CN104732215A - Remote-sensing image coastline extracting method based on information vector machine - Google Patents

Abstract

The invention discloses a remote-sensing image coastline extracting method based on an information vector machine. The remote-sensing image coastline extracting method includes the steps of firstly obtaining a remote-sensing image containing coastline information, calculating the normalization difference water body index according to wave band information of the remote-sensing image, and obtaining a NDWI image; then selecting sample points on the NDWI image, extracting the color characteristics, the textural characteristics and the types to construct a training sample, and training an IVM model; then automatically classifying all pixels of the image through the trained IVM model, and partitioning seawater from land of the remote-sensing image; finally extracting a coastline with the graying and binarization image processing technology. The remote-sensing image coastline extracting method has the advantages that the coastline extracting accuracy under noise pollution is remarkably improved, the coastline is rapidly and accurately extracted, and an efficient technological means is provided for measuring, identifying and analyzing the coastline.

Description

A kind of remote sensing image tidal saltmarsh method based on information vector machine

Technical field

The invention belongs to remote sensing image tidal saltmarsh technical field, relate to a kind of remote sensing image tidal saltmarsh method based on information vector machine.

Background technology

Shore line is the datum line dividing flood and field management area; be human research's sea-land interaction, use sea movable on the impact of coastal zone and the important content of ICM and safeguard of coast and marine ecological system; its change directly changes Tideland resources stock number and Coastal Zone Environment; affect the survival and development of the people, the dynamic change tool of therefore monitoring shore line fast and is accurately of great significance.

The method comparatively commonly used of tradition tidal saltmarsh adopts artificial field operation GPS to measure, but this method time and effort consuming, efficiency is low, the work period is long and precision is not high, is difficult to extract shore line fast and accurately.Remote sensing because having powerful data retrieval capabilities, large scale, round-the-clock, synchronous, frequent dynamically observe and obtain different scale space time information and the feature such as to reduce investment outlay, and overcomes the deficiency of traditional tidal saltmarsh method.Along with the development of remote sensing technology in recent years, the resolution of remote sensing image, to improve constantly with the subject such as digital image processing, artificial intelligence constantly perfect, and the tidal saltmarsh method based on remote sensing image is emerged in large numbers gradually.

The method that current remote sensing image extracts shore line conventional has thresholding method, edge detection method, neural network classification method and support vector cassification method etc.Thresholding method is also called density slice method, and it mainly utilizes the difference of background and prospect gray-scale value, is realized the segmentation of remote sensing image by suitable threshold values, thus extracts shore line.The method only utilizes the gray feature of remote sensing image, do not consider its contextual information, have ignored local space relation, and the gray-scale value on seawater and land does not have clear and definite differentiation, therefore the Image Segmentation effect for complexity is poor, extract shore line precision can not satisfy the demand.Edge detection method is the edge by detecting remote sensing image, using the edge detected as shore line.Common are Roberts operator, Canny operator and Sobel operator etc.This kind of extracting method make use of the local space relation of remote sensing image, but it is very responsive to noise, it is often greater than the response to image border to the response of noise, thus makes the Remote Sensing Image Edge that detects fuzzy, and the shore line finally extracted is difficult to obtain gratifying effect.Along with neural network and support vector machine are applied to the extraction in shore line, overcome the shortcoming of the classic methods such as threshold segmentation to a certain extent, the shore line precision of extraction makes moderate progress.But neural network parameter is selected do not have unified criterion, is easily absorbed in local minimum, extract result and be not easy to meet engineering practice requirement.Support vector machine lacks theoretic guidance for the selection of kernel function and nuclear parameter, obtain good model by experience and test mostly in realistic model, and the cost of computing time and consumption internal memory is relatively large, be also difficult to extract shore line fast and accurately.

Information vector machine (Informative Vector Machine, IVM) is a kind of novel statistical learning algorithm proposed in 2002 by Neil D.Lawrence.The method is based on the Bayesian statistics theories of learning and kernel method, have that parameter adaptive obtains, high-dimensional and complex nonlinear problem strong adaptability, prediction output possesses the plurality of advantages such as probability meaning, adopt the method based on information entropy theory simultaneously, informational sample composition active set (active set) of most of part is optimized from a large amount of training sample, by reaching the results of learning close with former training sample set to the study of active set, and represent in conjunction with rarefaction nuclear matrix, thus significantly reduce time complexity and the space complexity of study.In addition, assumed density is approached (assumed density filtering, ADF), also known as the introducing of match by moment (moment matching) approximation method and KL divergence (Relative Entropy), the method is made to have possessed the approximate processing ability stronger to non-gaussian distribution noise (e.g., two classification problems) situation.Therefore, information vector machine is applied to complicated remote sensing image process by the present invention, extracts quickly and accurately to realize complex outline shore line.

