CN103955926A - Method for remote sensing image change detection based on Semi-NMF - Google Patents

Abstract

The invention discloses a method for remote sensing image change detection based on the Semi-NMF. A processed object simultaneously comprises an optical remote sensing image and a synthetic aperture radar image. The method mainly solves the problems that when a strong change region is obtained through an existing method for remote sensing image change detection, weak and small change regions can not be detected, and more detail and marginal information can not be effectively kept. The realization process of the method comprises the steps that (1) a difference image is generated according to the type of a remote sensing image; (2) each feature vector based on neighborhood information and corresponding to one pixel in the difference image is obtained through PCA and a feature matrix X is established through the feature vectors; (3) the Semi-NMF algorithm is conducted on the feature matrix X, so that the feature matrix X is dissolved to a basis matrix F and a coefficient matrix G through iterative operation; (4) according to the coefficient matrix, a change type omega c and an unchanged type omega u are judged, the soft clustering function is achieved, and a binary change detection result is obtained. According to the method, loss of marginal information is reduced, the strong, weak and small change regions can be detected at the same time, the total error rate is reduced, more detail information is kept, and the change result is effectively obtained.

Description

Based on the method for detecting change of remote sensing image of Semi-NMF

Technical field

The invention belongs to digital image processing techniques field, relate generally to Multitemporal Remote Sensing Images and change detection research direction, specifically based on Semi-NMF (Semi-Nonnegative Matrix Factorization, half Non-negative Matrix Factorization) method for detecting change of remote sensing image, handling object has comprised multidate remote sensing image and synthetic-aperture radar (Synthetic Aperture Radar, SAR) image simultaneously.The method can be applicable to many practical problemss of earth observation.

Background technology

In recent years, along with the fast development of spationautics, sensor technology, computer technology and related discipline, remote sensing technology is also able to continuous progress, the remote-sensing flatform (as aircraft or satellite etc.) of various lift-launch optics and SAR sensor moves in succession, can large area, rapidly, dynamically carry out earth observation, obtain the remote sensing images of wide cut, multidate, for atural object perception and detection provide effective data source.Being limited by the movable aggravation of the mankind and disaster takes place frequently, utilize areal remote sensing images that different time obtains to carry out earth's surface and change that to detect be the focus of remote sensing application research always, and be widely used in all many-sides of national economy and national defense construction, such as soil utilization, environmental monitoring, forest inventory investigation, city planning, the condition of a disaster assessment etc.

Multitemporal Remote Sensing Images changes detection and refers to that two width or several remote sensing images by the same area, different times are obtained compare analysis, and then obtain the change information of interested atural object, scene or target according to difference between image, mainly comprise three basic steps: 1) image pre-service; 2) change detection extracts; 3) aftertreatment and Performance Evaluation.The present invention is mainly the innovation work extracting for the 2nd step change detection.

So far, Chinese scholars has proposed a lot of effectively change detecting methods, and summing up to be divided into has supervision to change detection and change and detect without supervision.Wherein, having supervision to change detection needs to obtain in advance the sample training about the true classification in ground, but in most of the cases, is difficult to maybe cannot obtain about the truth on ground, has restricted to a great extent its application in practice.Changing detection without supervision is directly to compare analysis for the image of different times, does not need out of Memory, relatively simple, directly perceived, has therefore obtained broad research and application.It is roughly divided into: (1) Multitemporal Remote Sensing Images based on Cluster Distribution Divergence changes detection; (2) Multitemporal Remote Sensing Images of analyzing based on differential image changes detection; And the Multitemporal Remote Sensing Images that merge based on Markov (3) changes detection.Particularly general to the research of Equations of The Second Kind algorithm, core concept is the binary classification/segmentation problem that variation test problems is considered as to image, can be subdivided into again the strategies such as cluster analysis, intelligent optimization, Threshold segmentation, Finite mixture model, markov random file, active profile and level set, relate to many basic theories and the method for image processing, pattern-recognition and field of machine vision.Wherein, cluster analysis simple by it, effectively and generally approved.

