CN101101234A - Independent ingredient analysis global search method for implementing high spectrum terrain classification - Google Patents

Abstract

The invention relates to independent component analysis global search method of realizing the high spectrum fine classification under no prior knowledge situation, the method including: reading in the high spectral data, the establishment of independent component analysis model based on the kurtosis, the center of the data, the ball of data, the iterative solution based on quantum genetic algorithm, independent component compositor, two value of the image, the feature classification. The invention method can established on the circumstance of no data background model, using the self high statistical data to achieve the feature fine unsupervised classification of the high spectral data; at the same time, avoided to plunge in the local best solution problem in the independent component analysis solution process, and compared with the traditional genetic algorithm, the invention used quantum genetic algorithm has less number of iterations, fast convergence, high search efficiency and the strong overall search capability and so on features.

Description

A kind of independent ingredient analysis global search method of realizing high spectrum terrain classification

Technical field

The present invention relates to the sensor information process field, a kind ofly specifically be applied to high-spectral data feature spy and get global search method with terrain classification, can effectively improve the terrain classification precision based on independent component analysis.

Background technology

Hyperspectral imager is a kind of novel remote sensing load, its spectrum has tight, continuous characteristics, can write down the spectrum and the spatial information feature of tested same atural object simultaneously, originally the material that can not survey in broadband remote sensing can be detected in high-spectrum remote-sensing, therefore, high-spectrum remote-sensing provides strong detection means for the sophisticated category of atural object.At present, use the high spectrum image data and carry out the method for terrain classification and mainly be divided into two big classes: have and supervise and do not have a supervised classification method, supervised classification method is can obtain preferable performance under there is the situation of reliable priori in hypothesis.Yet, because the actual spectrum data are easily affected by environment, have factors such as the spectra database true measurement in ground not well established and remote and the abominable area of natural conditions can realize hardly now, cause the priori spectral information to obtain to become the difficult point of target supervised classification algorithm.Therefore, no supervised classification becomes domestic and international many scholars' research focus

Independent component analysis (Independent Component Analysis, ICA) be a kind of analytical approach that just grew up in recent years based on signal higher order statistical characteristic, the purpose of this method is that observable data are carried out certain linear decomposition, it is resolved into add up independently composition.Just because of these characteristics, make ICA receive close concern in the signal Processing field,, it has also been had in many fields such as remote sensing very widely use in the increase of research interest aspect the ICA along with in recent years.At present, ICA research mainly roughly is divided into two big classes, based on the iteration method of estimation of information theory criterion with based on statistical algebraic manipulation method, in principle, these two kinds of methods all are to have utilized the independence and the non-Gauss (nongaussianity) of source signal.In based on information-theoretical method research, the various countries researcher has proposed a series of algorithm for estimating from minimum mutual information, maximum likelihood and maximization negentropy (negentropy) equal angles, and these algorithms are of equal value under given conditions in theory; Based in the statistical method mainly being the method for utilizing high-order semi-invariants such as second order semi-invariant (Second-order cumulant), fourth order cumulant.When utilizing above-mentioned various algorithm to adopt conventional numerical analysis method to find the solution, ubiquity is absorbed in the problem of locally optimal solution easily.At this problem, many scholars have also proposed the method that many neural learning algorithms combine with ICA, but the validity of algorithm directly has been subjected to depending on the influence of the neuronal activation function of signal probability distribution function influence.

(Quantum Genetic Algorithm QGA) is the product that quantum calculation combines with the evolutionary computation theory to quantum genetic algorithm, and it is a kind of probabilistic search algorithm based on quantal concept and principle.In QGA, adopt the chromosomal coded system of quantum bit, effect of utilization cervical orifice of uterus and quantum door update strategy are finished evolutionary search.With traditional genetic algorithm relatively, it is little and do not influence algorithm performance, have the strong characteristics of fast convergence rate and global optimizing ability concurrently simultaneously to have a population scale.

Summary of the invention

Be absorbed in locally optimal solution easily, be subjected to the influence of neuronal activation function based on the ICA of neural learning algorithm at the ICA algorithm, and high spectrum supervision terrain classification algorithm obtains problems such as priori difficulty, the present invention proposes and realize that high spectrum does not have the ICA method based on the improvement quantum genetic algorithm of supervision terrain classification, this method has overcome the precision that supervised classification method obtains the priori difficulty, effectively raises ICA ability of searching optimum and high spectrum terrain classification.

