CN101839980A - Unsupervised remote sensing image change detection method based on segmentation window - Google Patents

Abstract

The invention relates to an unsupervised remote sensing image change detection method based on a segmentation window, belonging to the technical field of remote sensing image processing. The method can segment a differential image into sub-images and determine the overall threshold of the differential image through solving the local thresholds of the sub-images. According to experimental results, the method can effectively solve the problem that a common change detection method cannot conduct accurate change detection when a change region is relatively large or small. Compared with the common change detection method, the detection precision is obviously improved. The method can be used in the remote sensing image change detection field, especially when the area scale of the change region of the remote sensing image is relatively large or small. The invention has the advantages that the method is of important practical application value on the further improvement of the precision of the change detection and has good application prospect in the remote sensing image change information detection field.

Description

A kind of Unsupervised remote sensing image change detection method based on segmentation window

Technical field

The invention belongs to technical field of remote sensing image processing, be specifically related to a kind of Unsupervised remote sensing image change detection method based on segmentation window.

Background technology

Remote Sensing Imagery Change Detection is meant to be analyzed the remote sensing images of areal different times, therefrom detect the information that changes in time between atural object, these information can be used for geosystem information updating, Monitoring of Resource and Environment, target dynamic supervision and military attack recruitment evaluation etc.The Remote Sensing Imagery Change Detection technology all is widely used in fields such as agricultural, environment, city planning and military affairs, therefore, becomes an importance in the sensor information scientific research ^[1]-[3]

Through the development of nearly decades, at various dissimilar remote sensing images, scholars have proposed the method for many change-detection ^[4]-[6]The method of these change-detection can be divided into change-detection and the unsupervised change-detection two big classes that supervision is arranged substantially.The change-detection that supervision is arranged is classified the image of different times earlier based on the image classification method that supervision is arranged, and determines the variation of class then.This method needs acquisition in advance to train about the sample of the true classification in ground.But in most of the cases, can't obtain about the truth on ground, therefore the application of this method is very limited.Unsupervised change-detection then is that the image at different times directly compares, and carries out change-detection, does not need other information, thereby has obtained widespread use.How to adopt the precision and the applicability of unsupervised method raising change-detection, become the focus of research.

The general step of unsupervised change-detection is: to two width of cloth remote sensing images of areal different times, (comprise geometrical registration through after the pre-service, steps such as gray correction), carry out the change vector analysis, produce a width of cloth differential image, determine the information of change information and non-variation by choosing appropriate threshold again, thereby obtain modified-image, obtain net result through aftertreatments such as filtering and noise reductions.The method of change vector analysis mainly is based on the radiation variation of two width of cloth images of different time, determines each pixel change direction and intensity, determines a width of cloth differential image according to the change information of pixel.

The result of differential image selection of threshold will directly influence the precision of change-detection, be the problem of most critical in the change-detection, also be the focal issue of research.At choosing of threshold value, a lot of scholars have proposed various method ^[7]-[11]Fuzzy entropy wherein ^[8]And histogram-fitting ^[11]Method be two kinds of the most frequently used methods.Fuzzy entropy is based on the dividing method of the ambiguity and the maximum entropy of image, and this method need be based on the hypothesis of histogram distribution, and has improved arithmetic speed, has robustness preferably, is a kind of the most frequently used method of selected threshold in the change-detection.2000, people such as Lorenzo Bruzzone ^[10]Proposed the differential image of change-detection is regarded as the image that two class Gauss models of change information and non-change information mix.People such as Li Yaping ^[11]Once having proposed thus to distribute with mixed Gaussian, (Gaussian Mixture Model GMM) simulates the method for difference image histogram, and adopts expectation maximum ^[12](Expectation Maximum, EM) algorithm comes the histogram of match differential image, obtains the parameter that mixed Gaussian distributes, and further determines the threshold value that changes.This method can obtain the quite good detecting result based on the histogram of differential image.

