CN106156787A - Multi-modal Wetland ecological habitat scene nuclear space source tracing method and device - Google Patents

Abstract

The present invention discloses a kind of multi-modal Wetland ecological habitat scene nuclear space source tracing method and device, and described method includes feature extraction: use space to enlist the services of the global and local feature in method and Scale invariant features transform method extraction wetlands ecosystems habitat image respectively；Feature pretreatment: the feature of extraction is normalized and centralization processes；Feature couples: use core principle component analysis method to be coupled to nuclear space by heterogeneous Projection Character；Trace to the source the stage: use large-spacing k-nearest neighbor to determine the ownership of Wetland ecological habitat scene coupling composition characteristics in nuclear space.Described device includes imaging device, movable flashing device, CF card reader and PC.The present invention is compared with the existing methods, its in the nuclear space effective integration isomery mode Wetland ecological habitat characteristics information of being particularly advantageous in that also determines the ownership of Wetland ecological habitat scene, improve the precision of source tracing method, have a good application prospect and considerable market value.

Description

Multi-modal Wetland ecological habitat scene nuclear space source tracing method and device

Technical field

The present invention relates to ecological calculating field, particularly to a kind of multi-modal Wetland ecological habitat scene nuclear space source tracing method and device.

Background technology

China is one of country that Wetland Biodiversity is the abundantest in the world, is also the country that Asia Wetland Type is the most complete, quantity is most, area is maximum.The research work of Chinese Wetland Ecological system protection is for protection Asia or even global range wetlands ecosystems and maintains region wetlands ecosystems balance will play an important role and meaning.The most international relevant academic society has carried out substantial amounts of relevant research work, but the feature due to wetlands ecosystems habitat diversification, complexity etc., causing part research work to carry out the most well, wherein wetlands ecosystems habitat scene is traced to the source to study and is relatively lagged behind the most always.

Research majority is traced to the source based on single mode image feature in traditional wetlands ecosystems habitat, the general effect considering the composite factors such as the plant in the image of habitats, biology, the hydrology, geography more, but ignore the effect of local factor, cause wetlands ecosystems Habitat Types precision of tracing to the source relatively low；Additionally, owing to wetlands ecosystems habitat scene information is relatively large with the change of observation place and angle, Wetland ecological scene global is caused to be prone to change with local message structure, the method of the combination of simple multiple features often makes the modal characteristics of these isomeries show relatively large dereferenced characteristic in luv space, and then cause traditional linear clustering algorithm cannot be accurately positioned the ownership of heterogeneous characteristic, thus reduce the precision of tracing to the source of wetlands ecosystems habitat information.Originally research and propose and utilize multi-modal wetlands ecosystems ecology habitat scene nuclear space source tracing method to build overall situation and partial situation's information efficient coupling mechanism, its detailed information is taken into full account while utilizing wetlands ecosystems habitat global information in nuclear space, feature to be traced to the source is input to large-spacing k-nearest neighbor again and compares with priori ecological characteristic in data base, Wetland ecological habitat nonlinear characteristic is classified after nuclear space linearisation by large-spacing nearest neighbour classification device, ensure the final effect identified, result according to grader output determines the ecological habitat ownership of the Wetland ecological image traced to the source.

This research with China 3 class representational wetlands ecosystems object, including: Yancheng, Jiangsu Province coastal tidal wetland, Area, Zoige, Sichuan highland and severe cold wetland and Hainan Dong Zhaigang seashore Mangrove Wetlands are example, verify the method proposed.The present invention is compared with the existing methods, it is particularly advantageous in that in nuclear space effective integration isomery mode Wetland ecological habitat characteristics information and determines the ownership of Wetland ecological scene, large-spacing k-nearest neighbor is used to be mapped in higher-dimension nucleus lesion by the wetlands ecosystems habitat scene coupling feature extracted, the semantic scene ownership of sample is divided in higher-dimension nucleus lesion, improve the precision of tracing to the source of wetlands ecosystems habitat scene, have a good application prospect and considerable market value.

