CN103955711B - A kind of mode identification method in imaging spectral target identification analysis - Google Patents

Abstract

A kind of mode identification method in imaging spectral target identification analysis, obtains the pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in testing image pixel；The pretreatment of spectral error is corrected to pixel light spectrum matrix to be predicted, pretreated pixel spectrum two-dimensional matrix is obtained；According to training sample feature weight, the weight Euclidean distance between forecast sample and training sample is calculated, obtain the n in each training sample classification of forecast sample respectively by weight Euclidean distance_cIndividual neighbour, according to these neighbours for being obtained, builds hyperplane respectively；Calculate forecast sample to the minimum range of each hyperplane for building；Judge that forecast sample is closest to which hyperplane, the result of pattern-recognition is exactly that forecast sample belongs to classification belonging to this hyperplane.The present invention solves the problems, such as that mode identification method complexity of the prior art is high low with accuracy of identification.

Description

A kind of mode identification method in imaging spectral target identification analysis

Technical field

The present invention relates to imaging spectral target identification analysis technical field, more particularly to a kind of imaging spectral target identification point Mode identification method in analysis.

Background technology

Imaging spectral is one of most important technological break-through in twentieth century earth observing system, and it overcomes traditional many ripples Section, the multispectral limitation at aspects such as wave band number, wavelength band, fine information expression, it is interval, more with narrower wave band Wave band quantity provides target information, target can be segmented and differentiated from spectral space, is obtained in multiple fields Extensive use.Obtain after imaging spectrometer data information, it is necessary to the spectrum of target is sorted out and differentiated, different materials being reacted Spectrum with component content is classified, and judges the constituent of target, and then judges destination properties.Due to spectrum point Resolution is increased substantially, imaging spectrometer data than multispectral with stronger potential target recognition capability, but in limited training sample Under the conditions of this, increasing for data dimension substantially reduces sample/dimensional ratio.Using conventional statistical pattern recognition method on the contrary Preferably classification results cannot be obtained.Widest method that pattern-recognition used is carried out to imaging spectrometer data at present for refreshing Through network and support vector machine method.During yet with the treatment recognized in imaging spectral target classification, " foreign matter is frequently encountered With spectrum " or " the different spectrum of jljl " phenomenon, i.e., the similar feature of different target optical spectral data presentations, same target is due to certain original Because showing different spectral signatures, this causes that multimodal normal distribution is presented the histogram of target more so that neural network classification Algorithm is difficult to restrain, and seriously reduces accuracy of identification.

Presently used SVMs function is typically directly to construct linear classification hyperplane in the input space, but Many problems are not linear separability in the input space, and the classifying face between classification can be more preferable with nonlinear curved surface Description.And the support vector machine method for being based on Non-linear Kernel function calculates complicated, parameter optimization sometimes can be absorbed in endless loop simultaneously And it is computationally intensive, there is limitation in the case where multi-class targets are recognized.

In sum, existing mode identification method, it is unsatisfactory on accuracy of identification and computational complexity, cause Limitation is larger in terms of host computer specific implementation, so as to cause the reduction of pattern-recognition precision.

The content of the invention

The technology of the present invention solve problem：Overcome the deficiencies in the prior art, there is provided in a kind of imaging spectral target identification analysis Mode identification method, to solve the problems, such as that mode identification method complexity of the prior art is high low with accuracy of identification.

The mode identification method that the present invention is provided calculate between forecast sample and training sample apart from when, it is contemplated that instruction Practice sample characteristics weight, and weight Euclidean distance is introduced when the neighbour of sample is calculated with hyperplane, weighted value is all instructions The ratio between difference of characteristic value in the difference and group of characteristic value, highlights the big wavelength characteristic value of contribution between white silk sample group.By weight Europe , closer to actual mathematical model, as a result precision is of a relatively high for the hyperplane that neighbour's sample that formula distance is selected builds.

