for environmental monitoring, earth resource surveying, military systems and the like.

Background

Machine Learning (ML) is a multi-domain cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. In the field of polarimetric SAR image classification, machine learning has been advanced with many breakthroughs, such as methods of Wishart Maximum Likelihood (WML), Support Vector Machines (SVM), and the like.

Most of the common machine learning methods use a method of manually extracting features, which is time-consuming and labor-consuming, and may not always obtain satisfactory features. Deep learning is a new field in machine learning research, and is a neural network simulating human brain to analyze and learn, and the neural network simulates the mechanism of human brain to interpret data. For polarized SAR image classification, the deep learning network can autonomously learn more abstract high-level representation attributes or characteristics from polarized SAR data, and the learned characteristics can be more effectively applied to the researches such as surface feature classification and environmental monitoring.

Whereas the existing deep learning model: the Stacked Autoencoder (SAE), the limiting boltzmann machine (RBM), the Deep Belief Network (DBN), the Convolutional Neural Network (CNN), etc. all require adjustment of many parameters. Such as learning rates (learning rates), momentums (momentums), sparsity penalties (sparsity penalties), etc., and the final determination of these parameters needs to be obtained by cross-validation, which takes a lot of time and effort. With the continuous development of the remote sensing field, the requirements of applications such as environmental monitoring, earth resource survey, military systems and the like on polarized SAR image processing are increased, and an ideal result is required to be obtained for the ground feature classification of the polarized SAR image.

Disclosure of Invention

The invention aims to provide a polarization SAR classification method based on a semi-supervised deep sparse filter network aiming at the defects of the prior art, and the accuracy of ground object classification is improved.

The invention relates to a polarization SAR classification method based on a semi-supervised deep sparse filter network, which is characterized by comprising the following steps of:

(1) inputting polarized SAR image data to be classified, namely a coherence matrix T of the polarized SAR image, obtaining a label matrix Y according to the ground feature distribution information of the polarized SAR image, wherein the distribution of the same ground feature is represented by the same class label, the ground feature distribution which can not determine the class is represented by 0 in the label matrix, and generating a sample matrix according to the coherence matrix T of the polarized SAR imageN is the total number of samples, x_iThe ith sample is represented.

(2) Extracting training samples and test samples, and randomly extracting L training samples and M test samples according to sample data X and a label matrix Y of the polarized SAR image, wherein L + M is N, randomly selecting 1% of samples in each class as the training samples according to the class information of all the samples, and the rest are the test samples.

(3) Obtaining Wishart neighbor samples of the training samples, and obtaining each training sample x from all sample data_iL) (i ═ 1,2.. L) for K Wishart neighbor samples x_j(j ═ 1,2.. K), so far, the basic data processing is complete.

(4) The method comprises the steps of starting to optimize a deep network, initializing basic parameters of the deep network, randomly initializing weight parameters W of the deep sparse filter network, setting the number of nodes of each layer of the deep sparse filter network, and determining the overall structure of the deep sparse filter network.

(5) Pre-training a deep sparse filter network, sending a training sample and a Wishart neighbor sample corresponding to the training sample into the deep sparse filter network for pre-training, adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, simultaneously adding a semi-supervised neighbor keeping regular term in the pre-training of each layer, jointly optimizing the weight of the network with a sparse filter, and primarily optimizing the weight of the deep sparse filter network.

(6) And (3) fine-tuning the deep sparse filter network, and fine-tuning the deep sparse filter network by utilizing the training samples and the label information thereof and combining a Softmax classifier, so as to further optimize the weight of the network, so that the network becomes more stable, and thus, the optimization of the deep sparse filter network is completed.

(7) Testing the performance of the deep sparse filter network, performing class prediction on the test samples, sending the test samples into the deep sparse filter network, and predicting class labels of the test samples by using a Softmax classifier to obtain the prediction class of each test sample.

(8) And (4) outputting a classification result graph and classification precision of the polarized SAR image to be classified, outputting a final classification result of the polarized SAR image to be classified according to the training sample and the test sample with the predicted class in the step (7), and calculating the precision of the classification.

The technical idea of the invention is as follows: a single-layer sparse filter is expanded into a novel deep sparse filter network, under the condition that a small number of marked samples are possessed, feature learning is carried out on polarized SAR data by combining a semi-supervision thought, ground object classification is realized through a classifier, and classification accuracy is improved.

