CN102609721B - Remote sensing image clustering method - Google Patents

Abstract

The invention discloses a remote sensing image clustering method and belongs to the technical field of image analysis. The remote sensing image clustering method comprises the following steps: A, determining the number of optimal clustering centers of an original image; B, acquiring the multi-scale expression of the original image through a Gaussian convolution function, and mapping the original image into a scale space thereof to produce a multilayer document set; C, establishing a dirichlet distribution model with invariant overlapping image semanteme according to the multilayer document set, and estimating the mixed proportion parameter of the theme of each document in the multilayer document set and the distribution parameters of the theme which produces visual words according to the probability; and D, obtaining the clustering category of each visual word according to the posteriori probability maximation method. The calculation complicity due to the generation of the document set in advance is avoided, the correlation of the documents can be kept and the detection efficiency of the geographical target of the remote sensing image is improved.

Description

The clustering method of remote sensing image

Technical field

The present invention relates to image analysis technology field, relate in particular to a kind of clustering method of remote sensing image.

Background technology

The Di Li Cray apportion model (Latent Dirichlet Allocation, LDA) of diving is a kind of probability topic model for text modeling being proposed in 2003 by people such as Blei.By means of the expression way of probability graph model, it can carry out modeling to the conditional probability relation between " word ", " document " and " theme ", fully excavates the probability semantic information of document and two aspects of word.

It is generally acknowledged, for the first time propose probability topic model be truly by Hoffmann in 1999 based on latent semantic analysis model (Latent Semantic Analysis, LSA), abandon the singular value decomposition analysis mode of original complexity, latent semantic analysis (the Probabilistic Latent Semantic Analysis of probability building from the angle of generation model, pLSA) model, it is successfully applied to text analyzing.

But, because pLSA model is not set up suitable probabilistic framework in document aspect, and all this one-level variablees are all regarded as to the parameter of model, there are how many documents so just to have the parameter of how many models corresponding with it, therefore, treat that the quantity of estimated parameter is along with increasing of number of documents is linear growth, thereby make model be prone to over-fitting and lack the processing power to new document.

As the significant improvement to pLSA model, LDA model is generation model completely, it is by introducing a super parameter, the blending ratio of theme in document is distributed and is considered as obeying the multinomial distribution of Dirichlet priori, but not in the set of the individual parameter of particular document direct correlation, therefore overcome the over-fitting problem that pLSA model exists.In addition,, for different practical application request, taking pLSA and LDA model as basis, also evolving development has gone out a series of other probability topic models.Although the otherness that may exist some to process, probability topic model generally possesses common basic theory hypothesis, thinks that document is the mixing by some themes, and each theme is a probability distribution about word.Not in the situation by any supervision message, subject information and the semantic information of this class model in can automatic mining data, this has opened up new thinking for the natural language understanding based on Statistical Learning Theory.

Due to the probability topic model statistical dependence relation between analytical documentation, theme and word preferably, this class model all has good application in the field such as computer vision, pattern-recognition, has many successful application cases in natural image identification, retrieval, scene analysis.Meanwhile, the modeling object of probability topic model is mapped to the analytic target of object-oriented high-resolution remote sensing image, i.e. " pixel clusters " of " word " correspondence " pixel ", " document " corresponding AD HOC, " theme " correspondence " atural object classification " center, in so each pixel clusters, the discrimination of pixel class ownership is converted into the theme attaching problem of differentiating visual word in each document very naturally.Therefore, the inherent feature of probability topic model and the application demand of high-resolution remote sensing image information extraction are very identical, use for reference it at natural picture processing and the successful example of application in analyzing, and it is feasible applying it to remote sensing images analysis field.

At present, probability topic model is cut apart or identifies in application at the image of text modeling or computer vision, and document is all in advance given, and in modeling process, is assumed to separate.In remote sensing image information extracts, people need from given remote sensing image, to produce in some way the document for probability topic model modeling, for example dividing body in image or little image block.Once produced these documents, in modeling process they remain be assumed to separate.But (between pixel or atural object) spatial coherence in order to embody in remote sensing image, necessarily requires these documents to have to a certain degree overlapping.In other words, between document, not separate, there is stronger correlativity.But, under current this pattern, same pixel in different document may be different semantic classess by probability topic model identification, conventionally need further aftertreatment to remove this ambiguity, and the quantity of document can form sharp increase along with the increase of overlapping degree.In addition, probability topic model general hypothesis document in the time of text analyzing is the set of unordered word, but the spatial coherence that may exist between visual word has been ignored in this supposition based on " word bag " model (bag of words), therefore also unreasonable when for remote sensing images analysis modeling.Therefore, how to incorporate word order, grammer or syntactic information between visual word, by the spatial context information that contributes to further to excavate in pixel aspect.