T.Celik is at document [T.Celik, " Unsupervised change detection in satellite images using principal component analysis and K-means clustering; " IEEE Geoscience and Remote Sensing Letters, vol.6, no.4, the method for detecting change of remote sensing image of cluster analysis is proposed pp.772-776,2009.] the earliest.On this basis, applicant produces the differential image of squelch by low order Fractional Fourier Transform, to reduce noise to detecting the impact of performance.Both all use the simplest hard clustering algorithm (being K-means cluster) to realize differential image feature clustering, produce corresponding variation class and do not change class, but being limited to the limitation of K-means clustering algorithm, are difficult to obtain reasonable testing result.Follow-up, [the A.Ghosh such as A.Ghosh, N.S.Mishra, and S.Ghosh, " Fuzzy clustering algorithms for unsupervised change detection in remote sensing images ", Information Sciences, vol.181, no.4, pp.699-715, 2011.] and [M.Volpi such as M.Volpi, D.Tuia, G.Camps-Valls, and M.Kanevski, " Unsupervised change detection with kernels, " IEEE Geoscience and Remote Sensing Letters, vol.9, no.6, pp.1026-1030, 2012] philosophy adopts fuzzy C-means clustering (Fuzzy C-means), Gustafson-Kesse fuzzy clustering and core K-means clustering algorithm carry out Multitemporal Remote Sensing Images and change detection, improve to a certain extent detection performance, but still there is higher probability of false detection, particularly for high-resolution remote sensing images.Therefore, be necessary to invent a kind of simple and can obtain the better Multitemporal Remote Sensing Images change detecting method of detection effect, to improve Remote Sensing Imagery Change Detection using value in practice.

Summary of the invention

The present invention proposes a kind of Remote Sensing Imagery Change Detection new method based on Semi-NMF.The method is extracted the eigenvector that compacts corresponding to the each pixel of differential image by principal component analysis (PCA) (Principal Component Analysis, PCA), with construction feature matrix; Then carry out Semi-NMF algorithm and be decomposed into basis matrix F and matrix of coefficients G.Wherein, matrix of coefficients is the degree of membership matrix of differential image pixel, and it is adjudicated to the soft cluster that realizes set of pixels, is finally changed testing result.Specific implementation process comprises the steps:

(1) utilize and obtain at areal different time that two width sizes are identical, the remote sensing images Y of mutual registration ₀and Y ₁to produce differential image Y _d, different with SAR image calculation mode for remote sensing image.

(2) construction feature matrix X.This step is mainly to utilize PCA to obtain the proper vector of each pixel in differential image, utilizes these proper vector construction feature matrixes X.Concrete implementation can be divided into following four steps.

(2a) by differential image Y _dbe divided into the non-overlapped sub-block of h × h, and h>=2.In a certain order differential image piece is arranged as to vector form, is expressed as y _d(x, y).

(2b) the corresponding vector set of non-overlapped sub-block is carried out to PCA to generate latent vector space.First calculate the average vector μ of above-mentioned column vector collection; Then each vector is deducted to average vector μ and obtain difference value vector collection Δ _a, and build covariance matrix C; Obtain the proper vector { e of covariance matrix C finally by svd _sand eigenwert { λ _s, and by eigenwert size descending sort corresponding with them proper vector foundation.

(2c) for each pixel coordinate (i, j) of differential image, shine upon y _d(x, y) arrives latent vector space generating feature vector,

V(i，j)＝[v ₁?v ₂...v _s] ^T

Wherein 1≤S≤h ², and parameter S determines the dimension at the proper vector V of locus (i, j) (i, j), y _d(i, j) is the overlapping block obtaining on differential image, according to the identical vector form that is regularly arranged in (2a).

(2d) building size using the proper vector V (i, j) of all pixels of differential image as column vector is the eigenmatrix X of S × HW.

(3) carry out Semi-NMF algorithm.This step is mainly to carry out Semi-NMF algorithm to obtaining eigenmatrix X in step (2), is decomposed into basis matrix F and matrix of coefficients G.