To reach above-mentioned purpose, the technical proposal of the invention is realized in this way：

Step 101, obtains the pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in testing image pixel；

Step 102, the pretreatment of spectral error is corrected to pixel spectrum two-dimensional matrix to be predicted, is pre-processed Prediction pixel spectrum two-dimensional matrix afterwards；

Step 103, is to also pass through in forecast sample, with library of spectra by pretreated prediction pixel spectrum two-dimensional matrix Above-mentioned pretreated known class spectrum is training sample, carries out pattern-recognition matching；Calculate training sample feature weight；

Step 104, according to training sample feature weight, calculate weight between forecast sample and training sample it is European away from From obtaining the n in each training sample classification of forecast sample respectively by weight Euclidean distance_cIndividual neighbour, according to being obtained These neighbours, hyperplane is built respectively；

Step 105, calculates forecast sample to the minimum range of each hyperplane for building；

Step 106, judges that forecast sample is closest to which hyperplane, and the result of pattern-recognition is exactly forecast sample category In the classification belonging to this hyperplane.

The pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in testing image pixel, tool are obtained in the step 101 Body includes：

Three-dimensional imaging spectroscopic data in the image picture elements to be predicted is expressed as the pixel light of imaging spectral reflectivity Spectrum two-dimensional matrix：

R_m×n=[p₁,p₂,...,p_y×i+j,...,p_x×y],0≤i≤x,0≤j≤y

Or R_m×n=[p₁,p₂,...,p_i+x×j,...,p_x×y],0≤i≤x,0≤j≤y

Wherein, R_m×nRepresent pixel spectrum two-dimensional matrix, [p₁,p₂,...,p_y×i+j,...,p_x×y] and [p₁,p₂,..., p_i+x×j,...,p_x×y] the pixel spectrum vector of testing image is represented, m represents wave band number, and n represents the sum of testing image pixel, X represents the line number of testing image pixel, and y represents the columns of testing image pixel, n=x × y.

The pretreatment of spectral error is corrected in the step 102 to pixel spectrum two-dimensional matrix to be predicted, is obtained The method of pretreated prediction pixel spectrum two-dimensional matrix uses orthonormal transformation method and Wavelet noise-eliminating method.

The step 103 is that the pretreated prediction pixel spectrum two-dimensional i.e. forecast sample of matrix is same with library of spectra It is that training sample carries out pattern-recognition matching by above-mentioned pretreated known class spectrum；Calculate training sample feature weight It is specific as follows：

If there is training sample set, comprising L sample, J classification, each sample includes d feature, is designated as：x_i =(x_i1,...,x_id)^TIts generic is y_i=c (i=1 ..., L；C=1 ..., J)；

Calculate the feature weight of training sample：

Wherein r_jIt is calculating process intermediate variable；It is the global average of whole training sample jth dimensional features；y_i=c (i= 1,...,L；C=1 ..., J) it is the corresponding classification of training sample；It is the average of c class training sample jth dimensional features；I (y_i) function is represented, work as y_iIt is 1 during=c, is otherwise 0；x_ijIt is training sample x_iJth dimensional feature value；w_jIt is training sample The weighted value of j dimensional features；D is sample characteristics number.

According to training sample feature weight in the step 104, the weight Europe between forecast sample and training sample is calculated Formula distance：

Wherein D (x_i, q) it is the weight Euclidean distance between forecast sample and the sample of given each classification；x_ijFor Training sample jth dimensional feature value；q_jIt is the spectral vector q=(q of the pixel of forecast sample₁,...,q_d)^TJth dimensional feature value；w_j It is the weighted value of training sample jth dimensional feature.

In the step 104 by weight Euclidean distance obtain respectively forecast sample in each training sample classification n_cIndividual neighbour, according to these neighbours for being obtained, builds hyperplane respectively, specific as follows：

V_·i=p_ci-m_c

Wherein LH_cQ () is the set of constructed hyperplane；S is hyperplane；m_cIt is the n for belonging to classification c of forecast sample_c The average value of individual neighbour；p_ciBelong to the neighbour of classification c for forecast sample；V_·iFor the feature between neighbour and neighbour's average value is poor Value；V_.jIt is the column vector of the feature difference between neighbour and neighbour's average value；α is Laplce selected when constructing hyperplane Operator vector；a_iIt is the component of a；n_cIt is artificial selection, n_c>=2, and n_cNo more than training sample number.

