CN105956622A - Polarized SAR image classification method based on multi-characteristic combined modeling - Google Patents

Abstract

The invention discloses a polarized SAR image classification method based on multi-characteristic combined modeling. The method comprises the steps of: (1) collecting pixel points of various kinds of ground objects from polarized SAR images which are manually marked, and using the pixel points as training samples; (2) calculating a polarized covariance matrix of each pixel point in the current kind of training samples, and constructing polarized characteristic vectors of the current kind of training samples; (3) constructing a (i) CoAS model of the current kind of training samples; (4) constructing polarized characteristic vectors of images to be classified; (5) constructing a combined posterior probability of the images to be classified; and (6), according to the combined posterior probability, carrying out classification on the imaged to be classified based on an (i) MRF model. The method has an excellent classification effect, and compared with a conventional polarized SAR image classification method based on a statistic model, the classification precision can be improved by 5%-10%.

Description

Polarization SAR image classification method based on multiple features combining modeling

Technical field

The invention belongs to Remote Sensing Image Processing Technology field, particularly relate to a kind of polarization based on multiple features combining modeling SAR image classification method.

Background technology

Synthetic aperture radar (Synthetic Aperture Radar, SAR) is a kind of Active Imaging Lidar system, has Round-the-clock, the feature of imaging round the clock.Polarization SAR system can be operated under different POLARIZATION CHANNEL integrated mode, can obtain more Ground object target information.In recent years, increasing polarization SAR system occurs, interpretation and the application of polarization SAR information demonstrate more Add the prospect of being widely applied.Image classification is an important content of polarization SAR image interpretation, in the face of magnanimity polarization SAR data, By polarization SAR image being classified automatically and identifying, it is favorably improved interpretation efficiency, increases people and polarize from large scene The ability of SAR image capturing information.

Existing polarization SAR image classification method substantially can be divided into 3 classes: 1) according to polarization SAR imaging characteristics, to data Carry out statistical model analysis, utilize Bayes method to classify；2) by Polarization target decomposition, based on polarization SAR scattering mechanism Classify；3) start with from processing method, on the basis of existing polarization characteristic collection, introduce more effective machine learning field Processing method, utilizes polarization information more fully, reaches purpose of classifying.

The focus of polarization SAR statistical modeling and classification always polarization SAR image interpretation, as back scattering matrix S2 The multiple Gaussian distribution of modeling.These distributions typically all meet product model, it is assumed that scattering object and coherent spot obey a certain point Cloth is derived, and is generally directed to covariance matrix C3 or coherence matrix T3 modeling, such as classical multiple Wishart distribution, K distribution, G Distribution and KummerU distribution etc..In recent years, along with the resolution of polarization SAR image is more and more higher, researchers propose multiple Mixture Distribution Model describes heterogeneous or extremely heterogeneous areas statistical property, if Doulgeris is in the multiple dimensioned Gauss of polarization SAR During mixed model, it is proposed that describe the K-Wishart mixed distribution of covariance matrix C3；Gao then polarizes for haplopia and regard more SAR data, it is proposed that multiple Gaussian Mixture distribution and multiple Wishart Mixture Distribution Model, achieves preferable classifying quality.

In existing polarization SAR image classification method based on statistical model, statistical model is both for for the survey that polarizes The back scattering matrix S2 of degree, polarization covariance matrix C3, polarization coherence matrix T3 etc..And these polarization to estimate matrix be all multiple Matrix, theoretical derivation and calculating are the most relative complex；The proposition of mixed model has been further exacerbated by this situation；It addition, these systems Meter model is all according to Polarization scattering theoretical derivation, does not take into full account the physical characteristic of ground object target, i.e. polarization SAR letter Cease and be underutilized.

Summary of the invention

The deficiency existed for prior art, the present invention is from polarization characteristic statistical property, it is provided that one can be abundant Use SAR information and the polarization SAR image classification method based on multiple features combining modeling of amount of calculation can be reduced.

Thinking of the present invention is as follows:

Polarization characteristic for the direct derivation of polarization covariance matrix C3 is modeled, and describes with Alpha-stable distribution The marginal distribution characteristic of polarization characteristic；Then, by Copula theory building Joint Distribution model；Finally, at MRF (Markov Random Field) under framework, in conjunction with Bayes criterion, it is achieved the classification of polarization SAR image.The present invention is with the covariance square that polarizes One group of simple polarization characteristic of battle array C3 direct derivation is modeling object, while considering the physical characteristic of ground object target, also fills Divide the polarization information that make use of image, thus polarization SAR image classification can be carried out exactly.

Matrix (S2, C3, T3) is estimated in the polarization that the modeling object of the present invention is no longer traditional, but by covariance matrix C3 Direct derivation and come one group of simple polarization characteristic vector v (C).When original polarization SAR image format is S2 or T3, need shadow As being converted to C3 form.

The one group of simple polarization characteristic come by the direct derivation of polarization covariance matrix C3 can form polarization characteristic vector v (C), in v (C), each dimension polarization characteristic is all real number, and modeling object is become real vector from complex matrix, and each dimension polarization characteristic is the most anti- Reflect the physical scatterers characteristic of typical feature.In the case of not considering reflection symmetry (Azimuthal symmetry), v (C) It is made up of 9 real polarization characteristics:

$v\left(C\right)=\{\sqrt{{\mathrm{\σ}}_{hh}^0},\sqrt{{\mathrm{\σ}}_{hv}^0},\sqrt{{\mathrm{\σ}}_{vv}^0},|{\mathrm{\ρ}}_{hhhv}|,{\mathrm{\ψ}}_{hhhv},|{\mathrm{\ρ}}_{hhvv}|,{\mathrm{\ψ}}_{hhvv},|{\mathrm{\ρ}}_{hvvv}|,{\mathrm{\ψ}}_{hvvv}\}---\left(1\right)$

