CN102592134B - Multistage decision fusing and classifying method for hyperspectrum and infrared data - Google Patents

Abstract

The invention discloses a multistage decision fusing and classifying method for hyperspectrum and infrared data. The multistage decision fusing and classifying method comprises the following steps of: firstly, carrying out noise suppression processing and spatial adjustment on the hyperspectrum and infrared data; secondly, establishing a hyperspectrum and infrared data combined characteristic space according to the characteristics of the hyperspectrum and infrared data; thirdly, monitoring and classifying the combined characteristic space established in the second step to obtain a ground object classification decision according to the ground object categories to be classified and a training sample; fourthly, determining ground object classes required to be subjected to small-object strengthening decision extraction according to the object size and carrying out small-object strengthening decision extraction by using the combined characteristic space established in the second step; fifthly, carrying out end member extraction and abundance estimation on hyperspectrum data subjected to noise suppression in the first step to obtain an abundance decision; and sixthly, designing a fusing rule and fusing the classification decision acquired in the third step, the small-object strengthening decision acquired in the fourth step and the abundance decision acquired in the fifth step to obtain a fusion and classifying result.

Description

A kind of high spectrum and infrared data multi-level decision-making integrated classification method

Technical field

The present invention relates to a kind of high spectrum and infrared data multi-level decision-making integrated classification method, belong to RS data disposal route and applied technical field, be applicable to theoretical method and the application technical research of the classification of high spectrum and infrared data high precision.

Background technology

Along with spacer remote sensing technical development and application, at many levels, multi-angle, comprehensive and round-the-clock Earth Information are obtained system and are set up gradually, and brought the rapid growth of earth observation data volume thus, how these information are effectively integrated application, RS data is complemented each other, and is one of hot issue of current Remote Sensing Study.

High spectrum resolution remote sensing technique has overcome traditional single band, multispectral remote sensing in the limitation of aspects such as wave band number, wavelength band, meticulous information representation, provide sensor information with narrower wave band interval, more wave band quantity, can be segmented and be differentiated atural object, obtained widespread use in fields such as resource, environment, city, ecologies.Infrared imagery technique can get access to the information of body surface heat radiation and internal heat dissipation thereof, have advantages such as antijamming capability is strong, climatic environment adaptability is strong, round-the-clock continuously, noncontact detection, high precision measurement of angle, in civilian and Military Application, brought into play great function.Merge high spectrum and infrared data, can be when utilizing the figure spectrum information of high spectrum, replenish the space heat distribution information with infrared image again, multi-source remote sensing information is replenished mutually, and then better service solves the problem that exists in high-spectrum remote-sensing and the infrared remote sensing in application such as target detection and the meticulous divisions of atural object.

Be subjected to the influence of factors such as instrument itself and environment, true remotely-sensed data ubiquity noise influence authenticity and the reliability of remotely-sensed data, so denoising is extremely important to remotely-sensed data.(Minimum Noise Fraction, MNF) conversion is a kind of good high-spectral data noise restraint method, and is effectively simple in actual applications in the minimal noise separation.Yet parameter--high-spectral data band noise covariance method of estimation still has problems in input as the core in the MNF conversion.Spectrum-space decorrelation method (Spectral and Spatial Decorrelation Method, SSDC) be that the adaptability of generally acknowledging present stage is strong, the noise extraction algorithm of stable accuracy, yet because the restriction of this theory of algorithm, it extracts effectively the noise of object image equably, and the noise precise decreasing that the situation that atural object on the image mixes is extracted.In fact, " homogeneity " supposed in noise is estimated most important, and homogeneous region extracts and will improve the noise results estimated to a great extent accurately.And in high-spectral data, homogeneous region and non-homogeneous zone generally can exist simultaneously, need take full advantage of homogeneous region and ask for noise, and reduce the non-homogeneous zone to asking for The noise.

Decision level fusion is very suitable for RS data and handles, and it can carry out conversion to each data source and estimates to obtain independently objective attribute target attribute, then with it as the decision-making fusion of making a strategic decision, be a kind of fusion method of flexible and efficient processing multi-source data.The major advantage that decision-making level merges is: it is low that fusion center is handled cost; Require very low to the information transmission bandwidth; When mistake appearred in one or several decision-making output, by suitable fusion, system can also obtain correct result; Widely applicable, original sensing data there is not special requirement, each sensor that raw data is provided can be foreign peoples's sensor, wherein even can comprise the information of being obtained by non-image sensor.

Summary of the invention

The objective of the invention is to propose a kind of high spectrum and infrared data multi-level decision-making integrated classification method, it has reduced the classification erroneous judgement that the deficiency of single data source and single sorting technique causes, and the mode that merges by a kind of multi-level decision-making realizes high precision, the high stability terrain classification of high spectrum and infrared data.

Technical solution of the present invention is: the mode that the present invention is merged with decision-making realizes the high precision atural object integrated classification of high spectrum and infrared data, taken into full account the influence of noise to the data authenticity, adopt a kind of method of the multidimensional gradient of surveying at the high-spectral data homogeneous region to realize the detection of high spectrum image homogeneous region for high-spectral data, solve accurate high spectral band noise estimated result, and then obtain accurate noise suppression high-spectral data; The mode that adopts multi-level decision-making to extract in the extraction of decision-making: high spectrum and the classification of infrared data association feature level, little target are strengthened the high-spectral data abundance method of estimation that decision-making is extracted, searched for based on variable end member optimization, wherein strengthening decision-making for little target extracts, designed the choice criteria of type of ground objects, extracted for the abundance decision-making and designed a kind of uncertain and probabilistic variable end member abundance method of estimation efficiently of end member of mixing that taken into full account; In addition, for better realizing the fusion of decision-making, designed a kind of fusion rule that combines obscure idea.The mode that the present invention is merged with a kind of multi-level decision-making has realized the high precision classification of high spectrum and infrared data.

