CN104021393A - Hyperspectral remote sensing image waveband selection method based on firefly optimization - Google Patents

Abstract

The invention discloses a hyperspectral remote sensing image waveband selection algorithm based on improving firefly algorithm, and object functions in the FA algorithm are improved. Optimization and improvement of waveband selection is characterized by carrying out random initialization on waveband index position, position matrix size being s=n*b (n being known parameters, b being user-input waveband selection number); selecting different spectrum type distance function as the object function, substituting the obtained initial position matrix into the object function for calculation, and obtaining a group of one-dimensional array corresponding to the fluorescence brightness values of fireflies; carrying out ranking (disadvantaged point approaching to advantaged point) according to the advantages and disadvantages of the brightness values, that is, the value of the object function value; updating the waveband which is subjected to feature selection, that is, the position information of the fireflies after movement; and recording the waveband selection results when according with maximum iterations or searching precision. According to the hyperspectral remote sensing image waveband selection method based on firefly optimization, the problems that a conventional hyperspectral remote sensing image waveband selection algorithm is not high in precision and time-consuming can be solved; and the method has the advantages of being good in waveband selection effect, and wide in adaptation and the like.

Description

The target in hyperspectral remotely sensed image band selection method of optimizing based on firefly

Technical field

The present invention relates to a kind of target in hyperspectral remotely sensed image band selection method of optimizing based on firefly, belong to high-spectrum remote sensing processing technology field.

Background technology

High-spectrum remote-sensing is also imaging spectroscopy (Imaging Spectroscopy), refer to and utilize a lot of narrow electromagnetic wave bands to obtain the technology of object relevant data, being one of major technological breakthrough of obtaining aspect earth observation of last 20 years of 20th century mankind, is also the remote sensing cutting edge technology in current and decades from now on.It utilizes the nano level spectral resolution of imaging spectrometer, obtains many very narrow and view data that spectrum is continuous, realizes synchronously obtaining of ground object space, radiation, spectral information; Thereby for each pixel provides tens of spectral informations to hundreds of narrow wave bands (waveband width is less than 10nm), generate a complete and continuous curve of spectrum.Since entering 21 century, high spectrum resolution remote sensing technique has been obtained major progress, is accompanied by the solution of a series of basic problems, and high-spectrum remote-sensing progressively turns to practical stage by the experimental study stage.And the emphasis of applying in this focus as high-spectrum remote-sensing is exactly the expansion of the raising of airborne-remote sensing treatment effeciency and the application being closely connected with it.

Band selection is one of important step of remote sensing images recognition and classification.At sample number, be not a lot of in the situation that, by a lot of features, classifying, is all unfavorable from the complexity calculated or performance.Therefore how research is compressed to low dimensional feature space high-dimensional feature space to be effectively treated as an important problem.In high-spectral data, each spectral band can be regarded a feature as, selects the process of some band subset that succeeding target is played a major role as image classification to be called band selection.By band selection, can from the Hyperspectral imaging of magnanimity, remove redundancy or noise wave band, thereby reduce the complexity of algorithm and improve the accuracy of classifying.

In general, select the principle of best band to have 3 points: the one, selected band class information amount should be maximum; The 2nd, the correlativity between selected wave band data is little; The 3rd, in study area, the spectral response feature of wish identification atural object can make the most easily to distinguish between some classification atural object.Therefore, those information contents are many, correlativity is little, object spectrum difference is large, wave band that separability is good is exactly the best band that select.Carried out in this respect at present serial research both at home and abroad, in multispectral application in early days, people have recognized that different spectral bands has diagnostic to different atural object, and by information divergence (Divergence), conversion divergence (Transformed Divergence), JM(Jeffreys-Matusita) distance and Ma Shi (Bahattacharyya) distance wait and be used for multispectral band selection; In addition, mutual information (Mutual Information) algorithm is also applied to the selection of the optimum wave band of TM.In recent years, development along with high-spectrum remote-sensing, not only above algorithm has expanded to high spectrum field, and some new algorithms also propose successively, as the algorithm based on statistic: entropy and combination entropy, the optimum index factor (OIF), band index (Band Index), Spectra Derivative etc., but these algorithms adopt a statistic to measure wave band with respect to the importance of follow-up classification substantially, can not eliminate the noise information being attached in data.Therefore some more complicated algorithms have been subject to attention, as the MVPCA based on PCA and Noise estimation and MSNRPCA algorithm, linear restriction minimum covariance (LCMV-BCC/BCM) scheduling algorithm based on least energy constraint.

