CN106203482A - Characteristics of The Remote Sensing Images dimension reduction method based on mRMR and KPCA - Google Patents

Abstract

The invention discloses a kind of Characteristics of The Remote Sensing Images dimension reduction method based on mRMR and KPCA, belong to technical field of remote sensing image processing.Characteristics of The Remote Sensing Images dimension reduction method of the present invention carries out feature selection first by mRMR method to the primitive character collection of remote sensing images, obtains the initial subset of Characteristics of The Remote Sensing Images；Then use KPCA method that the initial subset of Characteristics of The Remote Sensing Images is carried out further dimensionality reduction, obtain the majorized subset of Characteristics of The Remote Sensing Images.The invention also discloses a kind of Characteristics of The Remote Sensing Images dimensionality reduction device based on mRMR and KPCA and a kind of Classifying Method in Remote Sensing Image, device.Feature selection approach and eigentransformation method are carried out combination to carry out Characteristics of The Remote Sensing Images dimensionality reduction by the present invention, while effectively solving " dimension disaster " problem of remote sensing images, can be effectively improved the nicety of grading of remote sensing images.

Description

Characteristics of The Remote Sensing Images dimension reduction method based on mRMR and KPCA

Technical field

The present invention relates to technical field of remote sensing image processing, particularly relate to a kind of Characteristics of The Remote Sensing Images dimension reduction method.

Background technology

Along with the development of earth observation technology, remotely-sensed data diversification day by day, present that the obvious scale of construction is big, timeliness strong, " big data " feature such as type is miscellaneous, hard to distinguish between the true and false and potential value is big.ID information company (International Data Corporation, IDC) current research point out, the data that in the past few years whole world is newly-increased have 95% to be coarse, remote Unstructured data beyond normal data treatment scale.Further, at present the utilization rate of remotely-sensed data is not yet reached 10%, make The profligacy of resource and " dimension disaster " problem in pairs.Therefore, how while ensureing that Objects recognition rate does not declines, efficiently From high score remotely-sensed data, excavate the information with distinguishing ability, be the important topic being worth research.

As the means of solution " dimension disaster ", data mining (Data Mining, DM) just develops rapidly once proposition to be become Quite active field.DM solves the means of this problem and is to utilize the higher-dimension Combination of feature space to realize the yojan of dimension, It mainly includes feature selection approach and eigentransformation method.The former can effectively reject redundancy, reduces influence of noise.So And, even if the most remaining final two independent optimal characteristics, also may not be ensured of best feature combination, even can show worst. And there is the problems such as small sample, target characteristic subset dimension, new data type, additionally, the spectrum of high score remote sensing images, rectangular histogram Need just can be obtained by conversion etc. feature；The data of higher dimensional space are transformed into low-dimensional by modes such as mapping or conversion by the latter Space, it is possible to take into account the higher-order statistics of view data more, has the strongest in particular for non-overall situation linear data Classification capacity.But, the fixed structure that remote sensing images are the most unified, and there is ambiguity, high-resolution and image understanding The features such as polysemy.

In summary, it is used alone feature selection approach or eigentransformation method carries out dimensionality reduction to higher-dimension Characteristics of The Remote Sensing Images Mode all come with some shortcomings part.

Summary of the invention

The technical problem to be solved is to overcome prior art not enough, it is provided that a kind of based on mRMR (minimum Redundancy-Maximum Relevance, minimal redundancy maximal correlation) and KPCA (Kernel Principal Component Analysis, core principle component analysis) Characteristics of The Remote Sensing Images dimension reduction method, feature selection approach is become with feature The method of changing carries out combination to carry out Characteristics of The Remote Sensing Images dimensionality reduction, and effectively solving remote sensing images, (especially high-resolution is distant Sense image) " dimension disaster " problem while, the nicety of grading of remote sensing images can be effectively improved.