In the case of meeting reflection symmetry, there is ρ_hhhv=ρ_hvvv=0, cross polarization can be ignored for information about, v (C) The real vector of 5 elements can be reduced to, it may be assumed that

$v\left(C\right)=\{\sqrt{{\mathrm{\σ}}_{hh}^0},\sqrt{{\mathrm{\σ}}_{hv}^0},\sqrt{{\mathrm{\σ}}_{vv}^0},|{\mathrm{\ρ}}_{hhvv}|,{\mathrm{\ψ}}_{hhvv}\}---\left(2\right)$

Wherein:

It is respectively POLARIZATION CHANNEL hh, the amplitude characteristic of hv, vv；

|ρ_hhhv|、|ρ_hhvv|、|ρ_hvvv| it is respectively the complex phase relation between POLARIZATION CHANNEL hh and hv, between hh and vv, between hv and vv Number feature；

ψ_hhhv、ψ_hhvv、ψ_hvvvIt is respectively the phase contrast feature between POLARIZATION CHANNEL hh and hv, between hh and vv, between hv and vv.

Each polarization characteristic can be derived by simply from covariance matrix C3.

After extracting polarization characteristic vector v (C), the present invention is directed to v (C) propose a kind of general, flexibly statistical distribution build Mould method, mainly includes estimating marginal distribution and tectonic syntaxis two steps of distribution.In v (C), the theory of different polarization characteristics is divided Cloth form is entirely different, and the analytical expression of distribution is extremely complex, estimates that distributed constant is relatively difficult.The present invention uses Alpha-stable distribution describes the statistical property of all polarization characteristics in v (C), and estimates all kinds of atural object difference polarization characteristic Alpha-stable distributed constant, thus obtain distributed model.At Joint Distribution construction phase, utilize distributed model, calculate phase The probability density function (Probability Density Function, PDF) answered and cumulative density function (Cumulative Density Function, CDF), calculate Kendall ' s rank correlation coefficient, and then estimate to obtain the parameter of Copula function, from And obtain Joint Distribution model, it is designated as CoAS model.

Use training sample training to estimate CoAS model parameter, calculate the Posterior probability distribution of polarization SAR image to be sorted, In conjunction with MRF and Bayes taxonomy model, it is optimized by Graph Cuts, it is achieved polarization SAR image based on CoAS model divides Class.

For solving above-mentioned technical problem, the present invention adopts the following technical scheme that

A kind of polarization SAR image classification method based on multiple features combining modeling, including building CoAS model and treating point Class image carries out two steps of classifying, wherein:

Described structure CoAS model farther includes:

(1) from the polarization SAR image of artificial mark, the pixel of all kinds of atural object is gathered as training sample；

All kinds of atural object training samples are performed respectively step (2)～(3):

(2) according to polarization covariance matrix structure polarization characteristic vector v (C) of current class atural object training sample, wherein, It is respectively POLARIZATION CHANNEL hh, the amplitude of hv, vv；|ρ_hhhv|、|ρ_hhvv|、|ρ_hvvv| be respectively between POLARIZATION CHANNEL hh and hv, between hh and vv, hv And the multiple correlation coefficient between vv；ψ_hhhv、ψ_hhvv、ψ_hvvvIt is respectively the phase between POLARIZATION CHANNEL hh and hv, between hh and vv, between hv and vv Potential difference；

(3) the CoAS model of current class atural object is built, particularly as follows:

The 3.1 polarization characteristic vectors utilizing current class atural object training sample, estimate the Alpha-stable of each polarization characteristic Distributed constant；

3.2 calculate probability density function and the integral density letter of each polarization characteristic according to Alpha-stable distributed constant Number；

3.3 density of simultaneous distributions using the Copula function representation current class all polarization characteristics of atural object training sample, adopt Probability density function and cumulative density function with each polarization characteristic estimate Copula function parameter, obtain the CoAS of current class atural object Model；

Described image to be sorted carried out classification farther include:

(4) polarization covariance matrix according to image to be sorted builds the polarization characteristic vector v (C) of image to be sorted；

(5) according to polarization characteristic vector v (C) the tectonic syntaxis posterior probability of image to be sorted, particularly as follows:

5.1 pairs of each polarization characteristics are carried out respectively: use the CoAS model parameter of all kinds of atural object, estimate the spy that currently polarizes respectively Levy the probability density function about all kinds of atural objects, and obtain the cumulative density function of correspondence；

5.2 according to each polarization characteristic about the probability density function of all kinds of atural objects and accumulative density function, use all kinds ofly The CoAS model of thing, calculates the image to be sorted density of simultaneous distribution about all kinds of atural objects, i.e. combines posterior probability；

(6) according to associating posterior probability, in conjunction with MRF model, image to be sorted is classified.

Polarization characteristic vector described in step (2)Wherein, It is respectively POLARIZATION CHANNEL hh, the amplitude characteristic of hv, vv；|ρ_hhhv| for the phase relation between POLARIZATION CHANNEL hh and hv Number；ψ_hhhvFor the phase contrast feature between POLARIZATION CHANNEL hh and hv.

