CN101853506A - High optical spectrum image end member extraction method based on optimized search strategy - Google Patents

Abstract

The invention provides a high optical spectrum image end member extraction method based on an optimized search strategy. The invention is technically characterized in that based on feature space simplex maximum size conversion idea and combined with the optimized search strategy, the invention extracts end members in order by increasing simplex size step by step in a feature space. The invention estimates the actual ground feature number of a high optical spectrum remote sensing image with a virtual dimensionality method at first and reduces the dimensionality of the image with a PCA method. In an optional initial end member collection, the simplex size is amplified and converted step by step in the corresponding search area of each alternative unit and finally all end members are output. Experiments prove that the algorithm provision can reduce operation, save operation time and obtain stable and accurate end member extraction result. The invention overcomes the defects of huge operation, high time complexity, unstable final end member collection output and being vulnerable to the influence of the initial end member collection of the current high optical spectrum remote sensing image automatic end member extraction method based on the simplex maximum size conversion idea.

Description

A kind of high optical spectrum image end member extraction method based on the optimization searching strategy

Technical field

The present invention relates to a kind of high optical spectrum image end member extraction method, belong to a kind of end member extraction method of high-spectrum remote sensing, belong to high-spectrum remote-sensing Digital Image Processing and area of pattern recognition based on the optimization searching strategy.

Background technology

The face of land high-spectrum remote sensing that obtains by the high spectrum sensor of space has comprised the radiation intensity information of abundant atural object spatial information, a large amount of narrow-band spectrum information and face of land thing, simultaneously, still be the angle of microcosmic from macroscopic view no matter, high-spectrum remote sensing can both reflect that a certain regional complicated earth surface thing is launched, the electromagnetic wave information of reflection, scattering, absorption.Based on its extremely huge abundant packets of information content, high-spectrum remote sensing has become the powerful instrument in remote sensing field, be widely used in directions such as atural object discovery and mineral, rock, crops constituent analysis, simultaneously, the corresponding research focus that becomes of high spectrum digital image processing method.

In high-spectrum remote sensing, the existence of mixed pixel has become a kind of inevitable, and so-called mixed pixel promptly in a pixel of high-spectrum remote sensing, has comprised more than a kind of terrestrial object information.Cause the reason of mixed pixel to have a lot, wherein main cause has 2 points: the one, and the complicacy of atural object causes the ground table section of a pixel correspondence of high-spectrum remote sensing not have single atural object, and imaging unit must be mixed pixel thus; The 2nd, have to sacrifice a kind of compromise means of spatial resolution in order to obtain higher spectral resolution, make a pixel area---be the area of the corresponding actual ground of a pixel in the high-spectrum remote sensing, greater than single atural object area occupied, cause the image area coverage rate excessive, a pixel comprises a plurality of terrestrial object informations, mixed pixel occurs.Mixed pixel is prevalent in the high-spectrum remote sensing, analysis and application to high spectrum image have caused very big difficulty, pixel is separated and mixed is the prerequisite of many analytical applications, for example high-spectrum remote sensing classification, identification, Spectral matching etc., and previous work all needs mixed pixel is separated mixed.

End member extracts, and promptly extracts the pixel that only contains a kind of atural object in the high-spectrum remote sensing---end member, is that pixel is separated mixed important prerequisite, end member the unknown, and pixel is separated to mix and is not also known where to begin.Yet the mass data in the high-spectrum remote sensing, and the existence of complicated correlativity between the pixel make end member extraction work become a very thorny difficult problem.At present, the end member extraction algorithm of high-spectrum remote sensing has a lot, from the angle that whether needs the user to get involved, can be divided into that automatic end member extracts and the supervision end member extracts, and from the order that end member extracts, can be divided into that disposable end member extracts and order end member extraction etc.

In existing end member extraction algorithm, algorithm based on feature space monomorphous maximum volume conversion (MVT) thought occupies critical role, this thought is considered as a point (dimension of feature space equals the spectral band number) in the high-dimensional feature space with each the spectrum vector in the high-spectrum remote sensing, because each mixed pixel can be gone out by all end member linear lists, therefore in high-dimensional feature space, represent high-spectrum remote sensing spectrum vector a convex closure in the composition characteristic space is arranged, and this convex closure is positioned at by with the end member being the convex surface monomorphous that the summit constitutes.High-spectrum remote sensing is based on the theoretical foundation of the automatic extraction algorithm of monomorphous maximum volume transformation idea end member in this convex surface geometrical property of high-dimensional feature space.At present based on the automatic extraction algorithm of the end member of MVT thought in the process of calculating the monomorphous maximum volume owing to lack effective search strategy and corresponding search rule, have and expend time in longly, calculated amount is excessive, shortcomings such as output unstable result.