Above threshold value extracting method all is to be based upon region of variation and non-region of variation to distribute under the prerequisite of relatively equilibrium, is putting before this, and these class methods can obtain result preferably.If the area of region of variation is relatively large or hour with respect to the area ratio of view picture remote sensing images, because the spectral information great majority concentrate in wherein some gray scales territory, this moment, the boundary of region of variation and non-region of variation was also not obvious, no matter the method for fuzzy entropy at this moment, or the method for histogram-fitting, can accurately not detect change information, be easy to occur the situation of bigger flase drop or empty inspection.If region of variation is excessive, above-mentioned change detecting method is the part Changing Area Detection non-variation easily then; Otherwise,, then easily the zone of the non-variation of part is detected to changing if region of variation is too small.In actual conditions, occurring many is on the bigger remote sensing images of size, and the zone that atural object changes compares less, the method of general change-detection, can produce very high void inspection rate, the zone of non-variation is detected to changing, can accurately not detect region of variation.

List of references related to the present invention has:

[1]A.Singh，“Digital?change?detection?techniques?using?remotely-sensed?data”，Int.J.RemoteSensing，Vol.10，No.6，pp.989-1003，1989.

[2]L.Bruzzone，S.B.Serpico，“An?iterative?technique?for?the?detection?of?land-covertransitions?in?multitemporal?remote-sensing?images，”IEEE?Trans.Geosci.RemoteSensing，Vol.35，no.4，pp.858-867，July?1997.

[3]L.Bruzzone，S.B.Serpico，“Detection?of?changes?in?remotely?sensed?images?by?theselective?use?of?multi-spectral?information，”International?Journal?of?Remote?Sensing，Vol.18，no.18，pp.3883-3888，Dec.1997.

[4]Gonzalo，“A?Hopfield?Neural?Network?for?Image?Change?Detection”，IEEE?Trans.NeuralNetworks，Vol.17，No.5，pp.1250-1264，September?2006.

[5]Paolo?Gamba，Fabio?Dell’Acqua，，Gianni?Lisini，“Change?Detection?of?Multitemporal?SARData?in?Urban?Areas?Combining?Feature-Based?and?Pixel-Based?Techniques，”IEEE?Trans.Geosci.Remote?Sensing，Vol.44，No.10，pp.2820-2827，October?2006.

[6]Jordi?Inglada，Gregoire?Mercier“A?New?Statistical?Similarity?Measure?for?ChangeDetection?in?Mulitemporal?SAR?Images?and?Its?Extension?to?Multiscale?Change?Analysis，”IEEE?Trans.Geosci.Remote?Sensing，Vol.45，No.5，pp.1432-1445，May?2007.

[7]Fan?Jiulun，Xie?Winxin，“Minimum?error?thresholding：A?note，”Pattern?RecognitionLetters，pp.705-709，June?1997.

[8]Liang-Kai?Huang，Mao-Jiun?J.Wang，“Image?Thresholding?by?Minimizing?the?Measures?ofFuzziness，”Pattern?Recognition，Vol.28，No.1，pp.41-51，1995.

[9]Nobuyuki?Otsu，“A?Threshold?Selection?Method?from?Gray-Level?Histograms，”IEEETrans.Systems，Man，and?Cybernetics，Vol.9，No.1，pp.62-66，Jan.1979.

[10]L.Bruzzone，Diego?Fernadez?Prieto，“Automatic?Analysis?of?the?Difference?Image?forUnsupervised?Change?Detection，”IEEE?Trans.Geosci.Remote?Sensing，vol.38，pp.1171-1181，May?2000.

[11] Li Yaping, Yang Hua, Chen Xia, " Determination of Threshold in Change Detection onHistogram Approximation Using Expectation Maximization Algorithm and BayesInformation Criterion; " Journal of Remote Sensing, Vol.12, No.1, pp.85-90, the Jan.2008[Li Yaping, Yang Hua, Chen Xia, " being applied to remote sensing change-detection threshold value based on the histogram of EM and BIC by approximating method determines ", the remote sensing journal, Vol.12, No.1, pp.85-90, Jan.2008].