As preferably, the Alpha-stable of each polarization characteristic estimated by MCMC methodology correction sub-step 3.1 is used to divide Cloth parameter, particularly as follows:

1. to use the Alpha-stable distributed constant work of each polarization characteristic estimated based on the Fourier transformation estimation technique For Alpha-stable distributed constant primary quantityAnd default iterations；

2. to the l time iteration gained Alpha-stable distributed constantRespectively with normal distributionStochastic generation current candidate Alpha-stable is distributed Parameterδ_α、δ_β、δ_γ、δ_μTake the most all iteration or part time iteration gained respectively Alpha-stable profile parameter, the standard deviation of β, γ and μ；

3. probability of acceptance Accept of calculating current candidate Alpha-stable distributed constant:

$Accept=\underset{i=1}{\overset{M}{\Π}}\frac{S_{{\mathrm{new}}_{{\widehat{\mathrm{\α}}}_k^d}}\left(x_d^k{}_i\right|\widehat{\mathrm{\β}}_k^d{}_{new},\widehat{\mathrm{\γ}}_k^d{}_{new},\widehat{\mathrm{\μ}}_k^d{}_{new})}{S_{l_{{\widehat{\mathrm{\α}}}_k^d}}\left(x_d^k{}_i\right|\widehat{\mathrm{\β}}_k^d{}_l,\widehat{\mathrm{\γ}}_k^d{}_l,\widehat{\mathrm{\μ}}_k^d{}_l)}$

Wherein,Representing that in kth class training sample, the d of i-th sample ties up polarization characteristic, M is sample number；Represent and use Alpha-stable distributed constantGainedProbability Density function；Represent sampling candidate's Alpha-stable distributed constantGainedProbability density function；

4. stochastic generation obeys the random number u being uniformly distributed U (0,1), if u is ＞ Accept, by the l time iteration gained Alpha-stable distributed constant is as the l+1 time iteration gained Alpha-stable distributed constant；Otherwise, with current candidate Alpha-stable distributed constant is as the l+1 time iteration gained Alpha-stable distributed constant；

5., when iterations reaches to preset iterations, terminate, last iteration gained Alpha-stable is divided Cloth parameter, as revised Alpha-stable distributed constant, uses revised Alpha-stable to divide in sub-step 3.2 Cloth parameter calculates the probability density function of each polarization characteristic.

As preferably, Copula function uses Gumbel function.

Compared to the prior art, the invention have the advantages that and beneficial effect:

(1) modeling object is estimated matrix (S2, C3, T3) from traditional polarization and become simple polarization characteristic vector v (C), object of study becomes real number matrix from complex matrix, reduces the complexity of theoretical derivation, reduces amount of calculation, contributes to The quick interpretation of polarization SAR impact.

(2) modeling object is made up of polarization characteristic v (C), has reacted the physical scatterers characteristic of typical feature target, fully profit With the polarization information of image, it is favorably improved the nicety of grading of polarization SAR image.

(3) a kind of general polarization modeling method is proposed for each dimension polarization characteristic in polarization characteristic vector v (C), Describe the statistical property of each dimension polarization characteristic with Alpha-stable distribution, fitting effect is good, and avoids use theoretical distribution Time parameter estimation and combine modeling difficulty problem.

(4) a kind of flexible, general polarization SAR modeler model, i.e. CoAS mould are proposed for polarization characteristic vector v (C) Type.CoAS model uses Alpha-stable distribution to describe the marginal distribution of each dimension polarization characteristic, has various polarization characteristics Stronger capability of fitting；Then the marginal distribution of polarization characteristic, tectonic syntaxis distributed model is respectively tieed up by Copula functional link. CoAS model also can select sensitive polarization characteristic according to the scattering properties of ground object target, structure any dimension polarization characteristic Joint Distribution, and then carry out polarization SAR image interpretation.

(5) polarization SAR image method of the present invention has the classifying quality of excellence, with traditional polarization based on statistical model SAR image classification method is compared, and the inventive method nicety of grading can improve 5%～10%.

Accompanying drawing explanation

Fig. 1 is test sample in embodiment；

Fig. 2 is the artificial annotation results of Fig. 1 image；

Fig. 3 is 2 dimension polarization characteristicsStatistical modeling result, wherein, figure (a) be actual histogram, scheme (b) For CoAS models fitting result；

Fig. 4 is the inventive method framework；

Fig. 5 is the inventive method flow chart；

Fig. 6 CoAS of the present invention model modeling schematic flow sheet；

Fig. 7 is present invention polarization SAR based on CoAS model image classification schematic flow sheet；

Fig. 8 show CWishart distribution,Distribution and the classification results of the inventive method；

Fig. 9 shows the inventive method and additive method regional area classification results；

Figure 10 shows the sample size impact on sorting technique.

Detailed description of the invention

Understand and implement the present invention for ease of those of ordinary skill in the art, below in conjunction with the accompanying drawings and embodiment is to the present invention It is described in further detail, it will be appreciated that enforcement example described herein is merely to illustrate and explains the present invention, and need not In limiting the present invention.

In the present embodiment, with the RadarSAT-2 8m resolution pole that on April 2nd, 2008, Dutch Flevoland area obtained Change SAR image is as test sample, and polarization SAR image the most to be sorted, according to formula (1) and (2), by polarization SAR image by S2 Format conversion is C3 form, sees Fig. 1, and the Pauli that the figure illustrates these scape data decomposes pseudo-color composite diagram.Type of ground objects in Fig. 1 Artificial annotation results is shown in Fig. 2, uses the types of ground objects such as different colours difference waters, arable land, forest land, building, wherein visual interpretation The region of None-identified is not counted in final precision evaluation.

Flow process of the present invention is shown in Fig. 4～7, elaborates the detailed description of the invention of the present embodiment below in conjunction with accompanying drawing.

Fig. 6 shows CoAS model construction flow process, specifically comprises the following steps that

Step T1, gathers the pixel of all kinds of atural object as training sample from the polarization SAR image of artificial mark.

In the present embodiment, according to artificial annotation results, from Fig. 2, kth class ground object area randomly chooses the pixel of 10% As kth class training sample x^k, k=1,2 ... K, K are atural object classification number, in the present embodiment, and atural object classification number K=4.