The clustering algorithm of remote sensing image, according to analyzing primitive, can be divided into cluster and object-based cluster based on pixel.Because mainly utilizing the spectral information of pixel, the image clustering algorithm based on pixel analyzes, lack the introducing of spatial information, therefore in the cluster result of high-resolution remote sensing image, often there is significantly " spiced salt " phenomenon, thereby affect the effect of cluster result.Being unlike this, OO clustering algorithm is analyzed primitive toward being imaged object, the image patch obtaining as segmentation operators.Generally speaking, the obtaining of imaged object often depends critically upon partitioning algorithm and obtains the quality of cutting apart patch, and Image Segmentation is a more scabrous problem in current image processing field, not yet has at present good general Image Segmentation algorithm.Generally speaking, have at present in a lot of clustering algorithms and can utilize to a certain extent for spatial information, still, for the consideration of the semantic information between pixel, also seldom have at present this type of algorithm application in remote sensing image cluster analysis.

In sum, the clustering algorithm of existing remote sensing image, because needs generate immense document sets in advance, its computation complexity is high, storage overhead is large, and correlativity between document is poor, low to the detection efficiency of remote sensing image geography target.

Summary of the invention

(1) technical matters that will solve

The technical problem to be solved in the present invention is: a kind of clustering method of remote sensing image is provided, its reduced clustering algorithm complexity, reduced storage overhead, can keep the correlativity between document, improved the detection efficiency to remote sensing image geography target.

(2) technical scheme

For addressing the above problem, the invention provides a kind of clustering method of remote sensing image, comprise the following steps:

A: the Optimal cluster center number of determining raw video;

B: obtain the multi-scale expression of raw video by Gaussian convolution function, and raw video is mapped to its metric space generation multilayer document sets;

C: set up the semantic constant latent Di Li Cray apportion model of superimposed images according to described multilayer document sets, in the each document in estimation multilayer document sets, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability;

D: the cluster classification that obtains each visual word by the method for maximization posterior probability.

Wherein, described steps A further comprises: meet Gaussian Mixture according to the feature of minimum description length criterion hypothesis raw video and distribute, the MDL value of utilizing raw video and the correlationship of different cluster centre numbers are obtained hour Optimal cluster center number of MDL value that image is corresponding.

Wherein, in described step B, obtain the multi-scale expression of raw video by Gaussian convolution function, further comprise: by the convolution of the Gaussian function of changeable scale and raw video being obtained to the multi-scale expression of raw video.

Wherein, in described step C, setting up the semantic constant latent Di Li Cray apportion model of superimposed images according to described multilayer document sets further comprises: set up according to described multilayer document sets the Di Li Cray apportion model of diving and generate observation word, be built into the multilayer document sets being formed by observation word, make the same pixel that is under the jurisdiction of different document be assigned with same theme.

Wherein, set up according to described multilayer document sets the Di Li Cray apportion model of diving and generate observation word, specifically comprise: for described multilayer document sets

suppose to exist following generative process:

1) be β according to obeying parameter _sdirichlet Dirichlet distribute p (φ _k| β _s), sample out under each layer of yardstick, the distribution (φ of the visual word that K theme is corresponding _k) _s(N × K × S);

2) yardstick sampling: for t pixel, according to prior distribution p (s _t| γ) its yardstick coordinated indexing of sampling out s _t, show that this pixel should be from s _tlayer metric space distributes a theme;

3) document sampling: for t pixel, according to prior distribution p (d _t| σ, h) sampling obtains its document index d _t;

4) theme sampling: for t pixel, according to multinomial distribution

its subject categories of sampling out, wherein

the document d having sampled _tat yardstick s _tunder proportion coefficient;

5) visual word sampling: the visual word of corresponding t pixel is by theme Z _tdiscrete profile samples obtain.

Wherein, in described step C, in each document in estimation multilayer document sets, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability, further comprise: in the each document in the method estimation multilayer document sets of employing Gibbs sampling approximate resoning, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability.

