CN104680151A - High-resolution panchromatic remote-sensing image change detection method considering snow covering effect - Google Patents

Abstract

The invention discloses a high-resolution panchromatic remote-sensing image change detection method considering snow covering effect. According to high-resolution panchromatic remote-sensing image change detection method, the characteristics of the snow covering on a panchromatic image is adequately used, the change detection is performed by adopting a change detecting method on the basis of a textural feature, and a change area is extracted. Snow areas on the tested old and new images are extracted by using Chan-Vese partition method on the basis of a level set method, according to a change detecting result and a snow covering condition on the old and new images, a pseudo change caused by the snow covering change is removed, a change area is extracted, change detecting accuracy and the automation degree for map revision are improved, and a data updating cycle is shortened.

Description

A kind of panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account

Technical field

The invention belongs to image processing field, relate to a kind of remote sensing image process and change detecting method, particularly a kind of snow of taking into account based on high-resolution remote sensing image covers the high-resolution panchromatic remote sensing image region of variation extracting method affected.

Background technology

The Northeast's latitude is high, next-door neighbour dry monsoon seedbed, is one of the main accumulated snow district in China winter, large-area accumulated snow utilizes the water resource of Northeast Regional, diastrous weather and general circulation have important impact.When carrying out change to Northeast Regional and detecting, the covering of accumulated snow also creates great impact to the result that change detects.The change of Snow-Cover causes generating portion puppet change in change testing result.How to reduce the impact of snow cover, remove the puppet change of the change of snow cover situation and generation, extracting region of variation more accurately, is the difficult point that Northeast Regional change detects, and the research for very cold region has important meaning.

High-resolution remote sensing image widespread use, traditional change based on pixel detects has significant limitation on yardstick, a kind of object-based change detecting method arises at the historic moment, what first OO change detecting method will solve is the problem split, at present not for the universality algorithm of high-resolution remote sensing image.When change detection is carried out to the Northeast, get rid of accumulated snow impact and be even more important.How determining accumulated snow region, is get rid of accumulated snow to affect the problem that first will solve.Many scholars know method for distinguishing to snow remote sensing and have carried out useful exploration, at visible ray and near-infrared band, there is higher reflection characteristic according to accumulated snow, in the characteristic that the reflectivity at middle-infrared band place is lower, obtain snow cover information by certain digital image processing techniques.Dozier proposes the snow lid index concept utilizing accumulated snow reflectivity Characteristics, in order to distinguish accumulated snow and other atural object.Hall adopts snow lid index to develop complete snow-cover mapping SNOMAP algorithm.Shaded side accumulated snow, can accumulated snow in decipher shade according to the distribution of snow lid and the substantial connection of landform due to the None-identified that affects of the factors such as landform, has the accumulated snow researched and proposed in the method decipher shade correcting illumination change.Various algorithm respectively has its relative merits above, can be applicable to all changes detect case without any a kind of algorithm.The image that different sensors obtains has respective feature, and the scope of application is different, the complicacy of atural object self and the impact of environmental factor, makes to change the process detected and becomes particularly complicated.Use remote sensing image to carry out change to the Northeast to detect, the covering of accumulated snow makes to change the difficulty detected and increases, and single algorithm cannot extract region of variation accurately.

Summary of the invention

In order to overcome the deficiencies in the prior art, the present invention proposes a kind of panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account.The present invention uses the Chan-Vese dividing method based on level set, take into full account that pixel faces the relation in territory with it, be partitioned into the accumulated snow region on two phase remote sensing image, by itself and the results contrast detected based on group's strategy change, eliminate the puppet change that target to be detected produces due to Snow-Cover change, improve the precision that change detects.

