CN101571915B - Oil Spill Recognition Method Based on Eigenvalues in SAR Image - Google Patents

Abstract

The invention discloses a method for identifying oil spill of an SAR image based on a characteristic value, which comprises the following steps: a, selecting the SAR image to be identified for processing; b, performing edge detection on the image by using an improved C-V model so as to determine a target boundary; c, measuring the characteristic value after extracting a target; d, identifying an object by adopting a Mahalanobis distance method and a composite probability method; and e, judging whether the image in a dark region is the oil spill according to the two methods in the step d. The technical proposal uses the advantage that the synthetic aperture radar (SAR) can perform high-resolution oil spill monitoring day and night and under all-weather conditions; the selected characteristic value is small in quantity and has obvious effect of judging the oil spill compared with other conventional characteristic quantities (such as area, perimeter and the like); the method has a simplealgorithm and is easy to achieve by using the Mahalanobis distance method and the composite probability method; and the method is advantageous to be achieved by programming on computer.

Description

SAR image method for identifying oil spill based on eigenwert

Technical field

The present invention relates to a kind of method for identifying oil spill, relate in particular to a kind of SAR image method for identifying oil spill based on eigenwert.

Background technology

Can both find oil slick on all oceans, some extensive and small-scale phenomenons are all relevant with oil slick, and this ground contamination that the extra large gas balance of for example change fragility or marine ecosystems interaction cause has caused climate change.We utilize technological means such as satellite remote sensing and radar usually, realize to detect and the evaluation accident or deliberately the oil slick, the especially synthetic aperture radar (SAR) that cause of discharging can independently be used in the daytime under the weather condition.SAR uniquely is used for carrying out the daily task of tracking and monitoring marine oil overflow by deployment, the monitoring result of many oil spill accidents proved this remote sensor day and night, the ability of round-the-clock marine oil overflow monitoring.Unique phenomenon that but oil spilling is not SAR can monitor.In some cases, the spontaneous phenomenon on sea can be disturbed the form of expression of SAR image or produce the decoy of similar oil spilling phenomenon (be homologue, as: wind slickenside, surface stream, interior ripple, biological oil slick, low wind speed district, rain belt, hang down the dark areas that backward scattered Sha Po etc. causes).Therefore, need to propose a kind of method and can discern from the SAR image whether the dark space is oil spilling.

Summary of the invention

The present invention is directed to the proposition of above problem, and develop a kind of method for identifying oil spill, utilize the method can resolve from the SAR image and identify whether the dark space is oil spilling, be convenient to computer realization based on eigenwert.Technology of the present invention realizes that means are as follows:

A kind of SAR image method for identifying oil spill based on eigenwert is characterized in that comprising the steps:

A, choose SAR image to be identified, wait to declare the effective coverage of object in the intercepting SAR image, promptly the target area of cut-away view picture and near limited area background thereof are as pending object;

B, with improved C-V model image is carried out rim detection and determine object boundary, extract object boundary, promptly discern the dark space;

C, extraction target after object and background separation, are extracted and the amount of carrying out calculation the clarification of objective value, and the amount of concrete eigenwert is following:

1.. luminance standard difference ratio outside the border: the ratio of the average backscatter intensity of the radar outside the dark space and its standard deviation;

2.. radar back scattering outside the border: the radar interface back scattering average in the localized area outside the target area;

3.. dark space radar back scattering: average gray in the zone;

4.. dark space standard deviation: back scattering standard deviation in the dark space;

5.. standard deviation ratio: the ratio of the inside and outside back scattering standard deviation in dark space;

6.. form factor: the deviation of gray scale on its longitudinal axis in the zone;

7.. brightness ratio: the ratio of the back scattering average in inside, dark space and the outside localized area;

8.. dark space luminance standard difference ratio: the ratio of the inner backscatter intensity in dark space and gray standard deviation;

The above-mentioned eigenwert composition of vector x that tries to achieve, x=[x1, x2 ... xi] ^T(i=8);

D, employing Mahalanobis distance and probability of recombination method are discerned object:

At first the feature of object is compared the proper vector x=[x of known image target area with known oil spilling template ₁, x ₂..., x _i] ^T, i is the number of eigenwert among the step c, x _iBe the eigenwert that the amount of carrying out is calculated among the step c, proper vector m _j(j=1,2) are that two known classification are oil spilling and " homologue ", obtain the covariance matrix of proper vector x earlier

Then according to formula r ² _j=(x-m _j) ^TC ^-1(x-m _j) the proper vector x that obtains the image object district is the proper vector m of oil spilling and " homologue " with known classification _jMahalanobis distance between (j=1,2);

