CN104751176A - Method for selecting high-spectrum remote-sensing image wave band - Google Patents

Abstract

The invention discloses a method for selecting high-spectrum remote-sensing image wave band based on hybrid binary particle swarm optimization differential evolution (HBPSODE). The method comprises the steps of preprocessing an original high-spectrum remote-sensing image; initializing double swarm individual and algorithm parameters; iterating the double swarms in parallel by the HBPSODE; transferring the optimal solution information through the swarm; calculating the classification precision through an SVM classifier to be used as an adaptability value; updating evolution until reaching the specified evolution times or reaching the maximum precision.

Description

A kind of target in hyperspectral remotely sensed image band selection method

Technical field

The present invention relates to a kind of target in hyperspectral remotely sensed image band selection method, be specially a kind of based on mixing scale-of-two population differential evolution, with the target in hyperspectral remotely sensed image band selection method based on packaged type HBPSODE-SVM algorithm that to obtain optimal bands composite be target, belong to high-spectrum remote sensing processing technology field.

Background technology

Remote sensing (Remote Sensing) is one and utilizes electromagnetic wave principle to obtain distant signal and to make it imaging, and remotely can experience the technology of perception distant place things, be an emerge science.Along with the raising of computer technology and optical technology, remote sensing technology have also been obtained and develops rapidly.In recent years, remote sensing satellite miscellaneous constantly succeeds in sending up, and has promoted remotely-sensed data acquiring technology towards three height (high spatial resolution, high spectral resolution and high time resolution) and (multi-platform, multisensor, multi-angle) more than three future development.

High-spectrum remote-sensing has the high feature of spectral resolution, it by carrying EO-1 hyperion sensor on different spaces platform, thus can the visible ray of electromagnetic wave spectrum, near infrared, within the scope of infrared and Thermal infrared bands, with continuous print spectral band to earth surface area imaging simultaneously, wave band number can reach tens of so that hundreds of, and obtain atural object continuous print spectral information, thus achieve the synchronous acquisition of ground object space, radiation and spectral information.Compared with conventional remote sensing, the key distinction is that high-spectrum remote-sensing is narrow wave band imaging, and except the spatial information of two dimension, also add one dimension spectral information, and the application of remote sensing technology is expanded.

High-spectrum remote-sensing can detect more meticulous spectral characteristic, and high spectrum image has the spectral information that conventional remote sensing cannot be reached, and is conducive to the process such as terrain classification, identification and Decomposition of Mixed Pixels.But high spectrum image, while spectral information amount increases, too increases the dimension of data, makes the data volume of image increase sharply.Correlativity between its higher dimension and wave band not only can make computing become complicated, and processing speed declines greatly, and when finite sample, nicety of grading may be caused to reduce.After imaging spectrometer obtains hyperspectral image data, band selection seems particularly important.

Target in hyperspectral remotely sensed image band selection can be counted as the difficult combinatorial optimization problem of a NP, can select the band combination that the paired nicety of grading of several band group has better performance with intelligent suboptimum searching algorithm according to interpretational criteria function from all band.Common intelligent algorithm has been successfully applied in the middle of band selection, but its defect also comes out very soon, genetic algorithm effectively can not restrain within the limited time, ant group algorithm due to the deficient solving speed of its initial information element slow, the search time that general needs are longer, and easily there is precocious phenomenon.Although particle cluster algorithm fast convergence rate, the not high easy appearance of precision is precocious.PSO algorithm and DE algorithm all belong to the novel heuritic approach based on swarm intelligence, can apply to separately in the middle of target in hyperspectral remotely sensed image band selection.But two kinds of algorithms all also exist some defects, more much higher sample is remain at the initial stage individuality of population iteration, along with the increase of iterations, individuality in PSO algorithm is optimal particle in population progressively, DE algorithm adopts Greedy strategy when performing selection and operating, and namely only has and could participate in next iteration evolution when variation individuality is more outstanding than the fitness function value of current individual time.Although these evolutionary mechanisms can accelerate algorithm the convergence speed, but also make the difference between population at individual reduce gradually, the diversity of population also reduces thereupon, and now population is easy to be absorbed in locally optimal solution, and namely fitness function value changes slowly or has almost no change.For the defect of two kinds of algorithms, method of the present invention adopts two Evolution of Population strategy, allows two kinds of algorithm Simultaneous Iterations evolve and find optimal bands composite solution, helps population each other depart from locally optimal solution by a kind of Information exchange mechanism.