The technical solution adopted in the present invention is: a kind of panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account, is characterized in that, comprise the following steps:

Step 1: the two width remote sensing images inputting same place different times;

Step 2: adopt group's strategy to carry out change to two width remote sensing images of input and detect; The Chan-Vese dividing method based on level set is adopted to carry out accumulated snow region segmentation to two width remote sensing images of input, consider that pixel faces the relation in territory with it, be partitioned into the accumulated snow region on two phase remote sensing image, by itself and the results contrast detected based on group's strategy change, determine the change type of two width remote sensing images, described change type is divided into: change class, the class that do not change, pseudo-change class and suspiciously do not change class; Described change type determining standard is:

Change class: if two width remote sensing images are all without snow, then change testing result for change;

Do not change class: if two remote sensing width images are all without snow, then change testing result for not change;

Pseudo-change class: if the snowy width of two width remote sensing image one is without snow, then change testing result for change;

Suspiciously do not change class: if two width remote sensing images are all snowy, then change testing result for not change;

Step 3: the panchromatic image change taking snow into account detects, and when two width remote sensing images are all classified as suspicious region by during snow cover, does not process, determines in knock-out experiment process to detect target puppet change of generation due to the change of snow cover situation;

Step 4: adopt overall accuracy, Kappa coefficient, change class accuracy and do not change class accuracy to carry out precision evaluation.

As preferably, adopt group's strategies to carry out change detection described in step 2 to two width remote sensing images of input, its specific implementation comprises following sub-step:

Step 2.1.1: the textural characteristics figure of two width remote sensing images described in generation, calculate image greyscale co-occurrence matrix, textural characteristics value needed for window calculation is got by pixel, then by calculate textural characteristics value assignment to central point pixel, after completing, window is moved a pixel, form another one new window, recalculate, the like, constitute one with the eigenvalue matrix of the equal-sized gray level co-occurrence matrixes of former figure;

Step 2.1.2: adopt omnidirectional's texture value (i.e. directive mean value), calculate the average of the texture of 0 °, 45 °, 90 °, 135 ° four direction respectively, compared with the texture template image of single direction, omnidirectional's texture has rotational invariance;

Step 2.1.3: the textural characteristics choosing suitable pixel distance and window size acquisition needs;

Step 2.1.4: texture feature extraction, adopts following three kinds of method of discrimination to obtain difference image respectively, adopts the comprehensive three kinds of indexs of group's strategy to judge object variations; (i, j) for the gray-scale value of different two pixels, p (i, j) be gray level co-occurrence matrixes:

(4-a) second order is apart from (Eneryg), the quadratic sum of gray level co-occurrence matrixes element value, and it is formulated as:

$\mathrm{Eneryg}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}p{(i,j)}^2;$

(4-b) entropy (Entropy), image has the tolerance of the randomness of quantity of information, represents non-uniform degree or the complexity of texture in image; If texture is complicated, entropy is large; Otherwise if uniform gray level in image, in co-occurrence matrix, element size difference is large, and entropy is little; The computing formula of entropy is:

$\mathrm{Entropy}=-\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}p(i,j)\log (i,j);$

(4-c) homogeneity degree (Homogeneity) is also called contrast square, the homogeneity of reflection gradation of image; When pixel is to time even, the value of homogeneity degree is relatively high; Its computing formula is:

$\mathrm{Homogeneity}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}\frac{1}{1+{(i-j)}^2}p(i,j);$

Step 2.1.5: setting threshold value iterative, its algorithmic formula is f (i, j)=g _t1(i, j)-g _t2(i, j), wherein g _t1(i, j) and g _t2(i, j) for textural characteristics image being in the gray-scale value of the pixel of the i-th row jth row, f (i, j) is the gray-scale value of the pixel of the i-th row jth row on texture Difference image, choose process of iteration calculated threshold herein and carry out binaryzation, the process of process of iteration is:

Step 2.1.5.1: to choose on image the maximal value of gray scale and the average T of minimum value as initial threshold value;

Step 2.1.5.2: with T as Threshold segmentation image, grey scale pixel value be greater than T as prospect part, grey scale pixel value is less than the part as a setting of T;

Step 2.1.5.3: the gray average g calculating prospect part and background parts ₁and g ₂, calculate new threshold value T ', wherein: $T^{\′}=\frac{g_1+g_2}{2};$

When T ' and the difference of T are less than setting threshold values M, stop iteration, no person goes back to step 2.1.5.2 and continues to calculate, until satisfy condition, stops iteration, required by threshold value is now;

Step 2.1.6: group decision detects object variations, adopt single feature to carry out change respectively to detect, then the result that three changes detect is merged, obtain changing testing result more accurately, three kinds of testing result importance are consistent, what adopt is majority rule, judges that target to be detected there occurs change, then think that target to be detected there occurs change when there being two kinds or more in three kinds of testing results.