Step T2, calculates polarization covariance matrix C of all kinds of training sample₃。

RadarSAT-2 polarization SAR image to the present embodiment, its initial data passes through 2 × 2 complex matrix, i.e. back scattering Matrix S describes, as follows:

$S=\left[\begin{array}{cc}{S}_{hh}& S_{hv}\\ S_{vh}& S_{hh}\end{array}\right]---\left(1\right)$

In formula (1), S_pqRepresent that polarization SAR system is launched p-polarization ripple and accepted the multiple scattering system of the pixel that q polarized wave obtains Number, p can be h or v, q can be h or v.

In the case of single station, S meets heterogeneite, i.e. S_hv=S_vh.S can use 3-dimensional target vector to represent, is defined as κ=[S_hh, S_hv,S_vv]^T, T representing matrix transposition.It is achieved in that polarization covariance matrix C₃:

$C_3=<\mathrm{\&kappa;}\&CenterDot;{\mathrm{\&kappa;}}^H>=\left[\begin{array}{ccc}<S_{hh}{S}_{hh}^{*}>& <S_{hh}{S}_{hv}^{*}>& <S_{hh}{S}_{hh}^{*}>\\ <S_{hv}{S}_{hh}^{*}>& <S_{hv}{S}_{hv}^{*}>& <S_{hv}{S}_{hv}^{*}>\\ <S_{vv}{S}_{hh}^{*}>& <S_{vv}{S}_{hv}^{*}>& <S_{vv}{S}_{hv}^{*}>\end{array}\right]---\left(2\right)$

In formula (2), H represents complex conjugate transposition, and<>represents and regard average treatment, and * represents and takes complex conjugate.

Step T3, according to polarization covariance matrix C₃Construct the polarization characteristic vector v (C) of all kinds of training sample_train。

C₃Being ell rice spy's complex matrix, it describes also by 9 real numbers, and i.e. 3 amplitudes, 3 phase contrasts and 3 are relevant Range value, is defined as follows:

$C_3=\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{hh}^0& \sqrt{{\mathrm{\&sigma;}}_{hh}^0{\mathrm{\&sigma;}}_{hv}^0}{\mathrm{\&rho;}}_{hhhv}& \sqrt{{\mathrm{\&sigma;}}_{hh}^0{\mathrm{\&sigma;}}_{vv}^0}{\mathrm{\&rho;}}_{hhvv}\\ \sqrt{{\mathrm{\&sigma;}}_{hh}^0{\mathrm{\&sigma;}}_{hv}^0}{\mathrm{\&rho;}}_{hhhv}^{*}& {\mathrm{\&sigma;}}_{hv}^0& \sqrt{{\mathrm{\&sigma;}}_{hv}^0{\mathrm{\&sigma;}}_{vv}^0}{\mathrm{\&rho;}}_{hvvv}\\ \sqrt{{\mathrm{\&sigma;}}_{hh}^0{\mathrm{\&sigma;}}_{vv}^0}{\mathrm{\&rho;}}_{hhvv}^{*}& \sqrt{{\mathrm{\&sigma;}}_{hv}^0{\mathrm{\&sigma;}}_{vv}^0}{\mathrm{\&rho;}}_{hvvv}^{*}& {\mathrm{\&sigma;}}_{vv}^0\end{array}\right]---\left(3\right)$

In formula (3):

Represent POLARIZATION CHANNEL hh, the scattering strength of hv, vv respectively；

Represent POLARIZATION CHANNEL hh, the amplitude of hv, vv respectively；

ρ_hhhv、ρ_hhvv、ρ_hvvvRepresent the multiple correlation coefficient between POLARIZATION CHANNEL hh and hv, between hh and vv, between hv and vv respectively, Its exponential form is ρ=| ρ | e^jψ, ψ is phase contrast；

* represent and take complex conjugate.

So, complex matrix C of 3 × 3₃Can be converted into by 9 elementary composition realities vectorial:

$v\left(C\right)=\{\sqrt{{\mathrm{\&sigma;}}_{hh}^0},\sqrt{{\mathrm{\&sigma;}}_{hv}^0},\sqrt{{\mathrm{\&sigma;}}_{vv}^0},|{\mathrm{\&rho;}}_{hhhv}|,{\mathrm{\&psi;}}_{hhhv},|{\mathrm{\&rho;}}_{hhvv}|,{\mathrm{\&psi;}}_{hhvv},|{\mathrm{\&rho;}}_{hvvv}|,{\mathrm{\&psi;}}_{hvvv}\}---\left(4\right)$

Again due to, in the case of meeting reflection symmetry (Azimuthal symmetry), have ρ_hhhv=ρ_hvvv=0, hand over Fork polarization can be ignored, for information about such as city etc., therefore, in the case of meeting reflection symmetry, C₃Can be reduced to:

$C_3=\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{hh}^0& 0& \sqrt{{\mathrm{\&sigma;}}_{hh}^0{\mathrm{\&sigma;}}_{vv}^0}{\mathrm{\&rho;}}_{hhvv}\\ 0& {\mathrm{\&sigma;}}_{hv}^0& 0\\ \sqrt{{\mathrm{\&sigma;}}_{hh}^0{\mathrm{\&sigma;}}_{vv}^0}{\mathrm{\&rho;}}_{hhvv}^{*}& 0& {\mathrm{\&sigma;}}_{vv}^0\end{array}\right]---\left(5\right)$

The reality that v (C) can be reduced to 5 elements is vectorial:

$v\left(C\right)=\{\sqrt{{\mathrm{\&sigma;}}_{hh}^0},\sqrt{{\mathrm{\&sigma;}}_{hv}^0},\sqrt{{\mathrm{\&sigma;}}_{vv}^0},|{\mathrm{\&rho;}}_{hhvv}|,{\mathrm{\&psi;}}_{hhvv}\}---\left(6\right)$

Cross above-mentioned conversion and calculating, by complex matrix C of 3 × 3₃It is converted into real vector v (C).