(3) beneficial effect

The present invention is with Di Li Cray apportion model (the Latent Dirichlet Allocation that dives, LDA) be basis, adopt the document of implicit expression to generate scheme, do not need to generate in advance immense document sets, therefore the present invention can reduce clustering algorithm complexity, reduce storage overhead, realize effective detection of high-resolution remote sensing image geography target.The present invention can keep the correlativity between document, and the neighborhood space relation information between document can be considered; The present invention is under the jurisdiction of different document pixel by adding the semantic constant restrictive condition of superimposed images to ensure only can be assigned with same subject categories all the time, thereby the cluster process of image is fused into unified framework with the reasoning of model.In addition, the introducing of multi-scale expression, has obtained further considering fully to the spatial relationship between neighbor, makes the cluster result of image show very significant Object-oriented Features.

Brief description of the drawings

Fig. 1 is the process flow diagram of the clustering method of remote sensing image described in embodiment of the present invention;

In Fig. 2, a is the raw video of city QUICKBIRD described in embodiment of the present invention, and b utilizes MDL constraint criterion to detect the Optimal cluster center number schematic diagram of city QUICKBIRD image;

In Fig. 3, a is the raw video of rural area EROS-B described in embodiment of the present invention, and b utilizes MDL constraint criterion to detect the Optimal cluster center number schematic diagram of rural area EROS-B image;

In Fig. 4, a is the raw video of suburb QUICKBIRD described in embodiment of the present invention, and b utilizes MDL constraint criterion to detect the Optimal cluster center number schematic diagram of suburb QUICKBIRD image;

Fig. 5 is the probability graph model schematic diagram of msLDA model;

Fig. 6 adopts different clustering methods city QUICKBIRD image to be carried out to the result schematic diagram of quantitative comparison described in embodiment of the present invention; A is the true atural object distribution plan in earth's surface, and b is K-means clustering result, and c is traditional LDA clustering result, and d is msLDA cluster result;

Fig. 7 is the relatively schematic diagram of extensive overall entropy for corresponding each classification in the different clustering method results of city QUICKBIRD image in embodiment of the present invention;

Fig. 8 adopts different clustering methods rural area EROS-B image to be carried out to the result schematic diagram of quantitative comparison described in embodiment of the present invention; A is the true atural object distribution plan in earth's surface, and b is K-means clustering result, and c is traditional LDA clustering result, and d is msLDA cluster result;

Fig. 9 is the relatively schematic diagram of extensive overall entropy for corresponding each classification in the different clustering method results of rural area EROS-B image in embodiment of the present invention;

Figure 10 adopts different clustering methods suburb QUICKBIRD image to be carried out to the result schematic diagram of quantitative comparison described in embodiment of the present invention, a is the true atural object distribution plan in earth's surface, b is K-means clustering result, and c is traditional LDA clustering result, and d is msLDA cluster result;

Figure 11 is the relatively schematic diagram of extensive overall entropy for corresponding each classification in the different clustering method results of suburb QUICKBIRD image in embodiment of the present invention.

Embodiment

Below in conjunction with drawings and Examples, the specific embodiment of the present invention is described in further detail.Following examples are used for illustrating the present invention, but are not used for limiting the scope of the invention.

As shown in Figure 1, the clustering method of remote sensing image of the present invention, comprises the following steps:

A: the Optimal cluster center number of determining raw video;

In this step, meet Gaussian Mixture distribute according to the feature of minimum description length criterion hypothesis raw video, the MDL value of utilizing raw video and the correlationship of different cluster centre numbers are obtained hour Optimal cluster center number of MDL value that image is corresponding.

As: the present invention uses three width raw videos, and respectively as Fig. 2 a, 3a, shown in 4a.Carry out certain analysis and comparison according to MDL constraint criterion, the MDL value of three width images is from the correlationship figure of different cluster centre numbers as Fig. 2 b, and 3b, shown in 4b.

From Fig. 2-4, in the time that cluster centre number is set to 7, the MDL value that three width images are corresponding all obtains minimum value, thereby makes after cluster the complexity of image minimum.Therefore,, for three panel height resolution remote sense images of the present invention, the Optimal cluster center number that MDL criterion is selected is 7.

B: obtain the multi-scale expression of raw video by Gaussian convolution function, and raw video is mapped to its metric space generation multilayer document sets;

In this step, obtain the multi-scale expression of raw video by Gaussian convolution function, further comprise: by the convolution of the Gaussian function of changeable scale and raw video being obtained to the multi-scale expression of raw video.