The invention has the following advantages:

1. the method adopts a deep learning method, and utilizes the deep sparse filter network to learn the characteristics of the polarized SAR image autonomously, so that the complexity of ergonomics feature learning in the traditional method is avoided, and the deep sparse filter network can learn the more abstract and essential characteristics of the polarized SAR image, and the characteristics are more beneficial to ground feature classification of the polarized SAR image.

2. According to the method, the deep sparse filter network is obtained by expanding on the basis of the sparse filter, and the deep learning network has fewer parameters and stable performance, so that the defects that the parameters of the traditional deep learning network are complex and difficult to adjust are effectively overcome, and the condition that the classification result is not ideal due to improper parameter adjustment is properly avoided.

3. The method adopts a semi-supervised learning method, and adds a semi-supervised regular term in the pre-training process, thereby improving the problems of low classification accuracy of a classifier under the condition of less labeled samples and information waste caused by a large amount of unlabeled samples.

Drawings

FIG. 1 is a schematic flow chart of an implementation of the present invention;

fig. 2 is a graph of experimental results of a polarized SAR ground-object simulation image, in which fig. 2(a) is a Pauli exploded view of the polarized SAR simulation image, fig. 2(b) is a label graph of the simulation image, fig. 2(c) is a graph of classification results using a comparison method DNN, fig. 2(d) is a graph of classification results using a comparison method SAE, fig. 2(e) is a graph of classification results using a comparison method DSF, and fig. 2(f) is a graph of classification results using the method of the present invention.

Detailed Description

The invention is described in detail below with reference to the accompanying drawings

Example 1

Due to the development of remote sensing technology, the method is widely applied to the fields of environmental monitoring, earth resource surveying, military systems and the like, the requirement for polarized SAR image processing is continuously increased, deep learning has obvious advantages in a machine learning method, and a traditional deep learning network needs a large amount of parameter adjustment, consumes a large amount of time and can directly influence the performance of the deep learning network and a final classification result, so that the invention provides a polarized SAR classification method based on a semi-supervised deep sparse filter network, which is shown in figure 1 and comprises the following steps:

(1) inputting polarized SAR image data to be classified, namely a coherent matrix T of the polarized SAR image, obtaining a label matrix Y according to ground feature distribution information of the polarized SAR image, expressing the same ground feature in the label matrix by the same class label no matter how the same ground feature is distributed, expressing the ground feature distribution which can not determine the class in the label matrix by 0, and generating a sample matrix by the coherent matrix T of the polarized SAR imageN is the total number of samples, x_iThe ith sample is represented.

1a, taking the modulus values of 6 elements at the upper triangular position of a polarization coherent matrix T of a polarization SAR image as the original characteristic of each pixel point;

1b by MATLThe reshape function in the AB software converts the T matrix into a two-dimensional sample matrixX∈R^P×NP is 6 the dimension of the sample matrix, N is the total number of samples, each column represents one sample, which is the basic processing for polarising SAR data.

(2) Extracting training samples and test samples, and randomly extracting L training samples and M test samples according to sample data X and a label matrix Y of the polarized SAR image, wherein L + M is N, randomly selecting 1% of samples in each class as the training samples according to the class information of all the samples, and the rest are the test samples.

(3) Obtaining Wishart neighbor samples of the training samples, and obtaining each training sample x from all sample data_iL) (i ═ 1,2.. L) for K Wishart neighbor samples x_j(j ═ 1,2.. K), since the polarized SAR data themselves obey the Wishart distribution, two samples with smaller Wishart distances are more likely to belong to the same class, and thus, the basic data processing is completed.

(4) The method comprises the steps of starting to optimize the deep sparse filter network, initializing basic parameters of the deep sparse filter network, randomly initializing weight parameters W of the deep sparse filter network, and setting the number of nodes of each layer of the deep sparse filter network.

(5) Pre-training a deep sparse filter network, sending a training sample and a Wishart neighbor sample corresponding to the training sample into the deep sparse filter network for pre-training, adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, adding a semi-supervised neighbor keeping regular term in the pre-training of each layer, and optimizing the weight of the network together with sparse filtering. The regular term constraint of the training sample and the Wishart neighbor sample thereof can optimize a deep sparse filter network structure, so that the method is more favorable for carrying out ground feature classification on polarized SAR data which obey Wishart distribution, meanwhile, the use of a label-free sample is further enhanced in pre-training, and good ground feature classification precision can be obtained under the condition of a small amount of label samples. The sparse filtering has fewer parameters to be adjusted, the greedy algorithm is expanded into a deep sparse filtering network structure, the parameter adjustment is more advantageous, and the learned new sample characteristics for ground feature classification also have higher precision.