The step 105 calculates forecast sample to the minimum range of each hyperplane for building：

W=diag (w₁,...,w_d)

Wherein J_cQ () is Lagrangian minimum range operator；w_jIt is the weighted value of training sample jth dimensional feature；V_j·For near The jth dimensional feature value of the feature difference between adjacent and neighbour's average value；m_cjIt is m_cJth dimensional feature value；q_jIt is forecast sample Jth dimensional feature value；α is selected Laplace operator vector when constructing hyperplane, all of neighbours of V and neighbour's average value it Between feature difference composition vector, W be all training samples feature weight composition diagoned vector；S is hyperplane；Q is Forecast sample；w₁,...,w_dIt is characterized weight；λ is one is used to controlling the α values excessive parameter of possibility, value be usually 0-10 it Between.

Judge that forecast sample is closest to which hyperplane in the step 106, then the result of pattern-recognition is exactly Forecast sample belongs to the classification belonging to this hyperplane：

Label (q)=argmin_cJ_c(q)

Wherein label (q) is forecast sample generic；J_cQ () is Lagrangian minimum range operator；argmin_cJ_c Q () represents makes J_cClassification belonging to the minimum hyperplane of the value of (q).

Present invention advantage compared with prior art is：In imaging spectral target identification analysis provided by the present invention Mode identification method, carries out the judgement of material classification, so as to realize the identification to target constituent from spectrum dimension angle.This hair Bright advantage is that in mode identification procedure, preferably the feature of the known spectra in matching target optical spectrum and library of spectra is believed Breath, the feature of target optical spectrum between protrusion is different classes of, without complicated parameter setting, is that a kind of deterministic algorithm does not have at random Property；Feature is similar between the algorithm can solve sample, the problem of variable redundancy；Calculate forecast sample and hyperplane apart from when consider The weight of feature, and selection weight be characterized wavelength between classification group with group in the ratio between Euclidean distance, highlight contribution compared with Big wavelength characteristic value so that pattern recognition result precision is higher.Additionally, the present invention be directed to the pattern-recognition that spectrum dimension is carried out Operation, computing is carried out after can gathering a spectrum for pixel point immediately, and the target optical spectrum without waiting for whole region is all gathered Complete, so as to better meet requirement of the target identification in real-time.

Brief description of the drawings

Fig. 1 is the flow chart of mode identification method of the present invention；

Fig. 2 is the schematic diagram of the original image spectrum picture of the embodiment of the present invention；

Fig. 3 is the schematic diagram of the area image to be predicted of the embodiment of the present invention；

Fig. 4 shows for the pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in the testing image pixel of the embodiment of the present invention It is intended to.Wherein：Abscissa is the wavelength of not pretreated forecast sample, and ordinate is the anti-of not pretreated forecast sample Radiance rate value；

Fig. 5 is the pretreated prediction pixel spectrum two-dimensional matrix schematic diagram of the embodiment of the present invention.Wherein：Abscissa is By the wavelength of pretreated forecast sample, ordinate is the relative intensity by pretreated forecast sample.

Specific embodiment

The technical solution of the present invention is further elaborated with specific embodiment below in conjunction with the accompanying drawings.

The flow chart of the mode identification method in imaging spectral target identification analysis provided by the present invention, as shown in figure 1, Mainly include the following steps that：

Step 101, obtains the pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in testing image pixel.

The three-dimensional imaging spectroscopic data of the testing image pixel of acquisition is expressed as the pixel spectrum of imaging spectral reflectivity Two-dimensional matrix, it is as follows：

R_m×n=[p₁,p₂,...,p_y×i+j,...,p_x×y],0≤i≤x,0≤j≤y (1)

Or R_m×n=[p₁,p₂,...,p_i+x×j,...,p_x×y],0≤i≤x,0≤j≤y (2)

Wherein, (1) formula is the representation that image picture elements to be predicted are launched by row, and (2) formula is pressed for image picture elements to be predicted Arrange the representation for launching, R_m×nPixel spectrum two-dimensional matrix is represented, m represents wave band number, and n represents the total of image picture elements to be predicted Number, x represents the line number of image picture elements to be predicted, and y represents the columns of image picture elements to be predicted, n=x × y.[p₁,p₂,..., p_y×i+j,...,p_x×y] and [p₁,p₂,...,p_i+x×j,...,p_x×y] the pixel spectrum vector of image to be predicted is represented, by row exhibition The R for opening_m×nIn, p_y×i+jRepresent the spectrum vector of the i-th row jth row correspondence pixel in image to be predicted.The spectrum arrow of each pixel Amount includes reflectance value of the pixel at each wave band, for example：Assuming that h-th spectrum vector of pixel is p_h, then p_h= [p_h1,p_h2...p_hk...p_hm]^T, [p_h1,p_h2...p_hk...p_hm]^TRepresent [p_h1,p_h2...p_hk...p_hm] transposed matrix, wherein p_hkRepresent reflectance value of h-th pixel at k-th wave band.What is obtained due to imaging spectrometer is image picture elements point at each Reflectance value at wave band, therefore the reflectance value is known quantity.