Summary of the invention

The invention aims to solve the problem that existing wetlands ecosystems habitat scene tracing technology exists, especially for the effectiveness and the nonlinear characteristic nicety of grading problem that solve the coupling of present stage isomery mode wetlands ecosystems habitat characteristics information, it is provided that a kind of multi-modal Wetland ecological habitat scene nuclear space source tracing method and device.

The present invention multi-modal Wetland ecological habitat scene nuclear space source tracing method, the method comprises the steps:

Step 1, uses imaging device to obtain different classes of wetlands ecosystems habitat scene image sample, is stored in movable flashing device by image sample；

Step 2, by CF card reader, in the wetlands ecosystems habitat scene image sample of storage imports to the magnetic disk media of PC in step 1 movable flashing device, sets up wetlands ecosystems habitat scene database；

Step 3, uses the lower space of optimal resolution configuration in null tone territory to enlist the services of method and extracts wetlands ecosystems habitat scene global space networks roth's sign；

Step 4, uses Scale invariant features transform method to extract wetlands ecosystems habitat scene local scale invariant features transform characteristics；

Step 5, is normalized the wetlands ecosystems habitat scene global space networks roth's sign extracted in step 3 and 4 and local Scale invariant features transform feature respectively and processes with centralization；

Step 6, the global space after using core principle component analysis method step 5 to be processed is enlisted the services of feature with local Scale invariant features transform Projection Character to higher-dimension nucleus lesion linearisation, and the feature of this two classes different modalities is carried out linear coupling；

Step 7, uses large-spacing k-nearest neighbor the wetlands ecosystems habitat scene characteristic of linear coupling in higher-dimension nucleus lesion described in step 6 to be mapped in higher-dimension nucleus lesion, the semantic scene ownership of partition database sample in higher-dimension nucleus lesion；

Step 8, for wetlands ecosystems habitat scene to be traced to the source, use step 37 to treat sample of tracing to the source successively and carry out feature extraction, feature pretreatment and feature coupling, then it is mapped in the higher-dimension nucleus lesion of priori scene, in higher-dimension nucleus lesion, calculate the semantic similarity between priori scene coupling feature in wetlands ecosystems habitat scene database described in wetlands ecosystems habitat scene coupling feature to be traced to the source and step 7, and then infer the semantic attaching information of wetlands ecosystems habitat scene image.

In described step 1, the object of wetlands ecosystems habitat scene refers to: Yancheng, Jiangsu Province coastal tidal wetland, Area, Zoige, Sichuan highland and severe cold wetland and Hainan Dong Zhaigang seashore Mangrove Wetlands.

In described step 2, set up wetlands ecosystems habitat scene database and include 3 class Wetland ecological habitat scene image collection: Yancheng, Jiangsu Province coastal tidal wetland image set, Area, Zoige, Sichuan highland and severe cold wetland image set and Hainan Dong Zhaigang seashore Mangrove Wetlands image set.

In described step 3, use the lower space of optimal resolution configuration in null tone territory to enlist the services of method and extract the semantic feature of wetlands ecosystems habitat scene；In described null tone territory, the multiresolution 2D-Gabor bank of filters of the spatial frequency domain nearly orthogonal of wave filter employing " Flos Chrysanthemi shape " of method is enlisted the services of in the space under optimal resolution configuration.

In described step 4, Scale invariant features transform method uses Gaussian difference scale space method to carry out feature extreme point in the Wetland ecological habitat scene image of detection and positioning stablity, generates Wetland ecological habitat scene local feature description according to specific characteristic point directioin parameter.

In described step 5, by normalized, processing in the range of data are mapped to [-1,1], normalized algorithm is:

${\hat{x}}_i=2\·\left(\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}\right)-1$

Wherein, x_iRepresent that the i-th row treat normalization characteristic vector, x_minRepresent minima in column vector, x_maxRepresent maximum in column vector,Represent the column vector after the i-th row normalization.If maximum is equal to minima x in column vector_max=x_min, then column vector value is constant

In described step 5, centralization processes, and refers to that variable deducts its mathematical expectation: for sample data, refer to each observation of sample variableDeducting the meansigma methods of the sample of this sample variable, specific algorithm is as follows:

${\stackrel{\‾}{x}}_{\mathrm{ij}}={\hat{x}}_{\mathrm{ij}}-\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\hat{x}}_{\mathrm{ij}}$

Wherein, i correspondence sample, j represents total sample number to dependent variable, M.