The present invention solves above-mentioned technical problem the most by the following technical solutions:

Characteristics of The Remote Sensing Images dimension reduction method based on mRMR and KPCA, first by original to remote sensing images of mRMR method Feature set carries out feature selection, obtains the initial subset of Characteristics of The Remote Sensing Images；Then use KPCA method to Characteristics of The Remote Sensing Images Initial subset carry out further dimensionality reduction, obtain the majorized subset of Characteristics of The Remote Sensing Images.

Techniques below scheme can also be obtained according to identical invention thinking:

A kind of Classifying Method in Remote Sensing Image, in the training stage, first extracts the characteristics of image of remote sensing images training sample, and Using extraction characteristics of image as the input of disaggregated model, pre-using remote sensing images training sample generic as disaggregated model Phase exports, and is trained disaggregated model, and the disaggregated model after training is remote sensing image classification device；At sorting phase, First extract remote sensing images test sample characteristics of image, and using extraction characteristics of image as described remote sensing image classification device Input, the output of remote sensing image classification device is the classification of described remote sensing images test sample；Described characteristics of image be by with Characteristics of The Remote Sensing Images dimension reduction method described in upper technical scheme carries out Feature Dimension Reduction to the primitive image features of remote sensing images and obtains.

Characteristics of The Remote Sensing Images dimensionality reduction device based on mRMR and KPCA, including:

MRMR feature selection module, for using mRMR method that the primitive character collection of remote sensing images is carried out feature selection, Obtain the initial subset of Characteristics of The Remote Sensing Images；

KPCA Feature Dimension Reduction module, for the initial subset to the Characteristics of The Remote Sensing Images that mRMR feature selection module is exported Carry out nonlinear characteristic conversion, obtain the majorized subset of Characteristics of The Remote Sensing Images.

A kind of remote sensing image classification device, including feature extraction unit and remote sensing image classification device；Described feature extraction list Unit is for extracting the characteristics of image of remote sensing images, and the characteristics of image input remote sensing image classification device that will be extracted；Described remote sensing Image Classifier training in advance by the following method obtains: first extract the characteristics of image of remote sensing images training sample, and with institute Extract the characteristics of image input as disaggregated model, defeated using remote sensing images training sample generic as the expection of disaggregated model Going out, be trained disaggregated model, the disaggregated model after training is remote sensing image classification device；Described feature extraction unit Including primitive character extraction module and Characteristics of The Remote Sensing Images dimensionality reduction device described above, primitive character extraction module is used for extracting The primitive image features of remote sensing images, described Characteristics of The Remote Sensing Images dimensionality reduction device is for being extracted primitive character extraction module The primitive image features of remote sensing images carries out Feature Dimension Reduction.

Compared to existing technology, technical solution of the present invention has the advantages that

The present invention is directed to the feature of remote sensing images, feature selection approach is organically combined with eigentransformation method: be first First with mRMR method, the primitive character collection of remote sensing images is carried out feature selection, find out with Category Relevance maximum the most mutually Between the minimum initial subset of redundancy；Then utilize core principle component analysis method that initial subset is carried out further dimensionality reduction.This Invention, while effectively solving " dimension disaster " problem of remote sensing images (especially high-resolution remote sensing image), can be effectively improved The nicety of grading of remote sensing images.

Accompanying drawing explanation

Fig. 1 is the Characteristics of The Remote Sensing Images dimensionality reduction schematic flow sheet of the present invention in detailed description of the invention；

Fig. 2 is the remote sensing image classification accuracy rate contrast using different characteristic dimensionality reduction scheme；

Fig. 3 is the Feature Dimension Reduction Contrast on effect using different characteristic dimensionality reduction scheme.

Detailed description of the invention

Below in conjunction with the accompanying drawings technical scheme is described in detail:

Feature selection approach or eigentransformation method is individually used in order to solve existing Characteristics of The Remote Sensing Images dimensionality reduction technology Deficiency, feature selection approach and eigentransformation method are carried out combination and remote sensing images are carried out Feature Dimension Reduction by the present invention, The method mainly includes two steps: 1) take minimal redundancy maximal correlation (mRMR) method that primitive character collection is carried out dependency Analyzing, Preliminary screening also generates initial subset, to eliminate partial redundance and incoherent feature；2) core principle component analysis is utilized (KPCA) method carries out nonlinear transformation to initial subset, to obtain more conversion just getable characteristic information that needs, and then Character subset after being optimized.