Step 102, the pretreatment of spectral error is corrected to pixel light spectrum matrix to be predicted, obtains pretreated Prediction pixel spectrum two-dimensional matrix.

The purpose pre-processed to pretreated pixel spectrum two-dimensional matrix is corrected because atmospheric scattering etc. causes Noise and spectral error, the method for pretreatment is orthonormal transformation and wavelet transformation.

Wherein, the formula of orthonormal transformation treatment is as follows：

Wherein, p_hk,snvRepresent by after orthogonal transformation treatment in image to be predicted h-th pixel in the anti-of k-th wave band The value of rate is penetrated, m represents wave band number, and m-1 represents the free degree,H-th pixel is anti-at each wave band in representing image to be predicted Penetrate the average value of rate；p_hkFor in prognostic chart picture h-th pixel in k-th value of the reflectivity of wave band.

Small wave converting method principle used and step are as follows：

The wavelet transformation of discrete series is as follows:

F (t) is the discrete series expression formula of signal, ψ_J,kT () is wavelet basis function, c_J,kIt is that J layers of spectral signal is Number, is also low frequency coefficient, d_j,kIt is the high frequency coefficient of jth spectral signal；T is time series.

First, a wavelet function and decomposition scale are selected.

Then, the high frequency coefficient after wavelet decomposition is processed with the method for thresholding.The present invention is set using the method for soft-threshold Determine threshold value, it is specific as follows：

Wherein N is the level of wavelet decomposition；σ is the standard deviation of noise signal；w_j,kAt the beginning of the threshold value for just starting setting Value；It is the threshold value obtained after calculating；λ is soft-threshold；

Finally, spectrum samples spectrum p to be predicted is reconstructed according to formula (4)_hk,snv, remove the spectrum representated by high frequency coefficient Signal, retains the spectral signal representated by low frequency coefficient, obtains new spectral signal q_hk；.

After being pre-processed, accurate spectral information after correction spectrum can be obtained.It is pointed out that this hair Bright preprocess method is not limited solely to above two processing method, the spectrum that other any energy corrections cause by atmospheric scattering The method of error should also belong to protection scope of the present invention with denoising method.

Step 103, is to also pass through in forecast sample, with library of spectra by pretreated prediction pixel spectrum two-dimensional matrix Above-mentioned pretreated known class spectrum is training sample, carries out pattern-recognition matching；Calculate training sample feature weight；

If there is training sample set, comprising L sample, J classification, each sample includes d feature, is designated as：x_i =(x_i1,...,x_id)^TIts generic is y_i=c (i=1 ..., L；C=1 ..., J)；

Calculate the feature weight of training sample：

Wherein r_jIt is calculating process intermediate variable；It is the global average of whole training sample jth dimensional features；y_i=c (i= 1,...,L；C=1 ..., J), it is the corresponding classification of training sample；It is the average of c class training sample jth dimensional features；I (y_i) function is represented, work as y_iIt is 1 during=c, is otherwise 0；x_ijIt is training sample x_iJth dimensional feature value；w_jIt is training sample The weighted value of j dimensional features；D is sample characteristics number.

Step 104, according to training sample feature weight, calculate weight between forecast sample and training sample it is European away from From：

Wherein D (x_i, q) it is the weight Euclidean distance between forecast sample and the sample of given each classification；x_ijFor Training sample x_iJth dimensional feature value；q_jIt is the spectral vector q=(q of the pixel of forecast sample₁,...,q_d)^TJth dimensional feature value； w_jIt is the weighted value of training sample jth dimensional feature.

