CN104881867A - Method for evaluating quality of remote sensing image based on character distribution - Google Patents

Abstract

The invention discloses a method for evaluating the quality of a remote sensing image based on character distribution. The method comprises the following steps: sampling each category of an input remote sensing image, and extracting gray scales and characters of each category; calculating a mean value and a variance of each category in a characteristic sample, and taking the mean value and the variance as initial values of EM estimation (GMM estimation) of a Gauss model for performing GMM estimation of a Gauss mixed model; according to the variance and weight of characters, obtained according to GMM estimation, of each category, and calculating a kappa coefficient representing the quality of the remote sensing image by constructing a quality evaluation model based on the kappa coefficient. When the method is used for evaluating the quality of the remote sensing image, image characters of the remote sensing image are fully utilized, and verification of the evaluation method is carried out by utilizing an image sorting method, and objective evaluation of the quality of the image is effectively performed. The invention provides the Gauss mixed model with a fixed mean value, and the model is good in convergence.

Description

A kind of Remote Sensing Image Quality evaluation method of feature based distribution

Technical field

The invention belongs to Remote Sensing Image Quality and evaluate application, be specifically related to the Classification Precision of RS Images estimation model of the computing method of Remote Sensing Image Texture feature, the method for estimation of grain distribution and feature based distribution.

Background technology

At present, the development of various countries' remote sensing is in the ascendant, and China has also started the plan of high resolving power earth observation.Along with the enforcement of this plan, China will obtain the remotely-sensed data of a large amount of autonomous property right.In order to the ability utilizing these data to increase substantially the autonomous earth observation information of China, the image interpretation comprising Images Classification etc. is an indispensable sport technique segment.

Current image quality evaluating method mainly carries out from the angle of subjective assessment, and with regard to machine image classification, these methods cannot reflect the separability of machine sort.The subjective assessment of image is higher, has great help to artificial interpretation, but for machine sort, cannot measure the separability for remote sensing images terrain classification feature used and stability.

The present invention propose feature based distribution image quality evaluating method (wherein feature comprises gray scale, energy, contrast, unfavourable balance square, entropy, correlativity), from the angle research satellite full-colour image method for evaluating quality of remote sensing images terrain classification, this and traditional quality evaluation laying particular emphasis on image fidelity have far different implication, have important actual application value.

Summary of the invention

In machine sort, whether have good separability and stability for remote sensing images, namely how to remove to evaluate its quality in machine sort, the present invention proposes the image quality evaluating method of a kind of feature based distribution, its concrete steps are as follows:

S1: each atural object classification in image is sampled, extract gray scale and textural characteristics: be provided with remote sensing images I, the subregion choosing each classification in image as sample (according to), point centered by each pixel in sample, size of windowing is K × K (K gets 5 ~ 21), in this image block, add up the gray level co-occurrence matrixes of its central point, calculate textural characteristics, process is as follows:

(1) gray level co-occurrence matrixes statistics: get any point (x, y) in image block (K × K) and depart from its another point (x+a, y+b), if the right gray-scale value of this point is (g ₁, g ₂).Make point (x, y) move on whole picture, then can obtain various (g ₁, g ₂) value, if the progression of gray-scale value is k, then (g ₁, g ₂) combination have k ²kind.For whole picture, count each (g ₁, g ₂) value occur number of times, be then arranged in a square formation, then use (g ₁, g ₂) they are normalized to the probability P (g of appearance by the total degree that occurs ₁, g ₂), such square formation is called gray level co-occurrence matrixes.Range difference score value (a, b) gets different combinations of values, obtains the joint probability matrix under different combinations of values.