Summary of the invention

The defect that the present invention seeks to exist for prior art provides a kind of remote sensing image tidal saltmarsh method based on information vector machine, solves the problem that existing method is low for tidal saltmarsh efficiency, precision is low.

The technical solution used in the present invention is carried out according to following steps:

1) obtain the TM image of target area, and carry out band combination, geometry correction, cutting pre-service in remote sensing image processing software ERDAS IMAGINE 9.0, obtain the remote sensing image containing shore line information to be extracted;

2) calculate normalization difference water body index (normalized difference water index, NDWI) according to the band class information of the remote sensing image containing shore line information, obtain NDWI image, realize the preliminary differentiation on seawater and land;

3) in the NDWI image obtained, choose several Seawater Samples point with obvious seawater characteristics and land feature and land sample points respectively;

4) the Seawater Samples point obtained and land sample point are carried out to the extraction of color and textural characteristics, and get the desired value structure training sample database (x of color, textural characteristics value and correspondence _j, y _j) 1≤j≤2N, wherein x _j=[C _j ^r, C _j ^g, C _j ^b, asm _j, con _j, idm _j, ent _j] be training sample input vector, C _j ^r, C _j ^g, C _j ^brepresent the eigenwert of sample point j under R, G, B tri-Color Channels respectively, asm _j, con _j, idm _j, ent _jdistinguish the energy of corresponding grey scale co-occurrence matrix, contrast, the degree of correlation and entropy; y _jfor training sample exports desired value, get "-1 " when sample point is seawater, get "+1 " when sample point is land;

5) extraction of eigenwert carried out to the pixel of whole NDWI image and it can be used as the input vector X of test sample book _*, X _*=[x ₁, x ₂... x _m] ^t, M is whole image picture element number;

6) training sample database (x will be obtained _j, y _j) input information vector machine (Informative VectorMachine, IVM), and train, obtain the IVM sorter with better generalization ability;

7) by the input vector X of test sample book _*be input in the IVM sorter trained, obtain corresponding prediction and export target y* and according to the segmentation exporting target and complete remote sensing image, realize being separated of seawater and land;

8) gray processing, binaryzation are carried out to the segmentation result obtained, obtain the shore line of whole remote sensing image;

Further, described step 1) in TM image refer to and comprise 7 wave band: TM-1 ~ TM-7 by the multiband scan-image that Landsat 4 ~ No. 5 thematic mappers (thematic mapper) obtain.Band combination adopt TM4,3, the false color scheme of 2 (TM4 is near-infrared band, and TM3 is red spectral band, and TM2 is green light band) standard.

Further, described step 2) middle NDWI=(ρ _green-ρ _nIR)/(ρ _green+ ρ _nIR), ρ _greenand ρ _nIRrepresent green wave band and near-infrared band respectively, TM2, TM4 wave band of corresponding TM image.

Further, described step 4) described in the input vector x of training sample _jcarry out as follows:

A) obtain the eigenwert of sample point j under R, G, B tri-Color Channels, form the color feature vector C of sample point _j=[C _j ^r, C _j ^g, C _j ^b], 1≤j≤2N;

B) NDWI image is converted into grayscale image in business mathematics software Matlab, and by gray-scale compression to N _gobtain shadow gray scale as f;

C) in grayscale image f centered by sample point j, getting 3 × 3 pixels is the gray level co-occurrence matrixes P (r, t) that calculation window calculates 4 directions, and by its normalized, obtains Normalized Grey Level co-occurrence matrix p (r, t):

P(r,t)＝#{((k ₁,k ₂),(l ₁,l ₂))∈(L _x×L _y)×(L _x×L _y)|d,θ,f(k ₁,k ₂)＝r,f(l ₁,l ₂)＝t} (1)

