CN105608474A - High-resolution-image-based regional adaptive cultivated land extraction method - Google Patents

Abstract

The invention, which belongs to the technical field of remote sensing application, particularly relates to a high-resolution-image-based regional adaptive cultivated land extraction method. The method comprises: pretreatment is carried out on data; multi-scale segmentation is carried out on a high-resolution image, an optimal segmentation scale of cultivated land is determined, and an optimal segmentation layer of cultivated land classification is obtained; a plurality of cultivated land sample sets are constructed at the optimal segmentation layer to form an attribute feature space of each cultivated land sample set; optimal image feature space selection is carried out; according to the optimal image feature spaces, fuzzy classifier of cultivated land extraction are constructed; and on the basis of the cultivated land sampling sets and the fuzzy classifiers, cultivated land extraction is carried out. According to the invention, automatic identification of a cultivated land object is realized by combining the cultivated land sample sets and the fuzzy classifiers. The uniformity of the segmented object of the high-resolution image and the extraction precision of the fuzzy classifier are improved by selecting the optimal segmentation scale and the different cultivated land sample sets; and on the basis of construction of the cultivated land sample sets, the restriction requirements on the time phase and the number of the data source are reduced.

Description

Region adaptivity plant extraction method based on high resolution image

Technical field

The invention belongs to remote sensing application technical field, be specifically related to a kind of region adaptivity based on high resolution image and ploughGround extracting method.

Background technology

The quick, accurate of farmland information extracted cultivated land dynamic change monitoring, Soil fertility investigation of cultivated land and evaluation, cultivated land protection and baseThis farmland delimitation, land resources utilization degree analyzing, precision agriculture etc. are significant.

High resolution image has strengthened the internal diversity of ploughing, and the spectrum performance covering that makes to plough presents diversity,Strengthen the difficulty of ploughing and accurately extracting. Current high resolution image plant extraction method has a variety of, and it exists following problem:

1) general plant extraction method is the growth characteristic according to crops, the shadow of many phases of the growth cycle that is consistent when employingPicture is analyzed and researched, and the method is subject to time phase and the number quantitative limitation of data source on promoting;

2) adopt OO plant extraction method application more extensive, but it exists, application region is little, assorting process is multipleAssorted, automaticity is low, does not have universality and generalization.

Summary of the invention

The object of this invention is to provide a kind of plant extraction method based on high resolution image, described plant extraction methodBe not subject to time phase and the restricted number of data source, and there is universality and generalization.

For achieving the above object, the present invention proposes a kind of region adaptivity plant extraction side based on high resolution imageMethod, comprising:

S1, carries out pretreatment to data;

S2, carries out multi-scale division to high resolution image, determine plough optimum segmentation yardstick, obtain plough classificationOptimum segmentation layer;

S3 builds multiple arable lands sample set in optimum segmentation layer, forms the attributive character space of each arable land sample set;

S4, to each arable land sample set, carries out optimum image feature space and selects;

S5, to each arable land sample set, builds the Fuzzy Classifier of plant extraction according to optimum image feature space;

S6, carries out plant extraction based on arable land sample set and Fuzzy Classifier:

S61, according to the attributive character space of arable land sample set, carries out attribute to the object after step S2 optimum segmentation one by oneFeature calculation;

S62, the attributive character of the arable land sample set that the attributive character result of calculation respectively S61 being obtained and step S3 buildSpatial dimension value is mated, if the attributive character value of object, within the scope of the attributive character of a certain arable land sample set, is rememberedBe 1, otherwise be designated as 0, obtain mating matrix D_ij

$D_{ij}={\left[\begin{array}{c}{l}_1,l_2,l_4,\mathrm{...}{l}_i\\ p_1,p_2,p_3,p_4,\mathrm{...},p_i\\ n_1,n_2,n_3,n_4,\mathrm{...}{n}_i\\ b_1,b_2,b_3,b_4,\mathrm{...}{b}_i\\ \mathrm{...................}\\ w_1,w_2{w}_3,w_4,\mathrm{...}{w}_i\end{array}\right]}^T$

