A kind of feature selection approach based on the search of Memorability multiple point crossover gravitation
Technical field
The invention belongs to Remote Sensing Image Processing Technology fields, more particularly to a kind of Memorability multiple point crossover gravitation that is based on to search for
Feature selection approach.
Background technique
With the development of sensor technology, target in hyperspectral remotely sensed image after the 1980s, start environmental monitoring,
The fields such as survey of territorial resources assessment, urban planning play an important role.In the process, the nicety of grading of remote sensing image is straight
Connect the effectiveness for determining its application.Although the description that target in hyperspectral remotely sensed image spectral signature abundant can be more accurate
Face coverage information, but the spectral information of this magnanimity usually has very high redundancy, will cause in assorting process
Hughes phenomenon, and expend a large amount of storage and calculate space.Thus, how to be identified from high-dimensional spectral band most effective
A small amount of wave band, and realize the image classification of degree of precision, become a critical issue of Hyperspectral imaging application.
Method for Feature Selection is a kind of directly effectively wave band recognition methods.This method is according to certain objective function or comments
Valence criterion selects wave band as few as possible to form new character subset, appoints it to the analysis of the data such as such as image segmentation, classification
Business can achieve and quite or better effect before feature selecting.This back-and-forth method can be classified as the optimization of binary space
Algorithm.The research of early stage, exhaustive search algorithm achieve certain success, but the computation complexity of the algorithm, with iteration
It carries out in exponential increase.In practical applications, time-consuming too long due to calculating, it is difficult to meet the requirement of target in hyperspectral remotely sensed image.For
Reduction computation complexity, the art teaches branch-and-bound, sequential advancement, sequences to retreat scheduling algorithm, but these algorithms need
The wave band number for the predesignated subset that very important person is, so that obtained character subset is not usually optimal subset.
Gravitation search algorithm is that Rashedi E. et al. was inspired in 2009 by Newton's law of gravitation, one kind of proposition
Novel meta-heuristic searching algorithm.The algorithm thinks, attract each other between object among universe by gravitation, it is excellent
Change in algorithm it is considered that being attracted each other between each particle by gravitation in population.Due to gravitational size direct ratio
In the quality of two objects, be inversely proportional to the distance between object, the bigger object of quality to the attractions of other objects also more
Greatly.So if assigning the best particle of fitness in optimization algorithm to maximum quality, remaining particle will comply with newton second
Law is close to the best particle of fitness with certain acceleration, completes searching process.
In the past few years, which has successfully been applied to clustering, PID (Proportion
Integration Differentiation) parameter optimization, Economic Load Dispatch, pipeline schedule, multi-Level Threshold Image Segmentation etc.
In numerous optimization problems, series of advantages and the wide application prospects such as it is easy to apply, global optimizing ability is strong are shown.Phase
Than solving the problems, such as feature selecting using gravitation search algorithm in traditional the methods of exhaustive search, it is not necessary to wave band number is set, it can
With according to image feature it is adaptive obtain optimal feature subset, realize the feature selecting of greater efficiency and precision.
But the learning strategy of gravitation search algorithm and information interchange mode, it there is also some problems, such as: mode of learning
Single, the mode of all particles and other particle exchange of information is all by gravitation, this causes the development ability of algorithm
Deficiency, later period are easy to appear concussion problem, cause convergence rate slack-off;In addition, particle completes an iteration every time, searching before
Rope information is all lost, and algorithm does not have the ability to oneself empirical learning, and algorithm is caused to be easily trapped into local optimum, is occurred early
The problems such as ripe convergence.In order to more efficiently solve the problems, such as waveband selection using gravitation search algorithm, need to carry out algorithm
It improves and improves.
Summary of the invention
Deficiency present in demand and gravitation search algorithm for characteristics of remote sensing image selection, it is an object of the invention to
A kind of feature selection approach based on the search of Memorability multiple point crossover gravitation is provided, optimization is explored and developed by adaptive equalization
Character subset evaluation function, and make a variation to global history is optimal, it is insufficient, scarce to be effectively improved gravitation search algorithm development ability
Weary Memorability is easy the problem of Premature Convergence, finally obtains that wave band number is few and the optimal spectrum of available stable class result
Character subset.