Summary of the invention

Goal of the invention: in order to overcome the technical deficiency of existing high-spectrum remote sensing band selection, improve nicety of grading, the invention provides a kind of EO-1 hyperion band selection method, is a kind of with the band selection method based on packaged type mixing scale-of-two population differential evolution (HBPSODE-SVM) algorithm that to obtain optimal bands composite be target.

Technical scheme: a kind of target in hyperspectral remotely sensed image band selection method, first pre-service is carried out to original target in hyperspectral remotely sensed image, by two population at individual and algorithm parameter initialization, then application mix scale-of-two population differential evolution (HBPSODE) method, allow two population parallel iteration by transmitting optimum solution information between population, and utilize SVM classifier to calculate nicety of grading as fitness value, upgrade and evolve until reaching regulation evolution number of times or till reaching maximal accuracy.

Corresponding amendment is made to PSO algorithm and DE algorithm, proposes a kind of scale-of-two differential evolution algorithm of hybrid coding, can be extended in discrete domain; First adjuvant search space S '=[-a, a] is defined ^d, a is positive integer, solution space S={0,1} ^d, d is the dimension of problem; Then tie up real number vector X by adjuvant search space D and add solution space scale-of-two D dimensional vector B i.e. (X, B) as the hybrid coding representation of individual (or variant); Real number vector X still performs mutation operation and interlace operation according to differential evolution algorithm, and before operation is selected in execution, needing is developed real number vector X by epimorphism is mapped to binary vector B, and epimorphism evolution mapping function defines:

Wherein, h _ij(t+1) be each component value of variant after interlace operation, for ambiguity function, b _ij(t+1) be each component value of binary vector B, Dynamic gene μ can control b _ij(t+1) be set to the probability size of 1, get μ=0.5.

If (X _i(t), B _i(t)) and (X _i(t+1), B _i(t+1)) represent that t generation of population and t+1 are for individual i respectively, (H _i(t+1), E _i(t+1)) represent the variant of t+1 for individual i, f (x) represents fitness function.New selection behaviour is defined as follows:

EO-1 hyperion band selection method, specifically comprises the steps:

Step 1: original target in hyperspectral remotely sensed image pre-service.Reject jammr band, preliminary election type of ground objects, arranges the dimension D of search volume, algorithm greatest iteration evolution number of times MaxDT.

Step 2: population Ppso and the correlation parameter of HBPSO algorithm evolution are pressed in initialization.Arranging population number is Np, arranges Studying factors c ₁, Studying factors c ₂, maximum inertia weight coefficient w _max, minimum inertia weight coefficient w _mindeng.In order to improve the performance of population (PSO) algorithm, wherein inertia weight w upgrades according to following formula, and i represents i-th iteration.

$w=w_{\max }-\frac{w_{\max }-w_{\min }}{\mathrm{MaxDT}}\·i---\left(1\right)$

Step 3: population Pde and the correlation parameter of HBDE algorithm evolution are pressed in initialization.Arranging population number is Nd, zoom factor F, Crossbreeding parameters CR etc.In order to improve differential evolution (DE) algorithm performance, wherein zoom factor F upgrades according to following formula, and F0 is a constant, and i represents i-th iteration.

$F=F0\·2^{\exp \left(\frac{1-\mathrm{MaxDT}}{\mathrm{MaxDT}+1-i}\right)}---\left(2\right)$

Step 4: evolution iteration count t=0 is set.

Step 5:Ppso population carries out a position according to HBPSO algorithm and speed upgrades, and utilizes SVM classifier to implement classification to the band combination after renewal, and calculates nicety of grading as fitness value, record t for optimal adaptation angle value and band combination.

Step 6:Pde population makes a variation to all individualities according to HBDE algorithm, intersect, select operation.Utilize SVM classifier to calculate fitness value, record t for optimal adaptation angle value and band combination.

Step 7: compare Ppso and Pde t for the optimal adaptation angle value chosen, adjust the optimum solution of respective population.

Step 8: upgrade evolutionary generation counter t=t+1.If evolutionary generation reaches maximum evolution number of times or meets accuracy requirement, then termination algorithm, otherwise goes back to step 5.