(6) And (3) fine-tuning the deep sparse filter network, and fine-tuning the deep sparse filter network by utilizing the training samples and the label information thereof and combining a Softmax classifier, so that the weight of the network is further optimized, and the network is more stable.

(7) And performing class prediction on the test samples, sending the test samples into a deep sparse filter network, and predicting class labels of the test samples by using a Softmax classifier to obtain the prediction class of each test sample.

(8) And (4) outputting a classification result graph and classification precision of the polarized SAR image to be classified, outputting a final classification result of the polarized SAR image to be classified according to the training sample and the test sample of which the class is predicted in the step (7), referring to a graph (f) in FIG. 2, and calculating the precision of the classification.

The deep learning network has less parameter adjustment, and effectively overcomes the defects that the traditional deep learning network has complex parameters and is difficult to adjust; the invention has stable performance and properly avoids the condition of undesirable classification result caused by improper parameter adjustment.

Example 2

The polarized SAR classification method based on the semi-supervised deep sparse filtering network is the same as that in the embodiment 1, and the Wishart neighbor sample of each training sample is obtained by the following steps of (3):

3a, training sample matrix isThe Wishart distance between the training sample and the other samples is found using the following formula:

d(x_i,x_j)＝ln((x_i)^-1x_j)+Tr((x_j)^-1x_i)-q(x＝1,2,...,L,j＝1,2,...,N)，

wherein, Tr () represents the trace of the matrix, and q is 3 for the radar with integrated transmission and reception due to reciprocity; for radars where transmission and reception are not integral, q is 4;

3b, utilizing the sort function in MATLAB to compare the Wishart distance d (x) obtained in the step 3a_i,x_j) Arranging according to the ascending order of absolute values, taking the first K samples, finding the first K samples corresponding to the first K samples as training samples x_iThe wishirt neighbor sample of (a) is noted:

example 3

The polarized SAR classification method based on the semi-supervised deep sparse filtering network is the same as the embodiment 1-2, and the parameters of the initialized deep network in the step (4) are as follows:

4a, the number of hidden layers of the deep sparse filter network is 3, and the number of nodes of each layer is respectively as follows: 25, 100, 50;

4b, initializing weight W of deep sparse filter network₁,W₁∈R^D×QD is the dimension of the input signal and Q is the number of nodes of the first hidden layer.

The deep sparse filter network has fewer parameters to be adjusted, the parameters are convenient to initialize, only simple regular term parameters need to be adjusted in the training process of the network, and the time consumption is less compared with other deep learning methods.

Example 4

The polarized SAR classification method based on the semi-supervised deep sparse filter network is the same as the embodiment 1-3, and the pre-training process of the deep sparse filter network in the step (5) is as follows:

5a, inputting a pre-training sample of the deep sparse filter network, and inputting a training sample x_i(i ═ 1,2.. L) and its corresponding K Wishart neighbor samplesInputting the pre-training samples into the deep network;

5b, learning new sample characteristics by utilizing a first hidden layer, and taking a traditional nonlinear sigmoid function as an activation function: phi (z) ═ 1+ exp (-z)^-1，x∈R^D×1If the input vector is, the output of the first hidden layer is:the output of the ith sample is denoted as h₁(x_i) As new sample features;

5c, optimizing the network weight W, wherein each layer of sparse filter network utilizes an LBFGS algorithm to solve the following objective function, so that the weight W of the network is optimized, and the objective function is as follows:

$L_1\left(W\right)=\underset{W}{\min mize}\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\left|\right|\frac{{\tilde{h}}_1\left(x_i\right)}{\left|\right|{\tilde{h}}_1\left(x_i\right)|{|}_2}|{|}_1+\frac{\mathrm{\λ}}{2}\underset{i,j=1}{\overset{n}{\mathrm{\Σ}}}{A}_{ij}||h_1\left(x_i\right)-h_1\left(x_j\right)|{|}^2,$

${\tilde{h}}_1\left(x_i\right)=\frac{h_1\left(x_i\right)}{\left|\right|h_1\left(x_i\right)|{|}_2},$

wherein in the objective functionIn order to be a sparse filtering objective function,the method for reference manifold learning optimizes the structure of the network by using the neighbor relation among samples for keeping a regular term for semi-supervised neighbor, wherein lambda is a parameter of the regular term,U(x_i) Representing a training sample x_iA corresponding Wishart neighbor sample set;