The class that the object that is in the dark space is belonged in oil spilling district or " homologue " is carried out the calculating of the probability of recombination then, considers image common characteristic value x _iIndividual (i=1,2 ..., 8), wherein i eigenwert is oil spilling probability is p _i(x _i), be that the probability of " homologue " is q _i(x _i), utilize function then $p=\frac{1}{1+\underset{i}{\mathrm{\Π}}\frac{q_i\left(x_i\right)}{p_i\left(x_i\right)}}$ Draw the probability of recombination p of oil spilling class;

E, whether is that oil spilling is judged by two kinds of methods in the steps d to the image of dark space, at first feature is carried out the Mahalanobis distance discrimination, (r1 is that object is with the Mahalanobis distance between the known classification oil spilling as if r1＜r2, r2 is that object is with the distance of the Mahalanobis between the known classification " homologue ") judge that then x belongs to oil spilling, otherwise then be homologue, if the Mahalanobis of x and two classification distance is more or less the same, then utilizing above-mentioned function to obtain probability of recombination p does further to judge, for the be defined as oil spilling of the probability of recombination at 67%-100%, the probability of recombination below 33% is being " homologue ", and whether can not determine between 34% and 66% is oil spilling.

Described step b is implemented as follows:

Wherein, μ 〉=0, υ 〉=0, λ ₁＞0, λ ₂＞0 parameter, C one approaches the outline line at edge, dark space gradually, c1, c2 represents the average gray of curve C interior zone and perimeter respectively, the length operator calculates the girth of C curve, the inside operator calculates C curve inside, Area operator reference area, and (1) makes curve C keep certain regularity, (2) makes the curve C edge of close object gradually, utilize initial profile line C to set up the symbolic distance Jacobian matrix, inner at outline line, outside for negative for just, pass through function, outline line constantly keeps to the side, and finally stops in edge, and the point that point symbol is opposite is exactly a marginal point on every side.

The specifically process of trying to achieve of Mahalanobis distance is in the described steps d: the covariance matrix of obtaining the proper vector x of known image target area earlier

K oil spilling ml having added up and the sample of " homologue " m2 are extracted in the back from experience database, obtain sample average M ₁ ^(k)And M ₂ ^(k), obtain x and M then respectively ₁ ^(k)And M ₂ ^(k)The Mahalanobis distance $r_j^2={(x-{\stackrel{\‾}{M}}_j^{\left(k\right)})}^T{C}^{(-1)}(x-{\stackrel{\‾}{M}}_j^{\left(k\right)}),$ Wherein (j=1,2).

I described in the described steps d is an i eigenwert of gained among the step c, its probability that belongs to oil spilling is pi (xi), the frequency that the result of calculation that this feature is measured is occurred in known oil spilling feature database is in like manner obtained i eigenwert and is existed " homologue " in probability be qi (xi).

Owing to adopted technique scheme, utilized synthetic aperture radar (SAR) day and night reaching the advantage of carrying out the high resolving power spilled oil monitoring under all weather conditions; The eigenwert quantity of choosing is few and obvious than other traditional characteristic amount (as area, girth etc.) effects for taking a decision as to whether oil spilling; Utilize Mahalanobis distance and probability of recombination method, algorithm is simple, is easy to realize; Help the realization of programming on computers.

Description of drawings

Fig. 1. be the process flow diagram of recognition methods of the present invention;

Fig. 2. be target source image to be judged among the embodiment;

Fig. 3. be the image behind edge extracting among the embodiment;

Fig. 4. for respectively object and background area being carried out monochromatic synoptic diagram of filling among the embodiment;

Fig. 5. be the block diagram of Mahalanobis Furthest Neighbor realization among the embodiment;

Fig. 6. among the embodiment at the backward scattered probability distribution graph of measurement features dark space radar, wherein ■ represents oil spilling, ● the expression " homologue ".