The Characteristics of The Remote Sensing Images dimension reduction method of the present invention, as it is shown in figure 1, specifically include following steps:

Step 1, use mRMR method carry out feature selection to the primitive character collection of remote sensing images, obtain at the beginning of remote sensing images Beginning subset:

Remotely-sensed data has higher-dimension and destructuring, has a strong impact on the classification performance to image and efficiency, causes distant The bulk redundancy of sense information.And directly retain the various implication relations in small part primitive character can lose image or between image, It is unfavorable for that subsequent operation obtains the nonlinear characteristic of remotely-sensed data.It is to say, tentatively reject partial redundance feature and retain big Part primitive character is conducive to subsequent operation to obtain the nonlinear characteristic of remotely-sensed data.Due to mRMR side based on Mutual Information Theory Method can minimize the redundancy between feature while maximizing character subset and Category Relevance, therefore, the present invention uses MRMR method tentatively eliminates partial redundance and incoherent feature, so that it is determined that the initial subset of remote sensing images.

MRMR method [Peng H, Long F, Ding C.Feature selection based on mutual information criteria of max-dependency,max-relevance,and min-redundancy[J] .IEEE Transactions on pattern analysis and machine intelligence,2005,27(8): 1226-1238.] choose majorized subset based on maximum dependency, it is assumed that from initial characteristics, select minimal feature subset S={x₁, x₂,...,x_i,x_i+1,...,x_m, it is this sub-set size with m, f is candidate feature, uses S_xRepresent the feature set selected, | S_x| represent Select the Characteristic Number that feature set comprises, then, the interpretational criteria function expression of the method is:

$J\left(f\right)=I(c;f)-\frac{1}{\left|S_x\right|}\underset{s\∈S_x}{\mathrm{\Σ}}I(s;f)$

In formula, I (s；F) mutual information between s and f is represented.If feature x_iWith feature x_jInterdepend, then mean I (x_i；x_j) bigger, i.e. the dependency between them is bigger, thus casts out any one in them all without to classification Differentiation power has too much influence, now, claims x_iWith x_jMutually redundant.It addition, during feature selection, the spy every time selected Levy all close to optimal solution.Assume existing feature set S_m-1, then, select next feature x_mProcess be i.e. optimized process, The optimum tolerance of this process is:

$\underset{x_j\∈(X-S_{m-1})}{\max }\[I(x_j,c)-\frac{1}{m-1}\underset{x_j\∈S_{m-1}}{\mathrm{\Σ}}I(x_j,x_i)\]$

It is below the specific algorithm step of present embodiment:

Step 1. subset initializes.

Mutual information I (x with classification c_i, c) maximum candidate feature x_i, it being added into initial subset, formula is as follows, now, Initial subset S={y1}, size is 1, i.e. the primitive image features x maximum with Category Relevance_iS collection will be firstly added；

$\arg \underset{x_i}{\max }I(x_i,c)$

Step 2. progressively expands S.

In order to take into account dependency and the redundancy of initial subset, candidate feature is carried out correlation analysis and redundancy is divided Analysis, gradually candidate feature is concentrated performance optimum i.e. with S weak redundancy and with optimal characteristics x of c strong correlation_iAdd S collection, screening Formula is；

$\arg \underset{x_i}{\max }\[\frac{1}{\left|S\right|}\mathrm{\Σ}I(x_i,c)-\frac{1}{{\left|S\right|}^2}\underset{y_j\∈S}{\mathrm{\Σ}}I(x_i,y_j)\]$

Step 3. determines S according to subset evaluation result.