$(r,t)=\frac{P(r,t)}{\underset{r=0}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=0}{\overset{N_g}{\mathrm{\Σ}}}P(r,t)}---\left(2\right)$

In formula: the gray level co-occurrence matrixes that P (r, t) is Ng × Ng, (L _x× L _y) be the scope territory that gray level co-occurrence matrixes defines, d represents distance, and θ represents direction, f (k ₁, k ₂)=r, f (l ₁, l ₂)=t is the gray-scale value of the corresponding ranks of grayscale image f; # represents the pixel number set up in braces; P (r, t) for gray level co-occurrence matrixes P (r, t) be normalized after Normalized Grey Level co-occurrence matrix;

D) textural characteristics of Normalized Grey Level co-occurrence matrix is calculated: energy asm, contrast con, degree of correlation idm, entropy ent, and using the textural characteristics value of gained textural characteristics value as this window center pixel j:

$\mathrm{asm}=\underset{r}{\mathrm{\Σ}}\underset{t}{\mathrm{\Σ}}{\left(p(r,t)\right)}^2---\left(3\right)$

$\mathrm{con}=\underset{m=0}{\overset{N_g-1}{\mathrm{\Σ}}}{m}^2\left[\underset{r=1}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=1}{\overset{N_g}{\mathrm{\Σ}}}p(r,t)\right],|r-t|=m---\left(4\right)$

$\mathrm{idm}=\underset{r=0}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=0}{\overset{N_g}{\mathrm{\Σ}}}\frac{p(r,t)}{1+{(r-t)}^2}---\left(5\right)$

$\mathrm{ent}=-\underset{r}{\mathrm{\Σ}}\underset{t}{\mathrm{\Σ}}p(r,t)\log \left(p(r,t)\right)---\left(6\right)$

The input vector x of the eigenvalue cluster synthesis training sample e) above step obtained _j, x _j=[C _j ^r, C _j ^g, C _j ^b, asm _j, con _j, idm _j, ent _j].

Further, described step 6) in the structure of IVM sorter be utilize the informational sample data of Method of Sample Selection screening most based on information entropy theory to set up active set I (index), substitute the study of former state notebook data collection, obtain the IVM sorter with better generalization ability.

Further, described step 7) described in Remote Sensing Image Segmentation according to prediction export target y* complete.As y*=-1, this pixel is water body, and during y*=+1, this pixel is land.

The invention has the beneficial effects as follows the precision significantly enhancing tidal saltmarsh under noise pollution, achieve shore line and extract quickly and accurately, for coastlining, identification and analysis provide a kind of technological means efficiently.

Accompanying drawing explanation

Fig. 1, a kind of remote sensing image tidal saltmarsh method flow diagram based on information vector machine;

Fig. 2, remote sensing image containing shore line information;

Fig. 3, NDWI image;

Fig. 4, IVM segmentation result;

Fig. 5, binary map;

Fig. 6, tidal saltmarsh result.

Embodiment

Below in conjunction with embodiment, the present invention is described in detail.Be illustrated in figure 1 technical scheme steps flow process of the present invention.The present invention carries out according to following steps:

Step 1) obtain the TM image of Beihai, Guangxi and surrounding waters thereof, and the pre-service such as band combination, geometry correction, cutting are carried out in remote sensing image processing software ERDAS IMAGINE 9.0, obtain the remote sensing image (Fig. 2) containing shore line information to be extracted.

Band combination wherein adopt TM4,3, the false color scheme of 2 (TM4 is near-infrared band, and TM3 is red spectral band, and TM2 is green light band) standard.Band combination, geometry correction, cutting all complete in remote sensing image processing software ERDAS IMAGINE 9.0, and specific implementation method is technology conventionally known to one of skill in the art.The sensor of image is TM, and framing orbit number is 125/045, and imaging time is November 17 calendar year 2001, and spatial resolution is 30 meters.

Step 2) calculate normalization difference water body index (normalized difference water index according to the band class information of the remote sensing image containing shore line information, NDWI), obtain NDWI image (Fig. 3), realize the preliminary differentiation on seawater and land: NDWI=(ρ _green-ρ _nIR)/(ρ _green+ ρ _nIR), ρ _greenand ρ _nIRrepresent green wave band and near-infrared band respectively, TM2, TM4 wave band of corresponding TM image;

Because the NDWI value on seawater and land is notable difference, therefore NDWI image energy effectively distinguishes seawater and land.The calculating of remote sensing image NDWI completes in remote sensing image processing software ERDAS IMAGINE 9.0.