Obtain the n in each training sample classification of forecast sample respectively by weight Euclidean distance_cIndividual neighbour, according to These neighbours for being obtained, build hyperplane respectively：

V_·i=p_ci-m_c

Wherein LH_cQ () is the set of constructed hyperplane；S is hyperplane；m_cIt is the n for belonging to classification c of forecast sample_c The average value of individual neighbour；p_ciBelong to the neighbour of classification c for forecast sample；V_·iFor the feature between neighbour and neighbour's average value is poor Value；V_.jIt is the column vector of the feature difference between neighbour and neighbour's average value；α is Laplce selected when constructing hyperplane Operator vector；a_iIt is the component of a；n_cIt is artificial selection, n_c>=2, and n_cNo more than training sample number.

Step 105, calculates forecast sample to the minimum range of each hyperplane for building：

W=diag (w₁,...,w_d) (9)

Wherein J_cQ () is Lagrangian minimum range operator；w_jIt is the weighted value of training sample jth dimensional feature；V_jFor near The jth dimensional feature value of the feature difference between adjacent and neighbour's average value；m_cjIt is m_cJth dimensional feature value；q_jIt is forecast sample Jth dimensional feature value；α is selected Laplace operator vector when constructing hyperplane, all of neighbours of V and neighbour's average value it Between feature difference composition vector, W be all training samples feature weight composition diagoned vector；w₁,...,w_dIt is characterized Weight；S is hyperplane；Q is forecast sample；λ is one is used to controlling the α values excessive parameter of possibility, value be usually 0-10 it Between.

Step 106, judges that forecast sample is closest to which sample hyperplane, then the result of pattern-recognition is exactly pre- Test sample originally belongs to the classification belonging to this sample hyperplane, specific as follows：

Label (q)=argmin_cJ_c(q) (11)

Wherein label (q) is forecast sample q generics；J_cQ () is Lagrangian minimum range operator；argmin_cJ_c Q () represents makes J_cClassification belonging to the minimum hyperplane of the value of (q).

Mode identification method in being analyzed the imaging spectral target identification of foregoing invention with reference to specific embodiment enters One step is elaborated.Imaging spectrometer data used by this example derives from airborne imaging spectrum instrument, and airborne imaging spectrum instrument is to adopt With the spectrometer of push-scanning image mode.The imaging spectrum used in the present embodiment is as shown in Fig. 2 size 1354x 2030 is Pixel, each pixel 224 includes a wavelength, and wave-length coverage from 369.85 nanometers to 2506.81 nanometers, for 10 receive by wavelength interval Rice.Specific identification process is as follows：

The image for taking the pixel shown in Fig. 2 in square frame is the image picture elements to be predicted of the present embodiment, the picture in square frame First sum is n=20 × 20=400 pixel, and pixel spectrum in square frame had in library of spectra determine it is matching Classification, the pixel in square frame belongs to three classifications.Fig. 3 is the three-dimensional imaging schematic diagram of image to be predicted；Fig. 4 is the to be measured of acquisition The pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in image picture elements；

The pretreatment of spectral error is corrected to pixel spectrum two-dimensional matrix to be predicted, method is orthonormal transformation Treatment and Wavelet Denoising Method treatment, obtain pretreated prediction pixel spectrum two-dimensional matrix, as shown in Figure 5；

Pretreated pixel spectrum is identified in library of spectra.There are 15 kinds in the embodiment of the present invention in library of spectra Known different classes of spectrum, including the classification of pixel spectrum is predicted in embodiment.Table 1 is pattern recognition result of the present invention.

Table 1

Mode identification method in imaging spectral target identification analysis provided by the present invention, thing is carried out from spectrum dimension angle The judgement of matter classification, so as to realize the identification to target constituent.The advantage of the invention is that in mode identification procedure, more The characteristic information of the known spectra in good matching target optical spectrum and library of spectra, the spy of target optical spectrum between protrusion is different classes of Levy, be that a kind of deterministic algorithm does not have randomness without complicated parameter setting；Feature is similar between the algorithm can solve sample, The problem of variable redundancy；Calculate forecast sample and hyperplane apart from when consider the weight of feature, and the weight of selection is characterized Wavelength between classification group with group in the ratio between Euclidean distance, highlight the larger wavelength characteristic value of contribution so that pattern-recognition knot Fruit precision is higher.Additionally, the present invention be directed to the pattern-recognition operation that spectrum dimension is carried out, a spectrum for pixel point can be gathered Carry out computing immediately afterwards, the target optical spectrum without waiting for whole region all gathers completion, so as to better meet target identification Requirement in real-time.