In order to find basis matrix F and matrix of coefficients G, we must solve cost function (being minimum reconstructed) and make its convergence.In view of cost function is not that associating is protruding for F and G, the iterative algorithm that can use F and G alternately to upgrade solves above-mentioned minimization problem, comprises two steps below, before iteration, first Semi-NMF algorithm is carried out to random initializtion,

(3a) retention coefficient matrix G is constant, upgrades basis matrix F,

(3b) keep basis matrix F constant, upgrade matrix of coefficients G,

Then whether checking restrains.If reach the condition of convergence, the computing of Semi-NMF algorithm finishes.If do not reach the condition of convergence, return to (3a) and locate to continue iteration.

(4) obtain and change testing result.Semi-NMF algorithm can be realized soft the function of convergence, is better than K-means and core K-means clustering algorithm, can weaken noise and the impact of heterogeneous pixel, intactly from differential image, extracts change information.Eigenmatrix X obtains through Semi-NMF algorithm process the matrix G that size is 2 × HW ^t, this matrix is exactly that differential image pixel belongs to variation class ω _cdo not change class ω _udegree of membership oriental matrix.According to matrix G ^tin each row, degree of membership size is to judge that representative pixel belongs to variation class ω _cor do not change class ω _u,

$c(i,j)=\left\{\begin{array}{cc}0,& G^T(1,(i-1)*H+j)\&GreaterEqual;G^T(2,(i-1)*H+j)\\ 1,& G^T(1,(i-1)*H+j)<G^T(2,(i-1)*H+j)\end{array}\right.$

Then calculate respectively the mean value that two classes comprise differential image pixel size, the pixel class larger for average is set to change class ω _c, another one pixel class is set to not change class ω _u, the variation that obtains two-value detects figure.

For different image sources, differential image Y _dthere is different obtain manners: for remote sensing image, differential image Y _dfor input picture Y ₀and Y ₁the absolute value directly subtracting each other, Y _d=| Y ₁-Y ₀|; For synthetic-aperture radar (SAR) image, its differential image Y _dget input picture Y ₀and Y ₁make the absolute value of logarithm ratio, $Y_D=\left|\log \frac{Y_1}{Y_0}\right|=|\log \left(Y_1\right)-\log \left(Y_0\right)|.$

Beneficial effect

1, the present invention, in conjunction with the neighborhood information of each pixel, utilizes PCA to extract the compact proper vector corresponding with each pixel in differential image, and has the eigenmatrix of excellent judgement ability with its structure, can effectively weaken noise and the impact of heterogeneous pixel.

2, the present invention first by Semi-NMF algorithm application in Remote Sensing Imagery Change Detection, be mainly used in eigenmatrix to decompose, obtain the degree of membership matrix of differential image pixel, realize the function of soft cluster.This algorithm can effectively be distinguished and changes class and do not change class pixel, accurately realizes changing and detects.

3, the method can at the stronger region of variation of complete acquisition simultaneously, detect faint, less region of variation, reduces the loss of marginal information.Can take into account false detection rate and loss detection rate, total false rate is reduced, retain more details information, effectively obtain result of variations.

Brief description of the drawings

The method for detecting change of remote sensing image general frame of Fig. 1 based on Semi-NMF;

First group of remote sensing image data collection Bangladesh that Fig. 2 emulation experiment is used, (a) the ESA Envisat ASAR image obtaining on April 12nd, 2007, (b) the ESA Envisat ASAR image obtaining on July 26th, 2007, (c) changes and detects reference diagram;

Second group of remote sensing image data collection lake that Fig. 3 emulation experiment is used, (a) remote sensing image obtaining on August 5th, 1986, (b) remote sensing image obtaining on August 5th, 1992, (c) change and detect reference diagram, (d) color base (Color Key) illustraton of model;

Bis-groups of differential images corresponding to remote sensing image data collection difference of Fig. 4: (a) Bangladesh data set differential image, (b) lake data set differential image;

The variation testing result of Fig. 5 the present invention and control methods emulation SAR view data Bangladesh, (a) the variation testing result of control methods 1, (b) the variation testing result of control methods 2, (c) variation testing result of the present invention, (d) changes and detects reference diagram;

The variation testing result of Fig. 6 the present invention and control methods simulate optical remote sensing image data lake, (a) the variation testing result of control methods 1, (b) the variation testing result of control methods 2, (c) variation testing result of the present invention, (d) changes and detects reference diagram;

Embodiment

Embodiments of the invention are described with reference to the accompanying drawings.Should be appreciated that each embodiment of the present invention described here is only used to explain better principle of the present invention and concept, instead of will limit the present invention.After reading such description, those skilled in the art are easy to construct other amendments or replacement, and such amendment or replacement should be understood to fall into scope of the present invention.

