CN104794496A - Remote sensing character optimization algorithm for improving mRMR (min-redundancy max-relevance) algorithm - Google Patents

Abstract

The invention provides a remote sensing character optimization algorithm for improving an mRMR (min-redundancy max-relevance) algorithm. In an object-oriented high-resolution remote sensing ground surface classifying process, a certain algorithm needs to be used for dividing a high-resolution remote sensing image so as to obtain a series of primitive objects, various features of the primitive objects are calculated, properties of the primitive objects are determined according to the features of the different primitive objects, and a ground block object classifying process is finished. Because ground block features are numerous, the feature which is the most effective on determination of ground blocks needs to be selected from the various features, and the process is a feature optimization process. On the basis of a correlation theory of mutual information in an information theory, three methods comprising binary discretization, histograms and F statistics are used for implementing a calculation process of the mRMR algorithm, and the feature optimization process is implemented by the mRMR algorithm. The feature optimization process can be implemented efficiently. Moreover, the remote sensing character optimization algorithm can be widely used for applications such as similar feature optimization and feature validity check and evaluation of certain ground object types.

Description

A kind of remote sensing features optimization algorithm improving mRMR algorithm

Technical field

The present invention relates in process in remote sensing digital image processing field the feature optimization algorithm be applied between feature calculation and image classification, specifically, in remote sensing image processing and information extraction process, need to carry out on the basis of characteristics of objects calculating after Remote Sensing Image Segmentation, carry out the preferred work of a part of feature selectively, to realize the assorting process of image under remaining on the constant prerequisite of precision more expeditiously.The present invention is applicable to data volume more greatly or in more complicated image classification process, particularly for needing the preferred process after calculating a large amount of feature calculation.

Background technology

Remote Image Classification is one of gordian technique in process in remote sensing digital image processing, has been widely used in the numerous areas such as agricultural, military affairs at present, has all played important effect in all trades and professions.Carrying out in the object oriented classification process of earth's surface information based on high-resolution remote sensing image, need first to adopt certain image segmentation algorithm Image Segmentation to be become the primitive object of a series of inner homogeneous, again box counting algorithm is carried out to corresponding object, and as the attached attribute of this object, judge finally by the type of corresponding characteristic attribute to this object.Owing to relating to a large amount of features in the process, and the importance of different characteristic is different, and may redundant information be there is between different characteristic, the different characteristic contribution rate that class differentiates over the ground also to some extent may be different, therefore need to select wherein the most key feature representatively from numerous feature sets, while guarantee does not reduce classification relevant information, reduce feature quantity, thus reach the preferred object of feature.Feature preferred process is a signature search process in essence, is exactly by certain evaluation criterion, concentrates select an optimal subset and analyze as a whole and the representative studied from primitive character.Comprising method/algorithm have principal component analysis (PCA) (PCA), independent component analysis (ICA), manifold learning etc.Relevant list of references comprises: Sotoca J M, Pla F, S á nchez J S.Band selection in multispectral images byminimization of dependent information [C] .Systems, Man, and Cybernetics, Part C:Applications andReviews, IEEE Transactions on, 2007,37 (2): 258-267; Wang Lu, Gong Guanghong. based on the multiple features remote sensing image classification of ReliefF+mRMR Feature Dimension Reduction algorithm. Chinese stereology and graphical analysis, 2014,19 (3): 250-257; Zou Lidong, Pan Yaozhong, red legend spring, changes etc. the object-oriented in conjunction with neighborhood coherent video and maximum correlation minimum redundancy feature selecting and detects. Journal of Image and Graphics, 2014,19 (001): 158-166; Deng.

