CN104463222A - Polarimetric SAR image classification method based on feature vector distribution characteristic - Google Patents

Abstract

The invention discloses a polarimetric SAR image classification method based on the feature vector distribution characteristic. The method mainly solves the problem of lack of cognition to the feature vector distribution characteristic in the prior art. The method includes the steps that (1) filtering is conducted on a coherence matrix T by enhancing Lee, and a filtered coherence matrix T' is obtained; (2) a T' eigenvalue is decomposed, and the maximum feature vector e1 is obtained; (3) module values of all components of the e1 are taken respectively, and the converted maximum feature vector e1' is obtained; (4) different homogeneous regions are selected to serve as class representation regions, and the e1' of each region is extracted; (5) mean value and covariance parameter computing of a three-dimensional Gaussian probability model is conducted on all classes of homogeneous e1's, and the three-dimensional Gaussian probability-density function of all classes of homogeneous region e1's is solved; (6) Bayesian classification is conducted on an image, and the initial classification result is obtained; (7) local Wishart iteration is conducted on the initial result, and the final classification result is output. The method has the advantage that the good classification effect on the polarimetric SAR image is achieved and can be used for polarimetric SAR image target detection and target identification.

Description

The Classification of Polarimetric SAR Image method of feature based vector distribution characteristic

Technical field

The invention belongs to technical field of image processing, relate to Classification of Polarimetric SAR Image, can be used for Radar Targets'Detection and target identification, specifically a kind of Classification of Polarimetric SAR Image method of feature based vector distribution characteristic.

Background technology

Along with the development of the information processing technology and electronic technology, synthetic-aperture radar (SAR) also towards high resolving power, multipolarization, multiband, multi-mode, the future development such as multi-platform, its superior function in earth observation field can be embodied more fully.Multipolarization SAR, as one of SAR important developing direction, can provide more more fully scattered information amounts for the decipher of target and correct understanding.And these abundant information provide important foundation for the decipher of Polarimetric SAR Image, and then to the automatic detection of target with automatically identify there is very important directive function.Existing Classification of Polarimetric SAR Image method roughly can be divided into method based on scattering properties and Corpus--based Method Characterization method, as the sorting technique of decomposing based on Freeman and multiple Wishart sorting technique etc.

Wherein, the Classification of Polarimetric SAR Image method of Corpus--based Method characteristic is that the pixel in Polarimetric SAR Image with certain aspect similar features is divided into a class, has had a lot of ripe sorting algorithm to be used in Classification of Polarimetric SAR Image.As a new branch of science, research work main at present concentrates in the extraction of scattering signatures and the classifier design of applicable Classification of Polarimetric SAR Image.The extracting method of different Polarization scattering feature obtains different features, and different features is suitable for again the sorter of different performance.Existing feature extracting method mainly contains the methods such as Freeman decomposition, Cloude decomposition and the decomposition of Pauli base; Typical sorter has multiple Wishart sorter and H/alpha sorter etc.Wherein, the physical meaning representated by the eigenwert obtained about feature decomposition and the application in terrain classification thereof are approved by increasing people.But be compared to eigenwert, proper vector can comprise more abundant scattered information, but also proper vector is not carried out classification as feature in existing method.

Summary of the invention

The object of the invention is the deficiency overcoming above-mentioned prior art, the polarization information comprised in the proper vector after deep excavation coherent matrix eigenvalue decomposes, and it can be made to instruct polarization SAR terrain classification.For this reason, the present invention proposes a kind of Classification of Polarimetric SAR Image method of feature based vector distribution characteristic, with the distribution character of clear and definite feature, improve the precision of Classification of Polarimetric SAR Image.

Technical scheme of the present invention is: the Classification of Polarimetric SAR Image method of feature based vector distribution characteristic, comprises the following steps:

(1) coherence matrix T={T (i) of Polarimetric SAR Image to be sorted is read in | i=1 ..., M}, M are total pixel number of image;

(2) enhanced Lee filtering is carried out to Polarimetric SAR Image to be sorted, obtains filtered coherence matrix T'={T'(i) | i=1 ..., M};

(3) to the filtered coherence matrix T'(i of each pixel) carry out Eigenvalues Decomposition, obtain eigenvalue of maximum characteristic of correspondence vector, i.e. maximal eigenvector e ₁={ e ₁(i) | i=1 ..., M};