Techniques and methods involved for a better understanding of the present invention, is introduced the theory that the present invention relates at this.

1, population (PSO) algorithm

The implementation method of PSO algorithm makes particle i be that in population, any one is individual, the position X of the t time iteration _i(t)=[x ₁, x ₂..., x _d], speed V _i(t)=[v ₁, v ₂..., v _d], the wherein dimension of D problem of representation, pBest _irepresent the history optimum solution position of particle i, gBest (t) represents globally optimal solution position in the t time iteration.Particle i is according to following formula renewal speed and position:

V _i(t+1)＝w·V _i(t)+c ₁·rand(0,1)·(pBest _i-X _i(t))

+c ₂·rand(0,1)·(gBest(t)-X _i(t)) (3)

X _i(t+1)＝X _i(t)+V _i(t+1) (4)

Wherein, c ₁and c ₂for aceleration pulse represents that particle is subject to the influence degree of social recognition and individual cognition, rand (0,1) represents the random number of obeying [0,1] and distributing, and w is inertia weight, and it can with iterative process dynamic conditioning particle rapidity.

2, differential evolution (DE) algorithm

The implementation method of differential evolution algorithm is first stochastic generation initial population, makes X _it () represents that t is for the i-th individuality in population.Mutation operation is exactly be X _it () produces a new variant.First, from current population, three Different Individual X are selected arbitrarily _r1(t), X _r2(t), X _r3t (), then produces the individual X of variation according to (2.3) formula _i(t) ':

X _i(t)'＝X _r1(t)+F·(X _r2(t)-X _r3(t)) (5)

Wherein F is zoom factor, and its span is [0.1,1].

Interlace operation object allows current individual X _i(t) and the individual X of variation _it () ' is intersected, thus introduce the diversity of population at individual.Specific operation process is as follows:

First stochastic generation integer r ∈ [1, D], the dimension of D problem of representation, then operates according to (6) formula the every one dimension of individual vector:

Wherein CR represents crossing-over rate, and its span is [0,1].Integer r can ensure that the new individuality after intersecting has at least the value of one-component different from current individual.

Can the new individuality after interlace operation is determined in selection operation enter next round iterative evolution.Whether be better than current individual by the interpretational criteria function individual fitness value that can judge to make new advances, selection operation operates according to (2.5) formula:

In DE algorithm, variation mode generally can be represented as: DE/x/y/z, and wherein x represents base vector selection mode in mutation operation, and y represents the number of difference vector, and z represents the pattern of intersecting and adopting.The variation mode introduced above is expressed as: DE/rand/1/bin, bin represent employing binomial cross-mode.In addition some other classical variation mode is also had:

DE/best/1/bin。This pattern represents that variation individuality is from current population, select two random individuals and optimum individual to combine, and its expression formula is as follows:

X _i(t)'＝X _best(t)+F·(X _r1(t)-X _r2(t)) (8) DE/best/2/bin。This pattern represents that 4 random individuals need be selected to produce two difference vectors from current population forms variation individuality, and its expression formula is as follows:

X _i(t)'＝X _best(t)+F·(X _r1(t)+X _r2(t)-X _r3(t)-X _r4(t)) (9)

DE/rang-to-best/1/bin。This pattern represents and the optimal vector in current population is joined in difference vector to form variation individual, and its expression formula is as follows:

X _i(t)'＝X _i(t)+λ·(X _best(t)-X _r1(t))+F·(X _r2(t)-X _r3(t)) (10)

1, support vector machine (SVM) sorter general introduction

Support vector machine (support vector machine, SVM) is a kind of new machine learning method that the people such as Vapnik put forward on Statistical Learning Theory basis.It is on the basis of linear classifier, introduces structural risk minimization, Optimum Theory and kernel method and develops.It is solve convex quadratic programming problem in essence, has greater advantage and be successfully applied in the middle of the problems such as regression problem, Classification and Identification, discriminatory analysis in solution small sample, non-linear and high dimensional pattern identification problem.