A kind of high spectrum of the present invention and infrared data multi-level decision-making integrated classification method, its step is as follows:

(1), high spectrum and infrared data are carried out noise suppression processing and spatial registration;

(2), according to high spectrum and infrared data characteristics, set up high spectrum and infrared data association feature space;

(3), according to treating branch atural object classification and training sample, the associating feature space that step (2) the is set up classification that exercises supervision obtains terrain classification and makes a strategic decision;

(4), according to target size, determine to carry out little target and strengthen the ground species that decision-making is extracted, the associating feature space that utilizes step (2) to set up carries out little target and strengthens decision-making and extract;

(5), the high-spectral data behind the noise suppression that step (1) is obtained carries out end member and extracts with abundance and estimate, obtains the abundance decision-making;

(6), the design fusion rule, merge little target that categorised decision, the step (4) obtained by step (3) obtain and strengthen the abundance decision-making that decision-making and step (5) are obtained, obtain the integrated classification result.

Wherein, the noise restraint method that adopts in the step (1): adopt wavelet method to the infrared image noise suppression, adopt and improve minimal noise separation transfer pair high-spectral data noise suppression, adopt a kind of circulation spectrum-space decorrelation noise estimation method based on high spectrum multidimensional gradient piecemeal selection to estimate high-spectral data band noise covariance matrix Cov in improving minimal noise separation conversion, its concrete steps are as follows:

A) make circulation label T=1, K ^(T)For complete high-spectral data space, with K ^(T)The space dimension is divided into joining and nonoverlapping square of Len * Len size, Len ∈ [5,15], and at K ^(T)Mark piecemeal situation, the image section that can't be divided into piece is rejected K ^(T)

B) based on K ^(T), utilize the SSDC method to find the solution band noise

Be the noise variance of wave band i, 1≤i≤L, L are the spectrum dimension of high-spectral data, if T＞1 and | E ^(T)-E ^(T-1)|＜ε, ε=10 ^-4, according to current K ^(T)Utilize the SSDC method to calculate band noise covariance matrix Cov, end loop; Otherwise carry out next step c);

C) according to band noise E ^(T)Calculate the weight factor of high spectrum multidimensional weight gradient $w^{\left(T\right)}=\{w_1^{\left(T\right)},w_2^{\left(T\right)},\·\·\·,w_L^{\left(T\right)}\},$ Computing formula is as follows:

$w_i^{\left(T\right)}=\left(\frac{1}{{\mathrm{\σ}}_i^2}\right)/\left(\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\frac{1}{{\mathrm{\σ}}_i^2}\right);$

D) utilize w ^(T)Find the solution each pixel of high-spectral data (x, gradient G y) _{X, y}, computing formula is as follows:

$G_{x,y}=\underset{k=1}{\overset{L}{\mathrm{\Σ}}}{w}_k^{\left(T\right)}{\mathrm{\ρ}}_{\mathrm{\δ}}\left(z_k\right),$

ρ wherein _δ(z _k) be the morphology gradient of K-band, 1≤k≤L, δ are structural elements, z _kIt is the k ripple

Section pixel (x, the gray-scale value of y) locating, ρ _δ(z _k) computing formula is as follows:

ρ _δ(z _k)＝d _δ(z _k)-e _δ(z _k)，

D wherein _δAnd e _δFor expanding and erosion operator;

E) again original high-spectral data space dimension is divided into joining and nonoverlapping square of Len * Len size, rejecting can't be divided into the image section of piece, obtains Q ^(T), and at Q ^(T)Mark piecemeal situation is calculated the average gradient of every high-spectral data, and formula is as follows:

$\stackrel{\‾}{G^q}=\frac{1}{{\mathrm{Len}}^2}\underset{(x.y)\∈B^q}{\mathrm{\Σ}}{G}_{x,y},$

B wherein ^qBe the coordinate set of piece q at former high-spectral data;

F) set average gradient threshold parameter γ, γ ∈ [0.1,0.8], calculate the average gradient threshold value according to following formula:

${\mathrm{TH}}_G=G_{\min }^q+\mathrm{\γ}(G_{\max }^q-G_{\min }^q),$ $\left\{\begin{array}{c}{G}_{\min }^q=\underset{q}{\min }\left\{\stackrel{\‾}{G^q}\right\}\\ G_{\max }^q=\underset{q}{\max }\left\{\stackrel{\‾}{G^q}\right\}\end{array}\right.,$

If

Piece q is homogeneous region, otherwise is the non-homogeneous zone, from Q ^(T)Last rejecting is non-together

Q is upgraded in the matter zone ^(T), T=T+1 makes K ^(T)=Q ^(T-1), return step (b).