When S expands a certain size m (m Yu M is close) to, and select feature set S_xRedundancy and weak to classification c relevant Feature is the most remaining, now, is rejected, and the size of S is m.Then, this subset is evaluated, and and S_initialTest Effect is compared, if evaluation effect allows of no optimist, then turns Step 2, and increases the gap of m Yu M, i.e. reduce m.Otherwise, turn Step 4.During this, the quality of evaluation effect is that comprehensive m considers with test effect, it is allowed to sacrifices subset space storage and disappears Consumption, i.e. retain bigger m, and do not allow to retain poor test effect；

Now, subset S is the initial subset that we need to Step 4., is designated as S_first={ y₁,y₂,...,y_m, this calculation Method terminates.

Step 2, use KPCA method carry out further dimensionality reduction to the initial subset of remote sensing images, obtain the excellent of remote sensing images Beggar collects:

KPCA is the typical eigentransformation method that data are transformed into by the way of nonlinear mapping lower dimensional space, with Linear eigentransformation method is compared, and KPCA can obtain more needs and convert the getable characteristic information of, especially can Process nonlinear characteristic well.Especially, the ununified fixing space structure of high score remote sensing images, and, mRMR method Determined by initial subset be the Selecting operation to primitive character collection, it is impossible to obtain the various implication relations between image, be unfavorable for Obtain the nonlinear characteristic of remotely-sensed data.Therefore, in order to fully use the characteristic information of high score remote sensing images, the present invention adopts With KPCA, initial subset is carried out nonlinear eigentransformation.

Data set is mapped to a high-dimensional feature space by kernel function by KPCA, then uses PCA method real in this space The linear transformation of existing data set, thus realize the nonlinear transformation of raw data set, concrete algorithm steps is as follows.

Algorithm steps:

Step 1. calculates nuclear matrix K={K_ij}_n×n, by this matrix, initial data is mapped to high-dimensional feature space.

Formula is as follows:

$K(x_i,x_j)=\exp \left(\frac{-\left|\right|x_i-x_j|{|}^2}{2{\mathrm{\σ}}^2}\right)$

In formula, x_iWith x_jIt is all normalized sample data, represents that sample average, σ represent sample variance with μ, then standard Change formula as follows:

$x_i=\frac{x_i-\mathrm{\μ}}{\mathrm{\σ}}$

Step 2. centralization matrix K, obtains matrix K^c, to revise nuclear matrix.

Assume with 1_NRepresent that each element is the matrix of the N × N of 1/N, then centralization formula is as follows:

K^c=K-1_NK-K1_N+1_NK1_N

Step 3. is to K^cCarry out Eigenvalues Decomposition, and be characteristic value and corresponding basis according to the contribution rate size of characteristic value Levy vector descending.

The formula of Eigenvalues Decomposition is as follows:

K^cV=λ V

Step 4. determines the character subset after optimization:

According to the threshold value being previously set, choose the accumulation contribution rate feature set [λ more than this threshold value₁,λ₂,...,λ_q], corresponding Eigenvector collection [v₁,v₂,...,v_q] it being the character subset W after optimization, initial subset is comprised by character subset herein Main information has been compressed in principal component, is the spatial alternation to initial subset；

Subset W is evaluated by Step 5., and is tested effect and S_firstTest effect compare.If the effect of evaluation Fruit allows of no optimist, then turn Step 2, and reduce the value of threshold value.Otherwise, Step 6 is turned.Wherein, the quality of evaluation effect is Size q according to W considers with test effect, and in tolerance interval, this algorithm allows to sacrifice test effect, and does not permits Permitted to retain bigger subset space storage consumption, i.e. in the case of q is sufficiently small, it is allowed to the test effect of W compares S_firstTest Effect is slightly worse；

Now, subset W is the majorized subset that the present invention finally gives to Step 6., the majorized subset obtained is designated as S_best={ v₁,v₂,...,v_p, algorithm terminates.