Embodiment

At first the SAR image is carried out pre-service, around the target of needs identification, intercept a rectangular area (as shown in Figure 2).As Fig. 3, the image that obtains is carried out rim detection, the amount of extracting the target area then and carrying out eigenwert is calculated, and compares with the eigenwert composition of vector that obtains and with known features, judges according to result relatively whether target is oil spilling.Embodiments of the invention are chosen to be optimized to clarification of objective and are simplified the feature of having chosen 8 tool representatives and judge (below describe in detail).Method for identifying oil spill based on eigenwert shown in Figure 1, implementation step is as follows:

A, choose SAR image to be identified, wait to declare the effective coverage of object in the intercepting SAR image, promptly the target area of cut-away view picture and near limited area background thereof are as pending object;

B, with improved C-V model image is carried out rim detection and determine object boundary, extract object boundary, promptly discern the dark space; Specifically be calculated as follows:

Wherein, μ 〉=0, υ 〉=0, λ ₁＞0, λ ₂＞0 parameter when experiment, is generally got λ ₁=λ ₂=1, υ=0.C one approaches the outline line at edge, dark space gradually, and c1, c2 represent the average gray of curve C interior zone and perimeter respectively, and the length operator calculates the girth of C curve, and the inside operator calculates C curve inside, Area operator reference area.(1) makes curve C keep certain regularity, and (2) makes the curve C edge of close object gradually.Utilize initial profile line C to set up the symbolic distance Jacobian matrix, inner at outline line, outside for negative for just, by function, outline line constantly keeps to the side, and finally stops in edge, at this moment, the point that point symbol is opposite is exactly a marginal point on every side, thereby realizes image segmentation.

C, as Fig. 4, continue original image is handled, with target area and background normalization, extract target, with object and background separation.Set up the information in Fig. 4 image and the mapping of source images, read the feature and the amount of carrying out of source images by mapping relations and calculate (promptly the SAR image information amount of carrying out being calculated).The eigenwert that the present invention relates to comprises following:

1.. luminance standard difference ratio outside the border: the ratio of the average backscatter intensity of the radar outside the dark space (BR) and its standard deviation (Sp);

2.. radar back scattering outside the border: the radar interface back scattering average in the localized area outside the target area;

3.. dark space radar back scattering: average gray in the zone;

4.. dark space standard deviation: back scattering standard deviation in the dark space;

5.. standard deviation ratio: the ratio of the inside and outside back scattering standard deviation in dark space;

6.. form factor: the deviation of gray scale on its longitudinal axis in the zone;

7.. brightness ratio: the ratio of the back scattering average in inside, dark space and the outside localized area;

8.. dark space luminance standard difference ratio: the ratio of the inner backscatter intensity in dark space and gray standard deviation;

The above-mentioned eigenwert composition of vector x that tries to achieve, x=[x1, x2 ... xi] ^T(i=8);

D, adopt sorting algorithm here, contrast done in target area and known experience template, obtain the Mahalanobis distance of target area and template and the probability of recombination so that object is discerned:

At first utilize Mahalanobis the feature of object to be identified and known template characteristic can be compared, wherein x is the proper vector of a generic the unknown, and its component is a stack features of being tried to achieve among the step c, makes m _j(j=1,2) are two known classification---oil spilling and " homologue ", can obtain the feature of two classification by experience database, and as known template, x and m _jBetween matching degree by calculating their Mahalanobis distance.Proper vector x=[x by the known image target area ₁, x ₂..., x _i] ^T, i is the number of eigenwert among the step c, x _iBe the eigenwert that the amount of carrying out is calculated among the step c, proper vector m _j(j=1,2) are that two known classification are oil spilling and " homologue ", obtain the covariance matrix of proper vector x earlier

Then according to formula r ² _j=(x-m _j) ^TC ^-1(x-m _j) the proper vector x that obtains the image object district is the proper vector m of oil spilling and " homologue " with known classification _jMahalanobis distance between (j=1,2) is also referred to as the distance of proper vector x and j class; Because of generalized case m _jBe unknown, so will from experience database, extract the individual oil spilling class m of k (k＞1) that has added up ₁" homologue " m ₂Sample, and obtain sample average M ₁ ^(k)And M ₂ ^(k), obtain x and M then respectively ₁ ^(k)And M ₂ ^(k)The Mahalanobis distance $r_j^2={(x-{\stackrel{\‾}{M}}_i^{\left(k\right)})}^T{C}^{(-1)}(x-{\stackrel{\‾}{M}}_i^{\left(k\right)})(j=\mathrm{1,2})$ (as shown in Figure 5), because square root function is an increment function, consider (x-M usually _j ^(k)) ^TC ⁽¹⁾(x-M _j ^(k)) rather than square root; Judge that X belongs to the class of Mahalanobis apart from minimum, promptly X is from minimum overall of distance.Each element is the variance between each vector element among the covariance matrix C amount of the being x, and this is promoting naturally from the scalar stochastic variable to high-dimensional random vector, and x is with i (i=1,2 here ... 8) column vector of individual scalar stochastic variable composition, and μ _kBe the expectation value of its k element, that is, and μ _k=E (X _k), covariance matrix is defined as then:

$C=E\left[(x-E\left(x\right))(x-E{\left(x\right)}^T)\right]=$

(p, q) individual element is x in the matrix _pWith x _qCovariance. this notion is to promote for the vague generalization of scalar variance of random variable.