The Analysis On Multi-scale Features of view data can be simulated by raw video being mapped to its metric space.Gaussian convolution core is the unique linear kernel that realizes change of scale, so the metric space of a width two-dimensional image I (x, y) can be defined as:

L(x，y，δ)＝G(x，y，δ)*I(x，y)

Wherein,

be the Gaussian function of changeable scale, (x, y) is volume coordinate, and δ is yardstick coordinate, and * is convolution algorithm symbol.

Therefore a given specific yardstick coordinate δ, just can generate the convolution image of corresponding yardstick, thereby build the document sets of this layer if consider the expression of S-1 layer yardstick, can obtain S layer document sets

C: set up the semantic constant latent Di Li Cray apportion model of superimposed images according to described multilayer document sets, in the each document in estimation multilayer document sets, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability;

In this step, set up according to described multilayer document sets the Di Li Cray apportion model of diving and generate observation word, be built into the multilayer document sets being formed by observation word, make the same pixel that is under the jurisdiction of different document be assigned with same theme.

Wherein, while setting up latent Di Li Cray apportion model generation observation word according to described multilayer document sets, for described multilayer document sets

suppose to exist following generative process:

1) be β according to obeying parameter _sdirichlet Dirichlet distribute p (φ _k| β _s), sample out under each layer of yardstick, the distribution (φ of the visual word that K theme is corresponding _k) _s(N × K × S);

2) yardstick sampling: for t pixel, according to prior distribution p (s _t| γ) its yardstick coordinated indexing of sampling out s _t, show that this pixel should be from s _tlayer metric space distributes a theme;

3) document sampling: for t pixel, according to prior distribution p (d _t| σ, h) sampling obtains its document index d _t;

4) theme sampling: for t pixel, according to multinomial distribution

its subject categories of sampling out, wherein

the document d having sampled _tat yardstick s _tunder proportion coefficient;

5) visual word sampling: the visual word of corresponding t pixel is by theme Z _tdiscrete profile samples obtain.

In each document in the method estimation multilayer document sets of employing Gibbs sampling approximate resoning, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability.

If yardstick coordinated indexing and the document index of given each pixel, the joint probability of subject categories and visual word can be approached by the Gibbs method of sampling,

$p(w_t=j,z_t=k|d_t=i,s_t=s,\stackrel{\→}{\mathrm{\α}},\stackrel{\→}{\mathrm{\β}})=\frac{n_i^{\left(k\right)}+{\mathrm{\α}}_k}{\underset{k=1}{\overset{K}{\mathrm{\Σ}}}(n_i^{\left(k\right)}+{\mathrm{\α}}_k)}*\frac{n_k^{\left(j\right)}+{\mathrm{\β}}_j}{\underset{j=1}{\overset{V}{\mathrm{\Σ}}}(n_i^{\left(j\right)}+{\mathrm{\β}}_j)}$

Thereby the distribution parameter Φ (N × K × s-matrix) that in each document, the blending ratio parameter Θ (K × Metzler matrix) of theme and each theme produce visual word according to probability can obtain by following formula.

${\mathrm{\θ}}_{\mathrm{ik}}=\frac{n_i^{\left(k\right)}+{\mathrm{\α}}_k}{\underset{k=1}{\overset{K}{\mathrm{\Σ}}}(n_i^{\left(k\right)}+{\mathrm{\α}}_k)},$ ${\mathrm{\φ}}_{\mathrm{kj}}=\frac{n_k^{\left(j\right)}+{\mathrm{\β}}_j}{\underset{j=1}{\overset{V}{\mathrm{\Σ}}}(n_i^{\left(j\right)}+{\mathrm{\β}}_j)}$

The document index of each pixel can obtain by the posteriority profile samples shown in following formula,

$P\left(d_t\right|w_t,z_t,s_t,\stackrel{\→}{\mathrm{\α}},\stackrel{\→}{\mathrm{\β}},\mathrm{\γ},\mathrm{\σ},h)\∝P(d_t=j|\mathrm{\σ},h)*\frac{n_{\⫬i,k}^{\left(j\right)}+{\mathrm{\α}}_k}{{\∑}_{\⫬i,k}^K(n_{\⫬i,k^{\′}}^{\left(j\right)}+{\mathrm{\α}}_{k^{\′}})}$

The yardstick coordinated indexing of each pixel can obtain by the posteriority profile samples shown in following formula,

$P\left(s_t\right|S_{t\⫬},D,Z,W)\∝P\left(s_t\right|\mathrm{\γ})\frac{n_{t\⫬,w_t^s}^{\left(z_t\right)}+{\mathrm{\β}}_{w_t^s}}{{\∑}_{w=1}^N(n_{t\⫬,w}^{\left(z_t\right)}+{\mathrm{\β}}_w)}$

D: the cluster classification that obtains each visual word by the method for maximization posterior probability.