Wherein, step (2) is set up high spectrum and infrared data association feature space, comprises containing much information in the high-spectral data behind the noise suppression and the wave band that correlativity is little, high spectrum principal component analysis (PCA), minimal noise separate the infrared image behind conversion, independent component analysis and textural characteristics wave band and the noise suppression;

Wherein, step (3) is according to treating branch atural object classification and training sample, specifically is the classification that exercises supervision of the associating feature space that utilizes algorithm of support vector machine that step (2) is set up, obtains terrain classification and makes a strategic decision;

Wherein, selection need be carried out the method that little target is strengthened the ground species of decision-making extraction in the step (4), and detailed process is as follows: if the target real area is A, image spatial resolution is l, when the shared pixel number of target Satisfy

The time, this target type need be carried out little target and be strengthened the decision-making extraction, wherein C is line number and the columns that not sum, W and H to be sorted are respectively high-spectral data, through above-mentioned selection, inferior pixel target, account for the few target of image percentage and be selected and carry out strengthening decision-making based on the little target of orthogonal subspaces projection operator and extract.

Wherein, " high-spectral data behind the noise suppression that step (1) is obtained carries out end member and extracts with abundance and estimate, obtains the abundance decision-making " described in the step (5), its concrete steps are as follows:

A) utilize vertex component analysis, pure pixel index, sequence maximum angular convex cone method to extract the image end member, constitute initial end member storehouse Θ;

B) based on the understanding to the ground actual conditions, standard spectrum storehouse spectral information is incorporated among the Θ as replenishing the image end member;

C) adopt cluster analysis to remove redundant end member with the mode that spectrum angle criterion combines, simplify the operand of subsequent treatment, concrete grammar is as follows: at first adopt clustering method that end member collection Θ is divided into the C class, C is atural object classification number, to reject such with such cluster centre spectrum angle greater than the end member spectrum of α (α＞0.1) in all kinds of, to reject this class less than the end member spectrum of β (β＜0.1) with his class cluster center light spectral corner simultaneously, obtain redundant end member library of spectra Θ, Θ={ ψ ₁..., ψ _i..., ψ _C, ψ _iBe the set of i class end member, 1≤i≤C;

D) utilize crosscorrelation Spectral matching method (Cross Correlogram Spectral Matching, CCSM), be that each pixel of high-spectral data is sought an end member subclass that participates in this pixel mixing based on Θ, in the process of seeking the end member subclass, if i class end member is selected into the end member subclass of certain pixel correspondence by CCSM, then i class end member no longer participates in selecting in the selection subsequently of this pixel, finally determines the end member subclass of each pixel correspondence;

E) utilizing the definite end member subclass of step d) that corresponding pixel is carried out non-negative restriction least square solution mixes, calculate the solution of the corresponding end member subclass of each pixel and mix error (RMSE, θ), wherein RMSE is that root-mean-square error, θ are the spectrum angle, and RMSE and θ computing formula are as follows:

$\mathrm{RMSE}=\sqrt{\frac{{(R-R^{\′})}^T(R-R^{\′})}{L}},$ $\mathrm{\θ}=\arccos \left(\frac{R^T{R}^{\′}}{\sqrt{R^{T}R}\sqrt{R^{\′T}{R}^{\′}}}\right),$

Wherein R is pixel spectrum, and R ' is the modeling spectrum of R;

F) it is as follows that mixed error threshold is separated in setting:

$\left\{\begin{array}{c}{\mathrm{TH}}_{\mathrm{RMSE}}=0.001\\ {\mathrm{TH}}_{\mathrm{\θ}}=0.01\end{array}\right.,$

TH wherein _RMSEBe the root-mean-square error upper limit, TH _θBe the spectrum angle upper limit, verification step e) whether the solution of the end member subclass of each pixel correspondence of calculating mix error and satisfy simultaneously less than TH _RMSEAnd TH _θIf satisfying then, corresponding end member subclass is the optimum end member subclass of this pixel, then this pixel is not labeled as the pixel that needs carry out end member optimization search if do not satisfy, all pixels judgements finish in to high-spectral data, obtain carrying out the pixel set omega of end member optimization search;

G) adopt binary coding that Θ is encoded, code word size is N, N is end member sum among the Θ, coded system is as follows: selected a certain end member if participate in the end member subclass of pixel mixing, then this end member correspondence position code word value is " 1 ", otherwise is " 0 ", and C " 1 " is arranged at most in all significant model based coding vectors in the code set, and in the similar end member set one " 1 " is arranged at most, other coding form is insignificant solution;

H) utilize scale-of-two particle swarm optimization algorithm solution procedure f) the optimum end member subclass of each pixel correspondence in the set omega that obtains, wherein use non-negative restriction least square solution to mix square-error 1/RMSE reciprocal ²As fitness, the fitness of the insignificant solution of mentioning in the step g) is set to 0;

I) utilize the optimum end member subclass of each pixel to carry out the abundance estimation, obtain the abundance decision-making.

Wherein, " design fusion rule; the abundance decision-making that the little target reinforcement decision-making that categorised decision, the step (4) that fusion is obtained by step (3) obtained and step (5) are obtained; obtain the integrated classification result " described in the step (6), detailed process is as follows: design fusion rule, the little target that the categorised decision Cla that fusion steps (3) obtains, step (4) obtain are strengthened decision-making Objs, the abundance decision-making Fras that obtains with step (5), obtain final integrated classification result, concrete grammar is as follows:

$D(x_0,y_0,i)=D_1(x_0,y_0,i)+\underset{t=1}{\overset{3}{\mathrm{\Σ}}}{D}_2^t(x_0,y_0,i)+D_3(x_0,y_0,i),$