Technical solution of the present invention is: based on the independent composition analysis algorithm that improves quantum genetic algorithm to be core realization nothing supervision high spectrum terrain classification, quantum intersection and quantum variation have been increased in the quantum genetic algorithm, and with the fitness function of fourth order cumulant-kurtosis as quantum genetic algorithm, satisfying by data centerization and nodularization of constraint condition realizes.Concrete steps are as follows:

(1) reads the data X that observation obtains;

(2) foundation is based on the ICA model X=AS of kurtosis;

(3) X is carried out centralization and albefaction processing;

(4) initialization population and genetic algebra t=1, and chromosome in the population is carried out quantum bit encode;

(5) separate by the observation attitude generation scale-of-two of population behind the coding, obtain separation matrix, calculate and separate the back signal, and use fitness function that it is estimated, preserve optimum solution simultaneously;

(6) the method end condition is judged: if satisfy end condition, then method finishes, execution in step (10), and preserve current optimum solution, otherwise the continuation method;

(7) carry out the quantum interlace operation, quantum rotation door update strategy, quantum not gate mutation operation is to obtain new population;

(8) produce scale-of-two by the observation attitude of population and separate, obtain separation matrix, and calculate and separate the back signal, and carry out centralization and albefaction operation, and the utilization fitness function estimates it, preserve optimum solution simultaneously separating the back signal;

(9) genetic algebra increases by 1, execution in step (6);

(10) finishing iteration computing obtains separation matrix by optimum solution, and calculates separation signal.

(11) subsequent treatment of method mainly comprises: independent component ordering, image binaryzation, terrain classification.

Description of drawings

Fig. 1 realizes the process flow diagram of high spectrum terrain classification for the present invention.

Fig. 2 is for improving the concrete steps of quantum genetic operation.

Fig. 3 is a ground reference information, and reference identification C4 represents paddy rice, and V4 represents Ipomoea batatas, and V13 represents caraway.

Fig. 4 (b) is an Ipomoea batatas for the present invention realizes the result of high spectrum terrain classification wherein, (a) being paddy rice, (c) is caraway.。

Embodiment

As shown in Figure 1, specific implementation method of the present invention is as follows:

1. based on the ICA modelling of kurtosis

ICA is a kind of polytomy variable data analysing method that produces the statistics independent component, and when utilizing it to carry out terrain classification, each classification shows as independent component, has therefore obtained the separability maximization of classification.If observation signal X=[x ₁, x ₂..., x _m] ^TBe the random vector of m dimension, source signal S=[s ₁, s ₂..., s _n] ^TBe the independent random vector of n dimension, hybrid matrix A is that m * n ties up nonsingular matrix, and then their linear combination can be described as:

X＝AS (1)

(1) formula is called the ICA model.The essence of ICA is under source signal s and hybrid matrix A condition of unknown, utilizes the statistical property of source signal S according to known observation signal X, determines separation matrix W, makes Y=WX, and satisfies Y=[y ₁, y ₂..., y _n] ^TEach variable between separate as much as possible, then Y is exactly the estimation of S, simultaneously can determine hybrid matrix A=W ^-1

The main method of utilizing in estimating the ICA model has non-Gauss's maximization, has minimizing with maximum likelihood of information to estimate mutually.By the central limit theorem in the statistics opinion: one group of independent random variable and distribution more approach Gaussian distribution than any source signal, therefore, can be with non-Gauss as the separate tolerance of random signal.According to this theorem, adopt non-Gauss's maximization to estimate the ICA model among the present invention, utilize fourth order cumulant-kurtosis to carry out non-Gauss's tolerance, this mainly is because kurtosis theoretical and calculate all very simple, and is easy to realize.The kurtosis of stochastic variable y is defined as:

kurt(y)＝E{y ⁴}-3(E{y ²}) ² (2)

For the stochastic variable of a Gaussian distribution, its kurtosis equals " 0 ", but for most non-Gaussian random variable, kurtosis is not equal to " 0 ".Since kurtosis have just have negative, therefore, the absolute value of non-Gauss's measurement general using kurtosis or square, be worth for " 0 " be gaussian variable, greater than " 0 " is non-gaussian variable.

2. the centralization of data, nodularization

What adopt is that the non-Gauss of maximization estimates the ICA model, adopts the prerequisite assumed condition of kurtosis as non-Gauss's measure: the stochastic variable zero-mean, satisfy E{yy simultaneously ^T}=I.Therefore need carry out centralization and spheroidising for signal variable to each of pending signal variable and genetic manipulation, the purpose of centralization is to guarantee the stochastic variable zero-mean, and the signal after the centralization is obtained by the average that observation signal X deducts X; The purpose of nodularization is to guarantee that stochastic variable satisfies E{yy ^T}=I condition, (Principal Component Analysis, PCA) algorithm is realized by principal component analysis in spheroidising.