In described step 6, k-fold cross-validation method is utilized to obtain the highest accuracy of identification to determine wetlands ecosystems habitat scene core principle component number in the modelling phase.

In described step 7, when inferring the semantic attaching information of wetlands ecosystems habitat scene, use Mahalanobis measure sample estimates similarity in higher dimensional space.

In described step 7, when inferring the semantic attaching information of wetlands ecosystems habitat scene, the parameter of the similar number of large-spacing k-nearest neighbor Feature Semantics is selected by user to set or by system recommendation voluntarily.

The device of tracing to the source of multi-modal Wetland ecological habitat scene nuclear space source tracing method, including imaging device, movable flashing device, CF card reader and PC, imaging device is used for gathering wetlands ecosystems habitat scene image sample and storing to movable flashing device, and CF card reader is for being directed into PC by the wetlands ecosystems habitat scene image sample in movable flashing device；Described imaging device uses varifocal high-resolution Complimentary Metal-Oxide semiconductor device quiescent imaging device.

The present invention is compared with the existing methods, the present invention utilizes the lower space of optimal resolution configuration in null tone territory to enlist the services of method and extracts wetlands ecosystems habitat scene global semantic feature, and the frequency band eliminating nonopiate multi-scale filtering device overlaps the impact of the redundancy feature output caused and the impact that part minutia information is capped at the filtered band place of overlapping；Utilize feature extreme point in the Wetland ecological habitat scene image of the detection of Scale invariant features transform method and positioning stablity, generate Wetland ecological habitat scene local feature description according to specific characteristic point directioin parameter；Normalization processing method is used to remove space networks roth's sign and the impact of Scale invariant features transform feature dimension；Data center's processing method is used to eliminate space networks roth's sign and the impact of Scale invariant features transform characteristic mean drift；By core principle component analysis method by pretreated non-linear heterogeneous Projection Character to higher-dimension nucleus lesion linearisation, in the high dimension linear space of isomorphism, multi-source heterogeneous Wetland ecological habitat characteristics is carried out linear coupling；Use large-spacing nearest neighbor algorithm to be mapped by coupling feature in higher-dimension nucleus lesion and carry out Semantic Similarity Measurement, the impact of the similarity deviation that algorithm of tracing to the source causes due to the non-linearity of characterization factor is eliminated when calculating semantic similarity between sample in higher-dimension nucleus lesion, the ownership positioning wetlands ecosystems habitat characteristics in higher-dimension core Feature Semantics space of sample semantic similitude parameter that is that specify finally according to user or that generated by system recommendation, and then determine the semantic ownership of wetlands ecosystems habitat scene, improve the precision of tracing to the source of wetlands ecosystems habitat scene.

Accompanying drawing explanation

Fig. 1 is the flow chart of the inventive method；

Fig. 2 is that the inventive method realizes schematic diagram.Wherein, symbol " ■ " expression Scale invariant features transform feature, symbol " ▲ " representation space net for catching fish or birds feature, symbol "●" represents Yancheng wetland coupling feature, and symbol " ◆ " represents Zoige Wetland coupling feature, symbolRepresent Dong Zhai port wetland coupling feature.

Detailed description of the invention

In order to make the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and the present invention is described in more detail referring to the drawings.

The present invention utilizes the lower space of optimal resolution configuration in null tone territory to enlist the services of method and extracts wetlands ecosystems habitat scene global semantic feature；Scale invariant features transform method is utilized to generate Wetland ecological habitat scene local feature description；Normalization processing method is used to remove space networks roth's sign and the impact of Scale invariant features transform feature dimension；Data center's processing method is used to eliminate space networks roth's sign and the impact of Scale invariant features transform characteristic mean drift；By core principle component analysis method by pretreated non-linear heterogeneous Projection Character to higher-dimension nucleus lesion linearisation, in the high dimension linear space of isomorphism, multi-source heterogeneous Wetland ecological habitat characteristics is carried out linear coupling；Use large-spacing nearest neighbor algorithm to be mapped by coupling feature in high-dimensional feature space and carry out Semantic Similarity Measurement, the impact of the similarity deviation that algorithm of tracing to the source causes due to the non-linearity of characterization factor is eliminated when calculating semantic similarity between sample in high-dimensional feature space, the ownership positioning wetlands ecosystems habitat characteristics in high dimensional feature semantic space of sample semantic similitude parameter that is that specify finally according to user or that generated by system recommendation, and then determine the semantic ownership of wetlands ecosystems habitat scene.