Step 3) in the NDWI image obtained, choose Seawater Samples point and land sample point that N number of (N=100) has obvious seawater characteristics and land feature respectively;

Step 4) N number of (N=100) seawater, the land sample point that obtain are carried out to the extraction of color and textural characteristics, and get the output desired value structure training sample database (x of color, textural characteristics value and correspondence _j, y _j), 1≤j≤2N; Wherein x _j=[C _j ^r, C _j ^g, C _j ^b, asm _j, con _j, idm _j, ent _j] for training sample be input vector, C _j ^r, C _j ^g, C _j ^brepresent the color feature value of sample point j under R, G, B tri-Color Channels respectively, asm _j, con _j, idm _j, ent _jfor the textural characteristics value of sample point j, distinguish the energy of corresponding grey scale co-occurrence matrix, contrast, the degree of correlation and entropy; y _jfor training sample exports desired value, get "-1 " when sample point is seawater, get "+1 " when sample point is land;

Wherein textural characteristics is a kind of meticulous description of spatial distribution pattern in image regional area (locus, direction in space and spatial-intensity).Texture is a kind of local spatial information, reflect the mutual relationship of object and neighborhood, overcome the shortcoming that spectral signature does not consider spatial information, reduce the impact of noise on provincial characteristics, be the information compared with spectral information high level, be more conducive to extracting shore line accurately.The present invention gets the input vector that the spectral signature of sample point and textural characteristics construct training sample jointly.The texture analysis method of current remote sensing image is generally divided into two large classes by character: statistical analysis technique and structure analysis method.Statistical method studies a kind of more and comparatively ripe method, and the method considers the space distribution of gray level in texture, and calculate the local feature of in image often, some statistics of deriving from the distribution of feature are to portray texture.The texture analysis method of Corpus--based Method has, and auto-relativity function method, gray level co-occurrence matrixes method and markov random file method etc., be wherein most widely used with gray level co-occurrence matrixes method.The present invention adopts gray level co-occurrence matrixes method to analyze textural characteristics.

Realize step 4) detailed process as follows:

A) obtain the eigenwert of sample point j under R, G, B tri-Color Channels, form the color feature vector C of sample point _j=[C _j ^r, C _j ^g, C _j ^b], 1≤j≤200;

B) NDWI image is converted into grayscale image in business mathematics software Matlab, and by gray-scale compression to N _gobtain grayscale image f and (get N _g=16);

C) in image f centered by sample point j, getting 3 × 3 pixels is the gray level co-occurrence matrixes P (r, t) that calculation window calculates 4 directions, and by its normalized, obtains Normalized Grey Level co-occurrence matrix p (r, t):

P(r,t)＝#{((k ₁,k ₂),(l ₁,l ₂))∈(L _x×L _y)×(L _x×L _y)|d,θ,f(k ₁,k ₂)＝r,f(l ₁,l ₂)＝t} (1)

$p(r,t)=\frac{P(r,t)}{\underset{r=0}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=0}{\overset{N_g}{\mathrm{\Σ}}}P(r,t)}---\left(2\right)$

In formula: the gray level co-occurrence matrixes that P (r, t) is Ng × Ng, (L _x× L _y) be the scope territory that gray level co-occurrence matrixes defines, d represents distance (getting d=1), and θ represents direction (getting θ=0 °, 45 °, 90 °, 135 ° of four directions), f (k ₁, k ₂)=r, f (l ₁, l ₂)=t is the gray-scale value of the corresponding ranks of grayscale image f; # represents the pixel number set up in braces; P (r, t) for gray level co-occurrence matrixes P (r, t) be normalized after Normalized Grey Level co-occurrence matrix;

D) consider that direction is on the impact of texture, weakens the effect in direction by the mean value getting 4 directions.Calculate 4 direction Normalized Grey Level co-occurrence matrix p (r, t) textural characteristics value: energy asm, contrast con, degree of correlation idm, entropy ent, and using the textural characteristics value of the mean value in gained textural characteristics value 4 directions as this window center pixel n:

$\mathrm{asm}=\underset{r}{\mathrm{\Σ}}\underset{t}{\mathrm{\Σ}}{\left(p(r,t)\right)}^2---\left(3\right)$

$\mathrm{con}=\underset{m=0}{\overset{N_g-1}{\mathrm{\Σ}}}{m}^2\left[\underset{r=1}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=1}{\overset{N_g}{\mathrm{\Σ}}}p(r,t)\right],|t-t|=m---\left(4\right)$

$\mathrm{idm}=\underset{r=0}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=0}{\overset{N_g}{\mathrm{\Σ}}}\frac{p(r,t)}{1+{(r-t)}^2}---\left(5\right)$

$\mathrm{ent}=-\underset{r}{\mathrm{\Σ}}\underset{t}{\mathrm{\Σ}}p(r,t)\log \left(p(r,t)\right)---\left(6\right)$

The input vector x of the eigenvalue cluster synthesis training sample e) above-mentioned steps obtained _j, x _j=[C _j ^r, C _j ^g, C _j ^b, asm _j, con _j, idm _j, ent _j], and combine corresponding output desired value y _jstructure training sample (x _j, y _j);

Step 5) according to step 4) pixel of described principle to whole NDWI image carry out the extraction of eigenwert and it can be used as the input vector X of test sample book _*, X _*=[x ₁, x ₂... x _m] ^t, M is the number of whole NDWI image picture element;

Step 6) training sample (x will be obtained _j, y _j) input information vector machine (Informative VectorMachine, IVM) and train, obtain the IVM sorter with better generalization ability;

IVM is based on the Bayesian statistics theories of learning and kernel method, the method combines supposition density approaches ADF and the Method of Sample Selection based on information entropy theory, utilize the approximate Posterior distrbutionp (being approximately Gaussian distribution) increased each time after a sample data (information vector) of ADF recurrence, and obtain approximate likelihood distribution, guarantee trackability and the tractability of algorithm, simultaneously, in this recurrence approximation, the informational sample data of Method of Sample Selection screening most based on information entropy theory is utilized to set up active set I (index), substitute the study of former state notebook data collection, and adopt recurrence more new record intermediate variable obtain the re-treatment of method ingenious avoidance covariance matrix K (or Σ), realize the object significantly reducing algorithm time and space complexity.

Realize step 6) detailed process as follows:

A) suppose two sample index collection I and H, wherein I is active set, and H treats selected works, time initial, and at any time, (from initial 200 training samples, screening 100 information vectors).Information vector obtains in the mode of a kind of continuous print, similar on-line study: first, application supposition density approaches (assumed density filtering, ADF) approximate have i information vector (active set is the set of information vector), i.e. I _itime Posterior distrbutionp and likelihood distribution:

$\left\{\begin{array}{c}{q_I}_i\left(f\right)=N(f;{{\mathrm{\μ}}_I}_i,{{\mathrm{\Σ}}_I}_i)\≈p\left(f\right|X_{I_{i,;}}{y}_{I_i},\mathrm{\θ})\\ p\left(y_{n_{I_i}}\right|f_{n_{I_i}})\≈N(m_{n_{I_i}};f_{n_{I_i}},{\mathrm{\β}}_{n_{I_i}}^{-1})\end{array}\right.---\left(7\right)$

In formula: p represents probability distribution, q represents APPROXIMATE DISTRIBUTION, and f represents latent variable collection, and μ represents Gaussian distribution average, and Σ is covariance matrix, and m represents likelihood substitute variable, and β represents noise profile variance, for the input vector of i information vector, for the output vector of i information vector, θ represents covariance function hyper parameter, and its optimum solution is just by maximizing edge likelihood p (y _i| X _i:, θ) and self-adaptation obtains.