3. population quantum bit coding, initialization, fitness function is selected

(1) quantum bit coding

Information memory cell minimum in the quantum genetic algorithm is a quantum bit, and it is based on the probability encoding mode of quantum bit notion.A quantum bit chromosome is made up of a string quantum bit.A quantum bit can be represented one state, " 0 " state, perhaps the linear superposition state of two states, its state can be expressed as: | ＞=α | 0＞+β | 1＞, wherein, α, β is two plural numbers, and expression is in the probability amplitude of " 0 " state and one state respectively, | α | ², | β | ²Represent that respectively quantum bit is " 0 " state, the probability of one state, and satisfy normalizing condition: | α | ²+ | β | ²=1.

When a quantum bit chromosome length was m, it can be expressed as: $\left[\begin{array}{cccc}\begin{array}{c}{\mathrm{\α}}_1\\ {\mathrm{\β}}_1\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_2\\ {\mathrm{\β}}_2\end{array}\right|& \begin{array}{c}\·\·\·\\ \·\·\·\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_m\\ {\mathrm{\β}}_m\end{array}\right.\end{array}\right],$ Adopt the coded system of quantum bit, a quantum bit chromosome can be represented the linear superposition of a plurality of states, therefore with traditional genetic algorithm relatively, it has better population diversity, and it is little and do not influence algorithm performance to have a population scale.

Therefore, the quantum bit coded system of employing has advantages such as the diversity that keeps population and algorithm computation speed is fast.The ICA method based on quantum genetic algorithm that the present invention relates to adopts the quantum bit coded system to separation matrix W, and each chromosomal quantum bit length is 10.If m=n=4, then separation matrix W is 4 * 4 dimension square formations, and it is carried out the quantum bit coding, uses q _i ^tRepresent t generation i chromosome individuality, the specific coding mode is as follows:

$W=\left[\begin{array}{cccc}{w}_{11}& w_{12}& w_{13}& w_{14}\\ w_{21}& w_{22}& w_{23}& w_{24}\\ w_{31}& w_{32}& w_{33}& w_{34}\\ w_{41}& w_{42}& w_{43}& w_{44}\end{array}\right];$ $q_i^t=\left[\begin{array}{cccccccccc}\begin{array}{c}{\mathrm{\α}}_{111}^t\\ {\mathrm{\β}}_{111}^t\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_{112}^t\\ {\mathrm{\β}}_{112}^t\end{array}\right|& \begin{array}{c}\·\·\·\\ \·\·\·\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_{1110}^t\\ {\mathrm{\β}}_{1110}^t\end{array}\right|& \begin{array}{c}{\mathrm{\α}}_{121}^t\\ {\mathrm{\β}}_{121}^t\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_{122}^t\\ {\mathrm{\β}}_{122}^t\end{array}\right|& \begin{array}{c}\·\·\·\\ \·\·\·\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_{1210}^t\\ {\mathrm{\β}}_{1210}^t\end{array}\right|& \begin{array}{c}\·\·\·\\ \·\·\·\end{array}& \left|\begin{array}{c}{\mathrm{\α}}_{4410}^t\\ {\mathrm{\β}}_{4410}^t\end{array}\right.\end{array}\right]$

(2) initialization population

At first, establishing population scale is 5, and promptly population is made up of 5 separation matrixes.Initialization kind group time, all $q_i^t{|}_{t=0}=q_i^0$ (i=1,2 ..., 5) in α _Ij ⁰, β _Ij ⁰(j=1,2 ..., 10) should be initialized as all Like this in a quantum bit chromosome, q _i ⁰The probability of all states of expression is identical.

(3) fitness function is selected

Selection is to separating the estimation Y=WX=[y of back signal ₁, y ₂..., y _n] ^TEach variable kurtosis absolute value and as fitness function, expression formula is as follows:

$J\left(Y\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\mathrm{kurt}\left(y_i\right)\right|=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|E\right\{{\left(y_i\right)}^4\}-3{\left(E\right\{{\left(y_i\right)}^2\left\}\right)}^2|$

Therefore, by observation signal X and separation matrix W, make J (Y) maximization, thereby isolate each independent component.The constraint condition of this optimization problem is the stochastic variable zero-mean, and satisfies E{yy ^T}=I realizes by centralization and nodularization to data.Under this constraint condition, the maximization kurtosis realizes non-Gauss measurement, and promptly for a certain separation matrix W, J (Y) is big more, shows Y=WX=[y ₁, y ₂..., y _n] ^TEach variable between independence strong more.