(2) textural characteristics calculates: the feature of conventional description texture has energy (Energy), contrast (Contrast), unfavourable balance square (Inverse Difference Moment), entropy (Entropy) and is correlated with (Correlation), wherein, L is gray level for normalized gray level co-occurrence matrixes, energy is used for weighing the homogeneity of intensity profile, and its computing formula is:

$f_1=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}\hat{p}{(i,j)}^2$

Contrast reflects the change severe degree of local grain, and its computing formula is:

$f_2=\underset{|i-j|=0}{\overset{L-1}{\mathrm{\Σ}}}{(i-j)}^2\left\{\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}\hat{p}(i,j)\right\}$

Unfavourable balance apart from the homogeney of reflection image texture, tolerance image texture localized variation number.Lack change between the zones of different that its value greatly then illustrates image texture, local is very even, and its computing formula is:

$f_3=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}\frac{p(i,j)}{1+{(i-j)}^2}$

Entropy is the tolerance of the quantity of information had in image, and the complexity of texture is higher just means that amount of image information is larger, and its entropy is larger, and formula is as follows:

$f_4=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}-\hat{p}(i,j)\ln \left(\hat{p}(i,j)\right)$

Each value in correlation metric tolerance gray level co-occurrence matrixes is expert at and the similarity degree on row.Therefore, the size of correlativity reflects the correlativity of local gray level distribution in image.When matrix element value even equal time, relevant just large; On the contrary, if matrix pixel value differs greatly, correlation is little.Wherein μ _x, μ _yfor being respectively the average on gray level co-occurrence matrixes line direction and column direction, δ _x, δ _ybe respectively the variance on gray level co-occurrence matrixes line direction and column direction, its computing formula is as follows:

$f_5=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}\frac{(i*j)\hat{p}(i,j)-{\mathrm{\μ}}_x{\mathrm{\μ}}_y}{{\mathrm{\δ}}_x{\mathrm{\δ}}_y}$

In formula:

${\mathrm{\μ}}_x=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}\text{i*}\hat{p}(i,j)$

${\mathrm{\μ}}_y=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}i*\hat{p}(i,j)$

${\mathrm{\δ}}_x=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}{(i-{\mathrm{\μ}}_x)}^2*\hat{p}(i,j)$

${\mathrm{\δ}}_y=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}{(j-{\mathrm{\μ}}_y)}^2*\hat{p}(i,j)$

S2: GMM (gauss hybrid models) parameter estimation of average value constraint: in remote sensing image all kinds of atural object due to some atural object feature distribution extremely similar, the probability density curve aliasing of feature is more serious.So calculate the sample characteristics extracted based on the GMM algorithm of the feature proposition interpolation average value constraint of remote sensing images herein.Its computation process is as follows:

(1) set every class average of current signature in present image as m ₁... m _c, then initial parameter value is made to be θ ⁰={ a ₁..., a _c, m ₁... m _c, δ ₁... δ _c; Wherein a ₁..., a _cfor the class weight of every class, this weight is determined according to image atural object prior distribution by user, as without priori, is just defaulted as 1/C, is namely evenly distributed, and meets δ ₁... δ _cfor eigenwert variance, C is classification number;

(2) by θ ⁰iteration obtains θ t time ^t, t is iterations, utilizes parameter value θ ^tcalculate the posterior probability β that current pixel belongs to jth class _j(x), j=1,2 ... C, calculating formula is as follows:

${\mathrm{\β}}_j\left(x\right)=\frac{a_{j}g(x,{\mathrm{\μ}}_j,{\mathrm{\δ}}_j)}{{\mathrm{\Σ}}_{k=1}^{C}g(x,{\mathrm{\μ}}_k,{\mathrm{\δ}}_k)}$

Wherein g (x, μ _j, δ _j) for average be μ _j, variance is δ _jthe probability density function of Gaussian distribution, namely

$g(x,{\mathrm{\μ}}_j,{\mathrm{\δ}}_j)=\frac{1}{{\left(2\mathrm{\π}\right)}^{1/2}{\left|{\mathrm{\δ}}_j\right|}^{1/2}}\exp (-[\frac{1}{2}{(x-{\mathrm{\μ}}_j)}^T{{\mathrm{\δ}}_j}^{-1}(x-{\mathrm{\μ}}_j)\left]\right);$