Utilize features described above dimension reduction method can classify remote sensing images further, target recognition, retrieval etc. should With.Such as, when remote sensing images are classified, in the training stage, first extract the characteristics of image of remote sensing images training sample, and Using extraction characteristics of image as the input of disaggregated model, pre-using remote sensing images training sample generic as disaggregated model Phase exports, and is trained disaggregated model, and the disaggregated model after training is remote sensing image classification device；At sorting phase, First extract remote sensing images test sample characteristics of image, and using extraction characteristics of image as described remote sensing image classification device Input, the output of remote sensing image classification device is the classification of described remote sensing images test sample；Described characteristics of image be by with Characteristics of The Remote Sensing Images dimension reduction method described in upper technical scheme carries out Feature Dimension Reduction to the primitive image features of remote sensing images and obtains.

Above-mentioned disaggregated model can use the prior aries such as support vector machine, feedforward neural network, degree of depth learning neural network, Here is omitted.

In order to verify inventive feature dimensionality reduction effect, applied to remote sensing image classification, and with based on initial data The remote sensing image classification of collection, remote sensing image classification based on mRMR dimensionality reduction, remote sensing image classification based on KPCA dimensionality reduction are imitated Fruit contrast.Experimental subject is a large-scale remote sensing image classification common test data set, and this data set is from United States Geological Survey (USGS) place downloads, and has 21 classes, and every class includes 100 width images, is QuickBird high-resolution remote sensing image, resolution sizes is 256 × 256.These images include the field in multiple states of the U.S. Scape: farmland, aircraft, infield, beach, building, shrubbery, jungle, forest, highway, golf course, sea Port, crossroad, medium dense residential area, prefabricated house district, viaduct, parking lot, river, runway, sparse population district, oil storage tank And tennis court.These images reflect the typical case of every class scene, have preferable representativeness.Test with the ratio of 4:1 with Machine extraction training sample and test sample, after carrying out Feature Dimension Reduction, use LIBSVM to realize classification.

Represent the current number of times of sample set random packet with transverse axis, the longitudinal axis represents corresponding nicety of grading, broken line " Initial " represents employing primitive character collection S_initialClassifying quality, broken line " mRMR ", " KPCA ", " Sel_Ex " respectively table Show the classifying quality using mRMR method, KPCA method and context of methods (Sel_Ex) to carry out Dimensionality Reduction.The most several differences The classifying quality of method contrasts as shown in Figure 2.

As shown in Figure 2, in the accuracy rate of classification, compared with KPCA method, Sel_Ex can reach divide suitable with it Class effect, and the nicety of grading of sample set can be improved, and the performance of mRMR method is worst.Chief reason has: one, mRMR Method easily loses information.A lot of features the most little with Category Relevance on the contrary can be combined into resolving ability strong with further feature Feature, now, the discriminant score of this feature is less, the most disallowable.Therefore, mRMR method likely abandons and much need not The weak relevant information wanted；Its two, remote sensing images structure is complicated, and especially high score remote sensing images comprise abundant information, have spectrum, Rectangular histograms etc. need nonlinear transformation just getable feature.KPCA method realizes dimensionality reduction by nonlinear mapping, it is possible to retain More pivot information, this is extremely beneficial to the sort research of high score remote sensing images.Therefore, KPCA method can reach considerable effect Really, and dimension can be reduced；Its three, Sel_Ex integrated use correlation analysis and Kernel-Based Methods.It is to S_initial The preliminary of redundancy reject the redundancy that can reduce feature set to a certain extent, on the other hand, the inventive method obtains Initial subset S_firstNot only expression effect is good, and can retain the sample data information of abundance, is especially able to ensure that The structure of feature space is without damage, and this greatly facilitates the nonlinear organization information being extracted primitive character collection by kernel function. Therefore, the inventive method can improve S_initialNicety of grading.