Minimal redundancy maximal correlation (mRMR) algorithm is a kind of characteristic optimization method based on Mutual Information Theory, its principle makes the correlativity between selected character subset and classification maximum, ensure that the redundancy between selected feature is as far as possible little, this is also the preferred fundamental purpose of feature simultaneously.MRMR algorithm is low to data demand, and has higher counting yield, is mainly used in medicine and life science at present.In remote sensing image classification research field, current domestic existing a small amount of scholar starts by mRMR algorithm application in remote sensing image classification experiment, but practical research and apply is also less.Relevant list of references comprises: Yao Xu, Wang Xiaodan, Zhang Yuxi, etc. feature selection approach is summarized. controls and decision-making, 2012,27 (2): 161-166; Ding C, Peng H.Minimumredundancy feature selection from microarray gene expression data.Journal of bioinformatics andcomputational biology, 2005,3 (02): 185-205; Wu Bo, Zhu Qindong, high petrel, etc. based on the feature selecting maximizing mutual information in object-oriented classification. land resources remote sensing, 2009.21 (3): 30-34; Deng.

In the retrieval preferably carrying out present patent for feature, the patent can found at present is also few, 2 are only had with this area is related, as: the patent " a kind of remote sensing image information extracts and decomposition method and module thereof " (application number: CN201310392133, publication number: CN103500344A) of China Surveying and Mapping Research Academy; The patent " a kind of characteristic optimization method based on coevolution for pedestrian detection " (application number: CN200810101705, publication number: CN101246555A) of China Science & Technology University.But from content angle, these patents are not all specifically related to the method for optimizing aspect of multiple features, and are only carried out simple application, and relevant discussion is also less.

Summary of the invention

The object of this invention is to provide a kind of remote sensing features optimization algorithm improving mRMR algorithm.In high-resolution remote sensing image process and assorting process, need to adopt certain algorithm (as mean shift algorithm) carried out splitting by high-resolution remote sensing image and form a series of primitive object, these primitive object are as the elementary cell of follow-up judgement, its state and attribute can not divide again, and the various features gone out by this calculation and object is used for as its a series of subsidiary attribute the institute's possession class judging this object.By calculating the various features of primitive object, then differentiating according to institute's possession class of feature to this object of different object, completing the assorting process of ground block object.Due to ground block feature numerous (can reach tens kinds even hundreds of), the participation of crossing multiple features calculates the efficiency that not only can affect algorithm largely and differentiate, and also can produce certain impact to the discrimination precision of primitive object, therefore need to choose from various features and the most effective feature is differentiated to plot, namely feature is preferred, and the present invention then mainly solves for this problem.The correlation theory of mutual information in combining information opinion of the present invention, research mRMR algorithm principle and in feature is preferred feasibility basis on, propose and adopt two-value discretize, histogram and F to add up three kinds of methods to realize the computation process of mRMR algorithm, and adopt mRMR method to realize the preferred process of multiple features.

Thinking of the present invention is: the classification for high-resolution remote sensing image is right, need first to adopt certain image division method to split remote sensing image, method can select watershed segmentation, multi-resolution segmentation, mean shift segmentation etc., forms a series of plot vector units (object) corresponding with image.On this basis, the feature of each cutting object is calculated and forms the attached series of features attribute of this object, because different features is to judging that different ground classes has different contribution rates, namely different features is applicable to judging different atural object, therefore the feature as far as possible more calculating this object is needed, when all calculating the individual features of this imaged objects all, then need on the one hand to consider that object number is more, object type after also will considering feature calculation on the other hand judges still to need to spend certain hour, therefore carry out preferably just seeming necessary to these features.For the optimal selection problem of multiple features in the present invention, develop mRMR feature optimization algorithm, and to its three kinds of implementations, the implementation procedure of two-value discretize, histogram, F statistics etc. is described, and by case verification, this method for optimizing can be raised the efficiency under the prerequisite keeping subsequent classification precision, reaches preferred object.