(3) fixing average, with θ ⁰={ a ₁..., a _c, m ₁... m _c, δ ₁... δ _cbe initial value, each parameter of iterative computation, comprises atural object class weight, covariance, and computing formula is:

${\mathrm{\β}}_j\left(x\right)=\frac{a_{j}g(x,{\mathrm{\μ}}_j,{\mathrm{\δ}}_j)}{{\mathrm{\Σ}}_{k=1}^{C}g(x,{\mathrm{\μ}}_k,{\mathrm{\δ}}_k)},j=\mathrm{1,2}..C,$

$a_j^{\mathrm{new}}=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^N{\mathrm{\β}}_j\left(x_i\right),j=\mathrm{1,2}..C,$

${\mathrm{\δ}}_j^{\mathrm{new}}=\frac{{\mathrm{\Σ}}_{i=1}^N{\mathrm{\β}}_j\left(x_i\right)(x_i-{\mathrm{\μ}}_j){(x_i-{\mathrm{\μ}}_j)}^T}{{\mathrm{\Σ}}_{i=1}^N{\mathrm{\β}}_j\left(x_i\right)},j=\mathrm{1,2},...C,$

N is number of pixels in image;

Repeat (2) (3), if || θ ^new-θ || < ζ, ζ are that (choose according to accuracy requirement, the present invention gets and is less than 10 error amount ^-5), iteration stopping, being then fixed current average is m ₁... m _ctime weight and variance evaluation result.

S3: build the Environmental Evaluation Model based on kappa coefficient:

Kappa coefficient is a kind of Measure Indexes evaluating overall precision, and it can as the consistency check of the classification results of sample and real type of ground objects.Its formula is as follows:

$\mathrm{Kappa}=\frac{N\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}{n}_{\mathrm{ii}}-\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}{N}_{i+}{N}_{+i}}{N^2-\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}{N}_{i+}{N}_{+i}}=\frac{\frac{\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}{n}_{\mathrm{ii}}}{N}-\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\frac{N_{i+}{N}_{+i}}{N^2}}{1-\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\frac{N_{i+}{N}_{+i}}{N^2}}$ (formula 1)

Make P _kbe the class weight calculated by S2, as prior probability, and have p _kkbelong to kth class under representing truth, be also correctly classified as the pixel ratio of kth class; P _kaddrepresent reality and do not belong to kth class, and be divided into the pixel ratio of kth class by mistake.Utilize P _k, P _kkand P _kadd, Kappa coefficient can be rewritten become directly related with classification results, the univocal new model of each several part, as follows:

$\mathrm{kappa}=1-\frac{{\mathrm{\&Sigma;}}_{k=1}^C{P}_{\mathrm{kadd}}}{{\mathrm{\&Sigma;}}_{k=1}^C\left[P_k(P_{\mathrm{kk}}+P_{\mathrm{kadd}})\right]}$ (formula 2)

From above formula, estimating of Kappa coefficient depends on P _k, P _kkand P _kaddsolve.

Due to the distributional pattern that Gaussian distribution is common, suppose that all kinds of characters of ground object all obeys Multi-dimensional Gaussian distribution here.For one-dimensional characteristic, if μ ₁, μ ₂... μ _c(μ ₁< μ ₂< ... < μ _cbe respectively all kinds of averages; σ ₁, σ ₂... σ _cfor the standard deviation that correspondence is all kinds of; P ₁, P ₂..P _cfor the prior probability that correspondence is all kinds of; Function phi (x) represents the cumulative distribution function of standardized normal distribution; h _j(1≤j < C) represents the categorised decision face between jth class and jth+1 class, then optional k=1,2..C, P _kkand P _kaddcan be expressed as:

$P_{\mathrm{kk}}=\left\{\begin{array}{c}P\{x\&le;h_k|x\&Element;C_k\},k=1\\ P\{h_{k-1}<x\&le;h_k|x\&Element;C_k\},2\&le;k<C\\ P\{x>h_k|x\&Element;C_k\},k=C\end{array}\right.$ (formula 3)

$P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=1,j!=k}^C{P}_j*P\{h_{k-1}\&le;x<h_k|x\&Element;C_k\}$ (formula 4)

Therefore, different interphase h is obtained based on different decision-making techniques _j(1≤j < C), P _kkand P _kaddvalue change thereupon.The present invention is based on minimum distance criterion.