To solve the above problems, the present invention adopts the following technical scheme that:
It is a kind of based on Memorability multiple point crossover gravitation search feature selection approach the following steps are included:
Step S1, original remote sensing image HI of the selection to feature selecting, determines total wave band number D of remote sensing image HI, simultaneously
According to atural object practical on original remote sensing image HI distribution and classification number, determines sample areas and extract sample data SHI;
Step S2, initialization process: setting Population Size sets number of particles as N, each particle represents to be selected
Character subset, the position X of each particle in initialization populationi(t) and its corresponding speed, 1≤i≤N, and i are positive integer, t
For current iteration number;
Step S3, start iteration: being based on support vector machines (Support Vector Machine, SVM) classifier, utilize
Each particle Xi(t) character subset represented classifies to sample data SHI, obtains the corresponding nicety of grading of N number of particle
Accuracy=[Acc1(t),Acc2(t),...,Acci(t),...,AccN(t)], wherein Acci(t) ith feature subset
Nicety of grading;
Step S4, the fitness value fit of each particle is calculatedi(t),Wherein, ω is
The weight of balanced sort precision and wave band number, nsThe dimension for being 1 for particle i intermediate value;
Step S5, based on the fitness value fit of each particlei(t) the individual history optimal location Pb of each particle is determinedi
(t) with the global history optimal location Gb (t) of population;
Step S6, the fitness value fit based on each particlei(t) the gravitational acceleration a of each particle is calculatedi(t);
Step S7, the Evolving State for judging population, the gravitational acceleration a based on each particlei(t), described each
The individual history optimal location Pb of particleiIt (t) is each grain according to Evolving State with the global history optimal location Gb (t) of population
Son selects most suitable information exchange method, is updated to particle rapidity and position;
Step S8, the corresponding character subset of N number of particle is calculated using SVM classifier according to the position of particle after update
Nicety of grading Acci(t+1), the fitness value that each particle is then updated according to evaluation function, obtains fiti(t+1);
Step S9, it is based on fiti(t+1) the optimal position of global history of the individual history optimal location of more new particle and population
It sets;
Step S10, t=t+1 is enabled, judges whether t is greater than maximum number of iterations TmaxIf t is not more than Tmax, it is based on step
Fit obtained in S8i(t+1) and the global history of individual the history optimal location and population of more new particle obtained in step S9
Optimal location carries out the t+1 times interative computation according to step S6 to step S9;If t is greater than Tmax, by fitness in current population
Best particle output, exports the global history optimal location of Population Regeneration described in step S9, as sample data SHI is most
Excellent character subset;
Step S11, the spectral band for including according to optimal feature subset, using SVM classifier to original remote sensing image HI
Classify, obtains its classification results.
Preferably, step S5 specifically: as t=1, initialize the individual history optimal location Pb of each particlei(t)
=Xi(t), then from Pbi(t) global history optimal location Gb (t) of the maximum particle of fitness as population is selected in;Work as t >
When 1, particle position more each first updates the fitness value of front and back, if the position that new position fitness is bigger, new
It is optimal as individual history;Then from new individual history it is optimal in, the global history before selecting best particle and update is most
It is excellent to make comparisons, if the particle fitness is larger, it is assigned to the global history optimal location of new population.
Preferably, step S7 specifically:
Whether the standard deviation std (fit) and random number rand for comparing particle fitness value in current population meet std
(fit) < rand and compare crossing-over rate pcMultiplied by t/TmaxValue and random number rand whether meet pc×t/Tmax> rand,
If two conditions are not satisfied, then it is assumed that search process is in exploration phase, particle i renewal speed are as follows:
Wherein,For particle i in the t times iteration d dimension renewal speed,For grain in the t times iteration
The speed of son,For particle i d dimension gravitational acceleration;
If two conditions have one or two to be met, then it is assumed that search process enters the development phase, for every
One dimension d will first have to judge aberration rate pmWhether meet p with random number randm> rand, if conditions are not met, Gb (t) is kept
It is constant;If it is satisfied, then to the d dimension Gb of Gb (t)d(t) Gaussian mutation is carried out, i.e.,
It is then based on the speed of each dimension of multiple point crossover policy update particle of Memorability:
Wherein,For particle i in the t times iteration d dimension renewal speed,For in the t times iteration
The optimal position in d dimension of the history of particle i,It is particle i in the t times iteration in the position of d dimension, c '=c+
2t/Tmax, maximum number of iterations Tmax。
Particle position is updated based on new speed, particle i updates position are as follows:
Wherein,It is particle i in the t times iteration in the update position of d dimension.
Beneficial effects of the present invention: the problems such as present invention is high, processing is time-consuming for target in hyperspectral remotely sensed image data redudancy,
Based on Memorability multiple point crossover gravitation search algorithm, each particle is set as to the alternative solution an of optimal feature subset, is passed through
Band subset evaluation function assesses the quality alternatively solved, and particle is guided to carry out information interchange, completes fast convergence.It is best in quality
The corresponding disaggregation of particle be optimal spectral signature subset.The present invention can select that wave band number is few and available stable class
As a result optimal spectral signature subset, to solve to calculate complicated, classification caused by target in hyperspectral remotely sensed image data redudancy is high
The problems such as precision is low.