If b) treat that in selected works H, a sample h adds active set I _iafter can maximize and reduce the information entropy of Posterior distrbutionp, namely then sample h is as the i-th+1 information vector

$\mathrm{\Δ}{H}_{I_i,h}=-\frac{1}{2}\log \left|{\mathrm{\Σ}}_{I_{i-1}}\right|+\frac{1}{2}\log \left|{\mathrm{\Σ}}_{I_i}\right|=-\frac{1}{2}\log \left|{\mathrm{\Σ}}_{I_{i-1}}{\mathrm{\Σ}}_{I_i}^{-1}\right|---\left(8\right)$

C) maximize according to above-mentioned b) middle the principle circulation reducing Posterior distrbutionp information entropy, until complete the selection of 100 information vectors, realize the study substituting raw sample data collection (200 training samples) with active set I (100 information vectors), can obtain in conjunction with formula (7):

$\left\{\begin{array}{c}p\left(y_I\right|X_{I,:},\mathrm{\θ})\≈N(m_I;0,K_I+B_I^{-1})\\ p\left(f\right|y_I,X_{I,:},\mathrm{\θ})\≈N(f;{\mathrm{\μ}}_I,{\mathrm{\Σ}}_I)\\ {\mathrm{\μ}}_I={\mathrm{\Σ}}_I{B}_I{y}_I,{\mathrm{\Σ}}_I={(B_I+K_I^{-1})}^{-1}\end{array}\right.---\left(9\right)$

In formula: B represents noise profile variance, K or Σ represents Gaussian distribution covariance matrix, such as Σ _irepresent the covariance matrix of active set, each element be wherein in active set information vector between two between covariance.

Above-mentioned b) and c) in replace raw sample data collection to be adopt the method based on information entropy theory with active set I, concentrate from 200 training samples to optimize informational sample composition active set I (activeset) of part most, by reaching the results of learning close with former training sample set to the study of active set, and represent in conjunction with rarefaction nuclear matrix, thus significantly reduce time complexity and the space complexity of study, realize the rapid extraction in shore line.

D) obtain IVM according to formula (9) and predict Posterior distrbutionp:

$\left\{\begin{array}{c}p\left(f_{*}\right|y,X,x_{*},{\mathrm{\θ}}_l)\≈p\left(f_{*}\right|y_I,X_{I,:}{x}_{*},{{\mathrm{\θ}}_l}_I)\\ =\∫p\left(f_{*}\right|f,X,x_{*},{{\mathrm{\θ}}_l}_I)p\left(f\right|y_I,X_{I,:},{{\mathrm{\θ}}_l}_I)\mathrm{df}\\ =N(f_{*}:{\mathrm{\μ}}_{*},{\mathrm{\σ}}_{*}^2)\\ {\mathrm{\μ}}_{*}=K_{*I}^T{K}_I^{-1}{\mathrm{\Σ}}_I{B}_I{y}_I,{\mathrm{\σ}}_{*}^2=k_{*}+K_{*I}^T{K}_I^{-1}({\mathrm{\Σ}}_I-K_I)K_I^{-1}{K}_{*I}\end{array}\right.---\left(10\right)$

X=[x in formula ₁, x ₂, x ₃x _2N] ^trepresent training sample point input vector, y is the output object vector of training sample, X _i, y _ibe respectively the input of active set, output vector, x _*represent the input vector of forecast sample to be entered, f _*for the latent variable of predicted value.

E) because the present invention is binary classification, on the basis of formula (10), the IVM sorter with generalization ability is therefore obtained:

$\left\{\begin{array}{c}p(y_{*}=+1|X,y,x_{*})=\∫\mathrm{\Φ}\left(f_{*}\right)p\left(f_{*}\right|y,X,x_{*},{\mathrm{\θ}}_l){\mathrm{df}}_{*}=\mathrm{\Φ}\left(\frac{{\mathrm{\μ}}_{*}}{\sqrt{1+{\mathrm{\σ}}_{*}^2}}\right)\\ \mathrm{if\; p}\≥0.5,y_{*}=+1,\mathrm{otherwies},y_{*}=-1\end{array}\right.---\left(11\right)$

In formula: Φ (x) represents standardized normal distribution accumulated probability density function, certainly also can substitute with other response function, as sigmoid logical function, y _*for the output desired value of sample to be predicted.

Step 7) by the input vector X of test sample book _*(now X _*for the x in formula (11) _*) be input in the IVM sorter trained, obtain corresponding prediction and export target y _*, and the segmentation of NDWI remote sensing image is completed according to output target, realize extra large land and be separated (Fig. 4).When the prediction of pixel exports target y _*when=-1, this pixel is divided into water body; When the prediction of pixel exports target y _*when=+ 1, this pixel is divided into land.

Step 8) image split is converted into gray-scale map, binary map successively in business mathematics software Matlab, on the basis of binary map, shore line can be proposed.