As preferably, described in step 2.1.3 choose suitable pixel distance and window size obtains the textural characteristics needed, wherein pixel distance 1, window size 5x5.

As preferably, the setting threshold values M described in step 2.1.5.3 is 0.01.

As preferably, described in step 2, adopt the Chan-Vese dividing method based on level set to carry out accumulated snow region segmentation to two width remote sensing images of input, its specific implementation comprise for: by making the energy functional E in following formula ^cVminimum realization is split:

$\begin{array}{c}{E}^{\mathrm{CV}}(c_1,c_2,C)=\mathrm{\μ}\·\mathrm{Length}\left(C\right)+{\mathrm{\λ}}_1\·\underset{\mathrm{inside}\left(C\right)}{\∫}|u_0(x,y)-c_1|\mathrm{dxdy}+\\ {\mathrm{\λ}}_2\·\underset{\mathrm{outside}\left(C\right)}{\∫}|u_0(x,y)-c_2|\mathrm{dxdy}\end{array};$

Fixing λ ₁=λ ₂=0, μ is normal number; c ₁and c ₂the mean value of the inside and outside pixel gray scale of curve C, for making energy functional E ^cV(c ₁, c ₂, C) and minimum, use represent curve C, x, y are the coordinate of theorem in Euclid space, then above formula can be expressed as:

Wherein δ _εz () is the Di Lake function of regularization, H _εz () is the regularization form of Hai Shi function,

$H\left(z\right)=\left\{\begin{array}{ccc}1,& \mathrm{if}& z\&GreaterEqual;0\\ 0,& \mathrm{if}& z<0\end{array}\right.,\mathrm{\&delta;}\left(z\right)=\frac{d}{\mathrm{dz}}H\left(z\right);$

Then the EVOLUTION EQUATION of level set is:

with it is the gray average of inside or outside of curve partial pixel.

As preferably, the employing overall accuracy described in step 4, Kappa coefficient, change class accuracy and do not change class accuracy to carry out precision evaluation, wherein:

Overall accuracy (OA) is Baidu's proportion by subtraction that correct pixel number of classifying accounts for all total pixel numbers, it reflects the overall accuracy detected; The wherein pixel number of correct classification eye diagonal line distribution in confusion matrix, for M11's and M22 in following table 1 and; The change pixel number correctly detected in table is M11; The number of the change pixel that error-detecting goes out is M12; What error-detecting went out does not change pixel number is M12; The pixel number that do not change correctly detected is M22;

Table 1

The computing formula of overall accuracy OA is expressed as:

$\mathrm{OA}=\frac{M11+M22}{T}\&times;100\%;$

Kappa coefficient is the method for another one evaluating precision, according to the data calculation formula in confusion matrix is:

$\mathrm{Kappa}=\frac{T\&times;(M11+M22)-(A\&times;C+B\&times;D)}{T^2-(A\&times;C+B\&times;D)};$

Accuracy is defined as the number percent that the change class pixel that detects and non-changing class pixel account for real image unit number very separately; Its computing formula is:

$P_{\mathrm{tc}}=\frac{M11}{A}\&times;100\%;P_{\mathrm{tu}}=\frac{M22}{B}\&times;100\%.$

The present invention makes full use of the feature according to snow cover on panchromatic image, adopts the change based on textural characteristics to detect detection method and carries out change detection, extract region of variation.The Chan-Vese dividing method based on Level Set Method is adopted to extract the accumulated snow region of testing on new and old image, according to change detect result and new and old image on snow cover situation, the puppet change that removal part snow cover changes and causes, extract region of variation, improve the degree of accuracy of change detection and the automaticity of map revision, shorten the Data Update cycle.