According to above-mentioned msLDA model, can obtain the theme proportion coefficient θ of the document of pixel centered by each pixel _ik=p (z=k|d=i), the each theme under yardstick s and yardstick s that each location of pixels is corresponding produces the distribution (θ of visual word according to probability _kj) _s=p (w=j|z=k, scale=s).

Therefore, for given pixel w=j, suppose that the yardstick that its sampling obtains is scale=s, its corresponding subject categories can obtain by maximizing posterior probability:

${\mathrm{Topic}}_{w_j}=\underset{1\&le;k\&le;K}{\mathrm{<figure-callout\; id="1"\; label="Arg\; max"\; filenames="HDA0000133481660000042.png,HDA0000133481660000052.png"\; state="\{\{state\}\}">Arg</figure-callout>}\max }(p(z=k|d=i)*p(w=j|z=k,\mathrm{scale}=s))$

Cluster result analysis based on msLDA method

Completing on the basis of above-mentioned msLDA methods analyst and calculation process understanding, utilize three panel height resolution remote sense images to carry out cluster analysis, thereby prove to a certain extent the validity of msLDA method.Selected three width experiment images cover respectively city, rural area and suburb, have taken into full account the atural object distribution situation under different scenes, therefore have representativeness and typicalness.In the process of this interpretation, the difference of msLDA clustering result and K-means, traditional LDA clustering result will be compared from quantitative and qualitative analysis two aspects.

1.1 qualitative evaluating method brief introductions

The difference that the qualitative evaluation of Clustering Effect distributes by the true atural object of visual contrast cluster result and earth's surface on the one hand realizes, and analyzes by the Object Oriented Characteristic of assessment cluster result on the other hand.

In order better to embody more objectively the Object Oriented Characteristic of cluster image, the present invention adopts two kinds of landscape indexes to analyze and compares three kinds of image cluster results that clustering method is corresponding.

Generally speaking, landscape index is normally used for the real surface distributed data of the reflection such as quantitative test map or land-use map landscape character, thereby the view that reflection geographical space distributes forms and compositing characteristic.View refers to the synthesis that space on soil and soil and object form, and it is complicated natural process and the reflection of mankind's activity.In image cluster result, view is to be specifically made up of a series of geographical patch, wherein geographical patch be with the true ground object target of geography one to one.Adopt two kinds of landscape indexes to evaluate the Landscape Characteristics of image cluster result, thereby reflect more intuitively the difference of the Object Oriented Characteristic of different cluster results.The specific descriptions of two kinds of landscape indexes choosing are as follows:

(1) patch number (NP): this index is in order to describe the patch number of the different cluster type pixels formations that are separated from each other in cluster image result.In optimal situation, the patch number in cluster result should equate with the real geographic object number in earth's surface, and both can be corresponding one by one.Under normal conditions, the real geographic object number in earth's surface is relatively fixing, if a patch numerical value is larger, illustrates that some geographic object is divided into broken sub-patch, thereby weakens the Object Oriented Characteristic of image.

(2) area fractal dimension (PAFRAC): this index is in order to describe the complex-shaped degree characteristic of patch, and this index prevailing value is greater than 1.Along with the raising of complicacy degree, this index can corresponding increase.When the shape facility of patch is very simple, as be square or when circular, this exponential quantity is 1 by value.

1.2 method for quantitatively evaluating

Entirety entropy (Overall entropy) is analyzed and the overall Clustering Effect that compares three kinds of clustering methods as a kind of quantitative evaluation index, and it is generally made up of cluster centre entropy (Cluster entropy) and classification entropy (Class entropy) two parts.Generally speaking, the less unreasonable property of effect that cluster is described of overall entropy.But overall entropy can not the Clustering Effect of complete reaction clustering algorithm to a certain specific atural object.Therefore, the present invention further introduces extensive overall entropy (Generalized Overall entropy).