The mechanism of SVM is the optimal separating hyper plane that searching one meets classificating requirement, makes this lineoid while guarantee nicety of grading, the white space of plane both sides can be made to maximize.Illustrate with the classification of two class data, suppose that training sample set is (x _i, y _i), i=1,2 ..., l, x _i∈ R ⁿ, y _i∈ { ± 1}.General d dimension space neutral line discriminant function can be expressed as:

g(x)＝ω ^Tx+b (11)

The lineoid the Representation Equation of so classifying is:

ω ^Tx+b＝0 (12)

Allly correctly to be classified and the sample point possessing class interval all must meet:

y _i·(ω ^Tx _i+b)≥1,i＝1,2,...,l,y _i∈{±1} (13)

Wherein, y _ibe the class label of i-th sample, ω is weight coefficient.As shown in Figure 2, those just drop on point on Optimal Separating Hyperplane and are called support vector (support vector), the class interval size of two class samples

Be expressed as:

$m\arg \mathrm{in}=\frac{1}{2}\left|\right|\mathrm{\ω}\left|\right|---\left(14\right)$

Now, find optimal hyperlane problem and be converted under the constraint of (13) formula, solved function:

Introduce Lagrange factor-alpha _i, (15) formula is solved and can be obtained:

$\min :L(\mathrm{\ω},b,\mathrm{\α})=\frac{1}{2}{\mathrm{\ω}}^T\mathrm{\ω}-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i[y_i({\mathrm{\ω}}^T{x}_i+b)-1]---\left(16\right)$

Local derviation is asked to ω, b, can obtain:

$\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0,\mathrm{\ω}=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i{x}_i---\left(17\right)$

(17) formula is substituted into (16) formula, obtains:

$\max W\left(\mathrm{\α}\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right)---\left(18\right)$

Solve and can obtain optimum solution order training sample be support vector, the weight coefficient vector ω of optimal separating hyper plane ^*for the linear combination of support vector.B ^*can by constraint condition y _i(ω ^tx _i+ b)-1=0 solves and obtains.

Above-described is a linear Optimal Separating Hyperplane, but the classifying face in the middle of a lot of practical problems between classification is often certain nonlinear curved surface.SVM, in solution linearly inseparable problem, adopts kernel function that the Nonlinear Classification in lower dimensional space is mapped in higher dimensional space, constructs linear Optimal Separating Hyperplane in higher dimensional space.Main following 4 kinds of typical kernel function at present:

linearly (linear) kernel function: K (x _i, x _j)=(x _ix _j);

polynomial expression (polynomial) kernel function: K (x _i, x _j)=[(x _ix _j)+1] ^q;

radial basis (RBF) kernel function: K (x _i, x _j)=exp{-|x _i-x _j| ²/ σ ₂;

s shape (sigmoid) kernel function: K (x _i, x _j)=tanh (υ (x _ix _j)+c).

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the embodiment of the present invention;

Fig. 2 is svm classifier lineoid figure;

Fig. 3 is by 50, the image of 27,17 wave band synthesis;

Fig. 4 is AVIRIS original atural object calibration figure;

Fig. 5 is algorithm optimum individual fitness value change curve;

Fig. 6 is algorithm classification result figure, wherein (a) is BPSO-SVM classification results figure, b () is HBPSO-SVM classification results figure, (c) is HBDE-SVM classification results figure, and (d) is HBPSODE-SVM classification results figure.

Embodiment

Below in conjunction with specific embodiment, illustrate the present invention further, these embodiments should be understood only be not used in for illustration of the present invention and limit the scope of the invention, after having read the present invention, the amendment of those skilled in the art to the various equivalent form of value of the present invention has all fallen within the application's claims limited range.

Target in hyperspectral remotely sensed image band selection method, first pre-service is carried out to original target in hyperspectral remotely sensed image, by two population at individual and algorithm parameter initialization, then application mix scale-of-two population differential evolution (HBPSODE) method, allow two population parallel iteration by transmitting optimum solution information between population, and utilize SVM classifier to calculate nicety of grading as fitness value, upgrade and evolve until reaching regulation evolution number of times or till reaching maximal accuracy.

Corresponding amendment is made to PSO algorithm and DE algorithm, proposes a kind of scale-of-two differential evolution algorithm of hybrid coding, can be extended in discrete domain; First adjuvant search space S '=[-a, a] is defined ^d, a is positive integer, solution space S={0,1} ^d, d is the dimension of problem; Then tie up real number vector X by adjuvant search space D and add solution space scale-of-two D dimensional vector B i.e. (X, B) as the hybrid coding representation of individual (or variant); Real number vector X still performs mutation operation and interlace operation according to differential evolution algorithm, and before operation is selected in execution, needing is developed real number vector X by epimorphism is mapped to binary vector B, and epimorphism evolution mapping function defines:

Wherein, h _ij(t+1) be each component value of variant after interlace operation, for ambiguity function, b _ij(t+1) be each component value of binary vector B, Dynamic gene μ can control b _ij(t+1) be set to the probability size of 1, get μ=0.5.