Fig. 1 has provided the flow process framework of the embodiment of the present invention, and concrete enforcement comprises following steps:

(1) utilize Multitemporal Remote Sensing Images to produce differential image.Two width sizes are the remote sensing images Y of H × W and mutual registration ₀={ y ₀(i, j) | 1≤i≤H, 1≤j≤W} and Y ₁={ y ₁(i, j) | 1≤i≤H, 1≤j≤W}, they are at areal different time t ₀and t ₁the remotely-sensed data obtaining respectively.Generate differential image Y according to the type of remote sensing images _d.

For optical imagery (Optical Images), input picture Y ₀and Y ₁the absolute value directly subtracting each other is exactly differential image Y _d,

Y _D＝|Y ₁-Y ₀|

And for SAR (Synthetic Aperture Radar) image, consider the property taken advantage of coherent speckle noise, differential image Y _dinput picture Y ₁and Y ₀ratio images, simultaneously in order to improve the pixel value of low amplitude, we carry out a logarithm operation by correlative value image,

$Y_D=\left|\log \frac{Y_1}{Y_0}\right|=|\log \left(Y_1\right)-\log \left(Y_0\right)|$

Wherein log is natural logarithm operator, can make to change in differential image class ω _cdo not change class ω _udistribute more even.Make Y simultaneously ₀and Y ₁the inherent residual property taken advantage of (×) spot disturbs and converts additivity (+) interference to.

(2) construction feature matrix X.This step is mainly to utilize PCA (principal component analysis (PCA)) to obtain each pixel characteristic of correspondence vector in differential image, utilizes these proper vector construction feature matrixes X.Concrete implementation can be divided into following four steps.

(2a) by differential image Y _dbe divided into the non-overlapped sub-block of h × h, and h>=2.H × h differential image piece taking (x, y) as reference coordinate can be expressed as in the time that h is odd number, (x, y) be the center in image block just in time, wherein be the mathematical upper limit operational character of asking, obtain the smallest positive integral larger than a certain number, for example in a certain order by differential image piece Y _d(x, y) is arranged as vector form and is expressed as,

(2b) vector set corresponding to non-overlapped sub-block carried out to PCA to generate latent vector space.Use in order to simplify mathematical notation represent vectorial Y _d(x, y), wherein a represents a Numerical Index, it is the mathematical lower limit of operation symbol of asking.

First calculate the average vector of above-mentioned vector set,

$\mathrm{\&mu;}=\frac{1}{M}\underset{a=1}{\overset{M}{\mathrm{\&Sigma;}}}{y}_D^a$

Then each vector is deducted to average vector μ: 1≤a≤M herein, and build covariance matrix

$C=\frac{1}{M}\underset{a=1}{\overset{M}{\mathrm{\&Sigma;}}}{{\mathrm{\&Delta;}}_a\mathrm{\&Delta;}}_a^T$

Wherein T is transposition symbol, and Matrix C size is h ²× h ².In order better to describe these data, use svd to obtain the h of covariance matrix C ²individual proper vector { e _sand eigenwert { λ _s.Suppose the proper vector foundation eigenwert descending sort corresponding with them that Matrix C obtains herein, for example λ _s>=λ _s+1.

(2c) for each pixel coordinate (i, j) of differential image, shine upon y _d(i, j) arrives latent vector space generating feature vector,

V(i，j)＝[v ₁?v ₂...v _s] ^T

Wherein 1≤S≤h ², and parameter S determines the dimension at the proper vector V of locus (i, j) (i, j), y _d(i, j) is the overlapping block obtaining on differential image, according to the identical vector form that is regularly arranged in (2a).

(2d) building size using the proper vector V (i, j) of all pixels of differential image as column vector is the eigenmatrix X of s × HW.