Minimum distance criterion is a kind of common sorting criterion.It utilizes different classes of terrain object attribute difference comparatively large, and the principle that in same classification, terrain object attribute difference is less, classify to the distance at all kinds of center according to pixel or object.If pixel or object minimum to the class centre distance of a certain class, then this pixel or object are then marked as this type of.Such as, for gray level image, if x _ifor any one pixel, μ _jthe center that (j=1,2 ..C) is jth class, and then x _ibe classified t class.

Based on the hypothesis of above-mentioned normal distribution, according under minimum distance classification criterion, the interphase of jth class and jth+1 class can be expressed as:

$h_j=\frac{{\mathrm{\&mu;}}_j+{\mathrm{\&mu;}}_j+1}{2}(1\&le;j<C)$

Then to different k, utilize probability theory, (formula 3), (formula 4) can be calculated as follows:

If 1. k=1,

$\left\{\begin{array}{c}{P}_{\mathrm{kk}}=P_k\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)\\ P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=2}^C{P}_j\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)\end{array}\right.$ (formula 5)

If 2. 1 < k < C,

$\left\{\begin{array}{c}{P}_{\mathrm{kk}}=P_k[\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\\ P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=2}^C{P}_j[\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\end{array}\right.$ (formula 6)

If 3. k=C,

$\left\{\begin{array}{c}{P}_{\mathrm{kk}}=P_k[1-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\\ P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=1}^{C-1}{P}_j[1-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\end{array}\right.$ (formula 7)

(formula 2) namely can be used as the image quality evaluation model of Graph-Oriented picture classification, and (formula 5) ~ (formula 7) is the image quality evaluation model circular under all kinds of equal Normal Distribution.When sorting criterion and characteristic of division are determined, the classification performance of piece image can estimate according to (formula 5) calculating K appa coefficient completely.

Accompanying drawing explanation

Fig. 1: the basic flow sheet of the inventive method.

Fig. 2: experimental data

Fig. 3: the data in German KOB area, therefrom intercepts field, forest land and three kinds, city type of ground objects, builds the picture of 200*600 size

Fig. 4 is based on average value constraint GMM estimated result and GMM estimated result (gray feature) comparison diagram in the present invention;

Fig. 5 is the present invention and the true kappa coefficient calculations results contrast figure based on minimum distance classification.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with accompanying drawing and exemplary embodiment, the present invention is further elaborated.As shown in Figure 1, detailed process of the present invention is:

(1) texture feature extraction: the texture blending carrying out based on gray level co-occurrence matrixes to the size of the input remote sensing images that are 600 × 200, the present embodiment windows size for getting 21 × 21.

(2) average initial value is obtained: to original remote sensing images, with atural object classification in image for standard (the present embodiment image category C gets 3), choose each classification image sampling, count gray level co-occurrence matrixes, calculate average and the variance of each feature, select gray feature to be the feature of the present embodiment, estimate average initial value as GMM.