(4) to the maximal eigenvector e of each point ₁(i)=[e ₁₁(i), e ₁₂(i), e ₁₃(i)] ^hevery dimension data e _1ji (), j=1,2,3 difference delivery values, obtain the maximal eigenvector vector e after converting ₁'={ e ₁' (i) | i=1 ..., M}, wherein e ₁' (i)=[e' ₁₁(i), e' ₁₂(i), e' ₁₃(i)] ^h, H is transpose of a matrix symbol;

(5) according to atural object actual distribution figure, k class homogenous region C={C is manually chosen ₁..., C _l..., C _k, and extract the maximal eigenvector e after the conversion of k class homogenous region ₁' as training sample;

(6) adopt EM algorithm respectively to the maximal eigenvector e after the conversion of k class training sample ₁' every dimension data e' _1j, j=1,2,3 carry out Gauss's estimation respectively, obtain the maximal eigenvector e after converting ₁' mean vector mu and standard deviation vector sigma:

Mu=[μ ₁, μ ₂, μ ₃] ^h, μ _j, j=1,2,3, corresponding to e ₁' the average of each component;

Sigma=[σ ₁, σ ₂, σ ₃] ^h, σ _j, j=1,2,3, corresponding to e ₁' the standard deviation of each component;

(7) the maximal eigenvector e after conversion is calculated ₁' covariance matrix Σ, according to formula cov (X, Y)=E (XY)-E (X) E (Y) calculate conversion after maximal eigenvector e ₁' in the covariance of component, obtain covariance matrix Σ, wherein E (XY) draws according to EM algorithm, the maximal eigenvector e thus after tectonic transition ₁' density function:

$P\left(e_1^{\text{'}}\right|\mathrm{mu},\mathrm{\Σ})={\left(2\mathrm{\π}\right)}^{-3/2}{\left|\mathrm{\Σ}\right|}^{-1/2}\exp [-\frac{1}{2}{(e_1^{\text{'}}-\mathrm{mu})}^H{\mathrm{\Σ}}^{-1}(e_1^{\text{'}}-\mathrm{mu})]$

(8) according to MAP principle, e is solved ₁' (i) ∈ C _q, st.p (e ₁' (i) | C _q)=maxP (e ₁' (i) | e ₁' (i) ∈ C _q), q=1 ..., l ... k, i=1 ..., the whole pixels of M to homogenous region selected in Polarimetric SAR Image to be sorted are classified, and output category result;

(9) the preliminary classification result of step (8) is carried out local Wishart iteration optimization.

The filtered coherence matrix T'(i to each pixel described in above-mentioned steps (3)) carry out Eigenvalues Decomposition, obtain eigenvalue of maximum characteristic of correspondence vector, carry out in accordance with the following steps:

2a) using the filtered coherence matrix T'(i of Polarimetric SAR Image pixel to be sorted) as input data;

2b) utilize the eigs function in MATLAB software to the filtered coherence matrix T'(i of each pixel of input) carry out Eigenvalues Decomposition, obtain the maximal eigenvector e that eigenvalue of maximum is corresponding ₁(i).

Employing EM algorithm described in above-mentioned steps (6) is respectively to the maximal eigenvector e after the conversion of k class training sample ₁' every dimension data e' _1j, j=1,2,3 carry out Gauss's estimation respectively, specifically carry out in accordance with the following steps:

3a) k class homogenous region training data is concentrated the maximal eigenvector e after the conversion of pixel ₁' three component e' ₁₁, e' ₁₂, e' ₁₃respectively as the input data of EM algorithm;

3b) setting EM algorithm iteration end condition is that the iteration error of algorithm is less than 1.0E-06, initialization average μ _jand standard deviation sigma _jfor random value;

3c) by the expectation maximization flow process in EM algorithm to average μ _jand standard deviation sigma _jupgrade, each upgrade after computational algorithm iteration error and judge whether to meet stopping criterion for iteration, iteration stopping when meeting stopping criterion for iteration; Otherwise, carry out stop technology according to iterations, re-start iteration; Obtain the maximal eigenvector e after converting thus ₁' the average of three-dimensional component and standard deviation;

Maximal eigenvector e 3d) after conversion ₁' corresponding mean vector is: mu=[μ ₁, μ ₂, μ ₃] ^h;

Maximal eigenvector e 3e) after conversion ₁' three component e' ₁₁, e' ₁₂, e' ₁₃the standard deviation vector that standard deviation corresponding is respectively formed is: sigma=[σ ₁, σ ₂, σ ₃] ^h.