In the process that solves corresponding entropy, need to use the real geographic object distributed intelligence in earth's surface Ground truth.H _ckin expression cluster image, in cluster centre k, pixel belongs to the pixel number that in Ground truth, classification is c, and

represent the pixel sum that in cluster result image, in all Ground of belonging to truth, classification is c.In like manner, h _kcthe pixel that in expression Ground truth, classification is c belongs to the pixel number of cluster k in cluster image, and

the pixel that in expression Ground truth, classification is c belongs to the pixel sum of cluster centre k in cluster result image.K is the sum of image cluster centre, and C is the classification sum in Ground truth.Cluster classification in each classification and cluster image in Ground truth has certain incidence relation, specifically: in Ground truth, each classification is corresponding one by one with the cluster classification of proportion maximum in cluster image.In cluster result image, the judgement of each cluster type quality is that the pixel homogenieity degree that each classification pixel is corresponding in Ground truth by judging this cluster type realizes.This homogenieity degree is generally to carry out concentrated expression by cluster centre entropy and classification entropy, and the higher homogenieity degree of less entropy correspondence.

For the classification c in Ground truth image, classification entropy E _ccomputing formula be shown below

$E_c=-\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}\frac{h_{\mathrm{ck}}}{h_c}\log \frac{h_{\mathrm{ck}}}{h_c}$

For the cluster centre k in cluster result image, cluster centre entropy E _kcomputing formula be shown below

$E_k=-\underset{c=1}{\overset{C}{\mathrm{\&Sigma;}}}\frac{h_{\mathrm{ck}}}{h_k}\log \frac{h_{\mathrm{ck}}}{h_k}$

For a certain specific atural object classification c, comprehensive classification entropy E _cand corresponding cluster centre entropy E _kcan build its extensive overall entropy E _generalized, concrete computing formula is shown below:

E _generalized＝βE _c+(1-β)E _k

β ∈ [0,1] in above formula, this variable is that a weight is adjusted parameter, it is 0.5 that variable β is set in experiment.Generally speaking, the cluster result homogenieity of the less corresponding higher degree of overall entropy.

2, cluster result experimental evaluation

Before experiment, first determine several key input parameters of msLDA model.Dirichlet priori is initialized as symmetry, and is specifically set to α=50/K and β=100, in reasoning process according to actual learning to model can reappraise adjustment; The size of implicit expression document is set as 17 × 17 pixels based on experience value; Cluster class number is according to the judgement of MDL criterion, and three width images are 7.

1.1 experiment #1: city QUICKBIRD image

As shown in Figure 2 a, the data that use in this test are to be taken on February 11st, 2002, cover the QUICKBIRD panchromatic image of urban area of Beijing.Wherein image size is 500 × 500 pixels, and resolution is 0.6 meter.In observation image, main atural object comprises buildings, road, shade, water body and bare area.Pay particular attention to, in this image, exist a large amount of shades to distribute, and almost have same gray-scale value with water body.

The cluster result of the true atural object distribution plan in earth's surface and three kinds of clustering methods to be compared is respectively as shown in Fig. 6 (a)-(d).

Qualitative evaluation

Can be clear that from Fig. 6, in K-means clustering result, nearly all shade is all judged to be water body mistakenly, because they are almost in same gray level.On the contrary, traditional LDA method and msLDA method but can be separated water body and shadow region substantially.The gray scale that what the main cause that occurs this result was in K-means clustering process that only actually adopts is based on pixel is cut apart, do not consider the spatial coherence between atural object, but in traditional LDA method and msLDA method, the gray difference information of pixel and the neighborhood document information of pixel are all effectively used, the judgement of the final cluster type of each pixel comprehensively determines by gray scale and its cluster type two category informations in neighborhood document, thereby can realize to a certain extent effective differentiation of water body and shade.

In addition, compare the cluster result of three kinds of methods by visualization, can find intuitively: the cluster result of msLDA method is for the cluster result of other two kinds of clustering methods, compacter, independently pixel clusters is less, therefore the cluster result of msLDA method has certain Object Oriented Characteristic, can be more directly corresponding one by one with the real geographic object in earth's surface.Two kinds of landscape index features of experimental result calculate by FRAGSTATS software, and the landscape index information of three kinds of corresponding cluster results of clustering method is as shown in table 1.

Table 1

	Patch number	Area fractal dimension
			K-means method	13542	1.4673
LDA method	11425	1.4454
			MsLDA method	5112	1.3944

Table 1 shows, landscape index corresponding to cluster result that two kinds of landscape indexes of msLDA clustering result are all less than other two kinds of clustering methods.Can judge thus, patch complexity corresponding to msLDA clustering result is relatively low, patch number is less, the space distribution of more approaching and the true geographic object in earth's surface, and the image Object Oriented Characteristic degree that therefore cluster result of the method possesses is higher than other two classes clustering methods.