According to method described in research contents, extract polarization characteristic, according to formula (4) and (6), to all kinds of training samples, point Not building D dimension polarization characteristic vector v (C) of the most each pixel, D dimension polarization characteristic vector v (C) of each pixel is by rows Constitute D dimension polarization characteristic vector v (C) of such training sample_train.V (C) is by back scattering amplitudeComplex phase relation Amplitude and phase place ψ of number ρ are constituted, and can characterize C completely₃In all polarization informations.

Definition and the theoretical distribution model of these polarization characteristics are prior art, repeat no more here, only briefly describe.

Back scattering amplitudeIt meets rayleigh distributed, σ⁰Represent POLARIZATION CHANNEL intensity.

Multiple correlation coefficient and phase contrast thereof can pass through C₃Directly calculating, multiple correlation coefficient ρ is defined as follows:

$\mathrm{\&rho;}=\frac{\frac{1}{L}\underset{i=1}{\overset{L}{\&Sigma;}}{S}_p\left(i\right)S_q^{*}\left(i\right)}{\sqrt{<|S_p{|}^2><|S_q{|}^2>}}---\left(7\right)$

Phase contrast ψ is defined as follows:

$\mathrm{\&psi;}=Arg\left(\frac{1}{L}\underset{i=1}{\overset{L}{\&Sigma;}}{S}_p\right(i\left)S_q^{*}\right(i\left)\right)---\left(8\right)$

In formula (7)～(8):

L is many regarding average real number；

P, q represent the polarization mode of two passages, and p, q can be hh, hv, vv, p ≠ q, the i.e. polarization mode of two passages Can not be identical；

S_pWhen () represents multiple look processing i, i-th backscattering coefficient value under p-channel；

S_qWhen () represents multiple look processing i, i-th backscattering coefficient value under q passage；

* representing and take complex conjugate,<>represents and regards average treatment；

$<|S_p{|}^2>=\frac{1}{L}\underset{i=1}{\overset{L}{\&Sigma;}}{S}_p\left(i\right)S_p^{*}\left(i\right),$ $<|S_q{|}^2>=\frac{1}{L}\underset{i=1}{\overset{L}{\&Sigma;}}{S}_q\left(i\right)S_q^{*}\left(i\right).$

Finally estimate D dimension polarization characteristic vector v (C) of the training sample obtained_train, D=9 here.Consider scattering symmetry In the case of, D=5.Before polarization characteristic extracts, mean filter method can be used C₃Carry out processing to reduce coherent spot impact, window Size W=5.

Step T4, builds CoAS model.

The theoretical distribution model of multiple correlation coefficient and phase contrast relates to Basel, hypergeometric function etc., and form is complicated, parameter Estimate difficulty, and with real data matching bad.Here introduce Alpha-stable distribution and describe all polarization characteristic edges The statistical property of distribution.Alpha-stable distribution meets broad sense central limit theorem, is to be distributed widely than Gaussian Statistical distribution pattern.Alpha-stable is distributed four parameter alpha, β, γ and μ, represent respectively characteristic index, gradient parameter, from Divergence and location parameter, wherein, α ∈ (0,2], β ∈ [-1,1],Alpha-stable is distributed usually not Analysable probability density function, is typically been described by with following characteristic function (Characteristic Function, CF):

$\mathrm{\&phi;}\left(\mathrm{\&omega;}\right)={\&Integral;}_{-\mathrm{\&infin;}}^{\mathrm{\&infin;}}{S}_{\mathrm{\&alpha;}}\left(x\right|\mathrm{\&beta;},\mathrm{\&gamma;},\mathrm{\&mu;})e^{j\mathrm{\&omega;}x}dx=\left\{\begin{array}{c}{e}^{j\mathrm{\&mu;}\mathrm{\&omega;}-|\mathrm{\&gamma;}\mathrm{\&omega;}{|}^{\mathrm{\&alpha;}}\&lsqb;1-jsign\left(\mathrm{\&omega;}\right)\mathrm{\&beta;}\tan \left(\frac{\mathrm{\&pi;}\mathrm{\&alpha;}}{2}\right)\&rsqb;},\mathrm{\&alpha;}\&NotEqual;1\\ e^{j\mathrm{\&mu;}\mathrm{\&omega;}-\left|\mathrm{\&gamma;}\mathrm{\&omega;}\right|\&lsqb;1-jsign\left(\mathrm{\&omega;}\right)\mathrm{\&beta;}\frac{2}{\mathrm{\&pi;}}\ln \left(\right|\mathrm{\&omega;}\left|\right)\&rsqb;},\mathrm{\&alpha;}=1\end{array}\right.\left(9\right)$

In formula (9):

ω represents frequency domain variable, and j represents the imaginary part of symbol；

φ (ω) represents the characteristic function CF of Alpha-stable distribution；

Sign is sign function,

X is time domain variable, and in the present invention, x represents each dimension polarization characteristic；

S_α(x | β, γ, μ) is the probability density function PDF of the Alpha-stable distribution of x.

Although Alpha-stable is distributed the most analysable probability density function, characteristic function still can be passed through Fourier transformation calculating probability density function PDF:

$S_{\mathrm{\&alpha;}}\left(x\right|\mathrm{\&beta;},\mathrm{\&gamma;},\mathrm{\&mu;})=\underset{-\mathrm{\&infin;}}{\overset{+\mathrm{\&infin;}}{\&Integral;}}\mathrm{\&phi;}\left(\mathrm{\&omega;}\right)e^{-j\mathrm{\&omega;}x}d\mathrm{\&omega;}---\left(9\right)$

For ensureing fitting precision, using the Alpha-stable distributed constant estimated by fourier transform method as initially Amount, by Markov Chain-Monte Carlo (Markov Chain Monte Carlo, MCMC) method, is carried out distributed constant Optimizing, iterative estimate number of times is 2000.