Maximal eigenvector e after calculating conversion described in above-mentioned steps (7) ₁' covariance matrix Σ, carry out in accordance with the following steps:

E' ₁₁, e' ₁₂, e' ₁₃between covariance obtain according to the covariance formula in theory of probability below:

cov(X,Y)＝E(XY)-E(X)E(Y)

4a) constructed variable XY, i.e. e' ₁₁e' ₁₂, e' ₁₂e' ₁₃, e' ₁₃e' ₁₁;

4b) calculate e' according to step (6) ₁₁e' ₁₂, e' ₁₂e' ₁₃, e' ₁₃e' ₁₁the average μ of three variablees ₁₂, μ ₂₃, μ ₁₃;

4c) according to the e' that step (6) obtains ₁₁, e' ₁₂, e' ₁₃average μ ₁, μ ₂, μ ₃, calculate covariance

Cov (e' _1me' _1j)=μ _mj-μ _mμ _jm, j=1,2,3 and m ≠ j

4d) according to covariance matrix be the characteristic of symmetrical matrix, and the element on diagonal line is e' ₁₁, e' ₁₂, e' ₁₃variance structure covariance matrix Σ

$\mathrm{\Σ}=\left[\begin{array}{ccc}{\mathrm{\σ}}_1^2& \mathrm{cov}\left(e_{11}^{\text{'}}{e}_{12}^{\text{'}}\right)& \mathrm{cov}\left(e_{13}^{\text{'}}{e}_{11}^{\text{'}}\right)\\ \mathrm{cov}\left(e_{11}^{\text{'}}{e}_{12}^{\text{'}}\right)& {\mathrm{\σ}}_2^2& \mathrm{cov}\left(e_{12}^{\text{'}}{e}_{13}^{\text{'}}\right)\\ \mathrm{cov}\left(e_{13}^{\text{'}}{e}_{11}^{\text{'}}\right)& \mathrm{cov}\left(e_{12}^{\text{'}}{e}_{13}^{\text{'}}\right)& {\mathrm{\σ}}_3^2\end{array}\right].$

Local Wishart iterative optimization procedure described in above-mentioned steps (9), carries out in accordance with the following steps:

The cluster centre V of each class 5a) is calculated according to the classification results of step (8) _q

wherein, T _q' (i)=T'(i) | and T'(i) ∈ C _q,

N _qbe the pixel number in q class, q=1 ..., l ... k;

5b) using common as input data for the classification results of filtered coherence matrix T' and step (8), calculate the Wishart distance of each point to each homogenous area cluster centre of k class

d(T'(i),V _q)＝ln|V _q|+tr(V _q ^-1T'(i))

Wherein, tr () represents Matrix Calculating mark.If for all q ≠ p, meet relational expression

d(T'(i),V _q)≤d(T'(i),V _p)

So this point belongs to q class.Upgrade according to the class mark of Wishart distance classification result to current point;

5c) according to 5a) upgrade k cluster centre, when the difference of twice cluster centre meets accuracy requirement, stop iteration; Otherwise, carry out stop technology according to iterations; Export Wishart iteration result;

5d) using the classification results of step (8) as judgment basis, judge 5c) Output rusults in the classification of being divided by mistake, line item of going forward side by side;

5e) according to 5d) record re-start 5b)-5c), for the classification that mistake is divided, when carrying out Wishart iteration, artificial block pixel the initial category of this point and by the classification of mistake minute between carry out iteration.

Beneficial effect of the present invention: the present invention compared with prior art has the following advantages:

1, start with from the angle of proper vector, excavate the polarization information that proper vector comprises, and use it in polarization SAR terrain classification;

2, The present invention gives the three-dimensional Gaussian model of proper vector, specify that the distribution character of the proper vector modulus value of each homogenous region and the otherness of different homogenous region proper vector modulus value distribution, thus improve the precision of terrain classification.

The simulation experiment result shows, the Classification of Polarimetric SAR Image method of the distribution statistics of the feature based vector modulus value that the present invention proposes can be effectively applied to Classification of Polarimetric SAR Image, and is applied to Radar Targets'Detection and target identification further.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the RGB composite diagram that the present invention emulates the Polarimetric SAR Image of employing and the 3 class training samples marked on this figure;

Fig. 3 is the result figure adopting the present invention to classify to Fig. 2.

Concrete implementation step

With reference to Fig. 1, specific embodiment of the invention is as follows:

Step 1. reads in coherence matrix T={T (i) of Polarimetric SAR Image to be sorted | i=1 ..., M}, M are total pixel number of image, and T (i) is the multiple Hermit matrix of 3 × 3.