(3) carry out Semi-NMF algorithm.Carry out Semi-NMF algorithm to obtaining eigenmatrix X in step (2), be decomposed into basis matrix F and matrix of coefficients G, as follows,

X≈FG ^T

Known according to Semi-NMF algorithm, eigenmatrix X and basis matrix F do not have non-negative constraint, allow to exist negative value, but matrix of coefficients G must meet non-negative condition.

In order to find basis matrix F and matrix of coefficients G, we must solve following cost function (being minimum reconstructed),

arg?min _F，G||X-FG ^T|| _F?s.t.G≥0

Wherein, || || _frepresent Frobenius norm, each element in the representing matrix G of G>=0 is not less than 0.

In view of cost function is not that associating is protruding for F and G, the iterative algorithm that uses F and G alternately to upgrade below solves above-mentioned minimization problem, comprises two steps below, before iteration, first Semi-NMF algorithm is carried out to random initializtion,

(3a) retention coefficient matrix G is constant, upgrades basis matrix F,

F＝XG(G ^TG) ^-1

(3b) keep basis matrix F constant, upgrade matrix of coefficients G,

$G_{\mathrm{ik}}\&LeftArrow;G_{\mathrm{ik}}\sqrt{\frac{{\left(X^{T}F\right)}_{\mathrm{ik}}^{+}+{\left[G{\left(F^{T}F\right)}^{-}\right]}_{\mathrm{ik}}}{{\left(X^{T}F\right)}_{\mathrm{ik}}^{-}+{\left[{G\left(F^{T}F\right)}^{+}\right]}_{\mathrm{ik}}}}$

Then whether checking restrains.If reach the condition of convergence, the computing of Semi-NMF algorithm finishes.If do not reach the condition of convergence, return to (3a) and locate to continue iteration.Note, the positive and negative part of matrix A is expressed as,

$A_{\mathrm{ik}}^{+}=\left(\right|A_{\mathrm{ik}}|+A_{\mathrm{ik}})/2,A_{\mathrm{ik}}^{-}=\left(\right|A_{\mathrm{ik}}|-A_{\mathrm{ik}})/2$

(4) obtain and change testing result.Semi-NMF algorithm can be realized the function of soft cluster, is better than K-means and core K-means algorithm, can weaken noise and the impact of heterogeneous pixel, intactly from differential image, extracts change information.Eigenmatrix X obtains through Semi-NMF algorithm process the matrix G that size is 2 × HW ^t, this matrix is exactly that differential image pixel belongs to variation class ω _cdo not change class ω _udegree of membership oriental matrix.According to matrix G ^tin each row, degree of membership size is to judge that representative pixel belongs to variation class ω _cor do not change class ω _u,

$c(i,j)=\left\{\begin{array}{cc}0,& G^T(1,(i-1)*H+j)\&GreaterEqual;G^T(2,(i-1)*H+j)\\ 1,& G^T(1,(i-1)*H+j)<G^T(2,(i-1)*H+j)\end{array}\right.$

Then calculate respectively the mean value that two classes comprise differential image pixel size, the pixel class larger for average is set to change class ω _c, another one pixel class is set to not change class ω _u, the variation that obtains two-value detects figure.

Effect of the present invention can further illustrate by following experiment:

1, emulation experiment condition

The method for detecting change of remote sensing image based on Semi-NMF and control methods that the present invention proposes are all to realize on identical experiment simulation platform.Experimental calculation machine is configured to AMD Athlon (tm) 7750 dual core processors, dominant frequency 2.70GHz, and RAM is 4.00GB (3.25GB can use), 32-bit operating system, software platform is MATLAB7.13.

2, emulation experiment data

What emulation experiment of the present invention was used is real multi-temporal remote sensing data, comprises remote sensing image and SAR image.

First group of remote sensing image data collection Bangladesh is the ground real change reference diagram that two width SAR view data and a width produce by manual analysis in Fig. 2.Wherein (a) and (b) are identical ASAR (the Advanced Synthetic Aperture Radar) images of two width sizes, by ESAEnvisat satellite in Bangladesh area (India) respectively on April 12nd, 2007 and acquisition on July 26th, 2007.Piece image is the earth's surface information in this region before flood generation on April 12nd, 2007, and the second width image is the earth's surface information of the same area after flood generation on July 26th, 2007, and the size of two width images is 300 × 300, and gray level is 256.The 3rd width image is to change reference diagram, and it is the width bianry image obtaining in conjunction with real ground, locality change information by the known view data of manual analysis, and the black-pixel region in figure represents the situation of change on ground.