More than research makes the treatment effeciency of high-spectral data and range of application obtain great expansion.Yet the general efficiency of current most high-spectrum remote-sensing band selection algorithm is lower, this is because on the one hand algorithm itself is perfect not, calculation of complex, processing time are long; Be that high-spectrum remote sensing data amount is huge on the other hand, wave band number is many.Therefore, develop the simple method of some relative fast operatings and carry out band selection.

Summary of the invention

Goal of the invention: the problem and shortage existing for above-mentioned prior art, the object of this invention is to provide a kind of target in hyperspectral remotely sensed image band selection method of optimizing based on firefly, in the band selection of solution high-spectrum remote-sensing, precision is not high, time-consuming longer problem.

Technical scheme: a kind of target in hyperspectral remotely sensed image band selection method of optimizing based on firefly, comprises the following steps:

Step 1, selection need be carried out the Hyperspectral imaging S of dimensionality reduction;

Step 2, carries out parameter setting, maximum iteration time MaxGeneration=500, step factor α=0.5, light intensity absorption coefficient γ=1, maximum Attraction Degree β ₀=1;

Step 3, random initializtion firefly position, location matrix size is s, s=n*b, parameter n is namely known experimental group number of firefly number, b is the band selection number of user's input;

Step 4, arranges the number of times of iterative loop, for each iteration of FA, carries out following steps:

(a) for the concrete wave band of selecting, determine objective function;

(b) for location matrix s definite in step 3, the selected objective function going out of its substitution step (a) is calculated, obtaining numerical value is its corresponding firefly brightness value;

(c), according to numerical values recited, the brightness value of firefly is sorted;

(d) utilize the more all wave band positions of having selected of new formula renewal, firefly position;

(e) repeated execution of steps a)-d), until FA meets iterated conditional, s (1) exports as optimum wave band.

Further, in described step 4, objective function adopts Jeffries-Matusita distance as measure function:

$J_{i,j}=2(1-e^{-B_{i,j}})$

Wherein

$B_{i,j}=\frac{1}{8}{({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T{\left(\frac{{\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j}{2}\right)}^{-1}({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)+\frac{1}{2}\ln \left(\frac{|({\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j)/2|}{{\left|{\mathrm{\Σ}}_i\right|}^{\frac{1}{2}}{\left|{\mathrm{\Σ}}_j\right|}^{\frac{1}{2}}}\right)$

In formula, i, j is classification order label, J _i, _jfor classification ω i and ω _jbetween mahalanobis distance; μ _iand μ _jbe respectively the average of i class and j class; ∑ _iand ∑ _jbe respectively the variance matrix of i class and j class.While calculating classification average and class variance in this measure function, need to use a large amount of ground actual value sample datas.

Objective function in described step 4 adopts conversion divergence (Transformed Divergence) as measure function:

$D(x_i,x_j)=\frac{1}{2}\mathrm{tr}\left[({\mathrm{\σ}}_i-{\mathrm{\σ}}_j)({\mathrm{\σ}}_j^{-1}-{\mathrm{\σ}}_i^{-1})\right]+\frac{1}{2}\mathrm{tr}\left[({\mathrm{\σ}}_j^{-1}+{\mathrm{\σ}}_i^{-1})({\mathrm{\μ}}_i-{\mathrm{\μ}}_j){({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T\right]$

In formula, i, j is class label, tr[] and be matrix trace, σ _iand σ _jbe the variance of i class and j class, μ _iand μ _jit is the average of i class and j class.First, formula the right is variance poor of i class and j class, represents the distributional difference of two classes; Second is the normalization distance between two classes.