Step 2. carries out enhanced Lee filtering to Polarimetric SAR Image to be sorted, obtains filtered coherence matrix T'={T'(i) | i=1 ..., M}.

Step 3. is to the filtered coherence matrix T'(i of each pixel of Polarimetric SAR Image to be sorted) carry out Eigenvalues Decomposition.

Utilize the eigs function of MATLAB software to carry out Eigenvalues Decomposition to filtered coherence matrix T', decomposition expression formula is as follows:

$T^{\text{'}}=[e_1,e_2,e_3]\left[\begin{array}{ccc}{\mathrm{\λ}}_1& 0& 0\\ 0& {\mathrm{\λ}}_2& 0\\ 0& 0& {\mathrm{\λ}}_3\end{array}\right]{[e_1,e_2,e_3]}^{*}$

Wherein, T' is filtered coherence matrix, and * is conjugate transpose symbol, λ ₁, λ ₂, λ ₃the eigenwert that T' Eigenvalues Decomposition obtains respectively, and λ ₁>=λ ₂>=λ ₃, e ₁, e ₂, e ₃be respectively and λ ₁, λ ₂, λ ₃corresponding proper vector.Therefore, e ₁the eigenvalue of maximum characteristic of correspondence vector that we need, referred to as maximal eigenvector.

Step 4. couple maximal eigenvector e ₁each component respectively delivery value, obtain the maximal eigenvector e after converting ₁'.

The coherence matrix of Polarimetric SAR Image is complex matrix, therefore, decomposes the maximal eigenvector e obtained ₁be one 3 × 1 complex vector located, to maximal eigenvector e ₁in three data get its modulus value respectively, in MATLAB software, realize with abs function.By the maximal eigenvector e after conversion ₁' represent.

e ₁'＝[e' ₁₁e' ₁₂e' ₁₃] ^H

Step 5., according to the actual distribution situation of atural object, chooses the homogenous region of k class atural object as training sample.

Step 6. adopts EM algorithm respectively to e ₁' in three component e' ₁₁, e' ₁₂, e' ₁₃carry out Gauss's estimation respectively, obtain the standard deviation sigma of mean vector mu and three component ₁, σ ₂, σ ₃.

6a) all kinds of homogenous region training data is concentrated the maximal eigenvector e after the conversion of pixel ₁' three component e' ₁₁, e' ₁₂, e' ₁₃respectively as the input data of EM algorithm;

6b) setting EM algorithm iteration end condition is that the iteration error of algorithm is less than 1.0E-06, initialization average μ _jand standard deviation sigma _jfor random value;

6c) by the expectation maximization flow process in EM algorithm to average μ _jand standard deviation sigma _jupgrade, each upgrade after computational algorithm iteration error and judge whether to meet stopping criterion for iteration, iteration stopping when meeting stopping criterion for iteration; Otherwise, carry out stop technology according to iterations.Export average and the standard deviation of the three-dimensional component of the maximal eigenvector after the conversion obtaining inputting;

6d) vectorial e ₁' corresponding mean vector mu=[μ ₁, μ ₂, μ ₃] ^h;

6e) vectorial e ₁' three component e' ₁₁, e' ₁₂, e' ₁₃the standard deviation vector that standard deviation corresponding is respectively formed is:

sigma＝[σ ₁，σ ₂，σ ₃] ^H

Step 7. calculates the maximal eigenvector e after conversion ₁' covariance matrix Σ.

According to the knowledge of theory of probability, covariance can obtain according to formula cov (X, Y)=E (XY)-E (X) E (Y)

7a) constructed variable XY, i.e. constructed variable e' ₁₁e' ₁₂, e' ₁₂e' ₁₃, e' ₁₃e' ₁₁;

Three variable e' of structure 7b) are calculated according to step 6 ₁₁, e' ₁₂, e' ₁₃average μ ₁₂, μ ₂₃, μ ₁₃;

7c) obtain three variable e' according to step 6 ₁₁, e' ₁₂, e' ₁₃average μ ₁, μ ₂, μ ₃, calculate covariance

Cov (e' _1me' _1j)=μ _mj-μ _mμ _jm, j=1,2,3 and m ≠ j

7d) according to covariance matrix be the characteristic of symmetrical matrix, and the element on diagonal line is e' ₁₁, e' ₁₂, e' ₁₃variance therefore, covariance matrix Σ can be constructed