[12]Moon，T.K.，“The?Expectation-Maximization?Algorithm，”IEEE?Signal?ProcessingMagazine，Vol.13，pp.47-60，November?1996

[13]Weian?Deng，Iyengar，S.S.，“A?new?probabilistic?relaxation?scheme?and?its?application?toedge?detection，”IEEE?Trans.Pattern?Anal.Mach.Intell.，Vol.18，pp.432-437，April?1996.

Summary of the invention

The objective of the invention is to propose a kind of change detecting method based on segmentation window for overcoming the defective of prior art.Be specifically related to a kind of Unsupervised remote sensing image change detection method based on segmentation window.

This method combines the thought of general change detecting method and segmentation window, and main process is: at first differential image is handled, the information that changes and do not change with big probability is screened earlier, reduce the spectrum complexity of differential image.Then differential image is divided into a series of subimages, determines global threshold by the local threshold of asking subimage.Even change detecting method of the present invention under the smaller situation of region of variation, can obviously improve the precision of change-detection, obtain experimental result preferably.

Particularly, in actual applications, when the shared area of region of variation is less relatively or bigger with respect to the area of view picture figure, directly the view picture differential image is carried out Threshold Segmentation, because region of variation and non-region of variation are in the two, one side occupies an leading position, and the opposing party's feature is then obvious inadequately, thereby can not detect result of variations exactly.The present invention is directed to the problem that prior art exists, on general change detecting method basis, proposed the method by segmentation window, the method with local threshold replacement global threshold detects changing unit accurately.Thought of the present invention is that differential image is divided into little subimage, and with respect to entire image, the complexity of the spectrum of subimage descends.Choose the subimage that to represent general image change information and non-change information difference, this subimage is carried out change-detection, obtain the threshold value of this type of subimage, determine the threshold value of view picture differential image at last according to the threshold value of this type of subimage.

Threshold value extracting method related to the present invention such as following:

1. fuzzy entropy method

In the change-detection threshold selecting algorithm, method based on fuzzy set and information-theoretical fuzzy entropy [8] is one of most widely used method, it is regarded differential image by the part (target) that changes and the part (background) of non-variation as and mixes, and determines the region of variation and the non-region of variation of differential image by minimizing entropy function.We come the differential expression information content of image with fuzzy entropy function H (X), and M * N is the size of differential image, x _Ij(i, the gray-scale value of j) locating, μ are each pixel x to the expression point _IjThe degree of membership that belongs to its region is defined by following formula

$\mathrm{\&mu;}\left(x_{\mathrm{ij}}\right)=\left\{\begin{array}{c}\frac{1}{1+|x_{\mathrm{ij}}-{\mathrm{\&mu;}}_0|/C},x_{\mathrm{ij}}<T\\ \frac{1}{1+|x_{\mathrm{ij}}-{\mathrm{\&mu;}}_1|/C},x_{\mathrm{ij}}\&GreaterEqual;T\end{array}\right.,---\left(1\right)$

In the formula, C is a normalized factor, makes 0.5≤μ (x _Ij)≤1, μ ₀And μ ₁Be the average of object set and background collection, then the quantity of information H of differential image (X) expression formula is

$H\left(X\right)=\frac{1}{\mathrm{MN}\ln 2}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}S\left(\mathrm{\&mu;}\left(x_{\mathrm{ij}}\right)\right)---\left(2\right)$

Wherein, s (μ) is the Shannon function, and its expression formula is

S(μ)＝-μlnμ-(1-μ)ln(1-μ).???????????(3)

When the value of fuzzy entropy hour, the threshold value of this moment is optimal threshold T ^*, promptly

T ^*＝argminH(T)，0≤T≤255.???????????????(4)

2. based on the histogram-fitting method of mixed Gauss model:

In change-detection, can regard the histogram of differential image by the pixel that changes and the pixel two class Gauss models of non-variation as to mix, see accompanying drawing 1.With the histogram information of h (x) expression differential image, P (C _i) be the prior probability of i class, then the probability density function that calculates according to the gray-scale value of differential image is

$P\left(x\right)=\underset{i=1}{\overset{2}{\mathrm{\&Sigma;}}}P\left(C_i\right)P(x/C_i),---\left(5\right)$

Wherein

$P(x/C_i)=\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_i}\exp [-\frac{{(x-{\mathrm{\&mu;}}_i)}^2}{2{\mathrm{\&sigma;}}_i}].---\left(6\right)$

Adopt expectation maximum ^[12]Algorithm mixed Gauss model is separated mixed, shown in (7), (8) and (9), obtain the average and the variance of two Gauss models, thereby determine changing unit and non-changing unit.In fact, under two class Gauss models overlapped not serious situation, the optimal threshold between the two class Gauss models was exactly their intersection point.