Further, the wave band position in described step (d) more new formula by following formula, realized respectively:

$x_{\mathrm{new}}=x_i+\mathrm{\β}\×(x_i-x_j)+\mathrm{\α}(\mathrm{rand}-\frac{1}{2})$

In formula, x _newfor the firefly position after upgrading is current band group, x _iand x _jfor the i that selected only and j firefly namely i combine j group wave band, α and β are the constants on [0,1], α is step factor, the Attraction Degree that β is firefly; Rand is the equally distributed random factor of ［ 0,1 ］ upper obedience.In order to strengthen region of search, avoid being absorbed in too early local optimum, in position updating process, increased disturbance term α * (rand-1/2).

Beneficial effect: the present invention is directed to the feature of high-spectrum remote sensing data, improved the objective function of FA algorithm, adopt JM distance and conversion divergence TD as the similarity measure of objective function.Method dimensionality reduction of the present invention is effective, and the dimensionality reduction characteristic obtaining contains maximum quantity of information, and its follow-up classification is consuming time is significantly less than similar dimension reduction method.Method of the present invention has the features such as band selection is effective, wide adaptability.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the target in hyperspectral remotely sensed image band selection method of optimizing based on firefly of the present invention;

Fig. 2 (a) and Fig. 2 (b) are respectively and adopt DC Mall data, and FA band selection adopts respectively the nicety of grading result curve that JM and TD are objective function;

Fig. 3 (a) and Fig. 3 (b) are respectively and adopt DC Mall data, the nicety of grading case line chart of FA-JM and FA-TD band selection method.

Embodiment

Below in conjunction with specific embodiment, further illustrate the present invention, should understand these embodiment is only not used in and limits the scope of the invention for the present invention is described, after having read the present invention, those skilled in the art all fall within the application's claims limited range to the modification of the various equivalent form of values of the present invention.

Thinking of the present invention is: the problems such as, redundance large for target in hyperspectral remotely sensed image data volume is large, difficult treatment and existing dimension reduction method efficiency is not high, calculation of complex, introducing bionic is theoretical, simple to operate according to Swarm Intelligence Algorithm, be easy to the features such as parallel processing and strong robustness, based on Spectral divisibility criterion, on the basis of improving firefly algorithm, build band selection model, the band selection method of research based on improving firefly algorithm, realizes the dimensionality reduction of the high-spectrum remote sensing of easy operating relatively fast.The method that the present invention realizes has not only been expanded the range of application of optimized algorithm, and for advancing the research of Hyperspectral imaging dimensionality reduction aspect to there is most important theories and realistic meaning.

The Hyperspectral imaging of the Washington D.C. that embodiment: experimental data is by HYDICE(Hyperspectral Digital Imagery Collection Experiment) sensor obtains.This data cover from 210 wave bands of 0.4 to 2.5um spectrum range, its spatial resolution is about 2.8m; Rejected after water absorption bands and noise wave band, retained 191 wave bands for data analysis.The number of sub images of experimental data for cutting from DC Mall raw video.Wherein size of data is 266 * 304, comprises 7 classifications such as road (Road), meadow (Grass), water body (Water), path (Trail), trees (Tree), shade (Shadow) and building (Roof).

As shown in Figure 1, specific implementation step is:

(1) original target in hyperspectral remotely sensed image data are carried out to data pre-service, remove noise wave band, then determine exemplary spectrum data and training sample data.A random initializtion N firefly, makes N=30;

(2) random initializtion N group is with the band combination of number, and combinatorial matrix size is s, s=n*b, and parameter n is known experimental group number, b is the band selection number of user's input;

(3) number of times of iterative loop is set, for each iteration of FA, determines objective function; It is firefly brightness value that initial position matrix s substitution objective function is calculated to one group of one-dimension array, and the size of comparison brightness value defines inferior position point and shifts to advantageous point, upgrades the brightness value of location matrix and wave band.When meeting search precision or reaching maximum search number of times, carry out next step; Otherwise searching times increases by 1, searches for next time.