$\mathrm{\Σ}=\left[\begin{array}{ccc}{\mathrm{\σ}}_1^2& \mathrm{cov}\left(e_{11}^{\text{'}}{e}_{12}^{\text{'}}\right)& \mathrm{cov}\left(e_{13}^{\text{'}}{e}_{11}^{\text{'}}\right)\\ \mathrm{cov}\left(e_{11}^{\text{'}}{e}_{12}^{\text{'}}\right)& {\mathrm{\σ}}_2^2& \mathrm{cov}\left(e_{12}^{\text{'}}{e}_{13}^{\text{'}}\right)\\ \mathrm{cov}\left(e_{13}^{\text{'}}{e}_{11}^{\text{'}}\right)& \mathrm{cov}\left(e_{12}^{\text{'}}{e}_{13}^{\text{'}}\right)& {\mathrm{\σ}}_3^2\end{array}\right]$

Step 8. constructs the three-dimensional Gaussian probability density function of the maximal eigenvector after about conversion according to the mean vector calculated in step 6 and step 7 and covariance matrix:

$P\left(e_1^{\text{'}}\right|\mathrm{mu},\mathrm{\Σ})={\left(2\mathrm{\π}\right)}^{-3/2}{\left|\mathrm{\Σ}\right|}^{-1/2}\exp [-\frac{1}{2}{(e_1^{\text{'}}-\mathrm{mu})}^H{\mathrm{\Σ}}^{-1}(e_1^{\text{'}}-\mathrm{mu})]$

Known according to multivariate Gaussian distribution definition: if there is Y=(Y ₁y _m) ^h, wherein Y ₁... Y _mindependent of one another and have identical distribution N (0,1); The expression X that Y can be linear is made to be X ≡ mu+A ^hy.Wherein mu is that m ties up constant vector, and A is the constant matrices of m × m dimension, and has A ^ha=Σ (noting: ≡ represents that the right and left has identical distribution).Then claim stochastic variable X=(X ₁x _m) ^hobey multivariate normal distribution, and be often designated as X ~ N _m(mu, Σ).

Prove the maximal eigenvector e' after conversion ₁statistical distribution characteristic be meet three-dimensional Gaussian distribution step as follows:

8a) calculate the maximal eigenvector e' after conversion according to step 6 and step 7 ₁mean vector mu and covariance matrix Σ.

8b) according to cholesky theorem, decomposition is carried out to Σ and obtains matrix A, make A meet A ^ha=Σ.

8c) make BA ^h=I, wherein I is unit battle array, according to formula X ≡ mu+A ^hy then can release

$Y=\left[\begin{array}{c}{Y}_1\\ Y_2\\ Y_3\end{array}\right]=B=(e_1^{\text{'}}-\mathrm{mu})$

For different polarization SAR data, as long as three components that can prove Y are approximate independent same distribution, and all obey standard normal N (0,1), just can prove the maximal eigenvector e after change ₁' be obey three-dimensional Gaussian distribution, and distribution function is

P(e ₁'|mu,Σ)＝(2π) ^-3/2|Σ| ^-1/2exp[-21(e ₁'-mu) ^HΣ ^-1(e ₁'-mu)]

Step 9., according to MAP principle, is classified to the pixel that test data is concentrated, row labels of going forward side by side.

By the maximal eigenvector e' after the conversion of pixel to be sorted _1jbring the probability that this point of step 8 formulae discovery belongs to all kinds of atural object into, then according to MAP principle, according to expression formula: e ₁' (i) ∈ C _q, st.p (e ₁' (i) | C _q)=maxP (e ₁' (i) | e ₁' (i) ∈ C _q), i=1 ..., M, q=1 ..., l ... the whole pixels of k to homogenous region selected in Polarimetric SAR Image to be sorted carry out kind judging, and markup tags, output category result.

The classification results of step 9 is carried out local Wishart iteration as preliminary classification result by step 10..

In order to improve the region consistency of classification results, keeping preliminary classification result classification number constant simultaneously, the basis of step 9 adds local Wishart iteration.