Quantitative evaluation

The overall entropy that calculates the cluster result of three kinds of clustering methods, as shown in table 2, the overall entropy of msLDA method is significantly less than other two kinds of methods.

Table 2

Clustering algorithm	Entirety cluster entropy	Entirety classification entropy	Entirety entropy
				K-means method	0.84326	1.3399	1.0916
LDA method	0.77865	1.3514	1.065
				MsLDA method	0.75479	1.3001	1.0274

Further calculate extensive overall entropy, in different clustering method results, the extensive overall entropy of corresponding each classification as shown in Figure 7.In msLDA clustering result, shade, buildings, the corresponding extensive overall entropy of road and bare area is all less than its value in K-means and traditional LDA clustering result.In other words, msLDA method can more accurately be determined these four kinds of atural objects.In addition, in the geographical entity extraction and analysis process of answering at water body, the result precision that the result precision that msLDA method is extracted is extracted lower than K-means method with atomic weak inferior position, but will be higher than the precision of traditional LDA method.In general, msLDA method all can obtain higher precision in the time obtaining all kinds of geographical entity information.Especially, shade and water body can well distinguish by this method.

1.2 experiment #2: rural area EROS-B image

As shown in Figure 3 a, the data that use in this test are to be taken on June 18th, 2010, cover the EROS-B panchromatic image of Anhui Province's Fuyang City.Image size is 800 × 800 pixels, and resolution is 0.6 meter.Test block is positioned at typical rural areas.In observation image, be distributed with large stretch of farming land and forest land.In addition, in image, also comprise the atural objects such as water body, shade and road.

For this experimental analysis, the cluster result of the true atural object distribution plan in earth's surface and three kinds of clustering methods to be compared is respectively as shown in Fig. 8 (a)-(d).

Qualitative evaluation

As shown in Figure 8, msLDA clustering result has compacter atural object distribution results for other two kinds of methods.Especially, can see the water body region that msLDA method institute cluster goes out, very level and smooth, even if even water body region has presented different gray tones at diverse location, the level and smooth mechanism water body cluster that msLDA method also can incorporate by itself evenly level and smooth, this is that K-means and traditional LDA method are not available.The outstanding Object Oriented Characteristic that msLDA clustering result shows has more also obtained further confirmation by the listed landscape index of table 3.

Table 3

	Patch number	Area fractal dimension
			K-means method	22443	1.4951
LDA method	22107	1.4755
			MsLDA method	4133	1.4263

Quantitative evaluation

Calculate respectively overall entropy and the extensive overall entropy of three kinds of clustering results.The contrast of entirety entropy is as shown in table 4; Extensive overall entropy contrasts as shown in Figure 9.

Table 4

Clustering algorithm	Entirety cluster entropy	Entirety classification entropy	Entirety entropy
				K-means method	0.77045	1.3309	1.0507
LDA method	0.72353	1.341	1.0322
				MsLDA method	0.59921	1.0216	0.8104

Significantly can see, every entropy index of msLDA clustering result is all lower than other two kinds of methods.Therefore, this experiment shows the rural areas high-resolution remote sensing image that distributes more for farming land and forest land, and msLDA method can obtain better Clustering Effect equally.

1.3 experiment #3: suburb QUICKBIRD image

As shown in Fig. 4 a, a width resolution is that QUICKBIRD (900 × 900 pixel) panchromatic image of 0.6 meter is chosen for this experimental data.This image capturing was on April 22nd, 2006.Study area is expressly chosen for the suburb between city and rural area, and the ground species therefore distributing is more complicated, to assess three kinds of methods for the image clustering performance under complex scene.In observation image, major surface features comprises trees, buildings, road, water body, shade and farmland.

For this experimental analysis, the cluster result of the true atural object distribution plan in earth's surface and three kinds of clustering methods to be compared is respectively as shown in Figure 10 (a)-(d).

Qualitative evaluation

Although the distribution scene of atural object is more complicated in the image that experiment #3 chooses, from Figure 10 and table 5, can see that msLDA method still shows better clustering performance, especially aspect differentiation water body and shade.

Table 5

	Patch number	Area fractal dimension
			K-means method	22443	1.4951
LDA method	22107	1.4755
			MsLDA method	7457	1.4229

Quantitative evaluation

The overall entropy of three kinds of methods of contrast as shown in table 6, msLDA clustering result is all lower than other other two kinds of methods.