Fig. 1 is the flow chart of the inventive method, Fig. 2 be the inventive method realize schematic diagram, as depicted in figs. 1 and 2, a kind of multi-modal Wetland ecological habitat scene nuclear space source tracing method and device include following step:

Step 1, uses varifocal high-resolution Complimentary Metal-Oxide semiconductor device quiescent imaging device to obtain different classes of wetlands ecosystems habitat scene image sample, is stored by image sample in Large Copacity movable flashing device；

Step 2, by CF card reader, imports to, in the large capacity disc medium in PC, set up wetlands ecosystems habitat scene image sample database by the ecosystem habitat scene image sample in Large Copacity movable flashing device in step 1.Wherein, traced to the source wetlands ecosystems habitat image sample is 3 apoplexy due to endogenous wind state representational wetlands ecosystems objects, including: Yancheng, Jiangsu Province coastal tidal wetland, Area, Zoige, Sichuan highland and severe cold wetland and Hainan Dong Zhaigang seashore Mangrove Wetlands.

All very color RGB image are converted into the image of intensity level Lycoperdon polymorphum Vitt yardstick between 0～255 by step 3, and conversion formula is:

Gray=0.2989R+0.5870G+0.1140B (1)

Wherein: R represents that red light intensity, G represent that green intensity, B represent blue light strength.

Step 4, the wetlands ecosystems habitat scene global space networks roth's sign of the lower multi-scale filtering device of extraction given optimal resolution configuration；Wherein, described wave filter uses the band of the two dimension of Sine Modulated to lead to 2D Gabor function, and its expression formula is:

$g(x,y)=\exp [-\frac{1}{2}(\frac{x^2}{{\mathrm{\σ}}_x^2}+\frac{y^2}{{\mathrm{\σ}}_y^2})]\·\cos \left(2\mathrm{\π}{u}_{0}x\right)---\left(2\right)$

Wherein, σ_xAnd σ_yIt is the bandwidth of Gaussian function in the spatial domain, u₀It it is the modulating frequency of cosine function；Its Fourier domain expression formula is:

$G(u,v)=\mathrm{\λ}\·\exp [-\frac{1}{2}(\frac{{\left({u-u}_0\right)}^2}{{\mathrm{\σ}}_u^2}+\frac{v^2}{{\mathrm{\σ}}_y^2})]---\left(3\right)$

Wherein, u and v is spatial frequency variable；σ_u=1/2 π σ_x, σ_v=1/2 π σ_yWith λ=2 π σ_xσ_y.Above-mentioned 2D-Gabor function is tolerable small echo, but it can't produce Orthogonal Decomposition.Therefore the image after wavelet transformation based on 2D-Gabor function has redundancy.

The present invention utilizes the optimum multiresolution 2D-Gabor wave filter of one " Flos Chrysanthemi shape ", this bank of filters is nearly orthogonal, thus ensure can contact with each other at the upper half peak amplitude of spatial frequency domain but do not overlap each other so that the output amount of redundancy of the Gabor filter under multiresolution is minimum.Function g (x, y) as morther wavelet, builds the bank of filters of self similarity by suitable yardstick and rotation transformation:

g′_mn(x, y)=τ^mg(x′,y′) (4)

Wherein, x '=τ^m(xcos θ+ysin θ), y '=τ^m(-xsin θ+ycos θ), and θ=n π/K, n ∈ [0,1 ..., N], N is the direction sum rotated.M ∈ [0,1 ... M], M is the yardstick sum in same direction of rotation.