In this step, all kinds of training samples are carried out respectively as follows:

Sub-step T4_1, estimates the Alpha-of each polarization characteristic in current class training sample polarization characteristic vector respectively Stable distributed constant.

For ensureing fitting precision, the present invention proposes to use the preferred side of MCMC methodology correction Alpha-stable distributed constant Case, particularly as follows:

To divide as Alpha-stable based on the Alpha-stable distributed constant estimated by the Fourier transformation estimation technique Cloth parameter primary quantityIterations is set to 2000.

The Alpha-stable distributed constant the l time iteration obtained is designated asDivide with normal state respectively ClothStochastic generation current candidate Alpha-stable divides Cloth parameterδ_α、δ_β、δ_γ、δ_μRepresent Alpha-stable profile parameter, β, γ and μ respectively Standard deviation, its value take respectively the most all iteration or part time iteration gained Alpha-stable profile parameter, β, γ and The standard deviation of μ.

Probability of acceptance Accept of calculating current candidate Alpha-stable distributed constant:

$Accept=\underset{i=1}{\overset{M}{\&Pi;}}\frac{S_{{\mathrm{new}}_{{\widehat{\mathrm{\&alpha;}}}_k^d}}\left(x_d^k{}_i\right|\widehat{\mathrm{\&beta;}}_k^d{}_{new},\widehat{\mathrm{\&gamma;}}_k^d{}_{new},\widehat{\mathrm{\&mu;}}_k^d{}_{new})}{S_{l_{{\widehat{\mathrm{\&alpha;}}}_k^d}}\left(x_d^k{}_i\right|\widehat{\mathrm{\&beta;}}_k^d{}_l,\widehat{\mathrm{\&gamma;}}_k^d{}_l,\widehat{\mathrm{\&mu;}}_k^d{}_l)}---\left(10\right)$

In formula (10):

Representing that in kth class training sample, the d of i-th sample ties up polarization characteristic value, M is sample number, in the present invention, sample This refers to pixel；

Represent and use Alpha-stable distributed constantGainedProbability density function；

Represent sampling candidate's Alpha-stable distributed constantGainedProbability density function.

Stochastic generation is obeyed and is uniformly distributed the random number u of U (0,1):

If u is ＞ Accept, then refusal accepts candidate's Alpha-stable distributed constant I.e.

If u≤Accept, then accept candidate's Alpha-stable distributed constantI.e.

Represent the Alpha-stable distributed constant that the l+1 time iteration obtains.

Take last iteration gained Alpha-stable distributed constant as revised Alpha-stable distribution ginseng Number

Sub-step T4_2, makes x_dRepresent that the d of current class training sample ties up polarization characteristic, according to x_dAlpha-stable Distributed constant, calculates x_dProbability density function

In this step, all kinds of training samples all can obtain D probability density function f_1≤d≤D。

Step T5, estimates Copula function parameter.

Under normal circumstances, the form of Alpha-stable multivariate statistical model is more complicated.Assume each dimension polarization characteristic Marginal density obeys Alpha-stable distribution, by each marginal distribution of Copula function link, reaches to construct polynary combine point The purpose of cloth, considers the independence between different variable simultaneously.

Assume x_dIt is to obey a certain distribution f_d(x_d) stochastic variable, then polarization characteristic vector X=(x₁,…,x_d,…,x_D) Density of simultaneous distribution f (x₁,…,x_d,…,x_D) it is:

$f(x_1,...,x_d,...,x_D)=c\left(F_1\right(x_1),...,F_d(x_d),...,F_D(x_D\left)\right)\underset{d=1}{\overset{D}{\&Pi;}}{f}_d\left(x_d\right)---\left(11\right)$

In formula (11):

x₁、…x_d、…x_DRepresent that 1 ... d ... the D of current class training sample ties up polarization characteristic respectively；

f_d(x_d) and F_d(x_d) represent polarization characteristic x respectively_dProbability density function and cumulative density function；

C represents Copula function.

Copula function has multiple, and the present invention selects the Archimedean Copula function Gumbel of single parameter, Gumbel function has analysable expression formula, and parameter estimation is simple.This step utilizes D the probability density that step T4 obtains Function f_1≤d≤DEstimate the unknown parameter θ of Gumbel function, especially by the relation function of θ and Kendall ' s rank correlation coefficient τEstimate.The experience estimation value of Kendall ' the s rank correlation coefficient τ between any two variable R and TFor:

$\widehat{\mathrm{\&tau;}}=\frac{4}{M(M-1)}\underset{1\&le;i<j\&le;M}{\&Sigma;}\&lsqb;I(r_i\&le;r_j)I(t_i\&le;t_j)\&rsqb;-1---\left(12\right)$

In formula (12):

R and T refers to the probability density function of any two polarization characteristic；

r_iAnd r_jRepresent i-th and the jth example of variable R respectively；

t_iAnd t_jRepresent i-th and the jth example of variable T respectively；

M is variable instance number, i.e. sample number, 1≤i≤j≤M.

If r_i≤r_j, then I (r_i≤r_j)=1, otherwise, then I (r_i≤r_j)=-1；I(t_i≤t_j) in like manner.

During D >=2, between any two polarization characteristic R and T, i.e. can get oneAccording toAnd then obtain correspondence Gumbel parameter, take the meansigma methods of all Gumbel parameters as final Gumbel estimates of parametersThat is:

$\widehat{\mathrm{\&theta;}}={\left(\begin{array}{c}D\\ 2\end{array}\right)}^{-1}\underset{1\&le;j_1<j_2\&le;D}{\&Sigma;}{\mathrm{\&tau;}}^{-1}\left({\widehat{\mathrm{\&tau;}}}_{j_1,j_2}\right)---\left(13\right)$

Step T5 has obtained the Gumbel estimates of parameters of current class training sample

All kinds of atural object training samples are performed step T1～T5 respectively, until the CoAS model parameter of all class training samples Estimate complete, thus obtain the CoAS model of all kinds of atural object.Represent kth class training sample CoAS model parameter.In the present embodiment, 1≤k≤K=4,1≤d≤D, D=9 or D=5.Fig. 3 is with 2 dimension polarization characteristicsAs a example by show that the effect of CoAS model modeling, figure (a) they be actual rectangular histogram, figure (b) is CoAS model plan Close result.