Accompanying drawing explanation

Fig. 1: be the schematic flow sheet of the embodiment of the present invention.

Fig. 2: be the aviation image of the somewhere different times 1m spatial resolution of the embodiment of the present invention.

Fig. 3: be the second moment change testing result of the embodiment of the present invention.

Fig. 4: be the Entropy Changes testing result figure of the embodiment of the present invention.

Fig. 5: be the homogeneity degree change testing result figure of the embodiment of the present invention.

Fig. 6: be three Fusion Features change testing result figure of the embodiment of the present invention.

Fig. 7: be image accumulated snow in the early stage region of the same place different times of the embodiment of the present invention.

Fig. 8: be the later stage image accumulated snow region of the same place different times of the embodiment of the present invention.

Fig. 9: be the textural characteristics change testing result of the embodiment of the present invention.

Figure 10: for the snow of taking into account of the embodiment of the present invention covers the change testing result affected.

Embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with drawings and Examples, the present invention is described in further detail, should be appreciated that exemplifying embodiment described herein is only for instruction and explanation of the present invention, is not intended to limit the present invention.

The present embodiment mainly adopts entropy, homogeneity degree and second moment three textural characteristics to carry out change and detects, and the change testing result being merged each texture by the plurality rule of group decision extracts region of variation.Extract the accumulated snow region on two width remote sensing images of same place different times with the Chan-Vese dividing method based on level set, after the impact of removing snow, carry out change detect.Its flow process is shown in Fig. 1.

The resolution that the present embodiment chooses image is 1m, and the size of image is 1000 × 1000 pixels.Contain buildings, house, bare area, road in this group image, it is as follows that change detects concrete implementation step:

Step 1: the two width remote sensing images inputting same place different times;

Step 2: adopt group's strategy to carry out change to two width remote sensing images of input and detect; The Chan-Vese dividing method based on level set is adopted to carry out accumulated snow region segmentation to two width remote sensing images of input, consider that pixel faces the relation in territory with it, be partitioned into the accumulated snow region on two phase remote sensing image, by itself and the results contrast detected based on group's strategy change, determine the change type of two width remote sensing images, described change type is divided into: change class, the class that do not change, pseudo-change class and suspiciously do not change class; Described change type determining standard is:

Change class: if two width remote sensing images are all without snow, then change testing result for change;

Do not change class: if two width remote sensing images are all without snow, then change testing result for not change;

Pseudo-change class: if the snowy width of two width remote sensing image one is without snow, then change testing result for change;

Suspiciously do not change class: if two width remote sensing images are all snowy, then change testing result for not change;

Adopt group's strategy to carry out change to two width remote sensing images of input in step 2 to detect, its specific implementation comprises following sub-step:

Step 2.1.1: the textural characteristics figure of two width remote sensing images described in generation, calculate image greyscale co-occurrence matrix, textural characteristics value needed for window calculation is got by pixel, then by calculate textural characteristics value assignment to central point pixel, after completing, window is moved a pixel, form another one new window, recalculate, the like, constitute one with the eigenvalue matrix of the equal-sized gray level co-occurrence matrixes of former figure;

Step 2.1.2: adopt omnidirectional's texture value (i.e. directive mean value), calculate the average of the texture of 0 °, 45 °, 90 °, 135 ° four direction respectively, compared with the texture template image of single direction, omnidirectional's texture has rotational invariance;

Step 2.1.3: the textural characteristics choosing suitable pixel distance and window size acquisition needs, wherein pixel distance is 1, and window size is 5x5.