The class that the object that is in the dark space is belonged in oil spilling district or " homologue " is carried out the calculating of the probability of recombination then, considers image common characteristic value x _iIndividual (i=1,2 ..., 8), wherein i eigenwert is oil spilling probability is p _i(x _i), be that the probability of " homologue " is q _i(x _i), i eigenwert wherein, its probability that belongs to oil spilling be the frequency that in known oil spilling feature database, occurred of the result of pi (xi) calculation that this feature is measured here.Its probability that belongs to oil spilling is pi (xi), the frequency that the result of calculation that this feature is measured is occurred in known oil spilling feature database (for example: in step c, try to achieve target signature 3. (that is: dark space back scattering) value be 165, its occurrence number in the whole feature database that is known as oil spilling is 3, supposing to contain in this storehouse sample number is 100, then its probability is 3%), in like manner get qi (xi), i exists for feature " homologue " in, as the frequency that occurs in the illusions such as rain belt, biological oil slick.Utilize function then:

$p=\frac{1}{1+\underset{i}{\mathrm{\Π}}\frac{q_i\left(x_i\right)}{p_i\left(x_i\right)}}$

Draw the probability of recombination p of oil spilling class.

Drawing of above-mentioned formula is to obtain by following process: for i proper vector x1, and x2 ... xi, its deviation is respectively Δ x1, Δ x2 ... Δ xi (i=1,2 ..., 8), then, the probability that new pixel belongs to the oil spilling zone is:

ΔP＝∏pi(xi)Δxi (1)

Same, the probability that new pixel belongs to non-oil spilling district is:

ΔQ＝∏qi(xi)Δxi (2)

Suppose that Δ P+ Δ Q is no inclined to one side, just obtained the probability that image belongs to the oil spilling district:

P＝ΔP/(ΔP+ΔQ) (3a)

And image belongs to the probability in non-oil spilling district:

Q＝ΔQ/(ΔP+ΔQ) (3b)

With (1) and (2) formula substitution (3a)

$p=\frac{1}{1+\underset{i}{\mathrm{\Π}}\frac{q_i\left(x_i\right)}{p_i\left(x_i\right)}}$

E, whether is that oil spilling is judged by two kinds of methods in the steps d to the image of dark space, at first feature is carried out the Mahalanobis distance discrimination, (r1 is that object is with the Mahalanobis distance between the known classification oil spilling as if r1＜r2, r2 is that object is with the distance of the Mahalanobis between the known classification " homologue ") judge that then x belongs to oil spilling, otherwise then be homologue, if the Mahalanobis of x and two classification distance is more or less the same, then utilizing above-mentioned function to obtain probability of recombination p does further to judge, for the be defined as oil spilling of the probability of recombination at 67%-100%, the probability of recombination below 33% is being " homologue ", and whether can not determine between 34% and 66% is oil spilling.

Because the Mahalanobis distance has: the Mahalanobis distance is irrelevant with unit, and the advantage that relevant stochastic variable is not added up, so Mahalanobis distance is a kind of good differentiation means, the Probability p that new object belongs to the oil spilling class can be estimated with the distance of the Mahalanobis between the barycenter of this two class by non-classified object.In addition, multiple image is carried out Mahalanobis differentiate, so just obtain the sample of a plurality of oil spilling classes and analog, make the differentiation structure more accurate.

For further the present invention being made an explanation, here the Fig. 4 that with the gray probability is the acquisition after example is handled Fig. 2 judges, choose pi (xi) expression oil spilling district's gradation of image value probability distribution function (256 color shade figure), qi (xi) is a non-oil spilling district gray-scale value probability distribution function, with in be illustrated at the measurement features dark space backward scattered probability distribution graph of radar, obtain probability distribution function Fig. 6, this figure has chosen 19 samples to be discriminated and has added up and draw wherein that ■ represents oil spilling, ● expression " homologue ".Comprehensive these two kinds of algorithms can judge fully whether the target area is oil spilling.

The above; only be the preferable embodiment of the present invention; but protection scope of the present invention is not limited thereto; anyly be familiar with those skilled in the art in the technical scope that the present invention discloses; be equal to replacement or change according to technical scheme of the present invention and inventive concept thereof, all should be encompassed within protection scope of the present invention.