Second group of remote sensing image data collection lake, in Fig. 3 (a) and (b) be two width remote sensing images, is (c) to change reference diagram, (d) is a width color base (Color Key) illustraton of model.Wherein image (a), (b) and size (c) are 200 × 200, and front two width images respectively on August 5th, 1986 and on August 5th, 1992 at Lake Tahoe, Reno, Nevada obtains.This region has been subject to arid impact, causes lake surface to change, and also has land use and deforestation to cause ground table section to change simultaneously, and the 3rd width image is the variation reference diagram in this region.The variation reference diagram in this region is to utilize Landsat multispectral scanner (Multi Spectral Scaner, MSS) and thematic mapper (Thematic Mapper, TM) data acquisition, for identification and description earth's surface change information.

In Fig. 3, (d) is the Color Key illustraton of model that color is encoded, and detects reference diagram for explaining to change.In Color Key figure, the increase of image Green and minimizing use green and pinkish red (Magenta) to represent successively, and brightness weaken the blueness that is shown as the different depths, the enhancing of brightness is shown as orange/red tone, table 1 has provided these corresponding relations.Any brightness and the green combination changing can characterize by color corresponding in Color Key figure.Detect in reference diagram in the variation of Fig. 3 (c), canescence represents not change or unessential variation; The borderline region in lake is shown as Chinese red, means due to the dry remarkable enhancing that causes brightness of lake surface; Orange pocket may be the variation causing due to the activity of the forest lands such as soil utilization variation or timber felling.

Corresponding relation between table 1 image change and Color Key illustraton of model

The variation of image	Green increasing	Green minimizing	Brightness strengthens	Brightness deterioration
					Color Key illustraton of model	Green	Pinkish red	Orange/red	Blue

3, evaluation of result index

General in the situation that actual change is known, weighing change detection algorithm performance has following three indexs.

(1) P _fA: false detection rate, represents that error-detecting pixel count (False Alarms, FA) accounts for the number percent of total number of image pixels (N), i.e. P _fA=FA/N × 100.Error-detecting pixel count represents that reality does not change and is detected as the sum of all pixels of variation.

(2) P _mA: loss detection rate, represents that loss detection pixel count (Missed Alarms, MA) accounts for the number percent of total number of image pixels (N), i.e. P _mA=MA/N × 100.Loss detection pixel count represents actual change and is detected as unchanged sum of all pixels.

(3) P _tE: total false rate, represents that total erroneous pixel number (Total Errors, TE) accounts for the number percent of total number of image pixels (N), i.e. P _tE=(FA+MA)/N × 100.Total erroneous pixel number is error-detecting pixel count (FA) above and the summation of loss detection pixel count (MA).This is to weigh the most important index of change detection algorithm, is the main basis for estimation that proves result quality.

4, contrast simulation experiment

For validity of the present invention is described, we compare simulated effect of the present invention and following two methods:

Control methods 1, document [T.Celik, " Unsupervised change detection in satellite images using principal component analysis and K-means clustering; " IEEE Geoscience and Remote Sensing Letters, vol.6, no.4, pp.772-776,2009.] variation that utilizes principal component analysis (PCA) (PCA, Principal Component Analysis) and K-means clustering algorithm to realize remote sensing images in detects.First the method utilizes principal component analysis (PCA) to extract each pixel characteristic of correspondence vector, re-uses K-means algorithm proper vector is divided into two classes (change class and do not change class), obtains changing detection figure.

Control methods 2, document [Y.Q.Cheng, H.C.Li, T.Celik, and F.Zhang, " FRFT-based improved algorithm of unsupervised change detection in SAR images via PCA and K-means clustering ", in Proceeding of IEEE International Geoscience and Remote Sensing Symposium, Melbourne, Australia, July2013:1952-1955.] be to utilize the Fractional Fourier Transform of low order to improve the method that document proposes above.