If (X _i(t), B _i(t)) and (X _i(t+1), B _i(t+1)) represent that t generation of population and t+1 are for individual i respectively, (H _i(t+1), E _i(t+1)) represent the variant of t+1 for individual i, f (x) represents fitness function.New selection behaviour is defined as follows:

As shown in Figure 1, specifically comprise the steps:

Step 1: original target in hyperspectral remotely sensed image pre-service.Reject jammr band, preliminary election type of ground objects, arranges the dimension D of search volume, algorithm greatest iteration evolution number of times MaxDT.

Step 2: population Ppso and the correlation parameter of HBPSO algorithm evolution are pressed in initialization.Arranging population number is Np, arranges Studying factors c1, Studying factors c2, maximum inertia weight coefficient w _max, minimum inertia weight coefficient w _mindeng.In order to improve the performance of population (PSO) algorithm, wherein inertia weight w upgrades according to following formula, and i represents i-th iteration.

$w=w_{\max }-\frac{w_{\max }-w_{\min }}{\mathrm{MaxDT}}\·i---\left(22\right)$

Step 3: population Pde and the correlation parameter of HBDE algorithm evolution are pressed in initialization.Arranging population number is Nd, zoom factor F, Crossbreeding parameters CR etc.In order to improve differential evolution (DE) algorithm performance, wherein zoom factor F upgrades according to following formula, and F0 is a constant, and i represents i-th iteration.

$F=F0\·2^{\exp \left(\frac{1-\mathrm{MaxDT}}{\mathrm{MaxDT}+1-i}\right)}---\left(23\right)$

Step 4: evolution iteration count t=0 is set.

Step 5:Ppso population carries out a position according to HBPSO algorithm and speed upgrades, and utilizes SVM classifier to implement classification to the band combination after renewal, and calculates nicety of grading as fitness value, record t for optimal adaptation angle value and band combination.

Step 6:Pde population makes a variation to all individualities according to HBDE algorithm, intersect, select operation.Utilize SVM classifier to calculate fitness value, record t for optimal adaptation angle value and band combination.

Step 7: compare Ppso and Pde t for the optimal adaptation angle value chosen, adjust the optimum solution of respective population.

Step 8: upgrade evolutionary generation counter t=t+1.If evolutionary generation reaches maximum evolution number of times or meets accuracy requirement, then termination algorithm, otherwise goes back to step 5.

The simulation experiment result is analyzed

1. experimental image

By emulation experiment, analysis and inspection is carried out to the performance of algorithm.For the validity of checking HBPSODE-SVM algorithm, test for a certain standard target in hyperspectral remotely sensed image herein.The remote sensing image used is a part for the Indiana, USA northwestward agricultural bulk testing district target in hyperspectral remotely sensed image obtained by AVIRIS sensor in June, 1992.Its wavelength coverage is 0.4 ~ 2.5 μm, and image size is 145 × 145pixel, and spatial resolution is 25m.Get rid of by with serious pollution wave bands (wave band 1 ~ 4,78,80 ~ 86,103 ~ 110,149 ~ 165,217 ~ 224) such as steam noises from original wave band, retain remaining 179 wave bands and test.Fig. 3 is that test chooses the 50th, 27,17 wave band synthesis R, G, B false color images.Fig. 4 is that the original atural object of AVIRIS is fixed.

Have 17 class atural objects in image, choose wherein 7 class atural objects and participate in classification experiments.Training sample and test sample book are chosen by the ratio uniform of 1:1.The numbering of 7 class atural objects shown in table-1-, title, training and testing sample size.Experimental arrangement adopts Matlab (R2009b) programming realization, and SVM classifier adopts libsvm tool box.