Wherein, l, p, n, b ..., w is respectively the attributive character of arable land sample set, and i is the number of arable land sample set, and j is for belonging toThe number of property feature, the transposition of T representing matrix;

S63, according to definition arable land sample set in the shared weighted value of each attributive character with mate matrix D_ij, calculate respectivelyThe similarity of object and each arable land sample set, obtains similarity matrix S_i，

$S_i=D_{ij}\×{\mathrm{\δ}}_j={\left[\begin{array}{c}{l}_1,l_2,l_3,l_4,\mathrm{...}{l}_i\\ p_1,p_2,p_3,p_4,\mathrm{...}{p}_i\\ n_1,n_2,n_3,n_4,\mathrm{...}{n}_i\\ b_1,b_2,b_3,b_4,\mathrm{...}{b}_i\\ \mathrm{...................}\\ w_1,w_2,w_3{w}_4,\mathrm{...}{w}_i\end{array}\right]}^T\×\left[\begin{array}{c}{\mathrm{\δ}}_1\\ {\mathrm{\δ}}_2\\ {\mathrm{\δ}}_3\\ {\mathrm{\δ}}_4\\ \mathrm{...}\\ {\mathrm{\δ}}_j\end{array}\right]=\left[\begin{array}{c}{S}_1\\ S_2\\ S_3\\ S_4\\ \mathrm{...}\\ S_i\end{array}\right]$

Wherein, δ_iFor sample set attributive character weight matrix, δ_iCarry out self-defined according to region situation;

S64, the threshold decision condition of setting similarity matrix Si, if the similarity of object and certain arable land sample set meetsThreshold decision condition, object and certain arable land sample set match;

S65, according to the matching result of S64, carries out Fuzzy Classifier corresponding to arable land sample set matching to this object,Obtain class categories function E (i) and classification degree of membership ω (i),

S66, adopts the Nearest Neighbor with Weighted Voting method weighted calculation of ploughing, and computing formula is as follows:

$E\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}\left(E\right(i)\×\mathrm{\ω}(i\left)\right)$

Wherein, E (x) is arable land weighting discriminant function, and n is the Fuzzy Classifier sum matching with object x, and i is fuzzyThe number of grader, i=1 ..., n;

Whether S67, be to plough to judge to object, and Rule of judgment is as follows:

Further, data pretreatment comprises to be processed high resolution image and digital elevation model, wherein, and image placeReason comprises carries out radiation and geometric correction to image, and digital elevation model processing comprises the generation gradient.

Further, in described step S2, the optimum segmentation yardstick in arable land is that peak value, shape refer to for meet mean variance simultaneouslyWhat number variance was valley cuts apart yardstick;

Wherein, mean variance S_C ²Computing formula be

Wherein, C_LFor single object is in the brightness average of L-band,For bright at L-band of all objects in imageDegree average, m is the summation of object number in image, i is object number, i=1 ... m;

Shape index variance S_SI ²Computing formula be ${S_{SI}}^2=\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m{(SI-\stackrel{\‾}{SI})}^2,SI=\frac{b}{V}$

Wherein, the shape index value that SI is single object,For the shape index average of all objects in image, m is shadowThe summation of object number in picture, i is object number, i=1 ... m, the object girth that b is single object, V is single objectObject volume.

Further, in described step S3,

The step that builds arable land sample set is as follows:

S31, by visual collection arable land sample, the sample that requirement gathers is uniformly distributed and is representative, contains allThe farmland types in district, quantity is no more than 5% of whole section object number;

S32, according to sample region, arable land, terrain slope feature, phenology feature and image feature, the sample of ploughingClassification, build multiple arable lands sample set, and form the attributive character space of each arable land sample set;

S33, rejects arable land sample:

If the mean value of certain classification arable land each characteristics of objects of sample isStandard deviation isStd＝(std₁，std₂，std₃，……，std_n), for a certain arable land sample X=(x₁，x₂，x₃......，x_n), if With the plough rejecting of sample of 2 times of standard deviations.

