CN103268593A - Method for enabling signals and noise in hyperspectral remote sensing images to be separated - Google Patents

Abstract

The invention relates to the field of remote sensing technology, in particular to an image processing method for enabling signals and noise in hyperspectral remote sensing images to be separated. The method for enabling the signals and the noise in the hyperspectral remote sensing images to be separated specifically comprises the following steps that by means of lowpass filtering and homogeneous partitioning, the signals and the noise are estimated respectively; according to the Bayes principle, quantitative calculation is carried out, and the dimensionality of the estimated signals and the dimensionality of the estimated noise are obtained; signal subspace and noise subspace are built, and by means of inclined subspace projection, signal components and noise components in the original hyperspectral remote sensing images are separated. According to the method for enabling the signals and the noise in the hyperspectral remote sensing images to be separated, the dimensionality of the estimated signals and the dimensionality of the estimated noise can be obtained in a quantitative calculation mode, and therefore the fact that the obtained dimensionality are not affected by subjective factors is guaranteed; by means of the inclined subspace projection, the relevance between the noise and the signals is fully taken into consideration, and therefore the signals can be well separated from the noise.

Description

The separation method of signal and noise in a kind of target in hyperspectral remotely sensed image

Technical field

The present invention relates to the remote sensing technology field, relate in particular to the image treatment method that the signal in the target in hyperspectral remotely sensed image is separated with noise.

Background technology

Separating of signal and noise is the important technology that target in hyperspectral remotely sensed image is handled.But existing noise separation method is based on various irrational hypothesis more: ignore the noise diversity, only consider additivity or multiplicative noise; Suppose that noise and signal are separate; Suppose that each row (OK) signal obedience same mathematical distributes; Suppose that whole scape image is homogeneous or approximate homogeneous, etc.And the existing method of obtaining signal dimension and noise dimension is subject to the influence of subjective factor, has reduced the separation accuracy of signal and noise.Specifically can be referring to document N.Acito, M.Diani, and G.Corsini, " Subspace-Based Striping Noise Reduction in Hyperspectral Images; " IEEE Transactions on Geoscience and Remote Sensing, vol.49, pp.1325-1342, Apr2011.

Summary of the invention

(1) technical matters that will solve

The separation method that the purpose of this invention is to provide signal and noise in a kind of target in hyperspectral remotely sensed image, solving the correlativity of ignoring noise type diversity and noise and signal in the existing method, and the too qualitative and subjective problem of dimension.

(2) technical scheme

In order to solve the problems of the technologies described above, the invention provides the separation method of signal and noise in a kind of target in hyperspectral remotely sensed image, specifically comprise:

Utilize low-pass filtering and homogenieity piecemeal, respectively estimated signal and noise;

According to bayesian criterion, quantitatively calculate the dimension of estimated signals and noise;

Make up signal subspace and noise subspace, and by oblique subspace projection, separate signal component and noise component in the original target in hyperspectral remotely sensed image.

Preferably, set up data model at target in hyperspectral remotely sensed image:

X _l(i,j)=S _l(i,j)⊕N _l(i,j)

S wherein _l(i, j) expression signal; N _l(i, j) expression noise; L=1 ..., N _B, i=1 ..., N _L, j=1 ..., N _S.; Symbol ⊕ represents direct sum; N _BExpression wave band number, N _LAnd N _SRepresent the ranks number respectively.

Preferably, described estimated signal and noise specifically comprise:

The operator that low-pass filtering is selected for use is: $\left[\begin{array}{ccc}1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\end{array}\right],$ With low-frequency information as the estimation to signal;

According to space continuity and the wave spectrum correlativity of atural object, given less wave spectrum angle threshold value is carried out piecemeal with image;

The noise of l wave band n piecemeal is estimated as C _Nl(1-r _Nl), C wherein _NlBe the pixel average of l wave band n piecemeal, r _NlIt is the multiple correlation coefficient of l wave band n piecemeal and adjacent two wave band correspondence position piecemeals.

Preferably, the contained pixel number of selected piecemeal must not be less than 30.