5d, training the rest hidden layers by a greedy algorithm, and enabling the new sample characteristics h learned in the step 5b₁Sending the new input to the next layer, using the greedy learning algorithm, using the output of the previous layer as the input of the next layer, and using the objective function L in 5c₁(W) optimizing the weight W of each hidden layer in turn until the last hidden layer is trained, sample x_iThe output through the last hidden layer can be represented as h₃(x_i)。

The method carries out deep network pre-training by combining sparse filtering and semi-supervised regularization terms, optimizes the weight parameters of the network to a proper range, further optimizes the structure of the network, and is more favorable for carrying out ground feature classification on the polarized SAR data samples obeying Wishart distribution.

Example 5

The polarized SAR classification method based on the semi-supervised deep sparse filter network is the same as the polarized SAR classification method in the embodiment 1-4, and the fine tuning of the deep sparse filter network in the step (6) has the following target functions:

$L_2\left(W\right)=\underset{W}{\min mize}\frac{1}{L}\underset{i=1}{\overset{L}{\Σ}}\left(\frac{1}{2}\right||y_i-h(x_i\left)\right|{|}^2)+\frac{\mathrm{\β}}{2}\underset{i,j}{\Σ}\left|\right|W_{ij}|{|}^2,$

wherein,is a term of a mean-square error,is a weight attenuation term aimed at reducing the magnitude of the weight and preventing overfitting, y_iRepresenting a training sample x_iCorresponding class label, h (x)_i) Is to train sample x_iFeature h learned through the entire deep network₃(x_i) Sending the output to a Softmax classifier, wherein β is 3e-3 which is a weight attenuation parameter;

solving the objective function L by using a gradient descent algorithm₂And (W), further optimizing the weight W of the sparse filter network, and realizing fine adjustment of the whole network, thereby constructing a deep sparse filter network, wherein the features learned by the network are more favorable for realizing the ground feature classification of the polarized SAR image.

A more detailed example is given below to further illustrate the invention:

example 6

The polarized SAR classification method based on the semi-supervised deep sparse filtering network is the same as the embodiment 1-5, and referring to FIG. 1, the specific implementation steps of the invention are as follows:

step 1, inputting polarized SAR image data to be classified, referring to fig. 2(a), inputting a coherence matrix T of the polarized SAR image, obtaining a label matrix Y according to the ground feature distribution information of the polarized SAR image, referring to fig. 2(b), wherein fig. 2(b) is an image directly generated by the label matrix Y, different color blocks in the image represent different ground features, the distribution of the same ground feature is represented by the same class label in the label matrix, the distribution of the ground feature which can not determine the class is represented by 0 in the label matrix, and generating a sample matrix according to the coherence matrix T of the polarized SAR imageN is the total number of samples, x_iThe ith sample is represented.

This example uses polarized SAR terrain simulation images. Because the polarized coherent matrix T of the polarized SAR data is a Hamilton semi-positive matrix, the module values of 6 elements at the upper triangular position of the polarized coherent matrix T with the dimension of 3 multiplied by 3 can be extracted as the original characteristics of each pixel point, the T matrix is converted into a two-dimensional sample matrix by utilizing a reshape function in MATLAB software, each column represents a sample, the simulation data comprises 18000 samples, and the dimension of each sample is 6 dimensions. Each sample corresponds to a pixel point on the polarized SAR image.

And 2, extracting training samples and test samples, and randomly extracting L training samples and M test samples from the sample data according to the type according to the sample data X and the label matrix Y of the polarized SAR image, wherein L + M is equal to N. And randomly extracting the sample data into training samples and testing samples according to the respective classes of the sample data and the proportion of 1:99, wherein the testing samples of each class account for 1 percent of the total number of the class.

Step 3, solving Wishart neighbor samples of the training samples, and solving each training sample x in all sample data_iL) (i ═ 1,2.. L) for K Wishart neighbor samples x_j(j＝1,2...K)。

3a for training sample x_i(i ═ 1,2.. L), finding the Wishart distance between it and other samples according to the modified Wishart distance formula: d (x)_i,x_j)＝ln((x_i)^-1x_j)+Tr((x_j)^-1x_i)-q(x＝1,2,...,L,j＝1,2,...,N)，

Wherein, Tr () represents the trace of the matrix, q is 3 for the radar with integrated transmission and reception, and q is 4 for the radar with non-integrated transmission and reception, the polarized SAR ground object simulation image data used in this example is obtained and generated by the radar system with integrated transmission and reception, and q is 3.