The CoAS model parameter of known all kinds of atural object, can carry out polarization SAR image classification, see Fig. 7.

In the present embodiment, with the whole scape data in Holland Flevoland area as test sample, polarization SAR image the most to be sorted, Seeing Fig. 1, size is 1400*1200pixels, and concrete classifying step is as follows:

Step C1, calculates polarization covariance matrix C of each pixel in test sample₃。

Step C2, according to polarization covariance matrix C₃Extract polarization characteristic, the D dimension polarization characteristic vector of structure test sample v(C)_test。

Step C3, the associating posterior probability of structure test sample.

For simplifying, it is considered herein that the inhomogeneity atural object i.e. prior probability of inhomogeneity polarization characteristic is to be uniformly distributed.

This step farther includes sub-step:

Sub-step C3_1, uses the CoAS model parameter of all kinds of atural object, estimates v (C)_testD dimension polarization characteristic about The probability density function of all kinds of atural objects, will be designated as about the probability density function of kth class atural objectD= 1,2,...D。

Step C3_2, according to d dimension polarization characteristic about the probability density function of all kinds of atural objects, estimates v (C)_testD Dimension polarization characteristic, about the cumulative density function of kth class atural object, will be designated as about the cumulative density function of kth class atural object

Step C3_3, according to v (C)_testEach dimension polarization characteristic about the probability density function of all kinds of atural objects and accumulative close Degree function, uses the CoAS model of all kinds of atural object, calculates the test sample density of simultaneous distribution about all kinds of atural objects, i.e. after associating Test probability, test sample is designated as about the associating posterior probability of kth class atural object

Step C4, according to test sample about the associating posterior probability of all kinds of atural objects, enters test sample in conjunction with MRF model Row classification.

Image classification is converted into and minimizes energy problem by MRF model, uses Graph Cut optimization in the present invention, Potts Model Parameter λ=1.1.

Beneficial effect of the present invention will be further illustrated below by experimental result.

Test sample is the RadarSAT-2 8m resolution polarization that on April 2nd, 2008, Dutch Flevoland area obtained SAR image, is shown in that Fig. 1, the artificial annotation results of this area's type of ground objects are shown in Fig. 2.Software platform is MATLAB R2015b (8.6.0). In experiment with classical distributed model CWishart andAs a comparison, according to whether in the case of considering scattering symmetry, this Bright method has made characteristic dimension D=5 and D=9 two groups test respectively, above sorting technique be denoted as respectively CWishart, CoAS₅And CoAS₉.It is right that different sorting techniques are mainly carried out as index by overall accuracy (Overall Accuracy, OA) Ratio.

(a), (b) in Fig. 8, (c) and (d) corresponding CWishart respectively,CoAS₅And CoAS₉Classification results, intuitively From the point of view of, four kinds of method " waters " decomposition result are all fine, reach more than 90%；Generally, CoAS₅And CoAS₉Classification results is It is good,Taking second place, CWishart classification results is worst, especially " builds " region.

Fig. 9 evaluates the performance of different sorting technique further by regional area classification results, and wherein, figure (a) is former Beginning polarization SAR image, (b) is the Google Earth optical picture that figure (a) is corresponding, and what (c) was figure (a) manually marks figure, schemes (d) ～(g) respectively CWishart,CoAS₅And CoAS₉Classification results.From the available conclusion similar to Fig. 8 of Fig. 9, and Also show CoAS₉Compare CoAS₅More preferable to construction area classifying quality, this is also consistent with theory.

Table 1 show different sorting technique to the nicety of grading of all kinds of atural objects and overall classification accuracy, by objectively finger Mark further illustrates the advantage of the inventive method, and nicety of grading improves 5%～10%, especially manually builds " building " etc. The region made, classifying quality is more preferable.Table 2 and table 3 show the inventive method CoAS₅And CoAS₉The confusion matrix of classification results, Depicting the mistake point situation of different classes of, forest land and arable land, forest land and building mistake split-phase more meticulously to seriously, this is mainly Due to 1) plough and forest land belongs to vegetation pattern, owing to the crops of plantation are different, the scene making arable land is more complicated；2) build Near building, it will usually plantation trees.

Figure 10 shows the number of samples impact on different sorting techniques.Transverse axis is sample number, and the longitudinal axis is nicety of grading, often Secondary experiment randomly selects the sample of some as training sample, and number of samples, from 5～500, has carried out 7 groups altogether and repeated real Test.Result shows sorting technique stable performance of the present invention, nicety of grading ratioStably exceed 5%, higher by 10% than CWishart.

Table 1 distinct methods classification results contrast (unit: %)

Table 2 CoAS₅The confusion matrix (unit: %) of classification results

Table 3 CoAS₉The confusion matrix (unit: %) of classification results

Above in association with concrete example, the present invention is described.Embodiment shown in above-mentioned, is not intended to limit this Bright, it is also not limited to Radarsat-2 polarization SAR image, applies also for ALOS-2/PALSAR, TerraSAR-X, EVNISAT/ Other spaceborne or airborne polarization SAR data such as ASAR, SIR-C, AIRSAR, EMISAR, PISAR.