Technical scheme of the present invention provides a kind of remote sensing features optimization algorithm improving mRMR algorithm, it is characterized in that comprising following implementation step:

1), carry out object-oriented segmentation to high-resolution remote sensing image, Image Segmentation is become a series of primitive object, dividing method can select the dividing methods such as average drifting, watershed divide, multiresolution, SLIC;

2), in step 1) in Remote Sensing Image Segmentation is become a series of primitive object after, the various features of each object is calculated;

3), adopt method provided by the invention preferred to multiple feature of carrying out;

4), verify the preferred result of feature, realization can adopt the classification results after image classification differentiate, sorting technique can select SVM, decision tree, KNN etc.;

5), compare whether applying the preferred classification results of feature and sum up mRMR algorithm effect.

The feature of above-mentioned implementation step is:

Step 1) in need the multi-scale division that first needs to carry out remote sensing image, Remote Sensing Image Segmentation to be sorted is become a series of primitive object, both separate between each primitive object, also there is certain space correlation relation simultaneously, need the feature gone out by subsequent calculations to carry out comprehensive descision.Here dividing method can select multiple multi-scale division algorithm realization, as multi-resolution segmentation, watershed segmentation, average drifting, SLIC algorithm etc.

Step 2) be on the basis of step 1, travel through all primitive object be partitioned into, then carry out the feature calculation of object with each object for base unit, comprise its spectral signature, textural characteristics, shape facility etc.

Step 3) be series of features for having chosen, adopt the preferred process of mRMR algorithm realization feature herein to realize preferably, picking out character subset.

Step 4) be that the character subset optimized is verified, method selects certain sorter, as SVM, C5.0 decision tree, KNN etc., 2 experiments are carried out to corresponding remote sensing image primitive object, one is the classification based on full feature, and another is that the character subset optimized based on step 3 is classified.

Step 5) be to step 4) in 2 experiments compare, Analysis &Validation process: if these 2 experiments can obtain basically identical classification results, and adopt the preferred character subset of mRMR can increase in efficiency, prove that this characteristic optimization method is effective, otherwise invalid.

The present invention compared with prior art has following features: the characteristic optimization method achieving mRMR in the present invention, the ultimate principle of the method and preferred standard make the correlativity between selected character subset and classification maximum, ensures that the redundancy between selected feature is as far as possible little simultaneously.MRMR algorithm principle is simple, low to data demand, and has higher counting yield, is mainly used in medicine and life science at present.In remote sensing image classification research field, current domestic existing scholar starts by mRMR algorithm application in remote sensing image classification process, but research is also relatively immature.MRMR algorithm is utilized to carry out feature herein preferred, two-value discretize, histogram method, F is adopted to add up the computation process of three kinds of distinct methods implementation algorithms for algorithm calculation features, and utilize k nearest neighbor sorter to carry out Images Classification based on preferred result, prove by analysis mRMR algorithm can effectively be applied to feature in remote sensing image classification field preferably in, and in efficiency, algorithm to be greatly increased.

Accompanying drawing explanation

Fig. 1 is the remote sensing features preferred flow schematic diagram based on the mRMR algorithm improved

Fig. 2 is the feature list schematic diagram that primitive object may calculate

Fig. 3 is the asymptotic expression signature search procedure chart in mRMR implementation procedure

Fig. 4 is the Shang Yao town remote sensing image that the present invention tests employing

Fig. 5 is 27 eigenwerts calculated for Fig. 4 segmentation result

Fig. 6 is without feature preferred KNN classifying quality figure

The mRMR that Fig. 7 is through two-value discretize preferably after KNN classifying quality figure

The mRMR that Fig. 8 is through histogram probabilistic method preferably after KNN classifying quality figure

The mRMR that Fig. 9 is through F statistic law preferably after KNN classifying quality figure

The character subset that Figure 10 adopts 3 kinds of computing method to realize mRMR to optimize respectively