Realize step 8) detailed process as follows:

A) image split is converted into gray-scale map by rgb2gray function in business mathematics software Matlab;

B) because gray scale image is transformed on segmentation basis, its grey level histogram is obvious bimodal distribution, chooses the threshold values that gray-scale value 60 makes gray scale image binaryzation, obtains the binary map (Fig. 5) of Land-sea Distributions.On the basis of binary map, adopt the bwboundaries function in business mathematics software Matlab to complete the extraction in shore line, obtain meticulous shore line profile (Fig. 6).

The present invention be advantageous in that:

1, the present invention takes full advantage of the spectral signature of remote sensing image and the Li Tezheng based on gray level co-occurrence matrixes of reflection spatial gradation information, overcome classic method only utilize spectral signature and do not consider seawater and land textural characteristics cause the bottleneck problem that shore line identification error is larger, effectively reduce the impact of noise on provincial characteristics, significantly improve the precision of shore line Computer Automatic Recognition.

2, it is strong that a kind of remote sensing image tidal saltmarsh method based on IVM that the present invention uses has small-sample learning ability, the optimized parameter of forecast model can self-adaptation obtain, forecast model has the advantages such as good generalization ability, overcome optimal network topological structure that current application exists based on the remote sensing image tidal saltmarsh method of artificial neural network comparatively widely and hyper parameter is not easily determined, easily be absorbed in the deficiency of local minimum, solve the problem that reasonable kernel functional parameter that the remote sensing image tidal saltmarsh method based on support vector machine that occurs in recent years exists is difficult to determine, the shore line of complicated remote sensing image is identified to have stronger applicability, in the protection of coastal landform landforms, the development and utilization of Tidal Flat Resources, the association areas such as the orderly development of seashore land resources have good engineer applied and are worth.

The above is only to better embodiment of the present invention, not any pro forma restriction is done to the present invention, every any simple modification done above embodiment according to technical spirit of the present invention, equivalent variations and modification, all belong in the scope of technical solution of the present invention.

Claims (6)

1., based on a remote sensing image tidal saltmarsh method for information vector machine, it is characterized in that: the method comprises the steps:

1) obtain the TM image of target area, and carry out band combination, geometry correction, cutting pre-service in remote sensing image processing software ERDAS IMAGINE 9.0, obtain the remote sensing image containing shore line information to be extracted;

2) calculate normalization difference water body index (normalized difference water index, NDWI) according to the band class information of the remote sensing image containing shore line information, obtain NDWI image, realize the preliminary differentiation on seawater and land;

3) in the NDWI image obtained, choose several Seawater Samples point with obvious seawater characteristics and land feature and land sample points respectively;

4) the Seawater Samples point obtained and land sample point are carried out to the extraction of color and textural characteristics, and get the desired value structure training sample database (x of color, textural characteristics value and correspondence _j, y _j) 1≤j≤2N, wherein x _j=[C _j ^r, C _j ^g, C _j ^b, asm _j, con _j, idm _j, ent _j] be training sample input vector, C _j ^r, C _j ^g, C _j ^brepresent the eigenwert of sample point j under R, G, B tri-Color Channels respectively, asm _j, con _j, idm _j, ent _jdistinguish the energy of corresponding grey scale co-occurrence matrix, contrast, the degree of correlation and entropy; y _jfor training sample exports desired value, get "-1 " when sample point is seawater, get "+1 " when sample point is land;

5) extraction of eigenwert carried out to the pixel of whole NDWI image and it can be used as the input vector X of test sample book _*, X _*=[x ₁, x ₂... x _m] ^t, M is whole image picture element number;

6) training sample database (x will be obtained _j, y _j) input information vector machine, and train, obtain the IVM sorter with better generalization ability;

7) by the input vector X of test sample book _*be input in the IVM sorter trained, obtain corresponding prediction and export target y ^*and complete the segmentation of remote sensing image according to exporting target, realize being separated of seawater and land;

8) gray processing, binaryzation are carried out to the segmentation result obtained, obtain the shore line of whole remote sensing image.

2. according to the remote sensing image tidal saltmarsh method based on information vector machine a kind of described in claim 1, it is characterized in that: described step 1) in TM image refer to the multiband scan-image that Landsat 4 ~ No. 5 thematic mappers obtain, comprise 7 wave band: TM-1 ~ TM-7, band combination adopt TM4,3,2 (TM4 is near-infrared band, TM3 is red spectral band, and TM2 is green light band) the false color scheme of standard.