Wherein

Expression is sought and is made function f (i) maximal parametric i,

Mean i ^*Make f (i) maximum, (x ₀, y ₀) be the position coordinates of any pixel on the image, Lable (x ₀, y ₀) be (x ₀, y ₀) the pixel class label at coordinate position place, according to categorised decision Cla, little target strengthen decision-making Objs, abundance is not found the solution D ₁,

D ₃, concrete computing method are as follows:

A) utilize categorised decision Cla to calculate D ₁, D ₁Be of a size of W * H * C, with D ₁Middle all elements is initialized as 0, if (x, the pixel of y) locating is divided into the i class, then D in Cla ₁(x, y, m)=a _m, m=1,2 ..., C, a _mSatisfy:

$a_m=\left\{\begin{array}{c}{\mathrm{Pr}}_1^i,m=i\\ (1-{\mathrm{pr}}_1^i)/(C-1),m\≠i\end{array}\right.,$

Parameter wherein

Represented the accuracy and stability of class i among the categorised decision Cla,

B) utilizing little target to strengthen decision-making Objs calculates Be of a size of W * H * C, t=1,2 ..., n, n are the number that little target is strengthened decision-making, will

All elements is initialized as 0, if (x, the pixel of y) locating is identified as i class atural object, then in t Objs

M=1,2 ..., C, a _mSatisfy:

$a_m=\left\{\begin{array}{c}{\mathrm{Pr}}_2^i,m=i\\ (1-{\mathrm{pr}}_2^i)/(C-1),m\≠i\end{array}\right.,$

Parameter wherein

Represent little target and strengthened the accuracy and stability of class i among the decision-making Objs,

${\mathrm{Pr}}_2^i\∈\left[\mathrm{0.7,1}\right);$

C) utilize abundance decision-making Fras to calculate D ₃, each pixel abundance vector of Fras all divided by this vector sum, is obtained D ₃

The present invention's advantage compared with prior art is: use a kind of effective method more to realize the squelch of high-spectral data, make data true more, effective; Adopted a kind of mixing uncertainty and probabilistic efficient, high-precision abundance method of estimation of end member of having taken into full account; And utilize classification results, selectively little target reinforcement decision-making extraction result and high-precision abundance estimated result as decision-making, the mode that merges with decision-making realizes high-precision high spectrum and infrared data multi-level decision-making integrated classification.It has following advantage: (1) introduces high spectrum multidimensional weight gradient as the standard of evaluating data homogeneous region, has improved SSDC noise covariance estimated accuracy, and then improves the accuracy of high-spectral data denoising; (2) high spectrum and infrared data association classification results, selectively little target are strengthened decision-making extraction result, and high precision high-spectral data abundance estimated result is as decision-making, the mode that merges with decision-making has realized high-precision decision-making integrated classification, the error in classification of having avoided the shortcomings and deficiencies of single algorithm, single data source to cause improves accuracy and the stability of classification results.

Description of drawings

Fig. 1 is process flow diagram of the present invention.

Embodiment

For a kind of high spectrum and the infrared data multi-level decision-making integrated classification method that the present invention relates to better is described, (high-spectral data size 300 * 300, wave band number are 273 to utilize a real data; The thermal infrared data are 8～13 microns broadband thermal infrared images, and its corresponding ground scope has covered the ground scope of high-spectral data correspondence) carry out high precision atural object classification (8 classes: vegetation, concrete floor, Chi Shui, pond are along, white calibration cloth, black calibration cloth, metal well lid, shade).As shown in Figure 1, a kind of high spectrum of the present invention and infrared data multi-level decision-making integrated classification method, the specific implementation step is as follows:

(1) reading the reference mark from high spectrum and infrared data, is that reference picture, infrared image are image subject to registration with the high-spectral data, utilizes the polynomial expression method to realize high spectrum and infrared data spatial registration; Adopt small echo noise suppression mode for infrared image, adopt improvement MNF transform method to carry out noise suppression for high-spectral data, wherein high-spectral data band noise covariance matrix Cov adopts the circulation SSDC noise estimation method based on high spectrum multidimensional gradient piecemeal selection, and its concrete steps are as follows:

A) make circulation label T=1, K ^(T)Be complete high-spectral data space, its space dimension is of a size of 300 * 300, with K ^(T)The space dimension is divided into joining and nonoverlapping square of 8 * 8 sizes, and the image section that can't be divided into piece is rejected, then piecemeal high-spectral data space K ^(T)Be of a size of 296 * 296;

B) based on K ^(T), utilize SSDC to find the solution band noise

Be the noise variance of wave band i, 1≤i≤273, if T＞1 and | E ^(T)-E ^(T-1)|＜10 ^-4, according to current K ^(T)Utilize SSDC to calculate band noise covariance matrix Cov, end loop; Otherwise carry out next step c);

C) according to band noise E ^(T)Calculate the weight factor of high spectrum multidimensional weight gradient $w^{\left(T\right)}=\{w_1^{\left(T\right)},w_2^{\left(T\right)},\·\·\·,w_{273}^{\left(T\right)}\},$ Computing formula is as follows:

$w_i^{\left(T\right)}=\left(\frac{1}{{\mathrm{\σ}}_i^2}\right)/\left(\underset{i=1}{\overset{273}{\mathrm{\Σ}}}\frac{1}{{\mathrm{\σ}}_i^2}\right);$

D) utilize w ^(T)Find the solution each pixel of high-spectral data (x, gradient G y) _{X, y}, computing formula is as follows:

$G_{x,y}=\underset{k=1}{\overset{273}{\mathrm{\Σ}}}{w}_k^{\left(T\right)}{\mathrm{\ρ}}_{\mathrm{\δ}}\left(z_k\right),$