$P^t\left(C_i\right)=\frac{1}{\mathrm{xh}\left(x\right)}\underset{L}{\mathrm{\&Sigma;}}\frac{h\left(x\right)P^{t-1}\left(C_i\right)P^{t-1}(x/C_i)}{P^{t-1}\left(x\right)}---\left(7\right)$

${\mathrm{\&mu;}}_i^t=\frac{\underset{L}{\mathrm{\&Sigma;}}h\left(x\right)P^t\left(C_i\right)x}{\underset{L}{\mathrm{\&Sigma;}}h\left(x\right)P^t\left(C_i\right)}---\left(8\right)$

${\mathrm{\&sigma;}}_i^t=\frac{\underset{L}{\mathrm{\&Sigma;}}h\left(x\right)P^t\left(C_i\right){(x-{\mathrm{\&mu;}}_i^t)}^2}{\underset{L}{\mathrm{\&Sigma;}}h\left(x\right)P^t\left(C_i\right)}---\left(9\right)$

The method for detecting change of remote sensing image that the present invention proposes comprises the steps:

At first, differential image is carried out preliminary extracting change information, determine a tonal range with the method for selection of threshold, the point in this tonal range is the point that whether changes to be determined, and the outer point of this tonal range then is to have determined whether the point that changes.Though when region of variation is excessive or too small, general threshold method can not be partitioned into change information and non-change information exactly, but the threshold value T that adopts general threshold method to determine can be used as a rational reference point, the zone of coming determining section to change and do not change.The pixel value x of ordering such as i on the differential image _iIf, x _iMore approaching with T, then can't judge whether to change; If fall far short, then can determine that this point changes or determine that this point does not change with bigger probability with T.If x _iMuch larger than T, can think that then this point does not change; Otherwise, if x _iMuch smaller than T, can think that then this point changes.Promptly near threshold value T, determine a tonal range (T ₁, T ₂), if the pixel value x that i is ordered _iGreater than T ₂, be labeled as and do not change; If x _iLess than T ₁, then be labeled as variation, as shown in Figure 2.By such processing, further reduced the spectrum complexity of differential image, made things convenient for choosing of subimage threshold value.For T ₁And T ₂Choose, can adopt (10) and (11) formula.Determining that grey scale pixel value is at (T ₁, T ₂) outside point after, this part the point of mark from differential image, remove.

T ₁＝T-δ ₁＝T-(T-0)×30％??????????(10)

T ₂＝T+δ ₂＝T+(256-T)×30％????????(11)

Secondly, differential image is divided into a series of subimage, cuts apart, can obtain all types of subimages to greatest extent with the window that slides.Then subimage is resequenced from big to small by variance.Variance has reflected the degree of scatter of pixel value in the image, variance is big more, illustrates that then the overstepping the bounds of propriety of grey value profile of image pixel looses, in differential image, the distribution of also just representing change information and non-change information is balanced more, and the threshold value of this class subimage can be represented the threshold value of differential image integral body more.Variance is little, illustrates that then grey value profile is concentrated, and the change information of this image and non-change information skewness are unfavorable for carrying out change-detection.After pressing the variance size ordering of segmentation window, variance is bigger, is the most balanced subimage of change information and non-change information distribution, also can represent the difference of differential image overall variation information and non-change information.For choosing of window size, answer the size in reference change zone and decide, window choosing too big, the region of variation area is still because too small and can't accurately detect in subimage.Window selects too for a short time, and then the threshold value of subimage can lose representativeness for whole differential image.(behind the p * q), concrete way is: with the moving window of p * q, from being the subimage of p * q through taking out all sizes the pretreated differential image, can obtain the subgraph image set is I to have selected the size of suitable segmentation window _s={ I _S1, I _S2, L I _SL, subimage is sorted from big to small by variance, get { I _SM1, I _SM2, L I _SML.