Preferably, the described dimension that quantitatively calculates estimated signals and noise according to bayesian criterion specifically comprises:

The bayesian criterion that uses is integrated minimal error rate criterion and Neyman – Pearson (N-P) criterion;

P is set _MThe expression false dismissal probability, P _FThe expression false-alarm probability, λ represents a certain eigenwert of covariance matrix, ω ₁Expression noise component, ω ₂The expression signal component, and p (λ | ω ₁) and p (λ | ω ₂) represent the approximate condition probability density of correspondence when λ belongs to noise or signal, p (ω respectively ₁| λ) and p (ω ₂| the posterior probability density of correspondence when λ) representing respectively that λ belongs to noise or signal;

Setting make p (λ | ω ₁) and p (λ | ω ₂) the most approaching eigenvalue ₁For meeting the noise separation of minimal error rate criterion, the P that this moment is corresponding _FIgnore, the critical value of N-P criterion is:

${\mathrm{\δ}}_{\mathrm{NP}}=\frac{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_1)}{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_2)}$

For arbitrary wave band, if

Then $\mathrm{\λ}\∈\left\{\begin{array}{c}{\mathrm{\ω}}_1\\ {\mathrm{\ω}}_2\end{array}\right..$

Preferably, described signal and the noise of setting up signal subspace and noise subspace and separating in the original Hyperspectral imaging by oblique subspace projection specifically comprises:

The proper vector corresponding with belonging to signal or characteristics of noise value constitutes the basic S of corresponding signal subspace or noise subspace _BasisOr N _Basis, the orthogonal intersection cast shadow matrix of signal subspace is

S _orthproject=E-S _basis(S _basis ^HS _basis) ^-1S _basis ^H

Wherein, E representation unit matrix in like manner obtains the orthogonal intersection cast shadow matrix N of noise subspace _Orthproject

Parallel noise subspace to the oblique projection matrix of signal subspace is:

P _SN=S _basis(S _basis ^HN _orthprojectS _basis) ^-1S _basis ^HN _orthproject

Signal is:

S=P _SNX；

The parallel signal subspace to the oblique projection matrix of noise subspace is

P _NS=N _basis(N _basis ^HS _orthprojectN _basis) ^-1N _basis ^HS _orthproject

Noise is

N=P _NSX。

(3) beneficial effect

The separation method of signal and noise in the target in hyperspectral remotely sensed image provided by the invention can quantitatively calculate the dimension of estimated signals and noise, has guaranteed that the dimension that obtains is not influenced by subjective factor; Tiltedly the use of subspace projection has fully taken into account noise and correlation between signals, thereby signal can well be separated with noise.

Description of drawings

Fig. 1 is the process flow diagram of the separation method of signal in the target in hyperspectral remotely sensed image of the present invention and noise;

Fig. 2 (a)-Fig. 2 (c) is first embodiment of the present invention: certain AVIRIS data the 218th wave band (centre wavelength: the signal that original image 2439.81nm) and utilizing obtains after Fig. 1 method and the image as a result of noise;

Fig. 3 (a)-Fig. 3 (c) is second embodiment of the present invention: certain Hyperion data the 57th wave band (centre wavelength: the signal that original image 925.41nm) and utilizing obtains after Fig. 1 method and the image as a result of noise.

Embodiment

Following examples are used for explanation the present invention, but are not used for limiting the scope of the invention.

Fig. 1 is the process flow diagram of the separation method of signal in the target in hyperspectral remotely sensed image of the present invention and noise, may further comprise the steps:

Step1, from high spectrum raw video estimated signal and noise respectively;

Step2, pass judgment on the dimension that criterion quantitatively calculates estimated signals and noise according to Bayes.Among first embodiment, the dimension that signal is estimated is 9, and the dimension that noise is estimated is 23; Among second embodiment, the dimension that signal is estimated is 4, and the dimension that noise is estimated is 149;

Step3, set up signal subspace and noise subspace, calculate the oblique projection matrix of signal subspace and noise subspace, the raw video oblique projection to different subspace, is obtained signal and noise respectively.As calculated, the subspace angle of signal subspace and noise subspace is 86.77 ° among first embodiment, and the subspace angle of signal subspace and noise subspace is 72.26 ° among second embodiment, and signal and noise are also not exclusively independent.Fig. 2 (a)-Fig. 2 (c) and Fig. 3 a(a)-Fig. 3 (c) is respectively signal that the raw video and utilizing of certain wave band of first embodiment and second embodiment obtains after Fig. 1 method and the image as a result of noise.

Following mask body is introduced the separation method of signal and noise in this target in hyperspectral remotely sensed image, and it comprises:

Utilize low-pass filtering and homogenieity piecemeal, respectively estimated signal and noise;

According to bayesian criterion, quantitatively calculate the dimension of estimated signals and noise;

Make up signal subspace and noise subspace, and by oblique subspace projection, separate signal component and noise component in the original target in hyperspectral remotely sensed image.