3., according to the remote sensing image tidal saltmarsh method based on information vector machine a kind of described in claim 1, it is characterized in that: described step 2) middle NDWI=(ρ _green-ρ _nIR)/(ρ _green+ ρ _nIR), ρ _greenand ρ _nIRrepresent green wave band and near-infrared band respectively, TM2, TM4 wave band of corresponding TM image.

4., according to the remote sensing image tidal saltmarsh method based on information vector machine a kind of described in claim 1, it is characterized in that: described step 4) described in the input vector x of training sample _jcarry out as follows:

A) obtain the eigenwert of sample point j under R, G, B tri-Color Channels, form the color feature vector C of sample point _j=[C _j ^r, C _j ^g, C _j ^b], 1≤j≤2N;

B) NDWI image is converted into grayscale image in business mathematics software Matlab, and by gray-scale compression to N _gobtain shadow gray scale as f;

C) in grayscale image f centered by sample point j, getting 3 × 3 pixels is the gray level co-occurrence matrixes P (r, t) that calculation window calculates 4 directions, and by its normalized, obtains Normalized Grey Level co-occurrence matrix p (r, t):

P(r,t)＝#{((k ₁,k ₂),(l ₁,l ₂))∈(L _x×L _y)×(L _x×L _y)|d,θ,f(k ₁,k ₂)＝r,f(l ₁,l ₂)＝t} (1)

$p(r,t)=\frac{P(r,t)}{\underset{r=0}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=0}{\overset{N_g}{\mathrm{\Σ}}}P(r,t)}---\left(2\right)$

In formula: the gray level co-occurrence matrixes that P (r, t) is Ng × Ng, (L _x× L _y) be the scope territory that gray level co-occurrence matrixes defines, d represents distance, and θ represents direction, f (k ₁, k ₂)=r, f (l ₁, l ₂)=t is the gray-scale value of the corresponding ranks of grayscale image f; # represents the pixel number set up in braces; P (r, t) for gray level co-occurrence matrixes P (r, t) be normalized after Normalized Grey Level co-occurrence matrix;

D) textural characteristics of Normalized Grey Level co-occurrence matrix is calculated: energy asm, contrast con, degree of correlation idm, entropy ent, and using the textural characteristics value of gained textural characteristics value as this window center pixel j:

$\mathrm{asm}=\underset{r}{\mathrm{\Σ}}\underset{t}{\mathrm{\Σ}}{\left(p(r,t)\right)}^2---\left(3\right)$

$\mathrm{con}=\underset{m=0}{\overset{\mathrm{Ng}-1}{\mathrm{\Σ}}}{m}^2\left[\underset{r=1}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=1}{\overset{N_g}{\mathrm{\Σ}}}p(r,t)\right],|r-t|=m---\left(4\right)$

$\mathrm{idm}=\underset{r=0}{\overset{N_g}{\mathrm{\Σ}}}\underset{t=0}{\overset{N_g}{\mathrm{\Σ}}}\frac{p(r,t)}{1+{(r-t)}^2}---\left(5\right)$

$\mathrm{ent}=-\underset{r}{\mathrm{\Σ}}\underset{t}{\mathrm{\Σ}}p(r,t)\log \left(p(r,t)\right)---\left(6\right)$

The input vector x of the eigenvalue cluster synthesis training sample e) above step obtained _j, x _j=[C _j ^r, C _j ^g, C _j ^b, asm _j, con _j, idm _j, ent _j].

5. according to the remote sensing image tidal saltmarsh method based on information vector machine a kind of described in claim 1, it is characterized in that: described step 6) in the structure of IVM sorter be utilize the informational sample data of Method of Sample Selection screening most based on information entropy theory to set up active set I, substitute the study of former state notebook data collection, obtain the IVM sorter with better generalization ability.

6., according to the remote sensing image tidal saltmarsh method based on information vector machine a kind of described in claim 1, it is characterized in that: described step 7) described in Remote Sensing Image Segmentation be according to prediction export target y ^*complete, work as y ^*when=-1, this pixel is water body, y ^*when=+ 1, this pixel is land.