τ is scale factor, can be calculated by equation below:

$\mathrm{\τ}={\left(\frac{U_{\max }}{U_{\min }}\right)}^{\frac{1}{M-1}}---\left(5\right)$

Wherein, U_maxAnd U_minIt is respectively minimum and maximum radial center frequency.Frequency coordinate after conversion can calculate (u ', v ')=(ucos θ+vsin θ ,-usin θ+vcos θ) with following formula.Scale parameter σ in Fourier domain in equation_uAnd σ_vCan be calculated as follows:

$\left\{\begin{array}{c}{\mathrm{\σ}}_u{|}_m={\mathrm{\τ}}^{m-1}\·\frac{(a-1)}{(a+1)}\·\frac{U_{\min }}{\sqrt{2\ln 2}}\\ {\mathrm{\σ}}_v{|}_m=\tan (\mathrm{\π}/2N)\sqrt{\frac{{(U_{\min }\·{\mathrm{\τ}}^{m-1})}^2}{2\ln 2}-{{\mathrm{\σ}}_u}^2{|}_m}\end{array}\right.---\left(6\right)$

By image be divided into 16 pieces by the grid of 4 × 4, calculate the average statistic histogram of each block of image filter response on different directions different scale, wherein filtering its direction number is N=8, scale parameter is M=4, the final 512 dimension semantic feature vectors obtaining the naturalness of description wetlands ecosystems habitat scene information, openness, roughness, swelling degree, dangerously steep degree etc..

Step 5, extraction given wetlands ecosystems habitat image local scale invariant features transform characteristics；Wherein, described extraction gives the key step of wetlands ecosystems habitat image local feature method and includes: detects and positions yardstick spatial extrema point, specific characteristic point directioin parameter and generate characteristic point and describe son.Algorithm mesoscale space L is by image I and gaussian kernelConvolution obtains:

$L(x,y,\mathrm{\σ})=G(x,y,\mathrm{\σ})\⊗I(x,y)---\left(7\right)$

Algorithm have employed Gaussian difference scale space to detect stable characteristic point:

D (x, y, σ)=L (x, y, k σ)-L (x, y, σ) (8)

In formula: k is constant, the gradient direction distribution characteristic utilizing characteristic point neighborhood territory pixel is each characteristic point assigned direction parameter, makes operator possess rotational invariance.

Then calculate pixel (x, y) place's gradient modulus value m (x, y) and direction H (x, y):

$\left\{\begin{array}{c}m(x,y)=\sqrt{f_x{(x,y)}^2+f_y{(x,y)}^2}\\ \mathrm{\θ}(x,y)={\tan }^{-1}{f}_y(x,y)/f_x(x,y)\end{array}\right.---\left(9\right)$

In formula, f_x(x, y)=L (and x+1, y)-L (x-1, y), f_y(x, y)=L (x, y+1)-L (x, y-1).The coordinate axes of image is rotated the principal direction to characteristic point by described algorithm, takes the window of 16 × 16 pixel sizes, then this window is evenly divided into 4 × 4 subwindows centered by characteristic point.In each subwindow of 4 × 4, calculate the gradient orientation histogram in 8 directions, and add up the aggregate-value of each gradient direction, obtain a seed points.Each characteristic point is made up of 4 × 4 seed points, and each seed points has the vector value in 8 directions, material is thus formed the characteristic vector of 128 dimensions.

Step 6, is normalized respectively to the space networks roth's sign under optimal resolution configuration in the null tone territory extracted and Scale invariant features transform feature, removes the impact of heterogeneous characteristic dimension.As follows for the Processing Algorithm in the range of the feature normalization of each sample to [-1,1]:

${\hat{x}}_j=2\·\left(\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}\right)-1---\left(10\right)$

In formula, x_jRepresent that jth row treat normalization characteristic vector, x_minRepresent minima in column vector, x_maxRepresent maximum in column vector,Represent the column vector after jth row normalization.If maximum is equal to minima x in column vector_max=x_min, then column vector value is constant

Step 7, uses centralization Processing Algorithm to eliminate space networks roth's sign and the impact of Scale invariant features transform characteristic mean drift.For sample data, refer to each observation of a variableDeducting the sample mean of this variable, specific algorithm is as follows:

${\stackrel{\‾}{x}}_{\mathrm{ij}}={\hat{x}}_{\mathrm{ij}}-\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\hat{x}}_{\mathrm{ij}}---\left(11\right)$

Wherein, i correspondence sample, j represents total sample number to dependent variable, M.