Step 2.1.4: texture feature extraction, adopts following three kinds of method of discrimination to obtain difference image respectively, adopts the comprehensive three kinds of indexs of group's strategy to judge object variations; (i, j) for the gray-scale value of different two pixels, p (i, j) be gray level co-occurrence matrixes;

(4-a) second order is apart from (Eneryg), the quadratic sum of gray level co-occurrence matrixes element value, and it is formulated as:

$\mathrm{Eneryg}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}p{(i,j)}^2;$

(4-b) entropy (Entropy), image has the tolerance of the randomness of quantity of information, represents non-uniform degree or the complexity of texture in image; If texture is complicated, entropy is large; Otherwise if uniform gray level in image, in co-occurrence matrix, element size difference is large, and entropy is little; The computing formula of entropy is:

$\mathrm{Entropy}=-\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}p(i,j)\log (i,j);$

(4-c) homogeneity degree (Homogeneity) is also called contrast square, the homogeneity of reflection gradation of image; When pixel is to time even, the value of homogeneity degree is relatively high; Its computing formula is:

$\mathrm{Homogeneity}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}\frac{1}{1+{(i-j)}^2}p(i,j);$

Gray level co-occurrence matrixes is the common method being described texture by the spatial correlation characteristic of research gray scale, it is the matrix function of pixel distance and angle, it, by the correlativity in computed image between certain distance and 2 gray scales of certain orientation, reflects the integrated information of image on direction, interval, amplitude of variation and speed.On the image of N × N size, certain a bit (x, y), and departs 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) number of times that value occurs, be then arranged in a square formation, with (g ₁, g ₂) they are normalized to the probability P of appearance by the total degree that occurs, such square formation is called gray level co-occurrence matrixes.Range difference score value (b, b) gets different combinations of values, can obtain the joint probability matrix under different situations.In experiment, following normalization is carried out to gray level co-occurrence matrixes:

$p_{i,j}=\frac{p(i,j,d,\mathrm{\&theta;})}{\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}p(i,j,d,\mathrm{\&theta;})}$

P (i, j, d, θ) is gray scale is i and j, and distance is d and direction is that the picture point of θ is to the number of times occurred.

Step 2.1.5: setting threshold value iterative, its algorithmic formula is f (i, j)=g _t1(i, j)-g _t2(i, j), wherein g _t1(i, j) and g _t2(i, j) for textural characteristics image being in the gray-scale value of the pixel of the i-th row jth row, f (i, j) is the gray-scale value of the pixel of the i-th row jth row on texture Difference image, choose process of iteration calculated threshold herein and carry out binaryzation, the process of process of iteration is:

Step 2.1.5.1: to choose on image the maximal value of gray scale and the average T of minimum value as initial threshold value;

Step 2.1.5.2: with T as Threshold segmentation image, grey scale pixel value be greater than T as prospect part, grey scale pixel value is less than the part as a setting of T;

Step 2.1.5.3: the gray average g calculating prospect part and background parts ₁and g ₂, calculate new threshold value T ', wherein: $T^{\&prime;}=\frac{g_1+g_2}{2};$

When T ' and the difference of T are less than setting threshold values M=0.01, stop iteration, no person goes back to step 2.1.5.2 and continues to calculate, until satisfy condition, stops iteration, required by threshold value is now;

Step 2.1.6: group decision detects object variations, adopt single feature to carry out change respectively to detect, then the result that three changes detect is merged, obtain changing testing result more accurately, three kinds of testing result importance are consistent, what adopt is majority rule, judges that target to be detected there occurs change, then think that target to be detected there occurs change when there being two kinds or more in three kinds of testing results.

In step, two width remote sensing images of input are adopted and carry out accumulated snow region segmentation based on the Chan-Vese dividing method of level set, its specific implementation comprise for: by making the energy functional E in following formula ^cVminimum realization is split:

$\begin{array}{c}{E}^{\mathrm{CV}}(c_1,c_2,C)=\mathrm{\&mu;}\&CenterDot;\mathrm{Length}\left(C\right)+{\mathrm{\&lambda;}}_1\&CenterDot;\underset{\mathrm{inside}\left(C\right)}{\&Integral;}|u_0(x,y)-c_1|\mathrm{dxdy}+\\ {\mathrm{\&lambda;}}_2\&CenterDot;\underset{\mathrm{outside}\left(C\right)}{\&Integral;}|u_0(x,y)-c_2|\mathrm{dxdy}\end{array};$

Fixing λ ₁=λ ₂=0, μ is normal number; c ₁and c ₂the mean value of the inside and outside pixel gray scale of curve C, for making energy functional E ^cV(c ₁, c ₂, C) and minimum, use represent curve C, x, y are the coordinate of theorem in Euclid space, then above formula can be expressed as:

Wherein δ _εz () is the Di Lake function of regularization, H _εz () is the regularization form of Hai Shi function,

$H\left(z\right)=\left\{\begin{array}{ccc}1,& \mathrm{if}& z\&GreaterEqual;0\\ 0,& \mathrm{if}& z<0\end{array}\right.,\mathrm{\&delta;}\left(z\right)=\frac{d}{\mathrm{dz}}H\left(z\right);$

Then the EVOLUTION EQUATION of level set is:

with it is the gray average of inside or outside of curve partial pixel.

To the differentiate of level set imbedding function:

Wherein Δ t is time step, adopts central difference schemes to have:

Wherein k ₁and k ₂for the step-length on x and y direction.For there is following formula:

Wherein:

Substitution full scale equation obtains:

Compared with the result of gray level threshold segmentation, the accumulated snow region adopting the Chan-Vese method based on level set to split is more complete, has taken into full account that pixel faces the relation in territory with it, more meets ground distribution of Snow Cover Over situation.

Step 3: the panchromatic image change taking snow into account detects, and when two width remote sensing images are all classified as suspicious region by during snow cover, does not process, determines in knock-out experiment process to detect target puppet change of generation due to the change of snow cover situation;

Step 4: adopt overall accuracy, Kappa coefficient, change class accuracy and do not change class accuracy to carry out precision evaluation;

Wherein overall accuracy (OA) is Baidu's proportion by subtraction that correct pixel number of classifying accounts for all total pixel numbers, it reflects the overall accuracy detected; The wherein pixel number of correct classification eye diagonal line distribution in confusion matrix, for M11's and M22 in following table 1 and; The change pixel number correctly detected in table is M11; The number of the change pixel that error-detecting goes out is M12; What error-detecting went out does not change pixel number is M12; The pixel number that do not change correctly detected is M22;

Table 1

The computing formula of overall accuracy OA is expressed as

$\mathrm{OA}=\frac{M11+M22}{T}\&times;100\%;$

Kappa coefficient is the method for another one evaluating precision, according to the data calculation formula in confusion matrix is:

$\mathrm{Kappa}=\frac{T\&times;(M11+M22)-(A\&times;C+B\&times;D)}{T^2-(A\&times;C+B\&times;D)};$

Accuracy is defined as the number percent that the change class pixel that detects and non-changing class pixel account for real image unit number very separately; Its computing formula is:

$P_{\mathrm{tc}}=\frac{M11}{A}\&times;100\%;P_{\mathrm{tu}}=\frac{M22}{B}\&times;100\%.$

Table 2 is the precision index of the evaluation of the embodiment of the present invention:

Table 2

Result in table 2 is when carrying out change detection to this group data, each coefficient calculations result, wherein take the result overall accuracy of snow impact into account, the accuracy of Kappa coefficient and detection has small size raising, illustrates that taking snow covering change detecting method into account is improved to some extent to precision.

The reason that the raising of precision is limited is being gone in pseudo-process, eliminates the terrain object that the while that fraction itself changing, snow cover situation has also changed, exert a certain influence to the precision that change detects.

Should be understood that, the part that this instructions does not elaborate all belongs to prior art.

Should be understood that; the above-mentioned description for preferred embodiment is comparatively detailed; therefore the restriction to scope of patent protection of the present invention can not be thought; those of ordinary skill in the art is under enlightenment of the present invention; do not departing under the ambit that the claims in the present invention protect; can also make and replacing or distortion, all fall within protection scope of the present invention, request protection domain of the present invention should be as the criterion with claims.