The cluster centre V of each class 10a) is calculated according to the classification results of step 9 _q

wherein, T _q' (i)=T'(i) | and T'(i) ∈ C _q,

N _qbe the pixel number in q class, q=1 ..., l ... k

10b) using common as input data for the classification results of filtered coherence matrix T' and step 9, calculate the Wishart distance of each point to each homogenous area cluster centre of k class

d(T'(i),V _q)＝ln|V _q|+tr(V _q ^-1T'(i))

Wherein, tr () represents Matrix Calculating mark.If for all q ≠ p, meet relational expression

d(T'(i),V _q)≤d(T'(i),V _p)

So this point belongs to q class.Upgrade according to the class mark of Wishart distance classification result to current point

10c) according to 10a) upgrade k cluster centre, when the difference of twice cluster centre meets accuracy requirement, stop iteration; Otherwise, carry out stop technology according to iterations; Export Wishart iteration result

10d) using the classification results of step (9) as judgment basis, judge 10c) Output rusults in the classification of being divided by mistake, line item of going forward side by side

10e) according to 10d) record re-start 10b)-10c), for the classification that mistake is divided, when carrying out Wishart iteration, artificial block pixel the initial category of this point and by the classification of mistake minute between carry out iteration.

Effect of the present invention can be verified by following emulation experiment.

1, experiment condition

1.1) originally experimentally machine running environment is the Matlab2012b software under windows7 system.

1.2) as shown in Figure 2, this figure is the RGB composograph of Western China peace west area full-polarization SAR to the Polarimetric SAR Image chosen in experiment.Wherein R representative | HH-VV|, G represent | and HV|+|VH|, B represent | and HH+VV|, picture size size is 510 × 510.In this Polarimetric SAR Image, the increased surface covering in region is village, river and wasteland, and true atural object comprises the Weihe River, across the Weihe Bridge in the Weihe River, and railway etc.

1.3) choose 3 class regions as training sample in experiment, 3 class samples represent farmland, river and cities and towns respectively.Rectangle symbol region in Fig. 2 is the training sample chosen in the different areas.

2, experiment content

2.1) parameter calculating is carried out to the estimation distributed model of each maximal eigenvector of above-mentioned homogenous region, and meet three-dimensional Gaussian distribution probability model according to the maximal eigenvector after derivation checking conversion.

The mean vector of the three kinds of samples calculated according to implementation step 6 and step 7 and covariance matrix are as shown in Table 1.

Table one mean vector and covariance parameter

According to step 8, according to sample data and corresponding mean vector and covariance matrix, ask vectorial Y respectively ₁, Y ₂, Y ₃, wherein, Y ₁corresponding sample 1, Y ₂corresponding sample 2, Y ₃corresponding sample 3.Three vectorial correlation parameters are in table two.

Table two Y ₁, Y ₂, Y ₃correlation parameter

In table two, because other atural objects mixed in sample three are more, the Y therefore calculated ₃in data, there is larger deviation in mean vector, but on the whole, the every one dimension component in three vectors is all independently, and is similar to standardized normal distribution.Therefore, can think that this model is correct.

Fig. 3 is the final classification results using this method to obtain.

2.2) the present invention and H/alpha-Wishart method and SC-Wishart method are contrasted, nicety of grading as shown in Table 3:

Table three three kinds of algorithm classification accuracy comparison tables

As can be seen from Table III, average classification ratio of precision H/alpha-Wishart method of the present invention and SC-Wishart method high.H/alpha-Wishart method mainly utilizes eigenwert and scattering angle to carry out classifying, and SC-Wishart method is then the sorting technique for coherence matrix T.But due to sample 1 and sample 3 higher in the similarity in coherence matrix T space and scattering angle space, characteristic vector space be then easier to distinguish, therefore the present invention can obtain good classification results.

2.3) the inventive method is utilized to classify to Polarimetric SAR Image, result as shown in Figure 3, as can be seen from Figure 3, the classification results that the present invention obtains is while guarantee classification is correct, can have good maintenance to detail section, this just shows that the present invention effectively can improve the precision of Classification of Polarimetric SAR Image.

To sum up, compared with prior art, the present invention has the following advantages:

1, start with from the angle of proper vector, excavate the polarization information that proper vector comprises, and use it in polarization SAR terrain classification;

2, The present invention gives the three-dimensional Gaussian model of proper vector, specify that the distribution character of the proper vector modulus value of each homogenous region and the otherness of different homogenous region proper vector modulus value distribution, thus improve the precision of terrain classification.

The simulation experiment result shows, the Classification of Polarimetric SAR Image method of the distribution statistics of the feature based vector modulus value that the present invention proposes can be effectively applied to Classification of Polarimetric SAR Image, and is applied to Radar Targets'Detection and target identification further.

The part do not described in detail in present embodiment belongs to the known conventional means of the industry, does not describe one by one here.More than exemplifying is only illustrate of the present invention, does not form the restriction to protection scope of the present invention, everyly all belongs within protection scope of the present invention with the same or analogous design of the present invention.