Table 6

Clustering algorithm	Entirety cluster entropy	Entirety classification entropy	Entirety entropy
				K-means method	1.0923	1.1172	1.1048
LDA method	1.0478	1.0326	1.0402
				MsLDA method	1.0344	0.97891	1.0066

In addition, the extensive overall entropy calculating for each atural object in distinct methods as shown in figure 11.In msLDA clustering result, overall entropy corresponding to this three classification of buildings, farmland and trees is all less than its value in K-means and traditional LDA clustering result.MsLDA method extremely approaches the cluster result of precision the best to the cluster result of water body in addition, therefore can think a good cluster result.For the extensive overall entropy of shade and road, msLDA model is not obtained minimum value, but traditional LDA method has but been accomplished.The cluster result of more traditional LDA method and msLDA method, consider their model mechanism, we find that in fact msLDA result can be seen as the smoothness constraint that has carried out Space Consistency by the multi-scale expression strategy of introducing image on the cluster result basis of traditional LDA method simultaneously.Therefore, a more level and smooth cluster result often needs a compromise with the maintenance at edge.In fact, we can obtain better edge maintenance effect by the smoothing parameter reducing in msLDA model.In general, msLDA method all may obtain higher precision in the time obtaining all kinds of geographical entity information.

Above embodiment is only for illustrating the present invention; and be not limitation of the present invention; the those of ordinary skill in relevant technologies field; without departing from the spirit and scope of the present invention; can also make a variety of changes and modification; therefore all technical schemes that are equal to also belong to category of the present invention, and scope of patent protection of the present invention should be defined by the claims.

Claims (6)

1. a clustering method for remote sensing image, is characterized in that, comprises the following steps:

A: the Optimal cluster center number of determining raw video;

B: obtain the multi-scale expression of raw video by Gaussian convolution function, and raw video is mapped to its metric space generation multilayer document sets;

C: set up the semantic constant latent Di Li Cray apportion model of superimposed images according to described multilayer document sets, in the each document in estimation multilayer document sets, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability;

D: the cluster classification that obtains each visual word by the method for maximization posterior probability.

2. the clustering method of remote sensing image as claimed in claim 1, it is characterized in that, described steps A further comprises: meet Gaussian Mixture according to the feature of minimum description length criterion hypothesis raw video and distribute, the MDL value of utilizing raw video and the correlationship of different cluster centre numbers are obtained hour Optimal cluster center number of MDL value that image is corresponding.

3. the clustering method of remote sensing image as claimed in claim 1, it is characterized in that, in described step B, the multi-scale expression that obtains raw video by Gaussian convolution function, further comprises: by the convolution of the Gaussian function of changeable scale and raw video being obtained to the multi-scale expression of raw video.

4. the clustering method of remote sensing image as claimed in claim 1, it is characterized in that, in described step C, setting up the semantic constant latent Di Li Cray apportion model of superimposed images according to described multilayer document sets further comprises: set up according to described multilayer document sets the Di Li Cray apportion model of diving and generate the visual word can observe, be built into the multilayer document sets being formed by the visual word can observe, make the same pixel that is under the jurisdiction of different document be assigned with same theme.

5. the clustering method of remote sensing image as claimed in claim 4, is characterized in that, sets up the Di Li Cray apportion model of diving generate the visual word can observe according to described multilayer document sets, specifically comprises: for described multilayer document sets

suppose to exist following generative process:

1) be β according to obeying parameter _sdirichlet Dirichlet distribute p (φ _k| β _s), sample out under each layer of yardstick, the distribution (φ of the visual word that K theme is corresponding _k) _s(N × K × S);

2) yardstick sampling: for t pixel, according to prior distribution p (s _t| γ) its yardstick coordinated indexing of sampling out s _t, show that this pixel should be from s _tlayer metric space distributes a theme;

3) document sampling: for t pixel, according to prior distribution p (d _t| σ, h) sampling obtains its document index d _t;

4) theme sampling: for t pixel, according to multinomial distribution its subject categories of sampling out, wherein

the document d having sampled _tat yardstick s _tunder proportion coefficient;

5) visual word sampling: the visual word of corresponding t pixel is by theme Z _tdiscrete profile samples obtain.

6. the clustering method of remote sensing image as claimed in claim 1, it is characterized in that, in described step C, in each document in estimation multilayer document sets, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability, further comprise: in the each document in the method estimation multilayer document sets of employing Gibbs sampling approximate resoning, the blending ratio parameter of theme and each theme produce the distribution parameter of visual word according to probability.

Images