After subimage sorted by the variance size, the present invention can not select for use the threshold value of the subimage of variance maximum to be used as the threshold value of differential image integral body easily, because the subimage of variance maximum also probably is because the unusual subimage that noise or illumination or the influence of other situations produce selects for use the subimage of variance maximum to have bigger risk.Here, the present invention chooses a few width of cloth subimages that come the front with the method for statistics, then these a few width of cloth subimages is carried out Threshold Segmentation, adopts the selection of threshold method of general change-detection, obtains threshold value separately.Combine by the threshold value of certain method then, determine the threshold value of final view picture differential image top several number of sub images.Promptly from { I _SM1, I _SM2, L I _SMLIn choose the M width of cloth subimage of variance maximum, i.e. distribute the most balanced M width of cloth subimage, i.e. { I of region of variation and non-region of variation _SM1, I _SM2, L I _SMM.The method of utilization Threshold Segmentation is asked threshold value to these subimages, can obtain { T _M1, T _M2, L T _MM.

At last, according to obtaining next local threshold, determine global threshold

Determine global threshold

Method can have two kinds: averaging method and intermediate value method.Averaging method promptly calculates { T _M1, T _M2, L T _MMMean value, this method is the simplest method, though averaging method can provide reliable result from the statistical significance, its shortcoming is in selected a few sample in case exceptional sample occurs, will considerable influence be arranged to final result.And the intermediate value method is to adopt { T _M1, T _M2, L T _MMIntermediate value threshold value as a whole

Adopt the intermediate value method, the probability that influences global threshold because of exceptional sample is reduced greatly, so the preferred intermediate value method of the present invention is chosen final global threshold

The invention provides a kind of Unsupervised remote sensing image change detection method based on segmentation window, it possesses following advantage: when the changing unit of remote sensing images shared area in entire image is excessive or too small, the present invention still can detect the part that really changes very accurately, and general method can't accurately detect this variation usually.The present invention is significant for the unsupervised change-detection of multichannel remote sensing images.

Description of drawings:

Fig. 1 is the histogram-fitting synoptic diagram of differential image.

Fig. 2 for the histogram of differential image with obtain threshold value after determine to change and indeclinable area schematic.

Fig. 3 is the data of experiment 1: image (a) in the May, 1998 of the 4th wave band of somewhere Landsat5 TM, (b) in July, 2000.

Fig. 4 is the differential image and histogram (a) differential image thereof of experiment 1 data, (b) histogram.

Fig. 5 is for directly carrying out result (a) the fuzzy entropy method of Threshold Segmentation, (b) histogram-fitting to the differential image of testing 1 data.

The cut apart subimage of Fig. 6 for the 1 The data fuzzy entropy method of testing is obtained.

The cut apart subimage of Fig. 7 for the 1 The data histogram-fitting method of testing is obtained.

The method that Fig. 8 proposes for the present invention is carried out net result (a) fuzzy entropy-segmentation window method of change-detection in conjunction with Threshold Segmentation, (b) histogram-fitting-segmentation window method, and (c) ground is true.

Fig. 9 is the data of experiment 2: Landsat5 TM the 3rd band image (a) in the May, 1998 in somewhere, (b) in July, 2000.

Figure 10 is for carrying out change-detection result (a) fuzzy entropy method to testing 2 data, (b) fuzzy entropy-segmentation window method, and (c) histogram-fitting method, (d) histogram-fitting-segmentation window, (e) ground is true.

For the ease of understanding, below will the present invention be described in detail by concrete drawings and Examples.It needs to be noted, instantiation and accompanying drawing only are in order to illustrate, obviously those of ordinary skill in the art can illustrate according to this paper, within the scope of the invention the present invention is made various corrections and change, and these corrections and change are also included in the scope of the present invention.Below, be concrete embodiment and the superiority thereof of example explanation with actual remote sensing image data.