Figure 11 is efficiency comparative for several mode and analytical table

Embodiment

Fig. 1 illustrates main realization approach of the present invention.In high-resolution remote sensing image process and assorting process, need the segmentation (method can select average drifting, watershed divide, multiresolution, SLIC etc.) adopting remote sensing multi-scale division algorithm realization image, further various features calculating is carried out to the primitive object after segmentation, differentiate according to institute's possession class of feature to this object of different object again, complete the assorting process of ground block object.The feature that can calculate due to plot is a lot, common feature calculations list as shown in Figure 2, and generally when primitive feature calculates, the application of Object Spectra feature, textural characteristics and shape facility is more, secondly also has the spatial relationship feature etc. between multi-object.Different features is for judging that different atural object can play different effects, and therefore to get up to carry out the judgement of atural object classification more effective for multiple features combining, and this is also common realization means.But cross multiple features participation calculating and also might not produce better precision or effect, the efficiency that algorithm differentiates can be affected on the contrary largely, therefore need to choose from various features plot differentiation the most effective feature, i.e. feature preferred process.The correlation theory of mutual information in combining information opinion of the present invention, research mRMR algorithm principle and in feature is preferred feasibility basis on, propose and adopt two-value discretize, histogram and F to add up three kinds of methods to realize the computation process of mRMR algorithm, and adopt mRMR method to realize the preferred process of multiple features.

Concrete implementation procedure is:

1.mRMR algorithm ultimate principle

MRMR (minimal redundancy maximal correlation) algorithm is proposed in 2005 by Peng, and its basic thought is the correlation theory utilized in information theory, using the size of mutual information as the standard weighing feature and feature, correlativity between feature and classification.Based on Mutual Information Theory, the preferred final purpose of feature is in whole feature, select a character subset S containing m feature, makes the mutual information between feature set S and class label C maximum, that is:

maxI(S；C)＝maxI({x ₁，x ₂，…，x _m}；C) (1)

(1) formula gives the evaluation criterion of sub-set selection, but the mutual information directly calculated between character subset S and class label C is too complicated, Peng gives new evaluation criterion, namely dependence is represented by the mutual information average of each feature in character subset S and class label C, use the redundancy of selected feature mutual information representation feature subset S between any two, final preferred object maximizes correlativity minimizing redundant simultaneously simultaneously.

The dependence of character subset S and class label C is designated as D, and formula is:

$D=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}I(x_i;C)---\left(2\right)$

The redundancy of character subset S is designated as R, and formula is:

$R=\frac{1}{m^2}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}I(x_i,x_j)---\left(3\right)$

Final choice criteria need consider dependence and redundancy simultaneously, and both considerations have equal weight, and in conjunction with (2) (3) formula, the preferred standard of feature is as follows:

maxΦ(D，R)，Φ＝D-R (4)

Formula (4) describes cardinal principle and the implementation procedure of mRMR algorithm, i.e. minimal redundancy maximal correlation.

Wherein mutual information I (X; Y) (Mutual Information)

$I(X;Y)=\∫\∫p(x,y)\log \frac{p(x,y)}{p\left(x\right)p\left(y\right)}\mathrm{dxdy}---\left(5\right)$

Represent the correlativity between two stochastic variables, i.e. the mutual information of two stochastic variable X, Y.

The asymptotic expression search procedure of 2.mRMR

According to the subset evaluation criterion of (4) formula, feature selection approach adopts progressive search algorithm.Suppose that feature complete or collected works are X, carried out now m-1 time and selected, selected the character subset S with m-1 feature _m-1, will carry out now the m time and select, choice criteria is:

$\underset{x_j\∈X-S_{m-1}}{\max }[I(x_j;C)-\frac{1}{m-1}\underset{x_i\∈S_{m-1}}{\mathrm{\Σ}}I(x_j;x_i)---\left(6\right)$

Namely when selecting for the m time, at feature set X-S to be selected _m-1in each feature to be selected is searched for, calculate according to judgement schematics, select relatively large with the dependence of class label C and with select feature set S _m-1the relatively little feature x of redundancy _j, be the feature selected for the m time.When carrying out certain and once selecting, if when the value of (6) formula equals zero or is less than the threshold value r of a certain setting, then select to stop, having selected feature set to be feature preferred result.

As shown in Figure 3, this figure illustrates one and realizes manifold preferred process gradually based on iterative manner for progressive search algorithm and feature preferred process.