Above example is provided just for the sake of the description purpose of the present invention, and is not intended to limit the scope of the present invention.This The scope of invention is defined by the following claims.The various equivalents that do not depart from spirit and principles of the present invention and make and repair Change, all should cover within the scope of the present invention.

Claims (7)

1. the mode identification method during a kind of imaging spectral target identification is analyzed, it is characterised in that comprise the following steps：

Step 101, obtains the pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in testing image pixel；

Step 102, the pretreatment of spectral error is corrected to pixel spectrum two-dimensional matrix to be predicted, obtains pretreated Prediction pixel spectrum two-dimensional matrix；

Step 103, by pretreated prediction pixel spectrum two-dimensional matrix be also pass through in forecast sample, with library of spectra it is above-mentioned Pretreated known class spectrum is training sample, carries out pattern-recognition matching；Calculate training sample feature weight；

Step 104, according to training sample feature weight, calculates the weight Euclidean distance between forecast sample and training sample, leads to Cross the n in each training sample classification that weight Euclidean distance obtains forecast sample respectively_cIndividual neighbour, according to obtained these Neighbour, builds hyperplane respectively；

Step 105, calculates forecast sample to the minimum range of each hyperplane for building；

Step 106, judges that forecast sample is closest to which hyperplane, and the result of pattern-recognition is exactly that forecast sample belongs to this Classification belonging to individual hyperplane；

The step 103 is will to be also passed through in pretreated prediction pixel spectrum two-dimensional matrix i.e. forecast sample and library of spectra Above-mentioned pretreated known class spectrum is that training sample carries out pattern-recognition matching；Calculate training sample feature weight specific It is as follows：

If there is training sample set, comprising L sample, J classification, each sample includes d feature, is designated as：x_i= (x_i1,...,x_id)^T, d feature generic is y_i=c (i=1 ..., L；C=1 ..., J)；

Calculate the feature weight of training sample：

$r_j=\frac{{\mathrm{\Σ}}_i{\mathrm{\Σ}}_{c}I(y_i=c){({\stackrel{\‾}{x}}_{cj}-{\stackrel{\‾}{x}}_j)}^2}{{\mathrm{\Σ}}_i{\mathrm{\Σ}}_{c}I(y_i=c){(x_{ij}-{\stackrel{\‾}{x}}_{cj})}^2}$

$\begin{array}{cc}{w}_j=\frac{\exp \left(r_j\right)}{{\mathrm{\Σ}}_{j=1}^d\exp \left(r_j\right)}& \∀j=1,\mathrm{...},d\end{array}$

Wherein r_jIt is calculating process intermediate variable；It is the global average of whole training sample jth dimensional features；y_i=c (i= 1,...,L；C=1 ..., J) it is the corresponding classification of training sample；It is the average of c class training sample jth dimensional features；I (y_i) function is represented, work as y_iIt is 1 during=c, is otherwise 0；x_ijIt is training sample x_iJth dimensional feature value；w_jIt is training sample The weighted value of j dimensional features；D is sample characteristics number.

2. the mode identification method during imaging spectral target identification is analyzed according to claim 1, it is characterised in that：The step The pixel spectrum two-dimensional matrix of three-dimensional imaging spectroscopic data in testing image pixel is obtained in rapid 101, is specifically included：

Three-dimensional imaging spectroscopic data in the image picture elements to be predicted is expressed as the pixel spectrum two of imaging spectral reflectivity Dimension matrix：

R_m×n=[p₁,p₂,...,p_y×i+j,...,p_x×y],0≤i≤x,0≤j≤y

Or R_m×n=[p₁,p₂,...,p_i+x×j,...,p_x×y],0≤i≤x,0≤j≤y

Wherein, R_m×nRepresent pixel spectrum two-dimensional matrix, [p₁,p₂,...,p_y×i+j,...,p_x×y] and [p₁,p₂,..., p_i+x×j,...,p_x×y] the pixel spectrum vector of testing image is represented, m represents wave band number, and n represents the sum of testing image pixel, X represents the line number of testing image pixel, and y represents the columns of testing image pixel, n=x × y.