Wherein, set up data model at target in hyperspectral remotely sensed image:

X _l(i,j)=S _l(i,j)⊕N _l(i,j)

S wherein _l(i, j) expression signal; N _l(i, j) expression noise; L=1 ..., N _B, i=1 ..., N _L, j=1 ..., N _S.; Symbol ⊕ represents direct sum; N _BExpression wave band number, N _LAnd N _SRepresent the ranks number respectively.

Wherein, described estimated signal and noise specifically comprise:

The operator that low-pass filtering is selected for use is: $\left[\begin{array}{ccc}1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\end{array}\right],$ With low-frequency information as the estimation to signal;

According to space continuity and the wave spectrum correlativity of atural object, given less wave spectrum angle threshold value is carried out piecemeal with image;

The noise of l wave band n piecemeal is estimated as C _Nl(1-r _Nl), C wherein _NlBe the pixel average of l wave band n piecemeal, r _NlIt is the multiple correlation coefficient of l wave band n piecemeal and adjacent two wave band correspondence position piecemeals.

Wherein, the contained pixel number of selected piecemeal must not be less than 30.

Wherein, the described dimension that quantitatively calculates estimated signals and noise according to bayesian criterion specifically comprises:

The bayesian criterion that uses is integrated minimal error rate criterion and Neyman – Pearson (N-P) criterion;

P is set _MThe expression false dismissal probability, P _FThe expression false-alarm probability, λ represents a certain eigenwert of covariance matrix, ω ₁Expression noise component, ω ₂The expression signal component, and p (λ | ω ₁) and p (λ | ω ₂) represent the approximate condition probability density of correspondence when λ belongs to noise or signal, p (ω respectively ₁| λ) and p (ω ₂| the posterior probability density of correspondence when λ) representing respectively that λ belongs to noise or signal;

Setting make p (λ | ω ₁) and p (λ | ω ₂) the most approaching eigenvalue ₁For meeting the noise separation of minimal error rate criterion, the P that this moment is corresponding _FIgnore, the critical value of N-P criterion is:

${\mathrm{\δ}}_{\mathrm{NP}}=\frac{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_1)}{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_2)}$

For arbitrary wave band, if

Then $\mathrm{\λ}\∈\left\{\begin{array}{c}{\mathrm{\ω}}_1\\ {\mathrm{\ω}}_2\end{array}\right..$

Wherein, described signal and the noise of setting up signal subspace and noise subspace and separating in the original Hyperspectral imaging by oblique subspace projection specifically comprises:

The proper vector corresponding with belonging to signal or characteristics of noise value constitutes the basic S of corresponding signal subspace or noise subspace _BasisOr N _Basis, the orthogonal intersection cast shadow matrix of signal subspace is

S _orthproject=E-S _basis(S _basis ^HS _basis) ^-1S _basis ^H

Wherein, E representation unit matrix in like manner obtains the orthogonal intersection cast shadow matrix N of noise subspace _Orthproject

Parallel noise subspace to the oblique projection matrix of signal subspace is:

P _SN=S _basis(S _basis ^HN _orthprojectS _basis) ^-1S _basis ^HN _orthproject

Signal is:

S=P _SNX；

The parallel signal subspace to the oblique projection matrix of noise subspace is

P _NS=N _basis(N _basis ^HS _orthprojectN _basis) ^-1N _basis ^HS _orthproject

Noise is

N=P _NSX。

The separation method of signal and noise in the target in hyperspectral remotely sensed image provided by the invention can quantitatively calculate the dimension of estimated signals and noise, has guaranteed that the dimension that obtains is not influenced by subjective factor; Tiltedly the use of subspace projection has fully taken into account noise and correlation between signals, thereby signal can be separated well with noise.

The above only is preferred implementation of the present invention; should be pointed out that for those skilled in the art, under the prerequisite that does not break away from the technology of the present invention principle; can also make some improvement and replacement, these improvement and replacement also should be considered as protection scope of the present invention.

Claims (6)

1. the separation method of signal and noise in the target in hyperspectral remotely sensed image is characterized in that: specifically comprise:

Utilize low-pass filtering and homogenieity piecemeal, respectively estimated signal and noise;

According to bayesian criterion, quantitatively calculate the dimension of estimated signals and noise;

Make up signal subspace and noise subspace, and by oblique subspace projection, separate signal component and noise component in the original target in hyperspectral remotely sensed image.