(3) the GMM parameter estimation of fixing average: initial parameter value θ ⁰={ a ₁, a ₂, a ₃, m ₁, m ₂, m ₃, δ ₁, δ ₂, δ ₃}={ 0.33,0.33,0.33,65.6,14.8,117.75,20.4,7.34,61.39}, utilize parameter value θ ^tcalculate posterior probability β _jx (), j=1,2 ..C, t are iterations.Fixing average, calculates each parameter (weight, covariance):

Average is μ _j, variance is δ _jthe probability density function of Gaussian distribution be:

$g(x,{\mathrm{\&mu;}}_j,{\mathrm{\&delta;}}_j)=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{1/2}{\left|{\mathrm{\&delta;}}_j\right|}^{1/2}}\exp (-[\frac{1}{2}{(x-{\mathrm{\&mu;}}_j)}^T{{\mathrm{\&delta;}}_j}^{-1}(x-{\mathrm{\&mu;}}_j)\left]\right)$

Substitute into formula:

${\mathrm{\&beta;}}_j\left(x\right)=\frac{a_{j}g(x,{\mathrm{\&mu;}}_j,{\mathrm{\&delta;}}_j)}{{\mathrm{\&Sigma;}}_{k=1}^{C}g(x,{\mathrm{\&mu;}}_k,{\mathrm{\&delta;}}_k)},j=\mathrm{1,2},3,$

$a_j^{\mathrm{new}}=\frac{1}{n}{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&beta;}}_j\left(x_i\right),j=\mathrm{1,2},3,$

${\mathrm{\&delta;}}_j^{\mathrm{new}}=\frac{{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&beta;}}_j\left(x_i\right)(x_i-{\mathrm{\&mu;}}_j){(x_i-{\mathrm{\&mu;}}_j)}^T}{{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&beta;}}_j\left(x_i\right)},j=\mathrm{1,2},3$

Double counting β _j(x) and each parameter, if || θ ^new-θ || < ζ, ζ are that (choose according to accuracy requirement, the present embodiment is taken at 10 to error amount ^-5), iteration stopping, then obtain current average for { m ₁, m ₂, m ₃other parameter estimation result { a when }={ 65.6,14.8,117.75} ₁, a ₂, a ₃, δ ₁, δ ₂, δ ₃}={ 0.2992,0.3022,0.3985,21.45,7.82,63.00}

(4) image classification accuracy index k appa coefficient calculations: calculate corresponding kappa coefficient by the weight of the corresponding average obtained and variance,

$\mathrm{kappa}=1-\frac{{\mathrm{\&Sigma;}}_{k=1}^C{P}_{\mathrm{kadd}}}{{\mathrm{\&Sigma;}}_{k=1}^C\left[P_k(P_{\mathrm{kk}}+P_{\mathrm{kadd}})\right]}$

The kappa coefficient calculating present image is 0.56.

Claims (3)

1. a Remote Sensing Image Quality evaluation method for feature based distribution, is characterized in that comprising the steps:

(1) obtain remote sensing images, intuitively obtained the atural object classification number C of image by image naked eyes;

(2) atural object all kinds of in image is sampled respectively, according to sample, generate gray level co-occurrence matrixes; According to co-occurrence matrix, calculate the sample characteristics of all kinds of atural object, comprise energy, contrast, unfavourable balance square, entropy and correlativity;

(3) calculate average and the variance of all kinds of ground object sample feature, as the initial parameter value of parameter estimation, carry out gauss hybrid models parameter estimation, obtain the class weight α of all kinds of atural object of remote sensing images _k, covariance sigma _k, k=1,2 ... C;

(4) computed image nicety of grading index coefficient k appa, comprises following sub-step:

(4.1) categorised decision face h is set up _j: $h_j=\frac{{\mathrm{\&mu;}}_j+{\mathrm{\&mu;}}_{j+1}}{2},1\&le;j<C,$

Wherein, μ _j(j=1,2 ..C) is average, is the center of the corresponding Gaussian spatial of jth class atural object;

(4.2) P is calculated _kk, P _kadd:

If k=1, then

$\left\{\begin{array}{c}{P}_{\mathrm{kk}}=P_k\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)\\ P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=2}^C{P}_j\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)\end{array}\right.;$

If 1 < k < C,

$\left\{\begin{array}{c}{P}_{\mathrm{kk}}=P_k[\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\\ P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=2}^C{P}_j[\mathrm{\&Phi;}\left(\frac{h_k-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\end{array}\right.;$