3. the mode identification method during imaging spectral target identification is analyzed according to claim 1, it is characterised in that：The step The pretreatment of spectral error is corrected in rapid 102 to pixel spectrum two-dimensional matrix to be predicted, pretreated prediction is obtained The method of pixel spectrum two-dimensional matrix uses orthonormal transformation method and Wavelet noise-eliminating method.

4. the mode identification method during imaging spectral target identification is analyzed according to claim 1, it is characterised in that：The step According to training sample feature weight in rapid 104, the weight Euclidean distance between forecast sample and training sample is calculated：

$D(x_i,q)=\sqrt{\underset{j=1}{\overset{d}{\Σ}}{w}_j{(x_{ij}-q_j)}^2}$

Wherein D (x_i, q) it is the weight Euclidean distance between forecast sample and the sample of given each classification；x_ijIt is training Sample jth dimensional feature value；q_jIt is the spectral vector q=(q of the pixel of forecast sample₁,...,q_d)^TJth dimensional feature value；w_jIt is instruction Practice the weighted value of sample jth dimensional feature.

5. the mode identification method during imaging spectral target identification is analyzed according to claim 1, it is characterised in that：The step Obtain the n in each training sample classification of forecast sample in rapid 104 respectively by weight Euclidean distance_cIndividual neighbour, according to These neighbours for being obtained, build hyperplane respectively, specific as follows：

${\mathrm{LH}}_c\left(q\right)=\left\{s\right|s=\underset{i=1}{\overset{n_c}{\Σ}}{\mathrm{\α}}_i{V}_{\·j}+m_c\}$

$m_c=\frac{1}{n_c}\underset{i=1}{\overset{n_c}{\Σ}}{p}_{ci}$

V_·i=p_ci-m_c

$\mathrm{\α}={({\mathrm{\α}}_1,\mathrm{...},{\mathrm{\α}}_{n_c})}^T$

Wherein LH_cQ () is the set of constructed hyperplane；S is hyperplane；m_cIt is the n for belonging to classification c of forecast sample_cIt is individual near Adjacent average value；p_ciBelong to the neighbour of classification c for forecast sample；V_·iIt is the feature difference between neighbour and neighbour's average value；V_.j It is the column vector of the feature difference between neighbour and neighbour's average value；α is Laplace operator selected when constructing hyperplane Vector；a_iIt is the component of a；n_cIt is artificial selection, n_c>=2, and n_cNo more than training sample number.

6. the mode identification method during imaging spectral target identification is analyzed according to claim 1, it is characterised in that：The step Rapid 105 calculate forecast sample to the minimum range of each hyperplane for building：

$\begin{array}{c}{J}_c=\underset{\mathrm{\α}}{\min }\underset{j=1}{\overset{d}{\Σ}}{w}_j{(V_{j\·}\mathrm{\α}+m_{cj}-q_j)}^2+{\mathrm{\λ\α}}^T\mathrm{\α}\\ =\underset{\mathrm{\α}}{\min }{(s-q)}^{T}W(s-q)+{\mathrm{\λ\α}}^T\mathrm{\α}\end{array}$

W=diag (w₁,...,w_d)

Wherein J_cQ () is Lagrangian minimum range operator；w_jIt is the weighted value of training sample jth dimensional feature；V_j·For neighbour with The jth dimensional feature value of the feature difference between neighbour's average value；m_cjIt is m_cJth dimensional feature value；q_jFor the jth of forecast sample is tieed up Characteristic value；α is Laplace operator vector selected when constructing hyperplane, between all of neighbours of V and neighbour's average value The vector of feature difference composition, W is the diagoned vector of the feature weight composition of all training samples；S is hyperplane；Q is prediction Sample；w₁,...,w_dIt is characterized weight；λ is one to be used to control the possible excessive parameter of α values, and value is 0-10.

7. the mode identification method during imaging spectral target identification is analyzed according to claim 1, it is characterised in that：The step Judge that forecast sample is closest to which hyperplane in rapid 106, then the result of pattern-recognition is exactly that forecast sample belongs to this Classification belonging to individual hyperplane：

Label (q)=argmin_cJ_c(q)

Wherein label (q) is forecast sample generic；J_cQ () is Lagrangian minimum range operator；argmin_cJ_c(q) generation Table makes J_cClassification belonging to the minimum hyperplane of the value of (q).