Detailed description of the invention
Fig. 1 is that the present invention is based on the overall construction drawings of the multiple point crossover gravitation searching method of Memorability;
Fig. 2 is the flow diagram of the multiple point crossover gravitation searching method based on Memorability;
Fig. 3 is image feature and binary-coded schematic diagram.
Specific embodiment
The specific embodiment of the method for the present invention is described further with reference to the accompanying drawing.
Feature selection approach provided by the invention based on the search of Memorability multiple point crossover gravitation, is suitable for AVIRIS
The area Indian Pine that (Airborne Visible Infrared Imaging Spectrometer) sensor obtains is high
The processing of spectral remote sensing image.The spatial resolution and spectral resolution of original AVIRIS data be respectively 10m (rice) and
10nm (nanometer), spectral range cover 400-2400nm, share 224 spectral coverages.Since the value of wherein 4 wave bands is 0, so
It is filtered out.In addition, have 20 wave bands on Indian Pine image, including wave band [104-108], [150-163] and 220
Value is easy to be influenced by water suction wave band, these wave bands are also filtered out in present case.So the Indian that present case uses
The wave band that Pine image uses altogether shares 200, and image size is 145 × 145 (pixels), and type of ground objects can recognize including 16 classes
Atural object (mainly different crops), rest of pixels are considered as background.
As shown in Figure 1, 2, the present invention provides a kind of feature selection approach packet based on the search of Memorability multiple point crossover gravitation
Include following steps:
Step S1, original remote sensing image HI of the selection to feature selecting, determines total wave band number D of remote sensing image HI, wave band
For (b1,b2,b3,...,bd,...bD), 1≤d≤D, and d be positive integer, D=200, at the same according to original remote sensing image HI on reality
The distribution of border atural object and classification number, determine sample areas and extract sample data SHI, wherein 5% picture of the random every class of selection
Member is sample data SHI, and remaining data are as test set.
Step S2, initialization process: setting Population Size sets number of particles as N, preferably N=30, each particle generation
One alternative features subset of table, each particle using two attribute of position and speed indicate, by binary-coded method with
The position of each particle in machine initialization population
And its corresponding speed1≤d≤D, 1≤i≤
N, and d, i are positive integer, t is the number of iterations, 1≤t≤Tmax, and t is positive integer, each particle position initial value is that each particle rapidity initial value is Vi(1)=0;As shown in Fig. 2, each particle is at the beginning of the position of d dimension
Initial value is set as 0 or 1 at random, is not selected if the d wave band of 0 expression sample data SHI, if 1 expression d
Wave band is selected, and selected d wave band is known as d dimension;Set maximum number of iterations TmaxIt is 200, crossing-over rate pcFor
0.05, aberration rate pmIt is 0.5, initialization coordinating factor c is 0.5.
Step S3, start iteration: being based on support vector machines (Support Vector Machine, SVM) classifier, utilize
Each particle Xi(t) character subset represented classifies to sample data SHI, obtains the corresponding nicety of grading of N number of particle
Accuracy=[Acc1(t),Acc2(t),...,Acci(t),...,AccN(t)], wherein AcciPoint of ith feature subset
Class precision.
Step S4, the fitness value fit of each particle is calculatedi(t), fitness value fitiIt (t) is evaluating characteristic subset matter
The evaluation function of amount,Wherein, ω is the weight of a balanced sort precision and wave band number,
It is preferred that ω is 0.5, nsThe dimension for being 1 for particle i intermediate value, i.e. selected wave band number.
Step S5, the individual history optimal location Pb of each particle is determinedi(t) with the global history optimal location Gb of population
(t).As t=1, the individual history optimal location Pb of each particle is initializedi(t)=Xi(t), then from Pbi(t) selection in
Global history optimal location Gb (t) of the maximum particle of fitness as population.As t > 1, particle position more each first
The fitness value for updating front and back, if new position fitness is bigger, it is optimal that new position becomes individual history;Then from new
Individual history it is optimal in, select best particle with the global history before update is optimal makes comparisons, if the particle fitness
It is larger, then it is assigned to the global history optimal location of new population.
Step S6, the gravitational acceleration for calculating each particle, according to gravitation search algorithm, the calculation formula of gravitational acceleration
Are as follows:
Wherein,It is inhaled in d dimension (selected wave band) by other all particles for particle i in the t times iteration
Draw the resultant force of effect, MiIt (t) is the quality of particle i in the t times iteration,It is particle i in the t times iteration in d dimension
Gravitational acceleration, wherein
Wherein, G (t) is gravitational constant G (t)=100 × exp (- 20 × t/T in the t times iterationmax), Rij(t)
For the Euclidean distance in the t times iteration between particle i and particle j, ε is a minimum constant greater than 0, can be set to ε
=10-6, mjIt (t) is the particle fitness value after particle j normalization in the t times iteration, MjIt (t) is particle j in the t times iteration
Quality, miIt (t) is the particle fitness value after particle i normalization in the t times iteration.