(4) carry out circulation experiment, the optimum wave band that the each experiment of output is selected.

(5) obtain selected wave band, the dimensionality reduction data that obtain are classified, adopt classification overall accuracy to assess the performance of dimension-reduction algorithm.

Wherein, the objective function in step 3 can adopt Jeffries-Matusita distance as measure function:

$J_{i,j}=2(1-e^{-B_{i,j}})$

Wherein

$B_{i,j}=\frac{1}{8}{({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T{\left(\frac{{\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j}{2}\right)}^{-1}({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)+\frac{1}{2}\ln \left(\frac{|({\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j)/2|}{{\left|{\mathrm{\Σ}}_i\right|}^{\frac{1}{2}}{\left|{\mathrm{\Σ}}_j\right|}^{\frac{1}{2}}}\right)$

In formula, i, j is classification order label, J _i,jfor classification ω i and ω _jbetween mahalanobis distance; μ _iand μ _jbe respectively the average of i class and j class; ∑ _iand ∑ _jbe respectively the variance matrix of i class and j class.While calculating classification average and class variance in this measure function, need to use a large amount of ground actual value sample datas.

Wherein, the objective function in step 3 adopts conversion divergence (Transformed Divergence) as measure function:

$D(x_i,x_j)=\frac{1}{2}\mathrm{tr}\left[({\mathrm{\σ}}_i-{\mathrm{\σ}}_j)({\mathrm{\σ}}_j^{-1}-{\mathrm{\σ}}_i^{-1})\right]+\frac{1}{2}\mathrm{tr}\left[({\mathrm{\σ}}_j^{-1}+{\mathrm{\σ}}_i^{-1})({\mathrm{\μ}}_i-{\mathrm{\μ}}_j){({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T\right]$

In formula, i, j is class label, tr[] and be matrix trace, σ _iand σ _jbe the variance of i class and j class, μ _iand μ _jit is the average of i class and j class.First, formula the right is variance poor of i class and j class, represents the distributional difference of two classes; Second is the normalization distance between two classes.

Wherein, the wave band position in step 3 more new formula by following formula, realized respectively:

$x_{\mathrm{new}}=x_i+\mathrm{\β}\×(x_i-x_j)+\mathrm{\α}(\mathrm{rand}-\frac{1}{2})$

In formula, x _newfor the firefly position after upgrading is current band group, x _iand x _jfor the i that selected only and j firefly namely i combine j group wave band, α and β are the constants on [0,1], α is step factor, the Attraction Degree that β is firefly; Rand is the equally distributed random factor of ［ 0,1 ］ upper obedience.In order to strengthen region of search, avoid being absorbed in too early local optimum, in position updating process, increased disturbance term α * (rand-1/2).

Adopt band selection method of the present invention to carry out experimental analysis to target in hyperspectral remotely sensed image data, and compare with similar band selection method, control methods mainly contains raw data (all wave bands), PCA and SFS, two kinds of different search optimization methods of SFFS contrast.For the use of DC mall data, to the classification results of remote sensing image, as shown in Fig. 2 (a) and Fig. 2 (b), visible FA method provides best classification performance.In addition, Fig. 3 (a) has provided the case line chart comparing result of FA band selection method nicety of grading to Fig. 3 (b).Table 1 is FA algorithm and the working time comparing result of PSO algorithm under the identical wave band number condition of selection, and visible FA algorithm has been saved the time of band selection greatly.To sum up, the performance of the inventive method is better than other similar approach.