Claims (6)

1. take the panchromatic remote sensing image variation detection method of high-resolution that snow covers impact into account, it is characterized in that, comprise the following steps:

Step 1: the two width remote sensing images inputting same place different times;

Step 2: adopt group's strategy to carry out change to two width remote sensing images of input and detect; The Chan-Vese dividing method based on level set is adopted to carry out accumulated snow region segmentation to two width remote sensing images of input, consider that pixel faces the relation in territory with it, be partitioned into the accumulated snow region on two phase remote sensing image, by itself and the results contrast detected based on group's strategy change, determine the change type of two width remote sensing images, described change type is divided into: change class, the class that do not change, pseudo-change class and suspiciously do not change class; Described change type determining standard is:

Change class: if two width remote sensing images are all without snow, then change testing result for change;

Do not change class: if two remote sensing width images are all without snow, then change testing result for not change;

Pseudo-change class: if the snowy width of two width remote sensing image one is without snow, then change testing result for change;

Suspiciously do not change class: if two remote sensing width images are all snowy, then change testing result for not change;

Step 3: the panchromatic image change taking snow into account detects, and when two width remote sensing images are all classified as suspicious region by during snow cover, does not process, determines in knock-out experiment process to detect target puppet change of generation due to the change of snow cover situation;

Step 4: adopt overall accuracy, Kappa coefficient, change class accuracy and do not change class accuracy to carry out precision evaluation.

2. the panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account according to claim 1, it is characterized in that: adopt group strategies to carry out change detection to two width remote sensing images of input described in step 2, its specific implementation comprises following sub-step:

Step 2.1.1: the textural characteristics figure of two width remote sensing images described in generation, calculate image greyscale co-occurrence matrix,

Textural characteristics value needed for window calculation is got by pixel, then by calculate textural characteristics value assignment to central point pixel, after completing, window is moved a pixel, form another one new window, recalculate, the like, constitute one with the eigenvalue matrix of the equal-sized gray level co-occurrence matrixes of former figure;

Step 2.1.2: adopt omnidirectional's texture value (i.e. directive mean value), calculate the average of the texture of 0 °, 45 °, 90 °, 135 ° four direction respectively, compared with the texture template image of single direction, omnidirectional's texture has rotational invariance;

Step 2.1.3: the textural characteristics choosing suitable pixel distance and window size acquisition needs,

Step 2.1.4: texture feature extraction, adopts following three kinds of method of discrimination to obtain difference image respectively, adopts the comprehensive three kinds of indexs of group's strategy to judge object variations; (i, j) for the gray-scale value of different two pixels, p (i, j) be gray level co-occurrence matrixes;

(4-a) second order is apart from (Eneryg), the quadratic sum of gray level co-occurrence matrixes element value, and it is formulated as:

$\mathrm{Eneryg}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}p{(i,j)}^2;$

(4-b) entropy (Entropy), image has the tolerance of the randomness of quantity of information, represents non-uniform degree or the complexity of texture in image; If texture is complicated, entropy is large; Otherwise if uniform gray level in image, in co-occurrence matrix, element size difference is large, and entropy is little; The computing formula of entropy is:

$\mathrm{Entropy}=-\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}p(i,j)\log (i,j);$

(4-c) homogeneity degree (Homogeneity) is also called contrast square, the homogeneity of reflection gradation of image; When pixel is to time even, the value of homogeneity degree is relatively high; Its computing formula is:

$\mathrm{Homogeneity}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}\frac{1}{1+{(i-j)}^2}p(i,j)$

Step 2.1.5: setting threshold value iterative, its algorithmic formula is f (i, j)=g _t1(i, j)-g _t2(i, j), wherein g _t1(i, j) and g _t2(i, j) for textural characteristics image being in the gray-scale value of the pixel of the i-th row jth row, f (i, j) is the gray-scale value of the pixel of the i-th row jth row on texture Difference image, choose process of iteration calculated threshold herein and carry out binaryzation, the process of process of iteration is:

Step 2.1.5.1: to choose on image the maximal value of gray scale and the average T of minimum value as initial threshold value;

Step 2.1.5.2: with T as Threshold segmentation image, grey scale pixel value be greater than T as prospect part, grey scale pixel value is less than the part as a setting of T;

Step 2.1.5.3: the gray average g calculating prospect part and background parts ₁and g ₂, calculate new threshold value T ', wherein: $T^{\&prime;}=\frac{g_1+g_2}{2};$

When T ' and the difference of T are less than setting threshold values M, stop iteration, no person goes back to step 2.1.5.2 and continues to calculate, until satisfy condition, stops iteration, required by threshold value is now;

Step 2.1.6: group decision detects object variations, adopt single feature to carry out change respectively to detect, then the result that three changes detect is merged, obtain changing testing result more accurately, three kinds of testing result importance are consistent, what adopt is majority rule, judges that target to be detected there occurs change, then think that target to be detected there occurs change when there being two kinds or more in three kinds of testing results.