ρ wherein _δ(z _k) be the morphology gradient of K-band, 1≤k≤273, δ is that bulk is 3 * 3 square structure unit, z _kBe K-band pixel (x, the gray-scale value of y) locating, ρ _δ(z _k) computing formula is as follows:

ρ _δ(z _k)＝d _δ(z _k)-e _δ(z _k)，

D wherein _δOperator is the maximal value of getting in the structural elements, e _δOperator is the minimum value of getting in the structural elements;

E) again the space in original high-spectral data space dimension is divided into joining and nonoverlapping square of 8 * 8 sizes, rejecting can't be divided into the image section of piece, obtains Q ^(T), bulk is 296 * 296, and at Q ^(T)The subscript piece situation of keeping the score is calculated the average gradient of every high-spectral data, and formula is as follows:

$\stackrel{\‾}{G^q}=\frac{1}{8\×8}\underset{(x.y)\∈B^q}{\mathrm{\Σ}}{G}_{x,y},$

B wherein ^qBe the coordinate set of piece q at former high-spectral data.

F) set average gradient threshold parameter γ=0.5, calculate the average gradient threshold value according to following formula:

${\mathrm{TH}}_G=G_{\min }^q+\mathrm{\γ}(G_{\max }^q-G_{\min }^q),$ $\left\{\begin{array}{c}{G}_{\min }^q=\underset{q}{\min }\left\{\stackrel{\‾}{G^q}\right\}\\ G_{\max }^q=\underset{q}{\max }\left\{\stackrel{\‾}{G^q}\right\}\end{array}\right.,$

If

Piece q is homogeneous region, otherwise is the non-homogeneous zone, from Q ^(T)Last rejecting is non-together

Q is upgraded in the matter zone ^(T), T=T+1 makes K ^(T)=Q ^(T-1), return step b).

Reading the reference mark from image, is that reference picture, infrared image are image subject to registration with the high-spectral data, utilizes the polynomial expression method to realize high spectrum and infrared data spatial registration;

(2) set up high spectrum and infrared data association feature space, comprise in the high-spectral data behind the noise suppression contain much information, 40 wave bands that correlativity is little, high spectrum principal component analysis (PCA), minimal noise separate conversion, independent component analysis and textural characteristics totally 20 wave bands, and the infrared image behind the noise suppression, the feature space dimension is 61;

(3) determine that according to existing priori the atural object classification is 8 classes: vegetation, concrete floor, Chi Shui, edge, pond, white calibration cloth, black calibration cloth, metal well lid, shade, obtain the atural object training sample, the associating feature space that utilizes step (2) to set up, utilize algorithm of support vector machine to realize terrain classification, obtain the terrain classification decision-making;

(4) select to carry out little target and strengthen the ground species that decision-making is extracted, through calculating, black cloth, calico and three kinds of atural objects of metal well lid are selected, and utilize the orthogonal subspaces projection operator to carry out little target and strengthen the decision-making extraction, obtain the little target reinforcement of three width of cloth decision-making and extract the result;

(5) high-spectral data behind the noise suppression that step (1) is obtained carries out end member and extracts, and adopts end member extraction method to obtain the image end member, the noise suppression high-spectral data is carried out abundance estimate that obtain the abundance decision-making, its concrete steps are as follows:

A) utilize vertex component analysis, pure pixel index, sequence maximum angular convex cone method to carry out the image end member and extract, constitute initial end member storehouse Θ;

B) based on the understanding to the ground actual conditions, standard spectrum storehouse spectral information is incorporated among the Θ as replenishing the image end member;

C) adopt cluster analysis to remove redundant end member with the mode that spectrum angle criterion combines, simplify the operand of subsequent treatment, concrete grammar is as follows: at first adopt clustering method that end member collection Θ is divided into 8 classes, to reject such with such cluster centre spectrum angle greater than 0.15 end member spectrum in all kinds of, to reject this class less than 0.09 end member spectrum with his class cluster center light spectral corner simultaneously, obtain redundant end member library of spectra Θ, Θ={ ψ ₁..., ψ _i..., ψ ₈, ψ _iBe the set of i class end member, 1≤i≤8;

D) utilize CCSM, seek one based on Θ for each pixel of high-spectral data and participate in the end member subclass that this pixel mixes, in the process of seeking the end member subclass, if i class end member is selected into the end member subclass of certain pixel correspondence by CCSM, then i class end member no longer participates in selecting in the selection subsequently of this pixel, the specific implementation method is as follows: calculate the response between pixel spectrum and the end member, judge the similarity degree between two spectrum, select the end member the highest with pixel spectrum similarity degree, as the end member classification that most probable in this pixel comprises, the definition response coefficient

As follows:

$r_e^j=\frac{\mathrm{L\Σ}{R}_e{R}_j-\mathrm{\Σ}{R}_e\mathrm{\Σ}{R}_j}{\sqrt{[\mathrm{L\Σ}{R_e}^2-{\left(\mathrm{\Σ}{R}_e\right)}^2][\mathrm{L\Σ}{R_j}^2-{\left(\mathrm{\Σ}{R}_j\right)}^2]}},$

R wherein _eAnd R _jBe respectively end member spectrum and pixel spectrum, L is high-spectral data spectrum dimension, by the peak response coefficient

Determine and pixel R _jThe end member spectrum R ' that similarity is the highest _e, definition residue pixel spectrum R ' _jAs follows:

$R_j^{\′}=R_j-\mathrm{\η}{r}_{\max }^j{R}_e^{\′},$

Wherein adjust coefficient η=0.6, with R ' _jReplace R _jBring the response coefficient formula into, continue relatively to remain pixel spectrum and remaining non-similar end member, carry out iteration successively, until R ' _jCertain component right and wrong on the occasion of, carry out said method and determine each pixel by order

In comprise the end member subclass R ' _e;

E) the end member subclass of utilizing each pixel correspondence is carried out non-negative restriction least square solution to this pixel and is mixed, calculate separate mix error (RMSE, θ), wherein RMSE is that root-mean-square error, θ are the spectrum angle, RMSE and θ computing formula are as follows:

$\mathrm{RMSE}=\sqrt{\frac{{(R-R^{\′})}^T(R-R^{\′})}{L}},$ $\mathrm{\θ}=\arccos \left(\frac{R^T{R}^{\′}}{\sqrt{R^{T}R}\sqrt{R^{\′T}{R}^{\′}}}\right),$

Wherein R is pixel spectrum, and R ' is R modeling spectrum;

F) it is as follows that mixed error threshold is separated in setting:

$\left\{\begin{array}{c}{\mathrm{TH}}_{\mathrm{RMSE}}=0.001\\ {\mathrm{TH}}_{\mathrm{\θ}}=0.01\end{array}\right.,$

TH wherein _RMSEBe the root-mean-square error upper limit, TH _θBe the spectrum angle upper limit, verification step e) whether the solution of the end member subclass of each pixel correspondence of calculating mix error and satisfy simultaneously less than TH _RMSEAnd TH _θIf satisfying then, corresponding end member subclass is the optimum end member subclass of this pixel, then this pixel is not labeled as the pixel that needs carry out end member optimization search if do not satisfy, all pixels judgements finish in to high-spectral data, obtain carrying out the pixel set omega of end member optimization search;

G) adopt binary coding that Θ is encoded, code word size is N, N=6136, coded system is as follows: selected a certain end member if participate in the end member subclass of pixel mixing, then this end member correspondence position code word value is " 1 ", otherwise is " 0 ", in all significant model based coding vectors 8 " 1 " is arranged at most in the code set, and in the similar end member set one " 1 " is arranged at most, other coding form is insignificant solution;

H) utilize scale-of-two particle swarm optimization algorithm solution procedure f) the optimum end member subclass of each pixel correspondence in the set omega that obtains, wherein use non-negative restriction least square solution to mix square-error 1/RMSE reciprocal ²As fitness, the fitness of the insignificant solution of mentioning in the step g) is set to 0;

I) utilize the best end member subclass of each pixel to carry out the abundance estimation, obtain the abundance decision-making.

(6) 3 little targets that design fusion rule, the categorised decision Cla that fusion steps (3) is obtained, step (4) obtain are strengthened decision-making Objs, and 8 abundance decision-making Fras with step (5) is obtained obtain final integrated classification result, and concrete grammar is as follows:

$D(x_0,y_0,i)=D_1(x_0,y_0,i)+\underset{t=1}{\overset{3}{\mathrm{\Σ}}}{D}_2^t(x_0,y_0,i)+D_3(x_0,y_0,i),$

Wherein

Expression is sought and is made function f (i) maximal parametric i,

Mean i ^*Make f (i) maximum, (x ₀, y ₀) be the position coordinates of any pixel on the image, Lable (x ₀, y ₀) be (x ₀, y ₀) the pixel class label at coordinate position place, strengthen decision-making Objs, abundance decision-making Fras according to categorised decision Cla, little target and find the solution D ₁,

D ₃, concrete computing method are as follows:

A) utilize categorised decision Cla to calculate D ₁, set up one and be of a size of 300 * 300 * 8 cube D ₁, with D ₁Middle all elements is initialized as 0, if (x, y) pixel of position is divided into the i class, then D in Cla ₁(x, y, m)=a _m, m=1,2 ..., 8, a _mSatisfy:

$a_m=\left\{\begin{array}{c}0.8,m=i\\ 0.2/7,m\≠i\end{array}\right.;$

B) utilizing little target to strengthen decision-making Objs calculates

Foundation is of a size of 300 * 300 * 8 cube

T=1,2,3, will

All elements is initialized as 0, if (x, the pixel of y) locating is identified as i class atural object, then in t Objs

M=1,2 ..., 8, a _mSatisfy:

$a_m=\left\{\begin{array}{c}0.8,m=i\\ 0.2/7,m\≠i\end{array}\right.;$

C) utilize abundance decision-making Fras to calculate D ₃, the element of each pixel abundance vector all divided by this vector sum, is obtained D ₃

Claims (1)

1. a high spectrum and infrared data multi-level decision-making integrated classification method, it is characterized in that: it comprises following steps:

(1), high spectrum and infrared data are carried out noise suppression processing and spatial registration;

(2), according to high spectrum and infrared data characteristics, set up high spectrum and infrared data association feature space;

(3), according to treating branch atural object classification and training sample, the associating feature space that step (2) the is set up classification that exercises supervision obtains terrain classification and makes a strategic decision;

(4), according to target size, determine to carry out little target and strengthen the ground species that decision-making is extracted, the associating feature space that utilizes step (2) to set up carries out little target and strengthens decision-making and extract;

(5), the high-spectral data behind the noise suppression that step (1) is obtained carries out end member and extracts with abundance and estimate, obtains the abundance decision-making;