Further: sample region, described arable land, terrain slope feature, phenology feature and image feature are respectively sampleThis district title, the gradient, NDVI and brightness, the attributive character space of each arable land sample set comprise the district title of sample,The gradient, NDVI and brightness.

Further, in described step S4, the step that optimum image feature space is selected is as follows:

S41, arable land characteristics of objects space primary election:

Set up storehouse, characteristics of objects space, arable land for spectrum, texture, the shape facility of high resolution image:

S42, carries out the selection of optimal characteristics space based on feature variance:

Calculate the variance of each arable land each feature of sample set, 2～3 features of statistical variance minimum, form corresponding eachThe optimal characteristics space of arable land sample set;

Feature variance S_A ²Computing formula is as follows:

${S_A}^2=\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}{(C_A-\stackrel{\‾}{C_A})}^2$

Wherein, C_AFor this characteristic value of single sample,For this characteristic mean of all samples in the sample set of same arable land,M is the number of sample in sample set for this reason, i=1 ... m;

Feature variance is analyzed, selects the optimal characteristics space of each arable land sample set.

Further: while building the Fuzzy Classifier of plant extraction in described step S5, according to the difference arable land sample obtainingThe optimal characteristics space of collection, carries out Feature Combination by great many of experiments, adopts Fuzzy classification to carry out plant extraction experiment, structureBuild multiple Fuzzy Classifiers of plant extraction.

Further: the Fuzzy Classifier of i sample set f (i) is as follows:

Wherein, Max.diff is maximum heterogeneous, and NDVI is normalized differential vegetation index, and Brightness is brightness,GLCM_{_Contrast}For texture contrast, Blue_Mean is blue wave band average, and Slope is the gradient.

The mode that the present invention combines with Fuzzy Classifier by arable land sample set has realized the automatic identification of arable land object.By selecting the attributive character space of optimum segmentation yardstick and different arable lands sample set, improve cutting apart of high resolution image rightThe homogenieity of elephant and the extraction accuracy of Fuzzy Classifier; By the structure of arable land sample set, reduced data source to time phase and numberQuantitative limitation requirement; Utilize arable land sample set attributive character and Fuzzy Classifier degree of membership value to improve image from being divided into classificationAutomaticity, be not subject to the restriction of region and data source, there is generality, be beneficial to popularization.

Brief description of the drawings

Fig. 1 is the flow chart that the present invention is based on the region adaptivity plant extraction method of high resolution image;

Fig. 2 be in embodiment mean variance with cutting yardstick change curve;

Fig. 3 be in embodiment shape index variance yields with cutting yardstick change curve.

Detailed description of the invention

Be example taking Hebei province below, the present invention is further detailed explanation by reference to the accompanying drawings.

Embodiment 1

A region adaptivity plant extraction method based on high resolution image, as shown in Figure 1, comprising:

S1, carries out pretreatment to data

Selecting high resolution image is the multispectral and panchromatic remote sensing images of several SPOT6 of Hebei province, and major phase is OctoberPart, wherein multispectral image (comprising indigo plant, green, red and near red four wave bands) has 6m spatial resolution, panchromatic remote sensing image skyBetween resolution ratio be 1.5m, this high resolution image is carried out to overshoot, geometric correction and visual fusion, if single width high-resolutionRate image, does not need to carry out visual fusion.

Meanwhile, select the 30mDEM (digital elevation model) of study area to carry out subsidiary classification, this digital elevation model is enteredThe generation of the row gradient, projective transformation and cut processing (nonessential processing), ensure its projection and point block size and SPOT6 data oneCause.

S2, carries out multi-scale division to high resolution image, determines the optimum segmentation yardstick of ploughing

(1) high resolution image is carried out to multi-scale division

Multi-scale division adopts heterogeneous minimum region merging algorithm, carries out to cut apart and appoint under the partitioning parameters of settingBusiness, by given yardstick threshold value, according to the homogeneity criterion of color and shape, based on the heterogeneous minimum principle of object, by lightThe similar adjacent picture elements of spectrum information is merged into significant object.