Summary of the invention

The technical matters that solves

For fear of the deficiencies in the prior art part, the present invention proposes a kind of high optical spectrum image end member extraction method based on the optimization searching strategy, can have low time complexity, significantly reduces operation times and time, and the output result is not subjected to the different influence of initial value, and is relatively stable.

Thought of the present invention is: concentrate the alternative set of the initial end member of selection arbitrarily at the point that high-spectrum remote sensing spectrum vector forms in high-dimensional feature space, then in conjunction with the search strategy of optimizing, in the process that the monomorphous volume progressively increases, choose point near end member, finally select whole end members, they are positioned at the place, monomorphous summit of volume maximum.

Technical scheme

A kind of high optical spectrum image end member extraction method based on the optimization searching strategy is characterized in that step is as follows:

Step 1: according to the correlation matrix of mining height spectral remote sensing image and the eigenwert of covariance matrix:

{ λ ₀, λ ₁..., λ _N-1, adopt virtual dimension method Virtual Dimensionality to estimate the roughly atural object quantity l that original high-spectrum remote sensing comprises, wherein: n is the spectral band number, the l initial value is zero; On a certain wave band The time roughly atural object quantity l add 1; False alarm rate P in the described virtual dimension method _FBe 0.001;

Step 2: adopt principal component analysis (PCA) Principal Components Analysis that original high-spectrum remote sensing is carried out dimension-reduction treatment, the spectrum dimension in the principal component analysis (PCA) equals the roughly atural object quantity l-1 that step 1 obtains;

Step 3: the high-spectrum remote sensing behind the dimensionality reduction that step 2 is obtained carries out end member and extracts, and divides following steps to realize:

Step a: choose out l point wantonly as the first initial set { e of end member ₀, e ₁..., e _L-1;

Step b: at { e ₀, e ₁..., e _L-1In optional vectorial e _i

Step c: cross some e _iDo and be parallel to lineoid { e ₁..., e _L-1Lineoid

With the first half of lineoid cutting data cloud cluster for being directed to e _iRegion of search and direction;

Steps d: at e _iThe region of search in, select a vectorial e _kSubstitute e _i, and calculate

$V_{e_i\↔e_k}=\frac{\left|\det \left[\begin{array}{ccccccc}1& ...& 1& 1& 1& ...& 1\\ e_0& ...& e_{i-1}& e_k& e_{i+1}& ...& e_{l-1}\end{array}\right]\right|}{(l-1)!};$

Step e: according to Obtain

Step f: repeat step a～step e l-1 time, obtain final end member set

Beneficial effect

The high optical spectrum image end member extraction method based on the optimization searching strategy that the present invention proposes has adopted feature space optimization searching strategy in conjunction with MVT thought, and by progressively increasing the monomorphous volume of feature space, order is obtained the end member of high-spectrum remote sensing.For any one alternative end member, this method is at first determined its region of search and direction at feature space, searches among a small circle then, greatly reduce operand, alleviate the algorithm time complexity, and improved output stability, can be used for multiple high spectrum image application scenario.

Because this method has utilized maximum monomorphous volume conversion idea in the high spectrum image high-dimensional feature space to be combined with method to the optimization searching strategy, can carry out the automatic extraction of high spectrum end member, be applicable to work such as separating of high-spectrum remote sensing is mixed, classification, identification, Spectral matching.

Description of drawings

Fig. 1: basic skills process flow diagram of the present invention

Fig. 2: at alternative end member e ₀The region of search, be that example is at initial end member collection { e with the two-dimensional space ₀, e ₁, e ₂In, at alternative end member e ₀The region of search.

Fig. 3: this method is applied to the end member output result of real image, and original image derives from ENVI, and atmospheric correction rear space resolution is 400*350, and spectral resolution is 50.Mark the first five end member that the output end member is concentrated among the figure, be designated A (alunite), B (water ammonium feldspar), C (kalzit), K (smalite) and M (white mica) respectively.Through distributing relatively with the known mineral in this region, the end member coordinate position is accurate, and at different initial end member collection, the output result is stable.

Embodiment

Now in conjunction with the embodiments, accompanying drawing is further described the present invention:

The hardware environment of present embodiment is: Pentium-43.0G computing machine, 1G internal memory, 128M video card; The software environment of operation is: Window XP operating system, realized the method that the present invention proposes with the IDL7.0 programming language in conjunction with ENVI.