Step 8, uses core principle component analysis method to be coupled to higher-dimension nucleus lesion by the heterogeneous Projection Character after normalized；

Described algorithm is by scale feature X=(x in the input space₁,x₁,...,x_m) (wherein x_i, i=1,2 ..., m is d dimensional vector, and m is the total number of samples in data set) the Hilbert space Γ of higher-dimension (even infinite dimension) is mapped the data into by characteristic function φ:

Feature space calculating characteristic function φ covariance matrix mapping:

$C=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\φ}{\left(x_i\right)}^T\mathrm{\φ}\left(x_i\right)---\left(12\right)$

By defining its characteristic equation:

C ν=λ v (13)

Wherein,Characteristic vector and eigenvalue with λ corresponding covariance matrix respectively.By Eigenvalues Decomposition and a journey factor alpha can be solved.

Its core principle component can be sought finally, for any one test sample x:

$(v,\mathrm{\φ}\left(x\right))=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i\left(\mathrm{\φ}{\left(x_i\right)}^T\mathrm{\φ}\left(x\right)\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}k(x_i,x)---\left(14\right)$

Step 9, uses large-spacing nearest neighbor algorithm to infer the attaching information of Wetland ecological image.Wherein, the target that described large-spacing nearest neighbor algorithm first optimizes is to realize input sample x_iWith minimizing of the average distance of its target neighbor:

$\underset{M}{\min }\underset{i,j\∈N_i}{\mathrm{\Σ}}d(x_i,x_j)=\underset{i,j\∈N_i}{\mathrm{\Σ}}{(x_i-x_j)}^{T}M(x_i-x_j)---\left(15\right)$

In formula, M is positive semidefinite matrix.

Second target optimized is to make input sample x_iTo the distance of its target base and its at least keep the interval of 1 unit to the distance of invasion neighbour:

${\∀}_{i,j\∈N_i,l,y_l\≠y_i}d(x_i,x_j)+1\≤d(x_i,x_j)---\left(16\right)$

Therefore, this optimization problem can be with Integrative expression:

Finally, the ownership of sample is determined according to each nuclear space spacing inputting k neighbouring target of sample.

Embodiment

At present, invention software data base has 900 comprise Yancheng, Jiangsu Province coastal tidal wetland, Area, Zoige, Sichuan highland and severe cold wetland and Hainan Dong Zhaigang seashore Mangrove Wetlands scene image sample, in every each 300 of class sample.Data set is modeling and forecast set two parts by random division, and wherein every class sample collection comprises 240 samples, and collection of tracing to the source comprises 60 samples.

Whole process can be divided into study and trace to the source two stages, and the step in its learning stage is: use the space networks roth's sign method under optimal resolution configuration to extract the overall semantic feature of the naturalness of statement wetlands ecosystems habitat scene of 512 dimensions, openness, roughness, swelling degree, dangerously steep degree etc. in data base；Use Scale invariant features transform method to carry out feature extreme point in the Wetland ecological habitat scene image of detection and positioning stablity, generate Wetland ecological habitat scene local feature description of 128 dimensions according to specific characteristic point directioin parameter；The space networks roth's sign extracted and Scale invariant features transform feature are normalized method and and centralization process；Core principle component analysis method is used to be coupled to nucleus lesion by pretreated heterogeneous Projection Character, wherein nuclear space maps and uses radially gaussian basis kernel function, utilizes modeling to concentrate the highest accuracy of identification obtained to determine wetlands ecosystems habitat scene core principle component number；Use large-spacing nearest neighbor algorithm to be mapped by coupling feature in higher-dimension nucleus lesion and carry out Semantic Similarity Measurement, eliminate the impact caused due to the non-linearity of feature, in high-dimensional feature space, finally divide the semantic ownership of wetlands ecosystems habitat characteristics.The model prediction accuracy of modeling data collection uses k-fold proof method to estimate, it is thus achieved that modeling collection confusion matrix result of calculation as shown in table 1.User can understand data base's sample confidence level according to the precision of prediction returned.