4. improve the quantum genetic operation of quantum genetic algorithm

The quantum genetic algorithm that the present invention relates to is the improvement of Dr.Han quantum genetic algorithm, quantum interlace operation, quantum not gate mutation operation have been increased, carry out quantum and intersect in order to make full use of all chromosomal information to produce more new model, execution quantum mutation operation is for fear of precocious convergence occurring.Thereby, improve quantum genetic algorithm and both had convergence property fast, have good ability of searching optimum again.The quantum genetic operation of carrying out as shown in Figure 2.Be implemented as follows:

(1) quantum interlace operation

Intersection is replaced the part-structure of two parent chromosome individualities reorganization and is generated the operation of new chromosome individuality, its objective is in order in the next generation, to produce new chromosome individuality, by interlace operation, the raising that the search capability of genetic algorithm is leaped is in order to avoid be absorbed in Local Extremum.Utilize the coherence of quantum among the present invention, carry out the quantum coherent interlace operation, be i.e. quantum interlace operation.Quantum interlace operation specific implementation process is as follows:

(a) each chromosome individuality of rearrangement population.What this paper adopted is randomly ordered mode.

(b) determine first new chromosome individuality.Get first gene of first gene of first chromosome individuality as new chromosome individuality; Get second gene of second gene of second chromosome individuality as new chromosome individuality; Circulate repeatedly like this, the chromosome individuality individual and original up to new chromosome has identical gene number.

(c) by that analogy, determine second new chromosome individuality, the 3rd new chromosome individuality ..., have identical scale up to the new population that generates with original population.

(d) the quantum interlace operation finishes.

(2) the quantum rotation door upgrades operation

By forming suitable quantum door U, realize the more operation of new population, to obtain the higher pattern of fitness.The kind of quantum door has a lot, as quantum not gate, quantum controlled not-gate, quantum Hadamard door, quantum rotation door etc.The quantum door operation all will satisfy U ^*U=UU ^*, U wherein ^*It is the conjugation adjoint matrix of U.What adopt in the present invention is the renewal that the quantum rotation door carries out population, and it is as follows to define its form:

$U=\left[\begin{array}{cc}\cos \mathrm{\θ}& -\sin \mathrm{\θ}\\ \sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right],$ Then have: $\left[\begin{array}{c}{\mathrm{\α}}_i^{\′}\\ {\mathrm{\β}}_i^{\′}\end{array}\right]=\left[\begin{array}{cc}\cos {\mathrm{\θ}}_i& -\sin {\mathrm{\θ}}_i\\ \sin {\mathrm{\θ}}_i& \cos {\mathrm{\θ}}_i\end{array}\right]\left[\begin{array}{c}{\mathrm{\α}}_i\\ {\mathrm{\β}}_i\end{array}\right]$

Wherein, α _i', β _i' (i=1,2 ..., be to upgrade the chromosomal value of back quantum bit n); θ _iBe the angle of quantum rotation door rotation, and θ _i=s (α _i, β _i) Δ θ _i, its size delta θ _iWith direction s (α _i, β _i) definite method generally design in advance, its principle is the current x of separating _iTo optimum solution b _iApproach gradually.The Δ θ that adopts among the present invention _iAnd s (α _i, β _i) to adjust strategy as shown in table 1.In the table 1, f (x) and f (b) represent current fitness function value of separating with optimum solution respectively.

Definite method of table 1 rotation angle

x i	b i	f(x)≥f(b)	Δθ i	s(α i，β i)
								α iβ i＞0	α iβ i＜0	α i＝0	β i＝0
				0	0	false	0					0	0	0	0
0	0	true	0					0	0	0	0
				0	1	false	0					0	0	0	0
0	1	true	0.05π					-1	+1	±1	0
				1	0	false	0.01π					-1	+1	±1	0
1	0	true	0.025π					+1	-1	0	±1
				1	1	false	0.005π					+1	-1	0	±1
1	1	true	0.025π					+1	-1	0	±1

(3) quantum mutation operation

The effect of variation mainly is to stop the prematurity convergence and the algorithm local search ability is provided.In improving quantum genetic algorithm, designed a kind of quantum mutation operation by the quantum not gate.The quantum mutation operation is actually the state of having changed this quantum bit attitude stack, make tended to the state of converging to " 1 " originally become the state of converging to that tends to " 0 ", perhaps opposite.Obviously, this mutation operation to chromosomal all stack attitudes all simultaneously effectively.Concrete grammar is as follows:

(a) from population, choose several chromosome individualities with certain probability P m;

(b) to the chromosome individuality chosen by the probability definitive variation position of determining;

(c) probability amplitude of selected bit quantum bit is carried out the non-door operation of quantum, promptly finish the mutation operation of this quantum bit.