3. the panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account according to claim 2, it is characterized in that: the textural characteristics choosing suitable pixel distance and window size acquisition needs described in step 2.1.3, wherein pixel distance is 1, and window size is 5x5.

4. the panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account according to claim 2, is characterized in that: the setting threshold values M described in step 2.1.5.3 is 0.01.

5. the panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account according to claim 1, it is characterized in that: adopting two width remote sensing images of input described in step 2 carries out accumulated snow region segmentation based on the Chan-Vese dividing method of level set, its specific implementation comprise for: by making the energy functional E in following formula ^cVminimum realization is split:

$\begin{array}{c}{E}^{\mathrm{CV}}(c_1,c_2,C)=\mathrm{\&mu;}\&CenterDot;\mathrm{Length}\left(C\right)+{\mathrm{\&lambda;}}_1\&CenterDot;\underset{\mathrm{inside}\left(C\right)}{\&Integral;}|u_0(x,y)-c_1|\mathrm{dxdy}+\\ {\mathrm{\&lambda;}}_2\&CenterDot;\underset{\mathrm{outside}\left(C\right)}{\&Integral;}|u_0(x,y)-c_2|\mathrm{dxdy}\end{array};$

Fixing λ ₁=λ ₂=0, μ is normal number; c ₁and c ₂the mean value of the inside and outside pixel gray scale of curve C, for making energy functional E ^cV(c ₁, c ₂, C) and minimum, use represent curve C, x, y are the coordinate of theorem in Euclid space, then above formula can be expressed as:

Wherein δ _εz () is the Di Lake function of regularization, H _εz () is the regularization form of Hai Shi function,

$H\left(z\right)=\left\{\begin{array}{cc}1,& \mathrm{ifz}\&GreaterEqual;0\\ 0,& \mathrm{ifz}<0\end{array}\right.,$ $\mathrm{\&delta;}\left(z\right)=\frac{d}{\mathrm{dz}}H\left(z\right);$

Then the EVOLUTION EQUATION of level set is:

with it is the gray average of inside or outside of curve partial pixel.

6. the panchromatic remote sensing image variation detection method of high-resolution taking snow covering impact into account according to claim 1, it is characterized in that: the employing overall accuracy described in step 4, Kappa coefficient, change class accuracy and do not change class accuracy to carry out precision evaluation, wherein:

Overall accuracy (OA) is Baidu's proportion by subtraction that correct pixel number of classifying accounts for all total pixel numbers, it reflects the overall accuracy detected; The wherein pixel number of correct classification eye diagonal line distribution in confusion matrix, for M11's and M22 in following table 1 and; The change pixel number correctly detected in table is M11; The number of the change pixel that error-detecting goes out is M12; What error-detecting went out does not change pixel number is M12; The pixel number that do not change correctly detected is M22;

Table 1

The computing formula of overall accuracy OA is expressed as:

$\mathrm{OA}=\frac{M11+M22}{T}\&times;100\%;$

Kappa coefficient is the method for another one evaluating precision, according to the data calculation formula in confusion matrix is:

$\mathrm{Kappa}=\frac{T\&times;(M11+M22)-(A\&times;C+B\&times;D)}{T^2-(A\&times;C+B\&times;D)};$

Accuracy is defined as the number percent that the change class pixel that detects and non-changing class pixel account for real image unit number very separately; Its computing formula is:

$P_{\mathrm{tc}}=\frac{M11}{A}\&times;100\%;$ $P_{\mathrm{tu}}=\frac{M22}{B}\&times;100\%.$