2. the separation method of signal and noise in the target in hyperspectral remotely sensed image as claimed in claim 1 is characterized in that, sets up data model at target in hyperspectral remotely sensed image:

X _l(i,j)=S _l(i,j)⊕N _l(i,j)

S wherein _l(i, j) expression signal; N _l(i, j) expression noise; L=1, L, N _B, i=1, L, N _L, j=1, L, N _S.; Symbol ⊕ represents direct sum; N _BExpression wave band number, N _LAnd N _SRepresent the ranks number respectively.

3. the separation method of signal and noise in the target in hyperspectral remotely sensed image as claimed in claim 1 is characterized in that described estimated signal and noise specifically comprise:

The operator that low-pass filtering is selected for use is: $\left[\begin{array}{ccc}1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\\ 1/9& 1/9& 1/9\end{array}\right],$ With low-frequency information as the estimation to signal;

According to space continuity and the wave spectrum correlativity of atural object, given less wave spectrum angle threshold value is carried out piecemeal with image;

The noise of l wave band n piecemeal is estimated as C _Nl(1-r _Nl), C wherein _NlBe the pixel average of l wave band n piecemeal, r _NlIt is the multiple correlation coefficient of l wave band n piecemeal and adjacent two wave band correspondence position piecemeals.

4. the separation method of signal and noise in the target in hyperspectral remotely sensed image as claimed in claim 3 is characterized in that, the contained pixel number of selected piecemeal must not be less than 30.

5. the separation method of signal and noise in the target in hyperspectral remotely sensed image as claimed in claim 1 is characterized in that, the described dimension that quantitatively calculates estimated signals and noise according to bayesian criterion specifically comprises:

The bayesian criterion that uses is integrated minimal error rate criterion and Neyman – Pearson (N-P) criterion;

P is set _MThe expression false dismissal probability, P _FThe expression false-alarm probability, λ represents a certain eigenwert of covariance matrix, ω ₁Expression noise component, ω ₂The expression signal component, and p (λ | ω ₁) and p (λ | ω ₂) represent the approximate condition probability density of correspondence when λ belongs to noise or signal, p (ω respectively ₁| λ) and p (ω ₂| the posterior probability density of correspondence when λ) representing respectively that λ belongs to noise or signal;

Setting make p (λ | ω ₁) and p (λ | ω ₂) the most approaching eigenvalue ₁For meeting the noise separation of minimal error rate criterion, the P that this moment is corresponding _FIgnore, the critical value of N-P criterion is:

${\mathrm{\δ}}_{\mathrm{NP}}=\frac{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_1)}{p\left({\mathrm{\λ}}_1\right|{\mathrm{\ω}}_2)}$

For arbitrary wave band, if $\frac{p\left(\mathrm{\&lambda;}\right|{\mathrm{\&omega;}}_1)>}{p\left(\mathrm{\&lambda;}\right|{\mathrm{\&omega;}}_2)<}{\mathrm{\&delta;}}_{\mathrm{NP}},$ Then $\mathrm{\&lambda;}\&Element;\left\{\begin{array}{c}{\mathrm{\&omega;}}_1\\ {\mathrm{\&omega;}}_2\end{array}\right..$

6. the separation method of signal and noise in the target in hyperspectral remotely sensed image as claimed in claim 1 is characterized in that, described signal and the noise of setting up signal subspace and noise subspace and separating in the original Hyperspectral imaging by oblique subspace projection specifically comprises:

The proper vector corresponding with belonging to signal or characteristics of noise value constitutes the basic S of corresponding signal subspace or noise subspace _BasisOr N _Basis, the orthogonal intersection cast shadow matrix of signal subspace is

S _orthproject=E-S _basis(S _basis ^HS _basis) ^-1S _basis ^H

Wherein, E representation unit matrix in like manner obtains the orthogonal intersection cast shadow matrix N of noise subspace _Orthproject

Parallel noise subspace to the oblique projection matrix of signal subspace is:

P _SN=S _basis(S _basis ^HN _orthprojectS _basis) ^-1S _basis ^HN _orthproject

Signal is:

S=P _SNX；

The parallel signal subspace to the oblique projection matrix of noise subspace is

P _NS=N _basis(N _basis ^HS _orthprojectN _basis) ^-1N _basis ^HS _orthproject

Noise is

N=P _NSX。

Images