3. the preferred computing method of feature

According to the feature of remote sensing features data and the calculation features of mRMR algorithm, following three kinds of experiments are carried out to the preferred account form of feature, namely the methods such as two-value discretize, histogram, F statistics are adopted to calculate respectively, thus the preferred process of realization character, launch one by one below to discuss.

3.1 two-value discretizes

MRMR algorithm, in the process calculating mutual information, needs a large amount of estimated probability density and multivariate probability density, for the problems referred to above, adopts Method of Data Discretization, processes, make probability estimate become easy to data.The discretization method adopted herein is two-value discretization method (Binary Discretization), is classified as two class values by each sample characteristics by formula, specifically describes as follows:

If feature adds up to M, total sample number is N, then a jth eigenwert x of i-th sample _ijcan be expressed as through two-value discretize:

$x_{\mathrm{ij}}^{,}=\left\{\begin{array}{c}1,(x_{\mathrm{ij}}\&GreaterEqual;{\stackrel{\&OverBar;}{x}}_j)\\ -1,(x_{\mathrm{ij}}<{\stackrel{\&OverBar;}{x}}_j)\end{array}\right.---\left(7\right)$

X ' in above formula _ijrepresent a jth eigenwert of i-th sample after two-value discretize, and represent the average of a jth eigenwert of all samples, that is:

After two-value discretize, the eigenwert of all samples all becomes 1 or-1 two kind of value, then, when estimated probability density, the ratio of this value number of samples and population sample number can be adopted to estimate, also can take same computing method for joint probability density.

3.2 histogram

Histogram method (Histogram Method) is a kind of method of estimation of probability distribution, for the value of each stochastic variable, several spaced points are got according to a determining deviation, its codomain is divided into some parts, by adding up the number of samples within every two spaced points, thus drafting histogram frequency distribution diagram, reach the object that probability distribution is approximate.

In conjunction with research herein, the specific descriptions of method are as follows:

If feature adds up to M, total sample number is N, for a jth feature x _j, determine that its histogram interval is h _j, between marker space, number is K _j, then its approximate probability density function is:

$\hat{f}\left(x\right)=\frac{1}{N{h}_j}{v}_{\mathrm{kj}}(t_{\mathrm{kj}}<x<t_{k+1j})---\left(8\right)$

V in above formula _kjrepresent for a jth feature x _j, according to its histogram interval h _jdivide, fall into the number of samples in a kth interval, kth interval left and right two spaced points uses t respectively _k, t _k+1represent.

Wherein about the determination of interval h, provide experimental formula as follows:

h≈3.73σn ^-1/3(9)

In formula, σ is the standard deviation of this feature samples value, and n represents number of samples.

3.3F statistics

Because in mRMR algorithm, the calculating of mutual information is comparatively complicated, consider to adopt calculating F statistic and Pearson correlation coefficient to represent degree of correlation for this reason, replace mutual information as the standard weighing correlativity, thus reach the object reducing computation complexity.

Adopt F statistic and Pearson correlation coefficient to replace mutual information, then above the progressive search choice criteria of (6) formula also changes to some extent, new choice criteria as shown in the formula:

$\underset{x_j\&Element;X-S_{m-1}}{\max }[F(x_j,C)-\frac{1}{m-1}\underset{x_i\&Element;S_{m-1}}{\mathrm{\&Sigma;}}|r(x_j,x_i)\left|\right]---\left(10\right)$

F (x in formula _j, C) and representation feature x _jwith the F statistic of class label C, r (x _j, x _i) representation feature x _jwith feature x _ibetween Pearson correlation coefficient.

The computing formula of Pearson correlation coefficient:

$r=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(X_i-\stackrel{\&OverBar;}{X})(Y_i-\stackrel{\&OverBar;}{Y})}{\sqrt{{\mathrm{\&Sigma;}}_{i=1}^n{(X_i-\stackrel{\&OverBar;}{X})}^2}\sqrt{{\mathrm{\&Sigma;}}_{i=1}^n{(Y_i-\stackrel{\&OverBar;}{Y})}^2}}---\left(11\right)$

In formula, X and Y is two stochastic variables, is two features, with respectively representation feature X and characteristic Y are based on the average of whole sample, and n represents number of samples.