(6), the design fusion rule, merge little target that categorised decision, the step (4) obtained by step (3) obtain and strengthen the abundance decision-making that decision-making and step (5) are obtained, obtain the integrated classification result;

Wherein, the noise restraint method that adopts in the step (1): adopt wavelet method to the infrared image noise suppression, adopt and improve minimal noise separation transfer pair high-spectral data noise suppression, adopt a kind of circulation spectrum-space decorrelation noise estimation method based on high spectrum multidimensional gradient piecemeal selection to estimate high-spectral data band noise covariance matrix Cov in improving minimal noise separation conversion, its concrete steps are as follows:

A) make circulation label T=1, K ^(T)For complete high-spectral data space, with K ^(T)The space dimension is divided into joining and nonoverlapping square of Len * Len size, Len ∈ [5,15], and at K ^(T)Mark piecemeal situation, the image section that can't be divided into piece is rejected K ^(T);

B) based on K ^(T), utilize spectrum-space decorrelation method to find the solution band noise

Be the noise variance of wave band i, 1≤i≤L, L are the spectrum dimension of high-spectral data, if T〉1 and | E ^(T)-E ^(T-1)|＜ε, ε=10 ^-4, according to current K ^(T)Utilize spectrum-space decorrelation method to calculate band noise covariance matrix Cov, end loop; Otherwise carry out next step c);

C) according to band noise E ^(T)Calculate the weight factor of high spectrum multidimensional weight gradient $w^{\left(T\right)}=\{w_1^{\left(T\right)},w_2^{\left(T\right)},\·\·\·,w_L^{\left(T\right)}\},$ Computing formula is as follows:

$w_i^{\left(T\right)}=\left(\frac{1}{{\mathrm{\σ}}_i^2}\right)/\left(\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\frac{1}{{\mathrm{\σ}}_i^2}\right);$

D) utilize w ^(T)Find the solution each pixel of high-spectral data (x, gradient G y) _{X, y}, computing formula is as follows:

$G_{x,y}=\underset{k=1}{\overset{L}{\mathrm{\Σ}}}{w}_k^{\left(T\right)}{\mathrm{\ρ}}_{\mathrm{\δ}}\left(z_k\right),$

ρ wherein _δ(z _k) be the morphology gradient of K-band, 1≤k≤L, δ are structural elements, z _kBe K-band pixel (x, the gray-scale value of y) locating, ρ _δ(z _k) computing formula is as follows:

ρ _δ(z _k)=d _δ(z _k)-e _δ(z _k)，

D wherein _δAnd e _δFor expanding and erosion operator;

E) again original high-spectral data space dimension is divided into joining and nonoverlapping square of Len * Len size, rejecting can't be divided into the image section of piece, obtains Q ^(T), and at Q ^(T)Mark piecemeal situation is calculated the average gradient of every high-spectral data, and formula is as follows:

$\stackrel{\‾}{G^q}=\frac{1}{{\mathrm{Len}}^2}\underset{(x.y)\∈B^q}{\mathrm{\Σ}}{G}_{x,y},$

B wherein ^qBe the coordinate set of piece q at former high-spectral data;

F) set average gradient threshold parameter γ, γ ∈ [0.1,0.8], calculate the average gradient threshold value according to following formula:

${\mathrm{TH}}_G=G_{\min }^q+\mathrm{\γ}(G_{\max }^q-G_{\min }^q),\left\{\begin{array}{c}{G}_{\min }^q=\underset{q}{\min }\left\{\stackrel{\‾}{G^q}\right\}\\ G_{\max }^q=\underset{q}{\max }\left\{\stackrel{\‾}{G^q}\right\}\end{array}\right.,$

If

Piece q is homogeneous region, otherwise is the non-homogeneous zone, from Q ^(T)Q is upgraded in last rejecting non-homogeneous zone ^(T), T=T+1 makes K ^(T)=Q ^(T-1), return step b);

Wherein, selection need be carried out the method that little target is strengthened the ground species of decision-making extraction in the step (4), and its implication is described as follows: if the target real area is A, image spatial resolution is l, when the shared pixel number of target

Satisfy

The time, this target type need be carried out little target and be strengthened the decision-making extraction, wherein C is line number and the columns that sum, W and H to be sorted are respectively high-spectral data, through above-mentioned selection, inferior pixel target, accounts for the few target of image percentage and is selected and carries out little target and strengthen the decision-making extraction;

Wherein, " high-spectral data behind the noise suppression that step (1) is obtained carries out end member and extracts with abundance and estimate, obtains the abundance decision-making " described in the step (5), its concrete steps are as follows:

A) utilize vertex component analysis, pure pixel index, sequence maximum angular convex cone method to extract the image end member, constitute initial end member storehouse Θ;

B) based on the understanding to the ground actual conditions, standard spectrum storehouse spectral information is incorporated among the Θ as replenishing the image end member;

C) adopt cluster analysis to remove redundant end member with the mode that spectrum angle criterion combines, simplify the operand of subsequent treatment, concrete grammar is as follows: at first adopt clustering method that end member collection Θ is divided into the C class, C is atural object classification number, to reject such with such cluster centre spectrum angle greater than the end member spectrum of α (α〉0.1) in all kinds of, to reject this class less than the end member spectrum of β (β＜0.1) with his class cluster center light spectral corner simultaneously, obtain redundant end member library of spectra Θ, Θ={ Ψ ₁..., Ψ _i..., Ψ _C, Ψ _iBe the set of i class end member, 1≤i≤C;