Table-1-training sample and test sample book

2. experimental technique and relative parameters setting

In order to verify the superiority of HBPSODE-SVM method, design A, B, C, D, 4 groups of contrast simulation experiments.Consider the fairness of algorithm relative parameters setting, A group adopts based on scale-of-two population (BPSO) algorithm, B group is based on mixing binary population (HBPSO) algorithm, and C group contrasts based on the band selection method mixing scale-of-two population differential evolution (HBPSODE) algorithm based on mixing scale-of-two differential evolution (HBDE) algorithm and D group.4 groups of experiments all adopt support vector machine (SVM) as sorter, adopt RBF kernel function.The relative parameters setting of algorithm and sorter is as following table-2-:

Table-2-relative parameters setting

3. experimental result contrast

After determining training sample and test sample book, we respectively from following two aspects to (BPSO) algorithm, mix binary population (HBPSO) algorithm, the experimental result of mixing scale-of-two differential evolution (HBDE) algorithm and mixing scale-of-two population differential evolution (HBPSODE) algorithm compares.

1. error matrix

Test the optimal bands combined that obtains for 4 groups to participate in terrain classification and test the error matrix that obtains as shown in table 3 ~ 6.

Table-3-A organizes BPSO-SVM error matrix

Table-4-B organizes HBPSO-SVM error matrix

Table-5-C organizes HBDE-SVM error matrix

Table-6-D organizes HBPSODE-SVM error matrix

As can be seen from table 3 ~ 6, HBPSODE-SVM is better than other 3 kinds of algorithms generally in producer's precision and user's precision, this with regard to the leakage point error that means HBPSODE-SVM algorithm and produce and many points of errors relatively little.The error matrix of contrast HBPSO-SVM and BPSO-SVM algorithm, although both overall classification accuracies only differ 0.1%, the producer's precision and the user's precision that compare each class atural object can find out that HBPSO-SVM is better than BPSO-SVM.The error matrix of contrast HBPSO-SVM and HBDE-SVM algorithm, DE algorithm is at overall classification accuracy, and producer's precision and user's precision are all better than PSO algorithm.Meet differential evolution (DE) algorithm mentioned in document 39 on overall performance, be better than population (PSO) algorithm.

2. Kappa analyzes and overall accuracy

Kappa analyzes the consistance informixs such as the overall accuracy of error matrix, producer's precision and user's precision can got up quantitatively between classification of assessment result and ground reference information.Table 7 is overall accuracy and the Kappa value of 4 groups of experiment algorithms.

Table-7-overall classification accuracy and Kappa value

By showing the Kappa value of-7-, can to evaluate HBPSODE-SVM algorithm quantitatively better than the classification performance being used alone HBDE-SVM or HBPSO-SVM algorithm and obtaining.HBDE-SVM performance is better than HBPSO-SVM, and HBPSO-SVM is better than general BPSO-SVM algorithm again.

Figure-5-is the change curve of 4 groups of experiment algorithms optimum individual fitness value (nicety of grading) in whole iterative process.

Can be seen by figure-5-, the curve of d figure occurs that the length of " platform " is the shortest, illustrates that the diversity of HBPSODE-SVM algorithm population in an iterative process keeps better, occurs to flee from locally optimal solution in shorter iterations when evolving stagnation.And all there is evolving the situation stagnated and can not flee from locally optimal solution in the short time in other 3 kinds of algorithms.Although HBPSO-SVM in final nicety of grading higher than BPSO-SVM, but HBPSO-SVM algorithm is easier to be absorbed in locally optimal solution and cannot flees from the short time in whole evolution iterative process, BPSO-SVM remains very much higher sample at the Evolution of Population iteration initial stage, but is absorbed in precocity soon until arrive maximum evolution iterations.It is better that HBDE-SVM compares that above-mentioned 2 kinds of algorithms keep on population diversity, although be also absorbed in locally optimal solution, can flee from before evolution iteration ends.The classification results figure of 4 groups of experiment algorithms as shown in Figure 6.

Can find out that the nicety of grading of the optimal bands combined of HBPSODE-SVM algorithm is best more intuitively by figure-6-(a) ~ (d), HBPSO-SVM nicety of grading is better than BPSO-SVM nicety of grading, and HBDE-SVM classification is better than HBPSO-SVM again.

Claims (5)

1. a target in hyperspectral remotely sensed image band selection method, it is characterized in that: first pre-service is carried out to original target in hyperspectral remotely sensed image, by two population at individual and algorithm parameter initialization, then application mix scale-of-two population differential evolution (HBPSODE) method, allow two population parallel iteration by transmitting optimum solution information between population, and utilize SVM classifier to calculate nicety of grading as fitness value, upgrade and evolve until reaching regulation evolution number of times or till reaching maximal accuracy.