4. experiment and interpretation of result

Fig. 4 illustrates the present invention's experiment image data used, utilizes mean shift algorithm to split it, obtains cutting object, and position attribution and the artificial selection of foundation atural object obtain training sample, comprise the types of ground objects such as settlement place, vegetation, water body, bare area.Feature calculation is carried out to all cutting objects, extracts feature 27 altogether, comprise spectral signature, geometric properties, spatial relationship etc. as shown in Figure 5.

First analyze nicety of grading, Fig. 6 ~ Fig. 9 illustrates and adopts distinct methods to carry out classifying quality figure.As can be seen from the figure adopt the classification results impact after mRMR different calculation methods little, Figure 10 is the different characteristic subset that different calculation methods is selected, and in order to judge the generic of different object.By the contrast to atural object classification type situation, four kinds of methods based on KNN sorter can be described, little to the classification difference of vegetation and water body type of ground objects, prior difference is embodied in the identification situation for settlement place and bare area type of ground objects, this wherein especially with KNN+F statistical method embody the most obvious, but several method Kappa coefficient is slightly different, the Kappa coefficient that KNN+ two-value discretize and KNN+F add up two kinds of methods will exceed KNN method and KNN+ histogram method.This both may be wherein that the principle of classifier methods itself causes, also may be relevant with choosing of training sample.

Figure 11 then illustrates classification effectiveness contrast.The shortcoming of KNN classifier methods is that algorithm space complexity and time complexity are all higher, and this point embodies obviously in experimentation.Contrast diverse ways is analyzed, if number of features is fewer contained by feature preferred subset, then method is shorter for working time, KNN+F statistical method only considers 7 features preferably, than considering the KNN method of whole 27 features fast nearly a times (49%) in time efficiency.

Judge in conjunction with working time and nicety of grading two aspect, three kinds of KNN combined methods, all than only good by KNN method, improve counting yield on the one hand, and nicety of grading also slightly improves on the other hand.This wherein effect be KNN+F statistical method the most significantly, have efficiency the highest in several method and the highest model accuracy.

Example of the present invention realizes on a pc platform and is verified, and the experiment proved that, the present invention can realize manifold preferred process, for the classification of follow-up high-level efficiency provides the character subset that can represent complete or collected works.In the present invention, mentioned method can be widely used in other needs to carry out in the preferred related application of feature.

Claims (3)

1. improve a remote sensing features optimization algorithm for mRMR algorithm, it is characterized in that comprising following step:

Step 1, carries out object-oriented segmentation to high-resolution remote sensing image, and Image Segmentation is become a series of primitive object, and method can select average drifting, watershed divide, multiresolution, SLIC dividing method;

Step 2, after in step 1 Remote Sensing Image Segmentation being become a series of primitive object, calculates the various features of each object;

Step 3, adopts mRMR method provided by the invention to carry out feature preferred;

Step 4, verifies the preferred result of feature, and the classification results after image classification can be adopted to differentiate, sorting technique can select SVM, decision tree, KNN;

Step 5, compares whether applying the preferred classification results of feature and sums up mRMR algorithm effect.

2. the mRMR characteristic optimization method according to step 3 in claim 1, is characterized in that: in feature preferred process, and in combining information opinion, the correlation theory of mutual information, can adopt two-value discretization method, histogram probabilistic method, F statistical method to realize.

3. the result verification process according to step 5 in claim 1, it is characterized in that: select the calculation and object all features out after certain sorting technique in step 4 and segmentation to carry out object type judgement, adopt same sorting technique and a small amount of feature after improvement mRMR algorithm of the present invention is preferred to carry out same classification experiments again, corresponding result is compared.