Claims (5)

1. the Classification of Polarimetric SAR Image method of feature based vector distribution characteristic, is characterized in that, comprise the following steps:

(1) coherence matrix T={T (i) of Polarimetric SAR Image to be sorted is read in | i=1 ..., M}, M are total pixel number of image;

(2) enhanced Lee filtering is carried out to Polarimetric SAR Image to be sorted, obtains filtered coherence matrix T'={T'(i) | i=1 ..., M};

(3) to the filtered coherence matrix T'(i of each pixel) carry out Eigenvalues Decomposition, obtain eigenvalue of maximum characteristic of correspondence vector, i.e. maximal eigenvector e ₁={ e ₁(i) | i=1 ..., M};

(4) to the maximal eigenvector e of each point ₁(i)=[e ₁₁(i), e ₁₂(i), e ₁₃(i)] ^hevery dimension data e _1ji (), j=1,2,3 difference delivery values, obtain the maximal eigenvector vector e after converting ₁'={ e ₁' (i) | i=1 ..., M}, wherein e ₁' (i)=[e' ₁₁(i), e' ₁₂(i), e' ₁₃(i)] ^h, H is transpose of a matrix symbol;

(5) according to atural object actual distribution figure, k class homogenous region C={C is manually chosen ₁..., C _l..., C _k, and extract the maximal eigenvector e after the conversion of k class homogenous region ₁' as training sample;

(6) adopt EM algorithm respectively to the maximal eigenvector e after the conversion of k class training sample ₁' every dimension data e' _1j, j=1,2,3 carry out Gauss's estimation respectively, obtain the maximal eigenvector e after converting ₁' mean vector mu and standard deviation vector sigma:

Mu=[μ ₁, μ ₂, μ ₃] ^h, μ _j, j=1,2,3, corresponding to e ₁' the average of each component;

Sigma=[σ ₁, σ ₂, σ ₃] ^h, σ _j, j=1,2,3, corresponding to e ₁' the standard deviation of each component;

(7) the maximal eigenvector e after conversion is calculated ₁' covariance matrix Σ, according to formula cov (X, Y)=E (XY)-E (X) E (Y) calculate conversion after maximal eigenvector e ₁' in the covariance of component, obtain covariance matrix Σ, wherein E (XY) draws according to EM algorithm, the maximal eigenvector e thus after tectonic transition ₁' density function:

$p\left(e_1^{\text{'}}\right|\mathrm{mu},\mathrm{\Σ})={\left(2\mathrm{\π}\right)}^{3/2}{\left|\mathrm{\Σ}\right|}^{-1/2}\exp [-\frac{1}{2}{(e_1^{\text{'}}-\mathrm{mu})}^H{\mathrm{\Σ}}^{-1}(e_1^{\text{'}}-\mathrm{mu})]$

(8) according to MAP principle, e is solved ₁' (i) ∈ C _q, st.p (e ₁' (i) C _q)=maxP (e ₁' (i) e ₁' (i) ∈ C _q), q=1 ..., l ... k, i=1 ..., the whole pixels of M to homogenous region selected in Polarimetric SAR Image to be sorted are classified, and output category result;

(9) the preliminary classification result of step (8) is carried out local Wishart iteration optimization.

2. the Classification of Polarimetric SAR Image method of feature based vector distribution characteristic according to claim 1, it is characterized in that, the filtered coherence matrix T'(i to each pixel wherein described in step (3)) carry out Eigenvalues Decomposition, obtain eigenvalue of maximum characteristic of correspondence vector, carry out in accordance with the following steps:

2a) using the filtered coherence matrix T'(i of Polarimetric SAR Image pixel to be sorted) as input data;

2b) utilize the eigs function in MATLAB software to the filtered coherence matrix T'(i of each pixel of input) carry out Eigenvalues Decomposition, obtain the maximal eigenvector e that eigenvalue of maximum is corresponding ₁(i).

3. the Classification of Polarimetric SAR Image method of feature based vector distribution characteristic according to claim 1, is characterized in that, the employing EM algorithm wherein described in step (6) is respectively to the maximal eigenvector e after the conversion of k class training sample ₁' every dimension data e' _1j, j=1,2,3 carry out Gauss's estimation respectively, specifically carry out in accordance with the following steps:

3a) k class homogenous region training data is concentrated the maximal eigenvector e after the conversion of pixel ₁' three component e' ₁₁, e' ₁₂, e' ₁₃respectively as the input data of EM algorithm;

3b) setting EM algorithm iteration end condition is that the iteration error of algorithm is less than 1.0E-06, initialization average μ _jand standard deviation sigma _jfor random value;