Claims (3)

1. A SAR image oil spill recognition method based on eigenvalue, is characterized in that comprising the steps: a. Select the SAR image to be identified, and intercept the effective area of the target object to be judged in the SAR image, that is, the target area of the intercepted image and the background of the limited area near it as the object to be processed; b. Use the improved C-V model to perform edge detection on the image to determine the target boundary, and extract the target boundary, that is, to identify the dark area. The specific implementation is as follows:

Among them, μ≥0, υ≥0, λ ₁ >0, λ ₂ >0 parameters, C is a contour line gradually approaching the edge of the dark area, c1, c2 respectively represent the average gray level of the inner area and outer area of the curve C , the length operator calculates the perimeter of the C curve, the inside operator calculates the interior of the C curve, and the Area operator calculates the area, the item (1) keeps the curve C regular, and the item (2) makes the curve C gradually approach the object The edge of the initial contour line C is used to establish a signed distance function matrix, which is positive inside the contour line and negative outside. Through the function change, the contour line is constantly approaching the edge, and finally stops at the edge. The points with opposite signs around the point are edge points ; c. Extract the target. After separating the target object from the background, extract and calculate the feature value of the target. The specific feature value calculation is as follows: ①. Brightness standard deviation ratio outside the boundary: the ratio of the average backscattering intensity of the radar outside the dark area to its standard deviation; ②. Out-of-boundary radar backscatter: the average value of radar interface backscatter in a limited area outside the target area; ③.Dark area radar backscatter: average gray level in the area; ④.Standard deviation of dark area: standard deviation of backscattering in dark area; ⑤.Standard deviation ratio: the ratio of the standard deviation of backscattering inside and outside the dark area; ⑥. Shape factor: the dispersion of the gray scale in the area on its longitudinal axis; ⑦.Brightness ratio: the ratio of the mean value of backscattering inside the dark area and in the outer limited area; ⑧.Dark area brightness standard deviation ratio: the ratio of the backscattering intensity inside the dark area to the gray standard deviation; The eigenvalues obtained above form a vector x, x=[x1, x2,...xi] ^T , i=8; d. Use the Mahalanobis distance and compound probability method to identify the target: First compare the features of the target object with the known oil spill template, the feature vector x=[x ₁ , x ₂ ,…, x _i ] ^T of the known image target area, i is the number of feature values in step c , x _i is the eigenvalue calculated in step c, and the eigenvector m _j , where j=1, 2, is two known classifications namely oil spill and “similar”, firstly find the correlation of the eigenvector x variance matrix

Then, according to the formula r ² _j =(xm _j ) ^T C ^-1 (xm _j ), the feature vector x of the image target area and the feature vector m _j (j=1 , 2) the Mahalanobis distance between; Then the compound probability is calculated for the target object in the dark area belonging to the oil spill area or "similar objects", considering that the image has i characteristic values, i=1, 2, ..., 8, where the first The probability that the i eigenvalues are oil spills is p _i (xi ₎ , and the probability of being “similar” is q _i (xi ₎ , and then using the function

The composite probability p of the oil spill is obtained; e. Use the two methods in step d to judge whether the image in the dark area is an oil spill. First, the Mahalanobis distance is used to judge the feature. If r1<r2, it is judged that x belongs to the oil spill, otherwise it is a similar object, and r1 is the target The Mahalanobis distance between the object and the known classification oil spill, r2 is the Mahalanobis distance between the target object and the known classification "similar objects", if the difference between x and the Mahalanobis distance of the two classifications is not large, use the above function Calculate the composite probability p for further judgment. For those with a composite probability of 67%-100%, they are determined to be oil spills, those with a composite probability of less than 33% are "similar", and those between 34% and 66% cannot be determined. Is it an oil spill. 2. according to the said SAR image oil spill identification method based on eigenvalue of claim 1, it is characterized in that the concrete obtaining process of Mahalanobis distance among the described step d is: first seek the co-ordinate of the characteristic vector x of known image target area variance matrix

Then, extract k statistical samples of oil spill m1 and "similar objects" m2 from the empirical database, and calculate the sample mean

and

Then find x and

and

Mahalanobis distance

where j=1,2. 3. according to the described SAR image oil spill identification method based on eigenvalue of claim 1, it is characterized in that i described in the step d is the i eigenvalue obtained in the step c, and the probability that it belongs to the oil spill is p _i (x _i ), the frequency that the result calculated by this feature appears in the known oil spill feature database, similarly, the probability of the i-th feature value in the "similar object" is calculated as q _i (x _i ).

Images