Table 1 modeling collection confusion matrix

In the stage of tracing to the source, extract space networks roth's sign and the graphical rule invariant features transform characteristics of 128 dimensions of the Wetland ecological habitat image of 512 dimensions respectively；The space networks roth's sign extracted and Scale invariant features transform feature are normalized method and and centralization process；Core principle component analysis method is used to calculate normalization series connection feature core principle component；The core principle component of extraction is mapped to contextual data planting modes on sink characteristic apoplexy due to endogenous wind in wetlands ecosystems habitat in higher-dimension nucleus lesion again, calculate the scene characteristic semantic similarity with data base's sample characteristics of tracing to the source, then Scene Semantics Similarity value is ranked up, finally according to image in the top n corresponding wetlands ecosystems habitat scene database that maximum similarity value output user sets, wherein most types reference category as sample of tracing to the source occur according to number of times in specimen types, collection confusion matrix result of calculation of tracing to the source is as shown in table 2.

Table 2 is traced to the source collection confusion matrix

All elements on leading diagonal in modeling collection and collection confusion matrix of tracing to the source is averaged, i.e. can obtain the most total multi-modal Wetland ecological habitat scene nuclear space modeling and the precision of model of tracing to the source, result of calculation is respectively 75.6% and 70.0%, shown in the results list 3, modeling collection and the precision 71.8% and 69.4% of collection of tracing to the source that result of calculation obtains than single mode Wetland ecological habitat scene nuclear space source tracing method improve 3.8 and 0.6 percentage points respectively.

Table 3 single mode and multi-modal modeling collection and collection precision of tracing to the source

Can be seen that from the example above, the present invention is compared with the existing methods, it be particularly advantageous in that its effective integration isomery mode Wetland ecological habitat nonlinear characteristic information in nuclear space, make full use of Wetland ecological scene global attribute character and take into full account relevant local detailed information simultaneously, use large-spacing k-nearest neighbor eliminates the impact of the nonlinear characteristic of higher-dimension coupling feature further, and then improves trace to the source precision and the robustness of Wetland ecological habitat scene.The proposition of the method understanding of the coupled relation between the scene of isomery mode Wetland ecological habitat effectively, compares with traditional wetlands ecosystems ecology habitat scene source tracing method, and the present invention shows and widely uses prospect and bigger market value.

Particular embodiments described above; the purpose of the present invention, technical scheme and beneficial effect are further described; it is it should be understood that; the foregoing is only the specific embodiment of the present invention; it is not limited to the present invention; all within the spirit and principles in the present invention, any modification, equivalent substitution and improvement etc. done, should be included within the scope of the present invention.

Claims (10)

1. a multi-modal Wetland ecological habitat scene nuclear space source tracing method, it is characterised in that should Method comprises the steps:

Step 1, uses imaging device to obtain different classes of wetlands ecosystems habitat scene shadow Decent, image sample is stored in movable flashing device；

Step 2, by CF card reader, by the wetland of storage in step 1 movable flashing device Ecosystem habitat scene image sample imports in the magnetic disk media of PC, sets up Wetlands ecosystems habitat scene database；

Step 3, uses the lower space of optimal resolution configuration in null tone territory to enlist the services of method extraction wetland raw State system habitat scene global space networks roth's sign；

Step 4, uses Scale invariant features transform method to extract wetlands ecosystems habitat scene office Portion's Scale invariant features transform feature；

Step 5, by empty for the wetlands ecosystems habitat scene global extracted in step 3 and 4 Between enlist the services of feature and local Scale invariant features transform feature be normalized respectively and in The heartization processes；

Step 6, the global space after using core principle component analysis method step 5 to be processed enlists the services of spy It is linear that local scale invariant features transform characteristics of seeking peace projects to higher-dimension nucleus lesion Change, and the feature of this two classes different modalities is carried out linear coupling；

Step 7, uses large-spacing k-nearest neighbor by higher-dimension nucleus lesion described in step 6 The wetlands ecosystems habitat scene characteristic of linear coupling is mapped to higher-dimension nucleus lesion In, the semantic scene ownership of partition database sample in higher-dimension nucleus lesion；