Additionally, evaluate the quality of dimensionality reduction effect, it is that the size to subset dimension considers, especially with test effect The satisfaction of subset space storage consumption is exceeded the satisfaction to test effect, therefore, from the angle of subset space storage consumption The performance of degree examination the inventive method (Sel_Ex).S_initialAnd the dimension of majorized subset that various method obtains is respectively 4200 dimensions, 1000 dimensions, 50 dimensions, 30 dimensions, if regarding unit circle 1 as the dimensional space of primitive character collection, with ball " initial " table Show S_initialDimensional space, with ball " mRMR ", " KPCA ", " Sel_Ex " represent respectively employing " mRMR ", " KPCA " method and The majorized subset S that the inventive method " Sel_Ex " obtains_bestDimensional space, coordinate axes represents dimensional space, the most various methods Dimensionality reduction Contrast on effect schematic diagram as shown in Figure 3.In Fig. 3, numeral 0.238,0.119,0.007 represents that various method obtains respectively Majorized subset space relative to the ratio magnitude of the dimensional space of primitive character collection.

Fig. 3 shows: the dimensional space of the majorized subset that the inventive method obtains is minimum, is S_initialDimensional space 7/1000ths, it is possible to a greater degree reduce primitive character collection dimension.

Complex chart 2 and Fig. 3 understands, and the inventive method can reduce the dimension of feature set to a greater degree, simultaneously can one Determine to improve in degree the accuracy of classification, maintain the balance between classification performance and arithmetic speed well.

Claims (4)

1. Characteristics of The Remote Sensing Images dimension reduction method based on mRMR and KPCA, it is characterised in that first by mRMR method to remote sensing The primitive character collection of image carries out feature selection, obtains the initial subset of Characteristics of The Remote Sensing Images；Then use KPCA method to distant The initial subset of sense characteristics of image carries out further dimensionality reduction, obtains the majorized subset of Characteristics of The Remote Sensing Images.

2. a Classifying Method in Remote Sensing Image, in the training stage, the first characteristics of image of extraction remote sensing images training sample, and with Extracted characteristics of image is as the input of disaggregated model, using remote sensing images training sample generic as the expection of disaggregated model Output, is trained disaggregated model, and the disaggregated model after training is remote sensing image classification device；At sorting phase, first First extract the characteristics of image of remote sensing images test sample, and using defeated as described remote sensing image classification device of extraction characteristics of image Entering, the output of remote sensing image classification device is the classification of described remote sensing images test sample；It is characterized in that, described characteristics of image Obtain for the primitive image features of remote sensing images being carried out Feature Dimension Reduction by Characteristics of The Remote Sensing Images dimension reduction method described in claim 1 Arrive.

3. Characteristics of The Remote Sensing Images dimensionality reduction device based on mRMR and KPCA, it is characterised in that including:

MRMR feature selection module, for using mRMR method that the primitive character collection of remote sensing images is carried out feature selection, obtains The initial subset of Characteristics of The Remote Sensing Images；

KPCA Feature Dimension Reduction module, for carrying out the initial subset of the Characteristics of The Remote Sensing Images that mRMR feature selection module is exported Nonlinear characteristic converts, and obtains the majorized subset of Characteristics of The Remote Sensing Images.

4. a remote sensing image classification device, including feature extraction unit and remote sensing image classification device；Described feature extraction unit For extracting the characteristics of image of remote sensing images, and the characteristics of image input remote sensing image classification device that will be extracted；Described remote sensing figure As grader training in advance by the following method obtains: first extract the characteristics of image of remote sensing images training sample, and to be carried Take the characteristics of image input as disaggregated model, defeated using remote sensing images training sample generic as the expection of disaggregated model Going out, be trained disaggregated model, the disaggregated model after training is remote sensing image classification device；It is characterized in that, described Feature extraction unit includes primitive character extraction module and Characteristics of The Remote Sensing Images dimensionality reduction device as claimed in claim 3, original Characteristic extracting module is for extracting the primitive image features of remote sensing images, and described Characteristics of The Remote Sensing Images dimensionality reduction device is for original The primitive image features of the remote sensing images that characteristic extracting module is extracted carries out Feature Dimension Reduction.