2. target in hyperspectral remotely sensed image band selection method as claimed in claim 1, is characterized in that: in population at individual and algorithm parameter initialization procedure, and take two population parallel search strategy, initialization 2 populations, two population parallel iteration is evolved.After each iteration terminates, share between population and search optimum solution separately, realize the interchange of information, allow the optimum solution of population when next iteration is evolved not only with reference to its population also can consider the optimum solution of the other side population, guide population to depart from locally optimal solution.

3. target in hyperspectral remotely sensed image band selection method as claimed in claim 2, is characterized in that: make corresponding amendment to PSO algorithm and DE algorithm, proposes a kind of scale-of-two differential evolution algorithm of hybrid coding, can be extended in discrete domain; First adjuvant search space S '=[-a, a] is defined ^d, a is positive integer, solution space S={0,1} ^d, d is the dimension of problem; Then tie up real number vector X by adjuvant search space D and add solution space scale-of-two D dimensional vector B i.e. (X, B) as the hybrid coding representation of individual (or variant); Real number vector X still performs mutation operation and interlace operation according to differential evolution algorithm, and before operation is selected in execution, needing is developed real number vector X by epimorphism is mapped to binary vector B, and epimorphism evolution mapping function defines:

Wherein, h _ij(t+1) be each component value of variant after interlace operation, for ambiguity function, b _ij(t+1) be each component value of binary vector B, Dynamic gene μ can control b _ij(t+1) be set to the probability size of 1, get μ=0.5.

If (X _i(t), B _i(t)) and (X _i(t+1), B _i(t+1)) represent that t generation of population and t+1 are for individual i respectively, (H _i(t+1), E _i(t+1)) represent the variant of t+1 for individual i, f (x) represents fitness function.New selection behaviour is defined as follows:

4. EO-1 hyperion band selection method as claimed in claim 3, is characterized in that, describedly carries out pre-service to original target in hyperspectral remotely sensed image, two population at individual and algorithm parameter initialization is comprised:

Original target in hyperspectral remotely sensed image pre-service: reject jammr band, preliminary election type of ground objects, and dimension D and algorithm greatest iteration evolution number of times MaxDT that search volume is set;

Initialization is by the population Ppso of HBPSO algorithm evolution and correlation parameter: arranging population number is Np, arranges Studying factors c ₁, Studying factors c ₂, maximum inertia weight coefficient w _max, minimum inertia weight coefficient w _mindeng; In order to improve the performance of particle cluster algorithm, wherein inertia weight w upgrades according to following formula, and i represents i-th iteration;

$w=w_{\max }-\frac{w_{\max }-w_{\min }}{\mathrm{MaxDT}}\·i---\left(1\right)$

Initialization is by the population Pde of HBDE algorithm evolution and correlation parameter: arranging population number is Nd, zoom factor F, Crossbreeding parameters CR etc.; In order to improve differential evolution (DE) algorithm performance, wherein zoom factor F upgrades according to following formula, and F0 is a constant, and i represents i-th iteration;

$F=F0\·2^{\exp \left(\frac{1-\mathrm{MaxDT}}{\mathrm{MaxDT}+1-i}\right)}---\left(2\right).$

5. EO-1 hyperion band selection method as claimed in claim 3, it is characterized in that, application mix scale-of-two population Differential evolution, allow two population parallel iteration by transmitting optimum solution information between population, and utilize SVM classifier to calculate nicety of grading as fitness value, upgrade and evolve until reaching regulation evolution number of times or till reaching maximal accuracy, being specially:

Evolution iteration count t=0 is set.

Ppso population carries out a position according to HBPSO algorithm and speed upgrades, and utilizes SVM classifier to implement classification to the band combination after renewal, and calculates nicety of grading as fitness value, record t for optimal adaptation angle value and band combination;

Pde population makes a variation to all individualities according to HBDE algorithm, intersect, select operation; Utilize SVM classifier to calculate fitness value, record t for optimal adaptation angle value and band combination;

Relatively Ppso and Pde t is for the optimal adaptation angle value chosen, and adjusts the optimum solution of respective population;

Upgrade evolutionary generation counter t=t+1; If evolutionary generation reaches maximum evolution number of times or meets accuracy requirement.