5, emulation experiment content

Test 1 calculated difference image

Emulation experiment of the present invention adopts two groups of remote sensing image data collection, comprises SAR image and remote sensing image.According to different formula, generate differential image Y according to the type of remote sensing images _d, respectively as Fig. 4 (a) with (b).

Experiment 2SAR view data Bangladesh

That this emulation experiment is used is first group of remote sensing image data collection Bangladesh.Fig. 5 is the variation testing result of the present invention and control methods emulation SAR view data Bangladesh, wherein Fig. 5 (a) is the variation testing result of control methods 1, Fig. 5 (b) is the variation testing result of control methods 2, Fig. 5 (c) is variation testing result of the present invention, and Fig. 5 (d) changes to detect reference diagram.

Except observing and carry out qualitative contrast according to vision, also carry out quantitative comparison.Table 2 has been listed the accuracy evaluation data of emulation experiment output image.

Table 2 the present invention and control methods emulation Bangladesh data variation testing result accuracy evaluation data

Algorithm	P FA	P MA	P TE
				Control methods 1	0％	4.84％	4.84％
Control methods 2	0％	4.76％	4.76％
				This method	1.39％	0.57％	1.96％

Test 3 remote sensing image data lake

That this emulation experiment is used is first group of remote sensing image data collection lake.Fig. 6 is the variation testing result of the present invention and control methods simulate optical remote sensing image data lake, wherein Fig. 6 (a) is the variation testing result of control methods 1, Fig. 6 (b) is the variation testing result of control methods 2, Fig. 6 (c) is variation testing result of the present invention, and Fig. 6 (d) changes to detect reference diagram.

6, analysis of simulation result

Variation testing result and accuracy evaluation data by analysis the present invention and control methods emulation SAR view data Bangladesh are known, and the total false rate of simulation result of the present invention is 1.96%, has obtained good result than other control methods.And loss detection rate of the present invention is 0.57%, 4.84% and 4.76% obviously lower than control methods, detects performance better.The false detection rate of control methods is 0%, mainly because contrast algorithm can not cause in complete change detected region.

By analyzing the variation testing result of the present invention and control methods simulate optical remote sensing image data lake and image data set, to change reference diagram information known, the variation that this view data has mainly shown that lake surface dries up, land use and deforestation cause.Detect in reference diagram in variation, canescence represents not change or unessential variation; The borderline region in lake is shown as Chinese red, means due to the dry remarkable enhancing that causes brightness of lake surface; Orange pocket may be the variation causing due to the activity of the forest lands such as soil utilization variation or timber felling.Control methods can detect the dry variation causing of lake surface preferably, still cannot effectively detect the variation that land use and deforestation cause, and loses a large amount of faint region of variation.Simulation result of the present invention not only can more intactly detect the dry variation causing of lake surface than other control methods, has retained the slight change that land use and deforestation cause simultaneously.

In sum, the present invention is applicable to detect multidate remote sensing image and SAR image simultaneously, obtains by PCA the proper vector of compacting, and has the eigenmatrix of excellent judgement ability with its structure.Then carry out Semi-NMF algorithm, eigenmatrix is decomposed into basis matrix F and matrix of coefficients G.Matrix of coefficients is exactly the degree of membership matrix of differential image pixel, set of pixels is realized to the function of soft cluster, reduces the impact of noise and heterogeneous pixel, is preferably changed testing result.By the comparison of quantitative and qualitative analysis, can prove the method for detecting change of remote sensing image based on Semi-NMF proposed by the invention, can in the stronger region of variation of complete acquisition, faint, less region of variation be detected, reduce the loss of marginal information.And take into account loss detection rate and false detection rate, total false rate is reduced, retained more details information, effectively obtained result of variations.

Those of ordinary skill in the art is obviously clear and understand, the inventive method for above embodiment only for the inventive method is described, and be not limited to the inventive method.Although effectively described the present invention by embodiment, there are many variations and do not depart from spirit of the present invention in the present invention.Without departing from the spirit and substance of the case in the method for the present invention, those skilled in the art are when making various corresponding changes or distortion according to the inventive method, but these corresponding changes or distortion all belong to the protection domain that the inventive method requires.