3b, utilizing the sort function in MATLAB to compare the Wishart distance d (x) obtained in the step 3a_i,x_j) Arranging according to the ascending order of absolute values, taking the first K samples, finding the first K samples corresponding to the first K samples as training samples x_iThe wishirt neighbor sample of (a) is noted:

and 4, initializing basic parameters of the deep sparse filter network, randomly initializing a weight parameter W of the deep sparse filter network, and setting the number of nodes of each layer of the deep sparse filter network.

Setting the number of three hidden layer nodes of the deep sparse filter network, wherein the number of the hidden layer nodes is 25, 100 and 50; initializing the weight matrix W of the network₁,W₁∈R^D×ND is the dimension of the input signal and N is the number of nodes of the first hidden layer.

The deep sparse filter network has fewer parameters and stable performance, effectively overcomes the defects of complex parameters and difficult adjustment of the traditional deep learning network, and properly avoids the condition of non-ideal classification results caused by improper parameter adjustment.

And 5, pre-training the deep sparse filter network.

5a, inputting a pre-training sample of the deep sparse filter network, and inputting a training sample x_iAnd K Wishart neighbor samples corresponding to the sameInputting the training samples into a deep network, wherein i is 1,2.

5b, learning new sample characteristics by utilizing a first hidden layer, and taking a traditional nonlinear sigmoid function as an activation function: phi (z) ═ 1+ exp (-z)^-1，x∈R^D×1If the input vector is, the output of the first hidden layer is:the output of the ith sample is denoted as h₁(x_i) As new sample features;

5c, optimizing the network weight W, wherein each layer of sparse filter network utilizes an LBFGS algorithm to solve the following objective function, so that the weight W of the network is optimized, and the objective function is as follows:

$L_1\left(W\right)=\underset{W}{\min mize}\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\left|\right|\frac{{\tilde{h}}_1\left(x_i\right)}{\left|\right|{\tilde{h}}_1\left(x_i\right)|{|}_2}|{|}_1+\frac{\mathrm{\λ}}{2}\underset{i,j=1}{\overset{n}{\mathrm{\Σ}}}{A}_{ij}||h_1\left(x_i\right)-h_1\left(x_j\right)|{|}^2,$

${\tilde{h}}_1\left(x_i\right)=\frac{h_1\left(x_i\right)}{\left|\right|h_1\left(x_i\right)|{|}_2},$

wherein in the objective functionIn order to be a sparse filtering objective function,the method for reference manifold learning optimizes the structure of the network by using the neighbor relation among samples for keeping a regular term for semi-supervised neighbor, wherein lambda is a parameter of the regular term,U(x_i) Representing a training sample x_iThe corresponding Wishart neighbor sample set.

5d, training the rest hidden layers by a greedy algorithm, and enabling the new sample characteristics h learned in the step 5b₁Sending the new input to the next layer, using the greedy learning algorithm, using the output of the previous layer as the input of the next layer, and using the objective function L in 5c₁(W) optimizing the weight W of each hidden layer in turn until the last hidden layer is trained, sample x_iThe output through the last hidden layer can be represented as h₃(x_i)。

The method carries out deep network pre-training by combining sparse filtering and semi-supervised regularization terms, optimizes the weight parameters of the network to a proper range, further optimizes the structure of the network, and is more favorable for carrying out ground feature classification on the polarized SAR data samples obeying Wishart distribution.

And 6, fine adjustment of the depth sparse filter network.

After the pre-training is completed, the weight W of the network is converged to a reasonable range, and the network is further finely adjusted by using the training samples and the corresponding class label information thereof in combination with a Softmax classifier:

$L_2\left(W\right)=\underset{W}{\min mize}\frac{1}{L}\underset{i=1}{\overset{L}{\Σ}}\left(\frac{1}{2}\right||y_i-h(x_i\left)\right|{|}^2)+\frac{\mathrm{\β}}{2}\underset{i,j}{\Σ}\left|\right|W_{ij}|{|}^2,$

wherein,is a term of a mean-square error,is a weight attenuation term aimed at reducing the magnitude of the weight and preventing overfitting, y_iRepresenting a training sample x_iCorresponding class label, h (x)_i) Is to train sample x_iFeature h learned through the entire deep network₃(x_i) Sending the output to a Softmax classifier, wherein β is 3e-3 which is a weight attenuation parameter;

solving the objective function L by using a gradient descent algorithm₂And (W), further optimizing the weight W of the sparse filter network, and realizing fine adjustment of the whole network.