Embodiment

Suppose two width of cloth k channel image X that obtain for the areal different times ₁, X ₂, its size is m * n.The change detection algorithm based on segmentation window that this paper proposed can be expressed as follows.

1. couples two good multichannel image X of width of cloth pre-service of Step ₁, X ₂,, have for each pixel i

$X_i^1={\left(\begin{array}{cc}{x}_{i1}^1,x_{i2}^{1}L& x_{\mathrm{ik}}^1\end{array}\right)}^T$ With $X_i^2={\left(\begin{array}{cc}{x}_{i1}^2,x_{i2}^{2}L& x_{\mathrm{ik}}^2\end{array}\right)}^T,$ According to the change vector formula

${\mathrm{\&Delta;}}_i={\left(X_i^1\right)}^{T}g{X}_i^2/\left(\right|\left|X_i^1\right|\left|\right|\left|X_i^2\right|\left|\right),$ Obtain differential image I.

The method of Step 2. usefulness Threshold Segmentation is asked for threshold value to differential image, determines an initial threshold T.

Step 3. determines the scope (T of pixel undetermined according to initial threshold T ₁, T ₂), the value of removing pixel from differential image is at (T ₁, T ₂) outside the point.

Step 4. is cut apart remaining pixel by the p * q size windows of sliding, obtain subgraph image set I _s={ I _S1, I _S2, L I _SL, according to the variance of each number of sub images subimage is sorted from big to small, take out the M number of sub images { I of variance maximum in the window _SM1, I _SM2, L I _SMM.

The method of the M number of sub images utilization carrying out image threshold segmentation that 5. pairs of previous steps of Step are selected is obtained threshold value, as local threshold { T _M1, T _M2, L T _MM.

Step 6. is with the threshold value { T of definite subimage _M1, T _M2, L T _MMSort, choose intermediate value, as the global threshold of differential image

Step 7. uses global threshold

After image cut apart, with probabilistic relaxation image is carried out aftertreatment again, obtain final change-detection result.

Embodiment 1

1. test 1 data

In this experiment, the present invention has chosen two width of cloth Landsat5TM data in areal in May, 1998 and in July, 2000, and size is 200 * 200.Accompanying drawing 3 is the gray-scale map of this two width of cloth image the 4th wave band, and visible main region of variation is the zone of lake water enlarged areas.And the area of this region of variation is less relatively with respect to the view picture remote sensing images, differential image that obtains after its change vector analysis and histogram thereof are respectively as accompanying drawing 4 (a) with (b), because region of variation shared area in entire image is very little, therefore in the histogram of differential image, changing unit does not have tangible peak value, and spectral information concentrates on non-region of variation mostly.

2. the result of conventional method

Carry out direct Threshold Segmentation with the method for fuzzy entropy and the method for histogram-fitting respectively, the threshold value T that is obtained is respectively 249 and 246, its change-detection result after probabilistic relaxation is handled as shown in Figure 5, as seen these two kinds of methods are the actual regional false retrieval that does not change of part region of variation all, and very high void inspection rate is arranged.

3. the result of method of the present invention

Adopt method proposed by the invention to carry out the change-detection of image.At first the threshold value of determining according to general change detecting method is determined (T ₁, T ₂), with pixel value at (T ₁, T ₂) outside the some mark well the back from differential image, remove.We adopt size is 30 * 30 segmentation window, and M is taken as 5.For differential image, the moving window with 30 * 30 takes out subimage one by one, then subimage is arranged from big to small by variance, takes out preceding 5 number of sub images of variance maximum, shown in accompanying drawing 6 and accompanying drawing 7, wherein black the point of zone for having removed.To the subimage that obtains, ask for the threshold value of each subimage respectively with corresponding threshold value extracting method, and finally determine global threshold (seeing Table 1).With the final global threshold of determining Image is cut apart, then through getting result's (seeing accompanying drawing 8) to the end after the probability relaxation processes.