Table 1

Claims (4)

1. a target in hyperspectral remotely sensed image band selection method of optimizing based on firefly, is characterized in that, comprises the following steps:

Step 1, selection need be carried out the Hyperspectral imaging S of dimensionality reduction;

Step 2, carries out parameter setting, maximum iteration time MaxGeneration=500, step factor α=0.5, light intensity absorption coefficient γ=1, maximum Attraction Degree β ₀=1;

Step 3, random initializtion firefly position, location matrix size is s, s=n*b, parameter n is namely known experimental group number of firefly number, b is the band selection number of user's input;

Step 4, arranges the number of times of iterative loop, for each iteration of FA, carries out following steps:

(a) for the concrete wave band of selecting, determine objective function;

(b) for location matrix s definite in step 3, will in the selected objective function going out of its substitution step (a), calculate, obtaining numerical value is its corresponding firefly brightness value;

(c), according to numerical values recited, the brightness value of firefly is sorted;

(d) utilize the more all wave band positions of having selected of new formula renewal, firefly position;

(e) repeated execution of steps a)-d), until FA meets iterated conditional, s (1) exports as optimum wave band.

2. the target in hyperspectral remotely sensed image adaptive band selection method based on particle group optimizing according to claim 1, is characterized in that, in described step 4, objective function adopts Jeffries-Matusita distance as measure function:

$J_{i,j}=2(1-e^{-B_{i,j}})$

Wherein

$B_{i,j}=\frac{1}{8}{({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T{\left(\frac{{\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j}{2}\right)}^{-1}({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)+\frac{1}{2}\ln \left(\frac{|({\mathrm{\Σ}}_i+{\mathrm{\Σ}}_j)/2|}{{\left|{\mathrm{\Σ}}_i\right|}^{\frac{1}{2}}{\left|{\mathrm{\Σ}}_j\right|}^{\frac{1}{2}}}\right)$

In formula, i, j is classification order label, J _i, _jfor classification ω i and ω _jbetween mahalanobis distance; μ _iand μ _jbe respectively the average of i class and j class; ∑ _iand ∑ _jbe respectively the variance matrix of i class and j class; While calculating classification average and class variance in this measure function, need to use a large amount of ground actual value sample datas.

3. the target in hyperspectral remotely sensed image adaptive band selection method based on particle group optimizing according to claim 1, is characterized in that, the objective function in described step 4 adopts conversion divergence as measure function:

$D(x_i,x_j)=\frac{1}{2}\mathrm{tr}\left[({\mathrm{\σ}}_i-{\mathrm{\σ}}_j)({\mathrm{\σ}}_j^{-1}-{\mathrm{\σ}}_i^{-1})\right]+\frac{1}{2}\mathrm{tr}\left[({\mathrm{\σ}}_j^{-1}+{\mathrm{\σ}}_i^{-1})({\mathrm{\μ}}_i-{\mathrm{\μ}}_j){({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T\right]$

In formula, i, j is class label, tr[] and be matrix trace, σ _iand σ _jbe the variance of i class and j class, μ _iand μ _jit is the average of i class and j class.First, formula the right is variance poor of i class and j class, represents the distributional difference of two classes; Second is the normalization distance between two classes.

4. the target in hyperspectral remotely sensed image adaptive band selection method based on particle group optimizing according to claim 1, is characterized in that, the wave band position in described step (d) more new formula is realized by following formula respectively:

$x_{\mathrm{new}}=x_i+\mathrm{\β}\×(x_i-x_j)+\mathrm{\α}(\mathrm{rand}-\frac{1}{2})$

In formula, x _newfor the firefly position after upgrading is current band group, x _iand x _jfor the i that selected only and j firefly namely i combine j group wave band, α and β are the constants on [0,1], α is step factor, the Attraction Degree that β is firefly; Rand is the equally distributed random factor of ［ 0,1 ］ upper obedience; In order to strengthen region of search, avoid being absorbed in too early local optimum, in position updating process, increased disturbance term α * (rand-1/2).