And 7, performing class prediction on the test samples, sending the test samples into a deep sparse filter network, and predicting class labels of the test samples by using a Softmax classifier to obtain the prediction class of each test sample.

7a, testing sample x_jInputting the data into a constructed deep sparse filter network, and learning a final characteristic theta_j，j＝1,2,...,M；

7b, sending the final characteristics learned in the step 7a to a Softmax classifier for class prediction:

the output of the Softmax classifier is y ∈ R^P×1P is expressed as the number of classes, test sample x_jThe prediction categories of (a) may be expressed as:

$y=\arg \underset{\mathrm{\α}}{max}{\mathrm{\θ}}_j,$

where α is sample x_jThe prediction class confidence of.

And 8, outputting the classified bone-knitting jigsaw puzzle and the classification precision of the polarized SAR image to be classified, outputting the final classification result of the polarized SAR image to be classified according to the training sample and the test sample with the predicted class in the step 7, and calculating the precision of the classification.

(8a) And (f) using the sample category predicted by the classifier to correspond to each pixel point on the polarized SAR image, coloring each pixel point by using red, green and blue as three primary colors according to a three-primary-color coloring method, and outputting a result graph, which is shown in fig. 2 (f).

(8b) And comparing the class label predicted by the classifier with the real class label of the test sample to obtain the classification accuracy of the experiment.

According to the method, a deep learning method is adopted, and the deep sparse filter network is utilized to learn the characteristics of the polarized SAR image autonomously, so that the complexity of ergonomics feature learning in the traditional method is avoided, the deep sparse filter network can learn the more abstract and essential characteristics of the polarized SAR image, and the characteristics are more beneficial to ground feature classification of the polarized SAR image.

The effect of the invention can be specifically explained by simulation experiments:

example 7

The polarized SAR classification method based on the semi-supervised deep sparse filtering network is the same as the embodiment 1-6,

1. conditions of the experiment

The hardware platform is as follows: intel (R) core (TM) i5-2410M CPU @2.30GHz and RAM 4.00 GB;

the software platform is as follows: MATLAB R2011 b;

the experiment selects a 120 × 150 polarized SAR terrain simulation image for testing, 18000 sample points are provided, the dimension of each sample point is 6 dimensions, the category number is 9, and the mark is C_i1,2, 9. In the experiment, 1% of samples of each type are randomly selected as training samples, and the rest are testing samples.

2. Contents and results of the experiments

The method is combined with a Softmax classifier to classify the polarized SAR terrain simulation image, and is compared with other deep learning methods on the premise of the same experimental setup, wherein DNN is a deep neural network, and FIG. 2(c) is a result image of classifying FIG. 2(a) by the DNN; an SAE stacked self-encoder is also adopted in the experiment, and FIG. 2(d) is a result graph of classifying FIG. 2(a) by using an SAE method; FIG. 2(e) is a diagram of the results of classifying FIG. 2(a) using a DSF deep sparse filter network; SDSF is the process of the present invention. Table 1 shows the feature classification accuracy and the total classification accuracy of the polarized SAR feature simulation images obtained by the above 4 methods, respectively.

TABLE 1 Classification precision (%) and Total Classification precision (%) of the ground features on the simulation graphs for the respective methods

As can be seen from table 1, in the case that all the training samples are 1%, the present invention has higher classification accuracy compared to the existing deep learning method. Referring to the simulation experiment result, fig. 2, the invention has higher visibility.

In conclusion, the polarized SAR classification method based on the semi-supervised deep sparse filter network provided by the invention solves the problems that the traditional deep learning method has excessive parameter adjustment and has higher requirement on training samples, can effectively improve the classification precision of the polarized SAR image, and can obtain higher classification precision under the condition of less training samples.