In order to estimate the change-detection performance of method proposed by the invention quantitatively, the result that general change detecting method is obtained, and general change detecting method compares with the change-detection result who obtains after the segmentation window method combines, respectively they and ground are truly compared, from the accuracy that detects, the bit error rate, empty inspection rate and Kappa index several respects are assessed, the data assessment result who obtains is as shown in table 2, as seen with general change detecting method with after segmentation window combines, except the bit error rate is a little high a little, other detection accuracy rate has all had the raising of highly significant.

Threshold value in the table 1 segmentation window algorithmic procedure

The experimental result evaluation of general change detecting method of table 2 and segmentation window method

Method	Accuracy	Empty inspection rate	The bit error rate	The Kappa index
					Fuzzy entropy	????82.46％	????17.45％	????0.09％	????0.2160
Histogram-fitting	????90.08％	????9.73％	????0.19％	????0.3504
					Fuzzy entropy-segmentation window	????98.32％	????0.83％	????0.85％	????0.8698
Histogram-fitting-segmentation window	????98.09％	????1.15％	????0.76％	????0.8295

Embodiment 2.

Test 2 data and experimental result

In this experiment, adopted another width of cloth image, be two width of cloth Landsat5 TM data in somewhere in May, 1998 and in July, 2000, size is 230 * 230, accompanying drawing 9 is the gray-scale map of this two width of cloth image the 3rd wave band, still use 30 * 30 sliding window to get subimage, get the variance maximum preceding 5, result of experiment as shown in Figure 10, provided respectively direct with fuzzy entropy method and the result who obtains with the method for fuzzy entropy-segmentation window, the direct result who obtains with the method for the method of histogram-fitting and histogram-fitting-segmentation window, it is true that accompanying drawing 10 (e) has provided ground.Table 3 provided respectively the method for method with fuzzy entropy, histogram-fitting and they and segmentation window method in conjunction with after the accuracy of detection evaluation truly compared of result and ground, use the segmentation window method as can be seen after, the accuracy of change-detection also is significantly improved.

The evaluation of result of table 3 experiment 2

Method	Accuracy	Empty inspection rate	The bit error rate	The Kappa index
					Fuzzy entropy	??80.93％	??19.05％	??0.02％	??0.2085
Histogram-fitting	??85.90％	??14.06％	??0.04％	??0.2747
					Fuzzy entropy-segmentation window	??99.03％	??0.18％	??0.79％	??0.9677
Histogram-fitting-segmentation window	??98.82％	??0.70％	??0.48％	??0.8960

Claims (3)

1. the Unsupervised remote sensing image change detection method based on segmentation window is characterized in that comprising the steps:

1) differential image is carried out preliminary extracting change information,

To the hyperchannel remote sensing images of different times, obtain differential image by the change vector analysis, determine tonal range (T with the method for selection of threshold ₁, T ₂), the point in this tonal range is the point of variation to be determined, the outer point of this tonal range is the point of fixed variation; The threshold value T that adopts threshold method to determine, as the reference point, the zone that determining section changes and do not change;

(2) adopt the method for segmentation window to determine final threshold value

Above-mentioned differential image is divided into the subimage of series, then subimage is resequenced from big to small by variance, choose the subimage that comes the front, carry out Threshold Segmentation, threshold value with top subimage combines then, determines the threshold value of final view picture differential image;

(3) obtain final change-detection result through aftertreatment

With the global threshold that obtains previously differential image is cut apart, that obtain is exactly the result of change-detection, with probabilistic relaxation image is carried out aftertreatment again, eliminates assorted spot, obtains final change-detection result.

2. by the described Unsupervised remote sensing image change detection method of claim 1, it is characterized in that the tonal range (T of described step 1) based on segmentation window ₁, T ₂) determine near threshold value T described T ₁And T ₂Adopt following formula to choose:

T ₁＝T-δ ₁＝T-(T-0)×30％????????(10)

T ₂＝T+δ ₂＝T+(256-T)×30％??????(11)。

3. by the described Unsupervised remote sensing image change detection method of claim 1, it is characterized in that described step 2 based on segmentation window) the threshold value employing averaging method or the intermediate value method of definite final view picture differential image.

Images