Claims (5)

1. the method for detecting change of remote sensing image based on Semi-NMF, comprises the steps:

Step 1, utilizes and obtains at areal different time that two width sizes are identical, the remote sensing images Y of mutual registration ₀and Y ₁to produce differential image Y _d;

Step 2, construction feature matrix X, this step is mainly to obtain each pixel characteristic of correspondence vector in differential image by PCA, utilizes these proper vector construction feature matrixes X;

Step 3, carries out Semi-NMF algorithm to obtaining eigenmatrix X in step 2, is decomposed into basis matrix F and matrix of coefficients G;

Step 4, judges and changes class and do not change class according to matrix of coefficients G, changes class location of pixels and sets to 0, and does not change class location of pixels and puts 1, obtains final two-value and changes testing result.

2. method for detecting change of remote sensing image according to claim 1, is characterized in that the construction feature matrix X described in step 2, and concrete enforcement comprises:

(a) by differential image Y _dbe divided into the non-overlapped sub-block of h × h, and h>=2; In a certain order differential image piece is arranged as to vector form, is expressed as y _d(x, y);

(b) the corresponding vector set of non-overlapped sub-block is carried out to PCA to generate latent vector space; First calculate the average vector μ of above-mentioned vector set; Then each vector is deducted to average vector μ and obtain difference value vector collection Δ _a, and build covariance matrix C; Obtain the proper vector { e of covariance matrix C finally by svd _sand eigenwert { λ _s, and by eigenwert size descending sort corresponding with them proper vector foundation;

(c) for each pixel coordinate (i, j) of differential image, shine upon y _d(i, j) arrives latent vector space generating feature vector,

V(i，j)＝[v ₁?v ₂...v _s] ^T

Wherein 1≤S≤h ², and parameter S determines the dimension at the proper vector V of locus (i, j) (i, j), y _d(x, y) is the overlapping block obtaining on differential image, according to the identical vector form that is regularly arranged in (a);

(d) building size using the proper vector V (i, j) of all pixels of differential image as column vector is the eigenmatrix X of S × HW.

3. method for detecting change of remote sensing image according to claim 1, it is characterized in that eigenmatrix X being carried out to Semi-NMF algorithm described in step 3, be decomposed into basis matrix F and matrix of coefficients G, for finding basis matrix F and matrix of coefficients G, use the iterative strategy that F and G alternately upgrade to solve minimum reconstructed, comprise two steps below, before iteration, first Semi-NMF algorithm is carried out to random initializtion

(a) retention coefficient matrix G is constant, upgrades basis matrix F,

(b) keep basis matrix F constant, upgrade matrix of coefficients G,

Then whether checking restrains; If reach the condition of convergence, the computing of Semi-NMF algorithm finishes; If do not reach the condition of convergence, return to (a) and locate to continue iteration.

4. method for detecting change of remote sensing image according to claim 1, is characterized in that the acquisition described in step 4 changes testing result, is embodied as: eigenmatrix X obtains through Semi-NMF algorithm process the matrix G that size is 2 × HW ^t, this matrix is exactly that differential image pixel belongs to variation class ω _cdo not change class ω _udegree of membership oriental matrix; According to matrix G ^tin each row, degree of membership size is to judge that representative pixel belongs to variation class ω _cor do not change class ω _u,

$c(i,j)=\left\{\begin{array}{cc}0,& G^T(1,(i-1)*H+j)\&GreaterEqual;G^T(2,(i-1)*H+j)\\ 1,& G^T(1,(i-1)*H+j)<G^T(2,(i-1)*H+j)\end{array}\right.$

Then calculate respectively the mean value that two classes comprise differential image pixel size, the pixel class larger for average is set to change class ω _c, another one pixel class is set to not change class ω _u, the variation that obtains two-value detects figure.

5. method for detecting change of remote sensing image according to claim 1, is characterized in that described differential image Y _d: for remote sensing image, differential image Y _dfor input picture Y ₀and Y ₁the absolute value directly subtracting each other, Y _d=| Y ₁-Y ₀|; For synthetic-aperture radar (SAR) image, its differential image Y _dget input picture Y ₀and Y ₁make the absolute value of logarithm ratio,