1. use the VD method to estimate the roughly atural object quantity that original high-spectrum remote sensing comprises.At first calculate the eigenwert of the correlation matrix and the covariance matrix of high-spectrum remote sensing:

{ λ ₀, λ ₁..., λ _N-1, (establishing the spectral band number is n) is if on a certain wave band

Then representing has obvious spectral signal to occur on this wave band, if Illustrate that then the spectral signal on this wave band comes self noise probably.At predefined false alarm rate P _FRestriction under, by the eigenwert of correlation matrix and covariance matrix relatively, the actual atural object quantity l that can roughly estimate high-spectrum remote sensing and comprised.

2. with principal component analysis (PCA) the spectrum dimension of high-spectrum remote sensing is reduced to the l-1 dimension by initial n dimension.For a secondary initial high-spectrum remote sensing I (i, j n), with its each pixel with the spectrum vector representation are:

$x_k,k=0,...,M-1,M=i*j,x_k\∈R^n,\underset{k=0}{\overset{M-1}{\mathrm{\Σ}}}{x}_k=0$

Its covariance matrix is:

$C=\frac{1}{M}\underset{k=0}{\overset{M-1}{\mathrm{\Σ}}}{x}_k{x}_k^T$

Obtain the eigenwert of its covariance matrix, { λ ₀〉=λ ₁〉=... 〉=λ _N-1, select wherein l-1 eigenwert { λ from big to small ₀〉=λ ₁〉=... 〉=λ _L-2, its characteristic of correspondence vector is formed orthogonal matrix P={p as column vector ₀..., p _L-2, use P={p ₀..., p _L-2To each spectrum vector x _kCarry out orthogonal transformation, get x ' _k=x _k* P, x ' _k∈ R ^L-1, finally obtain behind the dimensionality reduction high-spectrum remote sensing I ' (i, j, l-1).

3. end member extracts.Realize by following flow process:

1.: choose out l point wantonly as the first initial set { e of end member ₀, e ₁..., e _L-1.(l is determined by first step VD method) each point all is a vector in the l-1 dimensional feature space, represents a high-spectrum remote sensing I ' (i, j, l-1) the spectrum vector in.

2.: at { e ₀, e ₁..., e _L-1In optional vector, be without loss of generality, with e ₀Be example.

3.: determine at e ₀Region of search and direction.In feature space, cross some e ₀Do and be parallel to lineoid { e ₁..., e _L-1Lineoid

Be directed to e ₀Region of search and direction be the first half of lineoid cutting data cloud cluster.(the hypothesis that is without loss of generality monomorphous { e ₀, e ₁..., e _L-1Be positioned at lineoid The below)

4.: at e ₀The region of search in, select a vectorial e _kSubstitute e ₀, and calculate

$V_{e_0\↔e_k}=\frac{\left|\det \left[\begin{array}{cccc}1& 1& ...& 1\\ e_k& e_1& ...& e_{l-1}\end{array}\right]\right|}{(l-1)!}$

5.: ask for

Satisfy:

6.: this moment, the end member collection was Optional vector in remaining l-1 initial end member, the step above repeating is until obtaining final end member set

Claims (1)

1. high optical spectrum image end member extraction method based on the optimization searching strategy is characterized in that step is as follows:

Step 1: according to the correlation matrix of mining height spectral remote sensing image and the eigenwert of covariance matrix:

{ λ ₀, λ ₁..., λ _N-1, adopt virtual dimension method Virtual Dimensionality to estimate the roughly atural object quantity l that original high-spectrum remote sensing comprises, wherein: n is the spectral band number, the l initial value is zero; On a certain wave band

The time roughly atural object quantity l add 1; False alarm rate P in the described virtual dimension method _FBe 0.001;

Step 2: adopt principal component analysis (PCA) Principal Components Analysis that original high-spectrum remote sensing is carried out dimension-reduction treatment, the spectrum dimension in the principal component analysis (PCA) equals the roughly atural object quantity l-1 that step 1 obtains;

Step 3: the high-spectrum remote sensing behind the dimensionality reduction that step 2 is obtained carries out end member and extracts, and divides following steps to realize:

Step a: choose out l point wantonly as the first initial set { e of end member ₀, e ₁..., e _L-1;

Step b: at { e ₀, e ₁..., e _L-1In optional vectorial e _i

Step c: cross some e _iDo and be parallel to lineoid { e ₁..., e _L-1Lineoid

With the first half of lineoid cutting data cloud cluster for being directed to e _iRegion of search and direction;

Steps d: at e _iThe region of search in, select a vectorial e _kSubstitute e _i, and calculate

$V_{e_i\↔e_k}=\frac{\left|\det \left[\begin{array}{ccccccc}1& ...& 1& 1& 1& ...& 1\\ e_0& ...& e_{i-1}& e_k& e_{i+1}& ...& e_{l-1}\end{array}\right]\right|}{(l-1)!};$

Step e: according to

Obtain

Step f: repeat step a～step e l-1 time, obtain final end member set

Images