3c) by the expectation maximization flow process in EM algorithm to average μ _jand standard deviation sigma _jupgrade, each upgrade after computational algorithm iteration error and judge whether to meet stopping criterion for iteration, iteration stopping when meeting stopping criterion for iteration; Otherwise, carry out stop technology according to iterations, re-start iteration; Obtain the maximal eigenvector e after converting thus ₁' the average of three-dimensional component and standard deviation;

Maximal eigenvector e 3d) after conversion ₁' corresponding mean vector is: mu=[μ ₁, μ ₂, μ ₃] ^h;

Maximal eigenvector e 3e) after conversion ₁' three component e' ₁₁, e' ₁₂, e' ₁₃the standard deviation vector that standard deviation corresponding is respectively formed is: sigma=[σ ₁, σ ₂, σ ₃] ^h.

4. the Classification of Polarimetric SAR Image method of feature based vector distribution characteristic according to claim 1, is characterized in that, the maximal eigenvector e after the calculating conversion wherein described in step (7) ₁' covariance matrix Σ, carry out in accordance with the following steps:

E' ₁₁, e' ₁₂, e' ₁₃between covariance obtain according to the covariance formula in theory of probability below:

cov(X,Y)＝E(XY)-E(X)E(Y)

4a) constructed variable XY, i.e. e' ₁₁e' ₁₂, e' ₁₂e' ₁₃, e' ₁₃e' ₁₁;

4b) calculate e' according to step (6) ₁₁e' ₁₂, e' ₁₂e' ₁₃, e' ₁₃e' ₁₁the average μ of three variablees ₁₂, μ ₂₃, μ ₁₃;

4c) according to the e' that step (6) obtains ₁₁, e' ₁₂, e' ₁₃average μ ₁, μ ₂, μ ₃, calculate covariance

Cov (e' _1me' _1j)=μ _mj-μ _mμ _jm, j=1,2,3 and m ≠ j

4d) according to covariance matrix be the characteristic of symmetrical matrix, and the element on diagonal line is e' ₁₁, e' ₁₂, e' ₁₃variance structure covariance matrix Σ

$\mathrm{\Σ}=\left[\begin{array}{ccc}{\mathrm{\σ}}_1^2& \mathrm{cov}\left(e_{11}^{\text{'}}{e}_{12}^{\text{'}}\right)& \mathrm{cov}\left(e_{13}^{\text{'}}{e}_{11}^{\text{'}}\right)\\ \mathrm{cov}\left(e_{11}^{\text{'}}{e}_{12}^{\text{'}}\right)& {\mathrm{\σ}}_2^2& \mathrm{cov}\left(e_{12}^{\text{'}}{e}_{13}^{\text{'}}\right)\\ \mathrm{cov}\left(e_{13}^{\text{'}}{e}_{11}^{\text{'}}\right)& \mathrm{cov}\left(e_{12}^{\text{'}}{e}_{13}^{\text{'}}\right)& {\mathrm{\σ}}_3^2\end{array}\right].$

5. the Classification of Polarimetric SAR Image method of feature based vector distribution characteristic according to claim 1, is characterized in that, the local Wishart iterative optimization procedure wherein described in step (9), carries out in accordance with the following steps:

The cluster centre V of each class 5a) is calculated according to the classification results of step (8) _q

wherein, T _q' (i)=T'(i) | and T'(i) ∈ C _q,

N _qbe the pixel number in q class, q=1 ..., l ... k;

5b) using common as input data for the classification results of filtered coherence matrix T' and step (8), calculate the Wishart distance of each point to each homogenous area cluster centre of k class

d(T'(i),V _q)＝ln|V _q|+tr(V _q ^-1T'(i))

Wherein, tr () represents Matrix Calculating mark; If for all q ≠ p, meet relational expression

d(T'(i),V _q)≤d(T'(i),V _p)

So this point belongs to q class; Upgrade according to the class mark of Wishart distance classification result to current point;

5c) according to 5a) upgrade k cluster centre, when the difference of twice cluster centre meets accuracy requirement, stop iteration; Otherwise, carry out stop technology according to iterations; Export Wishart iteration result;

5d) using the classification results of step (8) as judgment basis, judge 5c) Output rusults in the classification of being divided by mistake, line item of going forward side by side;

5e) according to 5d) record re-start 5b)-5c), for the classification that mistake is divided, when carrying out Wishart iteration, artificial block pixel the initial category of this point and by the classification of mistake minute between carry out iteration.