Step S7, selection Information exchange mechanism adaptive according to the Evolving State of population.First determine whether the evolution of population
State is in exploration or development phase, is then the most suitable information exchange method of each particle selection according to Evolving State,
Particle rapidity and position are updated, realize the adaptive equalization to exploration and development.Also, in order to prevent Premature Convergence,
Mutation operation is executed with position of certain probability to global history optimal particle.
Judging the method for Evolving State is: (1) in more current population the standard deviation std (fit) of particle fitness value with
Whether random number rand meets std (fit) < rand and (2) compare crossing-over rate pcMultiplied by t/TmaxValue be with random number rand
It is no to meet pc×t/Tmax>rand。
If two conditions are not satisfied, then it is assumed that search process is in the exploration phase, and particle should be allowed adequately to carry out
Extensive region is explored in information interchange, so particle i renewal speed are as follows:
Wherein,For particle i in the t times iteration d dimension renewal speed,For grain in the t times iteration
Speed of the sub- i in d dimension.
If two conditions have one or two to be met, then it is assumed that search process enters the development phase.It is developing
Stage should carry out the fine search around preparation solution, preferably be solved.The generation of Premature Convergence in order to prevent, for every
One dimension d will first have to judge aberration rate pmWhether meet p with random number randm> rand, if conditions are not met, Gbd(t) it protects
Hold it is constant, if it is satisfied, then to Gbd(t) Gaussian mutation is carried out:
Wherein, Gbd(t) ' it is the global history optimal particle through Gaussian mutation in the t times iteration in the position of d dimension.
If obtained new position is better than original position, original position is replaced with new position, is then executed
The speed of each dimension of multiple point crossover policy update particle based on Memorability:
Wherein,For particle i in the t times iteration d dimension renewal speed,For in the t times iteration
The optimal position in d dimension of the history of particle i,It is particle i in the t times iteration in the position of d dimension, c '=c+
2t/Tmax, i.e., with iterations going on, c ' value is increasing, so that crossover operation tends to Gbd(t) position is conducive to accelerate
Convergence.
Step S8, it updates particle position: after completing speed update, particle position being updated based on new speed, grain
Sub- i updates position are as follows:
Wherein,It is particle i in the t times iteration in the update position of d dimension.
Step S9, the corresponding character subset of N number of particle is calculated using SVM classifier according to the position of particle after update
Nicety of grading Acci(t+1), the fitness value that each particle is then updated according to evaluation function, obtains fiti(t+1)。
Step S10, it is based on fiti(t+1) according to the individual history optimal location and kind of the more new particle of the method in step S5
The global history optimal location of group.
Step S11, t=t+1 is enabled, judges whether t is greater than maximum number of iterations Tmax;If t is not more than Tmax, it is based on step
Fit obtained in S9i(t+1) it is gone through with the obtained individual history optimal location of more new particle of step S10, the overall situation of Population Regeneration
History optimal location carries out the t+1 times interative computation according to step S6 to step S10;If t is greater than Tmax, will be fitted in current population
The best particle output of response, exports the global history optimal location of Population Regeneration described in step S10, as sample data
The optimal feature subset of SHI.At this moment, it is believed that this feature subset is few with wave band number and classification results are excellent.
Step S12, the spectral band for including according to optimal feature subset, using SVM classifier to original remote sensing image HI
Classify, obtains its classification results.
In order to verify the classifying quality of waveband selection, this implementation has counted the wave under 10 random training sets and test
Average nicety of grading and wave band number after section selection.Table 1 gives the method for the present invention and two kinds of improved gravitation search algorithms
GGSA(Adaptive gbest-guided gravitational search algorithm)、PSOGSA(Hybrid
Particle swarm optimization and gravitational search algorithm) it is used for Indian
Pine image spy carries out the wave band number obtained after sign selection and the general classification by svm classifier.As shown in table 1, it compares
In GGSA algorithm and PSOGSA algorithm, the method for the present invention can obtain more excellent, more stable nicety of grading, and required wave
Number of segment is also much smaller than two kinds of comparison algorithms.
1 feature selecting of table and classification results
It should be noted that described herein, the specific embodiments are only for explaining the present invention, is not intended to limit the present invention
Implementation and interest field, all technical solutions identical or equivalent with content described in the claims in the present invention should all be included in this
In invention protection scope.