The preferred embodiment of the present invention is described in detail above in association with accompanying drawing, but, the present invention is not limited to above-mentioned reality Execute the detail in mode, in the technology concept of the present invention, technical scheme can be carried out multiple letter Monotropic type, these simple variant belong to protection scope of the present invention.

Any the technical staff in the technical field of the invention, without departing from the spirit and scope of the present invention, it should Various changes can be made and amendment.Therefore protection scope of the present invention should with appended claims defined in the range of Accurate.

Claims (4)

1. a polarization SAR image classification method based on multiple features combining modeling, is characterized in that, including:

Build CoAS model and two steps that image to be sorted is classified, wherein:

Described structure CoAS model farther includes:

(1) from the polarization SAR image of artificial mark, the pixel of all kinds of atural object is gathered as training sample；

All kinds of atural object training samples are performed respectively step (2)～(3):

(2) according to polarization covariance matrix structure polarization characteristic vector v (C) of current class atural object training sample, wherein, It is respectively POLARIZATION CHANNEL hh, the amplitude of hv, vv；|ρ_hhhv|、|ρ_hhvv|、|ρ_hvvv| be respectively between POLARIZATION CHANNEL hh and hv, between hh and vv, hv And the multiple correlation coefficient between vv；ψ_hhhv、ψ_hhvv、ψ_hvvvIt is respectively the phase between POLARIZATION CHANNEL hh and hv, between hh and vv, between hv and vv Potential difference；

(3) the CoAS model of current class atural object is built, particularly as follows:

The 3.1 polarization characteristic vectors utilizing current class atural object training sample, estimate the Alpha-stable distribution of each polarization characteristic Parameter；

3.2 calculate probability density function and the cumulative density function of each polarization characteristic according to Alpha-stable distributed constant；

3.3 density of simultaneous distributions using the Copula function representation current class all polarization characteristics of atural object training sample, use each The probability density function of polarization characteristic and cumulative density function estimate Copula function parameter, obtain the CoAS mould of current class atural object Type；

Described image to be sorted carried out classification farther include:

(4) polarization covariance matrix according to image to be sorted builds the polarization characteristic vector v (C) of image to be sorted；

(5) according to polarization characteristic vector v (C) the tectonic syntaxis posterior probability of image to be sorted, particularly as follows:

5.1 pairs of each polarization characteristics are carried out respectively: use the CoAS model parameter of all kinds of atural object, estimate that current polarization characteristic closes respectively In the probability density function of all kinds of atural objects, and obtain the cumulative density function of correspondence；

5.2 according to each polarization characteristic about the probability density function of all kinds of atural objects and accumulative density function, use all kinds of atural object CoAS model, calculates the image to be sorted density of simultaneous distribution about all kinds of atural objects, i.e. combines posterior probability；

(6) according to associating posterior probability, in conjunction with MRF model, image to be sorted is classified.

2. the polarization SAR image classification method modeled based on multiple features combining as claimed in claim 1, is characterized in that: step (2) the polarization characteristic vector described inWherein, Point Wei POLARIZATION CHANNEL hh, the amplitude characteristic of hv, vv；|ρ_hhhv| for the correlation coefficient between POLARIZATION CHANNEL hh and hv；ψ_hhhvLogical for polarization Phase contrast feature between road hh and hv.

3. the polarization SAR image classification method modeled based on multiple features combining as claimed in claim 1 or 2, is characterized in that:

The Alpha-stable distributed constant of the employing each polarization characteristic estimated by MCMC methodology correction sub-step 3.1, particularly as follows:

1. using use based on the Fourier transformation estimation technique estimate each polarization characteristic Alpha-stable distributed constant as Alpha-stable distributed constant primary quantityAnd default iterations；

2. to the l time iteration gained Alpha-stable distributed constantRespectively with normal distributionStochastic generation current candidate Alpha-stable is distributed Parameterδ_α、δ_β、δ_γ、δ_μTake the most all iteration or part time iteration gained respectively Alpha-stable profile parameter, the standard deviation of β, γ and μ；

3. probability of acceptance Accept of calculating current candidate Alpha-stable distributed constant:

$\mathrm{Accept}=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}\frac{S_{{\widehat{\mathrm{\&alpha;}}}_k^d_{\mathrm{new}}}\left(x_d^k_i\right|{\widehat{\mathrm{\&beta;}}}_k^d_{\mathrm{new}},{\widehat{\mathrm{\&gamma;}}}_k^d_{\mathrm{new}},{\widehat{\mathrm{\&mu;}}}_k^d_{\mathrm{new}})}{S_{{\widehat{\mathrm{\&alpha;}}}_k^d_l}\left(x_d^k_i\right|{\widehat{\mathrm{\&beta;}}}_k^d_l,{\widehat{\mathrm{\&gamma;}}}_k^d_l,{\widehat{\mathrm{\&mu;}}}_k^d_l)}$

Wherein,Representing that in kth class training sample, the d of i-th sample ties up polarization characteristic, M is sample number；Represent and use Alpha-stable distributed constantGainedProbability Density function；Represent sampling candidate's Alpha-stable distributed constantGainedProbability density function；

4. stochastic generation obeys the random number u being uniformly distributed U (0,1), if u is ＞ Accept, by the l time iteration gained Alpha- Stable distributed constant is as the l+1 time iteration gained Alpha-stable distributed constant；Otherwise, with current candidate Alpha- Stable distributed constant is as the l+1 time iteration gained Alpha-stable distributed constant；

5. when iterations reaches to preset iterations, terminate, by last iteration gained Alpha-stable distribution ginseng Number, as revised Alpha-stable distributed constant, uses revised Alpha-stable distribution ginseng in sub-step 3.2 Number calculates the probability density function of each polarization characteristic.

4. the polarization SAR image classification method modeled based on multiple features combining as claimed in claim 1 or 2, is characterized in that:

Described Copula function is Gumbel function.