Step 8, for wetlands ecosystems habitat scene to be traced to the source, uses step successively 37 treat sample of tracing to the source carries out in feature extraction, feature pretreatment and nuclear space ecological Information couples, and is then mapped in the higher-dimension nucleus lesion of priori scene, at higher-dimension Nucleus lesion calculates wetlands ecosystems habitat scene to be traced to the source and step 7 institute State the language coupled between priori scene characteristic in the scene database of wetlands ecosystems habitat Justice similarity, and then infer the semantic ownership of wetlands ecosystems habitat scene image Information.

Optimal resolution the most according to claim 1 configuration lower Wetland ecological habitat scene is traced back Source method, it is characterised in that in described step 1, wetlands ecosystems habitat scene Object refer to: Yancheng, Jiangsu Province coastal tidal wetland, Area, Zoige, Sichuan highland and severe cold wetland With Hainan Dong Zhaigang seashore Mangrove Wetlands.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 2, sets up wetlands ecosystems habitat scene Data base includes 3 class Wetland ecological habitat scene image collection: Yancheng, Jiangsu Province coastal tidal Wetland image set, Area, Zoige, Sichuan highland and severe cold wetland image set and Hainan Dong Zhaigang sea Bank Mangrove Wetlands image set.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 3, uses optimal resolution in null tone territory to join Put lower space and enlist the services of the semantic feature of method extraction wetlands ecosystems habitat scene；Institute The space stated in null tone territory under optimal resolution configuration is enlisted the services of the wave filter of method and is used The multiresolution 2D-Gabor wave filter of the spatial frequency domain nearly orthogonal of " Flos Chrysanthemi shape " Group.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 4, Scale invariant features transform method uses Gaussian difference scale space method carries out the Wetland ecological habitat of detection and positioning stablity Feature extreme point in scene image, generates wetland according to specific characteristic point directioin parameter raw State habitat scene local feature description.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 5, by normalized, reflects data Processing in the range of being mapped to [-1,1], normalized algorithm is:

${\hat{x}}_i=2\·\left(\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}\right)-1$

Wherein, x_iRepresent that the i-th row treat normalization characteristic vector, x_minRepresent minima in column vector, x_maxRepresent maximum in column vector,Represent the column vector after the i-th row normalization.As Really in column vector, maximum is equal to minima x_max=x_min, then column vector value is constant In described step 5, centralization processes, and refers to that variable deducts its mathematical expectation: for Sample data, refers to each observation of sample variableDeduct this sample to become The meansigma methods of the sample of amount, specific algorithm is as follows:

${\stackrel{\‾}{x}}_{\mathrm{ij}}={\hat{x}}_{\mathrm{ij}}-\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\hat{x}}_{\mathrm{ij}}$

Wherein, i correspondence sample, j represents total sample number to dependent variable, M.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 6, utilizes k-fold to intersect in the modelling phase Proof method obtains the highest accuracy of identification and determines the wetlands ecosystems habitat main one-tenth of scene core Mark.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 7, is inferring field, wetlands ecosystems habitat During the semantic attaching information of scape, use Mahalanobis measure sample estimates at height Similarity in dimension space.

The scene nuclear space side of tracing to the source, multi-modal Wetland ecological habitat the most according to claim 1 Method, it is characterised in that in described step 7, is inferring field, wetlands ecosystems habitat During the semantic attaching information of scape, the similar number of large-spacing k-nearest neighbor Feature Semantics Parameter is selected by user to set or by system recommendation voluntarily.

10. one kind is traced back based on the multi-modal Wetland ecological habitat scene nuclear space described in claim 1 The device of tracing to the source of source method, it is characterised in that include imaging device, movable flashing device, CF card reader and PC, imaging device is used for gathering wetlands ecosystems habitat scene Image sample also stores to movable flashing device, and CF card reader is for filling movable flashing Put interior wetlands ecosystems habitat scene image sample and be directed into PC；Described imaging Device uses varifocal high-resolution Complimentary Metal-Oxide semiconductor device quiescent imaging Device.