D) utilize crosscorrelation Spectral matching method, seek one based on Θ for each pixel of high-spectral data and participate in the end member subclass that this pixel mixes, in the process of seeking the end member subclass, if i class end member is selected into the end member subclass of certain pixel correspondence by crosscorrelation Spectral matching method, then i class end member no longer participates in selecting in the selection subsequently of this pixel, finally determines the end member subclass of each pixel correspondence;

E) utilizing the definite end member subclass of step d) that corresponding pixel is carried out non-negative restriction least square solution mixes, calculate the solution of the corresponding end member subclass of each pixel and mix error (RMSE, θ), wherein RMSE is that root-mean-square error, θ are the spectrum angle, and RMSE and θ computing formula are as follows:

$\mathrm{RMSE}=\sqrt{\frac{{(R-R^{\′})}^T(R-R^{\′})}{L}},$ $\mathrm{\θ}=\arccos \left(\frac{R^T{R}^{\′}}{\sqrt{R^{T}R}\sqrt{R^{\′T}{R}^{\′}}}\right),$

Wherein R is pixel spectrum, and R' is the modeling spectrum of R;

F) it is as follows that mixed error threshold is separated in setting:

$\left\{\begin{array}{c}{\mathrm{TH}}_{\mathrm{RMSE}}=0.001\\ {\mathrm{TH}}_{\mathrm{\θ}}=0.01\end{array}\right.,$

TH wherein _RMSEBe the root-mean-square error upper limit, TH _θBe the spectrum angle upper limit, verification step e) whether the solution of the end member subclass of each pixel correspondence of calculating mix error and satisfy simultaneously less than TH _RMSEAnd TH _θIf satisfying then corresponding end member subclass is the optimum end member subclass of this pixel, then this pixel is not labeled as the pixel that needs carry out end member optimization search if do not satisfy, all pixels judgements finish in to high-spectral data, obtain carrying out the pixel set omega of end member optimization search;

G) adopt binary coding that Θ is encoded, code word size is N, N is end member sum among the Θ, coded system is as follows: selected a certain end member if participate in the end member subclass of pixel mixing, then this end member correspondence position code word value is " 1 ", otherwise is " 0 ", and C " 1 " is arranged at most in all significant model based coding vectors in the code set, and in the similar end member set one " 1 " is arranged at most, other coding form is insignificant solution;

H) utilize scale-of-two particle swarm optimization algorithm solution procedure f) the optimum end member subclass of each pixel correspondence in the set omega that obtains, wherein use non-negative restriction least square solution to mix square-error 1/RMSE reciprocal ²As fitness, the fitness of the insignificant solution of mentioning in the step g) is set to 0;

I) utilize the optimum end member subclass of each pixel to carry out the abundance estimation, obtain the abundance decision-making;

Wherein, " design fusion rule; the abundance decision-making that the little target reinforcement decision-making that categorised decision, the step (4) that fusion is obtained by step (3) obtained and step (5) are obtained; obtain the integrated classification result " described in the step (6), its implication is described as follows: the design fusion rule, the little target that the categorised decision Cla that fusion steps (3) obtains, step (4) obtain is strengthened decision-making Objs, the abundance decision-making Fras that obtains with step (5), obtain final integrated classification result, concrete grammar is as follows:

$D(x_0,y_0,i)=D_1(x_0,y_0,i)+\underset{t=1}{\overset{3}{\mathrm{\Σ}}}{D}_2^t(x_0,y_0,i)+D_3(x_0,y_0,i),$

Wherein Expression is sought and is made function f (i) maximal parametric i,

Mean i ^*Make f (i) maximum, (x ₀, y ₀) be the position coordinates of any pixel on the image, Lable (x ₀, y ₀) be (x ₀, y ₀) the pixel class label at coordinate position place, strengthen decision-making Objs, abundance decision-making Fras according to categorised decision Cla, little target and find the solution D respectively ₁,

D ₃, concrete computing method are as follows:

A) utilize categorised decision Cla to calculate D ₁, D ₁Be of a size of W * H * C, with D ₁Middle all elements is initialized as 0, if (x, the pixel of y) locating is divided into the i class, then D in Cla ₁(x, y, m)=a _m, m=1,2 ..., C, a _mSatisfy:

$a_m=\left\{\begin{array}{c}{\mathrm{Pr}}_1^i,m=i\\ (1-{\mathrm{Pr}}_1^i)/(C-1),m\≠1\end{array}\right.,$

Parameter wherein Represented the accuracy and stability of class i among the categorised decision Cla,

B) utilizing little target to strengthen decision-making Objs calculates

Be of a size of W * H * C, t=1,2 ..., n, n are the number that little target is strengthened decision-making, will

All elements is initialized as 0, if (x, the pixel of y) locating is identified as i class atural object, then in t Objs M=1,2 ..., C, a _mSatisfy:

$a_m=\left\{\begin{array}{c}{\mathrm{Pr}}_2^i,m=i\\ (1-{\mathrm{Pr}}_2^i)/(C-1),m\≠1\end{array}\right.,$

Parameter wherein Represent little target and strengthened the accuracy and stability of class i among the decision-making Objs,

${\mathrm{Pr}}_2^i\∈\left[\mathrm{0.7,1}\right);$

C) utilize abundance decision-making Fras to calculate D ₃, each pixel abundance vector of Fras all divided by this vector sum, is obtained D ₃

Images