If k=C,

$\left\{\begin{array}{c}{P}_{\mathrm{kk}}=P_k[1-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\\ P_{\mathrm{kadd}}={\mathrm{\&Sigma;}}_{j=1}^{C-1}{P}_j[1-\mathrm{\&Phi;}\left(\frac{h_{k-1}-{\mathrm{\&mu;}}_k}{{\mathrm{\&sigma;}}_k}\right)]\end{array}\right.;$

In formula, Φ is gauss of distribution function; P _kit is the class weight calculated by step (3); P _kkbelong to kth class under representing truth, be also correctly classified as the pixel ratio of kth class; P _kaddrepresent reality and do not belong to kth class, and be divided into the pixel ratio of kth class by mistake;

(4.3) nicety of grading index k appa is calculated:

$\mathrm{kappa}=1-\frac{{\mathrm{\&Sigma;}}_{k=1}^C{P}_{\mathrm{kadd}}}{1-{\mathrm{\&Sigma;}}_{k=1}^C\left[P_k(P_{\mathrm{kk}}+P_{\mathrm{kadd}})\right]}.$

2. method according to claim 1, is characterized in that, carries out gauss hybrid models parameter estimation and comprise following sub-step in described step (3):

(2.1) every class average in present image is established to be respectively m ₁... m _c, then initial parameter value is made to be θ °={ a ₁..., a _c, m ₁... m _c, δ ₁... δ _c; Wherein a ₁..., a _cfor the class weight of every class atural object, this weight is determined according to image atural object prior distribution by user, as without priori, is just defaulted as and is evenly distributed, a _i=1/C, meets δ ₁... δ _cfor eigenwert variance, C is classification number; G (x, μ _j, δ _j) be the probability density function of Gaussian distribution:

$g(x,{\mathrm{\&mu;}}_j,{\mathrm{\&delta;}}_j)=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{1/2}{\left|{\mathrm{\&delta;}}_j\right|}^{1/2}}\exp (-[\frac{1}{2}{(x-{\mathrm{\&mu;}}_j)}^T{{\mathrm{\&delta;}}_j}^{-1}(x-{\mathrm{\&mu;}}_j)\left]\right),$

In formula, μ _jfor average, δ _jfor variance;

(2.2) μ is made _j=m _j, j=1,2..C, iterative computation

${\mathrm{\&beta;}}_j\left(x\right)=\frac{a_{j}g(x,{\mathrm{\&mu;}}_j,{\mathrm{\&delta;}}_j)}{{\mathrm{\&Sigma;}}_{k=1}^{C}g(x,{\mathrm{\&mu;}}_k,{\mathrm{\&delta;}}_k)},j=\mathrm{1,2},...C,$ (formula 1)

$a_j=\frac{1}{n}{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&beta;}}_j\left(x_i\right),j=\mathrm{1,2}..C,$ (formula 2)

${\mathrm{\&delta;}}_j=\frac{{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&beta;}}_j\left(x_i\right)(x_i-{\mathrm{\&mu;}}_j){(x_i-{\mathrm{\&mu;}}_j)}^T}{{\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&beta;}}_j\left(x_i\right)},j=\mathrm{1,2},...C,$ (formula 3)

N is the number of pixels of all kinds of atural object in sample; X is the variable representing pixel, β _jx () represents that current pixel belongs to the posterior probability of jth class;

By (formula 1)-(formula 3) iterative computation, until reach predetermined iteration precision ζ, iteration stopping, being fixed average is m _jtime class weight α _jwith variance δ _jestimated result.

3. method according to claim 2, is characterized in that, in described step (2.2), iteration stopping condition is: || θ ^t+1-θ ^t|| < ζ,

Wherein, θ ^tbe the t time iteration result, t is iterations, and ‖ ‖ represents and asks Euclidean distance, and predetermined iteration precision ζ is 10 ^-5-10 ^-6, according to computational accuracy and speed of convergence choosing comprehensively.