5. independent component ordering

When the ICA based on quantum genetic algorithm that application the present invention relates to carries out high spectrum terrain classification,, need the independent component that be obtained be sorted in order to obtain to contain interesting target information to some extent.The ordering index that adopts is k ₃ ², wherein, k ₃Three rank semi-invariant-degrees of bias (skewness) of expression data.Because the degree of bias can directly reflect target and the departure degree of background on spectral space in the high-spectral data.

6. image binaryzation

Obtained to adopt suitable setting threshold strategy with image binaryzation, thereby realize suppressing background after the independent component of reflection target atural object, appeared target suddenly.The method of threshold setting is a lot, if setting is improper, will cause the reduction nicety of grading, adds problems such as serious mistake branch rate.In the present invention, threshold setting is [m-3 σ, m+3 σ], and wherein m and σ represent the average and the standard deviation of single composition respectively.

For the ICA method based on the improvement quantum genetic algorithm of the realization high spectrum terrain classification that the present invention relates to better is described, utilize PHI aviation high-spectral data to carry out area, Fang Lu tea plantation, Jiangsu crops sophisticated category, Fig. 3 has provided ground reference information, wherein, reference identification C4 represents paddy rice, V2 represents Ipomoea batatas, and V13 represents caraway.Fig. 4 has provided the unsupervised classification result who the present invention relates to algorithm, wherein, (a) is paddy rice, (b) is Ipomoea batatas, is caraway (c), has realized the high-spectral data crops sophisticated category under the situation of any terrestrial information of the unknown.By the comparison and analysis of Fig. 3 and Fig. 4, the present invention relates to as can be seen not have misclassification in the classification results of method, nicety of grading is higher; And this method has speed of convergence faster.

Claims (4)

1. the high spectrum terrain classification method based on the global search algorithm of independent component analysis comprises the steps:

(1) reads in high-spectral data;

(2) foundation is based on the independent component analysis model of kurtosis;

(3) data are carried out centralization and nodularization;

(4) based on the independent component analysis iterative computation of quantum genetic algorithm;

(5) independent component ordering;

(6) image binaryzation;

(7) obtain the terrain classification result.

2. high spectrum terrain classification method according to claim 1 based on the independent component analysis model of kurtosis is:

If observation signal X=[x ₁, x ₂..., x _m] ^TBe the random vector of m dimension, source signal S=[s ₁, s ₂..., s _n] ^TBe the independent random vector of n dimension, hybrid matrix A is that m * n ties up nonsingular matrix, and then the independent component analysis model can be described as:

X＝AS

The essence of independent component analysis is under source signal S and hybrid matrix A condition of unknown, utilizes the statistical property of source signal S according to known observation signal X, determines separation matrix W, makes: Y=WX, and satisfy Y=[y ₁, y ₂..., y _n] ^TEach variable between separate as much as possible, then Y is exactly the estimation of S, simultaneously can determine hybrid matrix A=W ^-1In the independent component analysis model parameter estimation, adopt the non-Gauss's of maximization method, utilization be the fourth order cumulant-kurtosis (Kurtosis) of data, stochastic variable y _iKurtosis be defined as follows:

$\mathrm{kurt}\left(y_i\right)=E\left\{{y_i}^4\right\}-3{\left(E\right\{{y_i}^2\left\}\right)}^2$

Wherein, E represents the average of variable.

3. high spectrum terrain classification method according to claim 1, the flow process of performing step (4) is as follows:

(1) initialization population and genetic algebra t=1, and chromosome in the population is carried out quantum bit encode;

(2) separate by the observation attitude generation scale-of-two of population behind the coding, obtain separation matrix, calculate and separate the back signal, and use fitness function that it is estimated, preserve optimum solution simultaneously;

(3) end condition is judged: if satisfy end condition, then finish, and execution in step (7), and preserve current optimum solution, otherwise continue;

(4) carry out the quantum interlace operation, quantum rotation door update strategy, quantum not gate mutation operation is to obtain new population;

(5) produce scale-of-two by the observation attitude of population and separate, obtain separation matrix, and calculate and separate the back signal, and carry out centralization and albefaction operation, and the utilization fitness function estimates it, preserve optimum solution simultaneously separating the back signal;

(6) genetic algebra increases by 1, execution in step (5);

(7) calculate separation matrix and separation signal, finish.

4. high spectrum terrain classification method according to claim 1, the independent component that adopts in the process of performing step (5) ordering index is: k ₃ ², wherein, k ₃Be three rank semi-invariant-degrees of bias (Skewness).

Images