The invention discloses a semi-supervised deep learning method based on semi-supervised sparse filtering. The technical problems that the traditional deep learning method is complex in parameter adjustment and low in classification precision when the traditional deep learning method has lower label data are solved, and the method comprises the following steps: 1. inputting polarized SAR image data to be classified, namely a coherent matrix T of a polarized SAR image, obtaining a label matrix Y according to ground feature distribution information of the polarized SAR image, and generating a sample matrix X by the coherent matrix T of the polarized SAR image; 2. extracting training samples and test samples, and randomly extracting L training samples and M test samples according to sample data X and a label matrix Y of the polarized SAR image; 3. obtaining Wishart neighbor samples of the training samples, and obtaining each training sample x from all sample data_iL) (i ═ 1,2.. L) for K Wishart neighbor samples; 4. initializing parameters of a deep sparse filter network, randomly initializing weight parameters W of the deep sparse filter network, and setting the number of nodes of each layer of the deep sparse filter network; 5. pre-training a deep sparse filter network, namely sending a training sample and a Wishart neighbor sample corresponding to the training sample into the deep sparse filter network for pre-trainingAdopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, simultaneously adding a semi-supervised neighbor keeping regular term in the pre-training of each layer, and optimizing the weight of the network together with the sparse filtering; 6. fine adjustment of the depth sparse filter network, namely fine adjustment of the depth sparse filter network is performed by utilizing training samples and label information thereof in combination with a Softmax classifier, and the weight of the network is further optimized; 7. performing class prediction on the test samples, sending the test samples into a deep sparse filter network, and predicting class labels of the test samples by using a Softmax classifier to obtain the prediction class of each test sample; 8. and outputting the classification image and the classification precision of the polarized SAR image to be classified, outputting the final classification result of the polarized SAR image to be classified according to the test sample of which the class is predicted by the training sample, and calculating the classification precision.

According to the method, the complexity of adjusting the deep learning network parameters is reduced by constructing a novel deep sparse filter network model, a semi-supervised neighbor preserving regular term method is added in the pre-training process, the number of training samples is reduced, and the accuracy of the ground feature classification of the polarized SAR image is improved by means of a small number of training samples. The method can be used for environmental monitoring, earth resource surveying, military systems and the like.

Claims (5)

1. The invention relates to a polarization SAR classification method based on a semi-supervised deep sparse filter network, which is characterized by comprising the following steps of:

(1) inputting polarized SAR image data to be classified, namely a coherent matrix T of the polarized SAR image, obtaining a label matrix Y according to ground feature distribution information of the polarized SAR image, and generating a sample matrix by the coherent matrix T of the polarized SAR imageN is the total number of samples, x_iRepresents the ith sample;

(2) extracting training samples and test samples, and randomly extracting L training samples and M test samples according to sample data X and a label matrix Y of the polarized SAR image, wherein L + M is N, randomly selecting 1% of samples in each class as the training samples according to the class information of all the samples, and the rest are the test samples;

(3) obtaining Wishart neighbor samples of the training samples, and obtaining each training sample x from all sample data_iL) (i ═ 1,2.. L) for K Wishart neighbor samples x_j(j＝1,2...K)；

(4) Initializing basic parameters of the deep sparse filter network, randomly initializing a weight parameter W of the deep sparse filter network, and setting the number of nodes of each layer of the deep sparse filter network;

(5) pre-training a deep sparse filter network, sending a training sample and a Wishart neighbor sample corresponding to the training sample into the deep sparse filter network for pre-training, adopting a layer-by-layer greedy pre-training method, taking the output of the previous layer as the input of the next layer until the last hidden layer is trained, simultaneously adding a semi-supervised neighbor keeping regular term in the pre-training of each layer, and jointly optimizing the weight of the network with a sparse filter;

(6) fine-tuning the depth sparse filter network, and utilizing training samples and label information thereof in combination with a Softmax classifier to fine-tune the depth sparse filter network so as to further optimize the weight of the network;

(7) performing class prediction on the test samples, sending the test samples into a deep sparse filter network, and predicting class labels of the test samples by using a Softmax classifier to obtain the prediction class of each test sample;

(8) and outputting a classification result graph and classification precision of the polarized SAR image to be classified, outputting a final classification result of the polarized SAR image to be classified according to the training sample and the test sample with the predicted class, and calculating the precision of the classification.

2. The semi-supervised depth sparse filtering network-based polarimetric SAR classification method according to claim 1, wherein the Wishart neighbor sample of each training sample obtained in the step (3) is obtained by the following steps:

3a, training sample matrix isThe Wishart distance between the training sample and the other samples is found using the following formula:

d(x_i,x_j)＝ln((x_i)^-1x_j)+Tr((x_j)^-1x_i)-q(x＝1,2,...,L,j＝1,2,...,N)，

wherein, Tr () represents the trace of the matrix, and q is 3 for the radar with integrated transmission and reception due to reciprocity; for radars where transmission and reception are not integral, q is 4;

3b, utilizing the sort function in MATLAB to compare the Wishart distance d (x) obtained in the step 3a_i,x_j) Arranging according to the ascending order of absolute values, taking the first K samples, finding the first K samples corresponding to the first K samples as training samples x_iThe wishirt neighbor sample of (a) is noted:

3. the semi-supervised deep sparse filtering network-based polarimetric SAR classification method according to claim 1, wherein the parameters of the initialized deep network in the step (4) are as follows:

4a, the number of hidden layers of the deep sparse filter network is 3, and the number of nodes of each layer is respectively as follows: 25, 100, 50;

4b, initializing weight W of deep sparse filter network₁,W₁∈R^D×QD is the dimension of the input signal and Q is the number of nodes of the first hidden layer.

4. The semi-supervised deep sparse filtering network-based polarimetric SAR classification method according to claim 1, wherein the pre-training process for the deep sparse filtering network in the step (5) is as follows:

5a, inputting a pre-training sample of the deep sparse filter network, and inputting a training sample x_iAnd K Wishart neighbor samples corresponding to the sameInputting the training samples into a deep network, wherein i is 1,2.

5b, learning new sample characteristics by utilizing a first hidden layer, and taking a traditional nonlinear sigmoid function as an activation function: phi (z) ═ 1+ exp (-z)^-1，x∈R^D×1If the input vector is, the output of the first hidden layer is: h is₁＝φ(W₁ ^Tx), the output of the ith sample is denoted as h₁(x_i) As new sample features;

5c, optimizing the network weight W, wherein each layer of sparse filter network utilizes an LBFGS algorithm to solve the following objective function, so that the weight W of the network is optimized, and the objective function is as follows:

$L_1\left(W\right)=\underset{W}{\min mize}\underset{i=1}{\overset{L}{\Σ}}\left|\right|\frac{{\tilde{h}}_1\left(x_i\right)}{\left|\right|{\tilde{h}}_1\left(x_i\right)|{|}_2}|{|}_1+\frac{\mathrm{\λ}}{2}\underset{i,j=1}{\overset{n}{\Σ}}{A}_{ij}||h_1\left(x_i\right)-h_1\left(x_j\right)|{|}^2,$

${\tilde{h}}_1\left(x_i\right)=\frac{h_1\left(x_i\right)}{\left|\right|h_1\left(x_i\right)|{|}_2},$

wherein in the objective functionIn order to be a sparse filtering objective function,the method for reference manifold learning optimizes the structure of the network by using the neighbor relation among samples for keeping a regular term for semi-supervised neighbor, wherein lambda is a parameter of the regular term,U(x_i) Representing a training sample x_iThe corresponding Wishart neighbor sample set.

5d, training the rest hidden layers by a greedy algorithm, and enabling the new sample characteristics h learned in the step 5b₁Sending the new input to the next layer, using the greedy learning algorithm, using the output of the previous layer as the input of the next layer, and using the objective function L in 5c₁(W) optimizing the weight W of each hidden layer in turn until the last hidden layer is trained, sample x_iThe output through the last hidden layer can be represented as h₃(x_i)。

5. The semi-supervised depth sparse filter network-based polarimetric SAR classification method according to claim 1, wherein the fine tuning of the depth sparse filter network in the step (6) has the following objective functions:

$L_2\left(W\right)=\underset{W}{\min mize}\frac{1}{L}\underset{i=1}{\overset{L}{\Σ}}\left(\frac{1}{2}\right||y_i-h(x_i\left)\right|{|}^2)+\frac{\mathrm{\β}}{2}\underset{i,j}{\Σ}\left|\right|W_{ij}|{|}^2,$

wherein,is a term of a mean-square error,is a weight attenuation term aimed at reducing the magnitude of the weight and preventing overfitting, y_iRepresenting a training sample x_iCorresponding class label, h (x)_i) Is to train sample x_iFeature h learned through the entire deep network₃(x_i) Sending the output to a Softmax classifier, wherein β is 3e-3 which is a weight attenuation parameter;

solving the objective function L by using a gradient descent algorithm₂(W) further optimizing the sparse filter networkAnd the weight W of the network realizes fine adjustment of the whole network.