Adopt many characteristic weighings criterion to carry out the merging of cut zone, formula is:

F＝w₁×H_color+w₂×H_shape+w₃×H_texture

Wherein, F is heterogeneous, w₁、w₂And w₃Be respectively the feature weight of spectrum, shape and texture, w₁+w₂+w₃＝1；H_color、H_shapeAnd H_textureBe respectively the characteristic value of spectrum, shape and textural characteristics.

Through pretreated high resolution image, in conjunction with expertise and visualization, utilize easy health software to carry out imageMulti-scale division, wherein having selected spectrum and shape criteria ratio is 9: 1, (both are all smoothness parameter with degree of compacting parameterTextural characteristics) ratio be 1: 1, select 15 cut apart yardstick threshold value (10,20,30,40,50,60,70,80,90,100,110,120,130,140,150) carry out multi-scale division, formed the imaged object layer of different scale. Multi-scale division parameterSelection principle be: meeting under the prerequisite of necessary shape criteria, adopt as far as possible color standard, this be because in imageImportantly spectral information, the weight of shape criteria is too high, can reduce on the contrary the quality of cutting apart. If the border of ground object target ratioMore level and smooth, just set higher smoothness parameter, the type of ground objects compacting for shape, should arrange larger degree of compacting ginsengNumber.

(2) determine the optimum segmentation yardstick of ploughing, obtain the optimum segmentation layer of ploughing and classifying;

Suitable scale selection, is defined as after making to cut apart the inside heterogeneity of imaged object as far as possible little, and differentHeterogeneity between classification object is as far as possible large.

The present embodiment adopts a kind of optimal scale computational methods based on object mean variance and the combination of shape index variance,Its principle is spectral signature and the shape facility characteristic that considers arable land, adds up all object wave band averages in different dividing layersWith the variance of shape index, then, respectively taking the mean variance of object and shape index variance as Y-axis, Image Segmentation yardstick is XAxle, rendered object mean variance and shape index variance are along with cutting apart the curve of dimensional variation, from two curve distribution formsJudge the optimum segmentation yardstick of ploughing.

Mean variance S_C ²Computing formula be

${S_C}^2=\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}{(C_L-\stackrel{\‾}{C_L})}^2$

Wherein, be mean variance, C_LFor single object is in the brightness average of L-band,For all objects in image existThe brightness average of L-band, m is the summation of object number in image, i is object number, i=1 ... m.

Shape index variance S_SI ²Computing formula be

${S_{SI}}^2=\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}{(SI-\stackrel{\‾}{SI})}^2$

$SI=\frac{b}{V}$

Wherein, be shape index variance, the shape index value that SI is single object,For the shape of all objects in imageMean value of index, m is the summation of object number in image, i is object number, i=1 ... m, the object girth that b is single object,V is the object volume of single object.

15 multi-scale division layers calculate mean variance and the shape index of SPOT6 high resolution image near infrared bandVariance as shown in Figure 2.

It is relatively even that arable land has spectral signature, and regular shape/pattern of farming compares the features such as rule, according to theseFeature, the partition value that the reference value of the optimum segmentation in arable land occurs in mean variance peak value and shape index variance valley intersectsPlace, meet mean variance is that the yardstick of cutting apart that peak value, shape index variance are valley is arable land optimum segmentation yardstick simultaneously. ThisBe because when pure object increases, and spectral differences opposite sex increase between adjacent object, and blending objects is while increasing, with phase adjacency pairThe heterogeneous reduction of spectrum between resembling, mean variance diminishes; And shape index is the smoothness that characterizes object surface, shape is rule more, and the boundary light slippage opposite sex between adjacent object is less, and the variance of shape index just diminishes.

Obtain by Fig. 2 tracing analysis, arable land optimum segmentation yardstick is 70, can find out on 70 yardsticks by visualGood carries out arable land cutting apart of optimum.

On the optimum segmentation layer that all image processing are below is all 70 at arable land optimum segmentation yardstick, carry out.

S3 builds multiple arable lands sample set in optimum segmentation layer, forms the attributive character space of each arable land sample set

Consider that same class atural object causes that because its size, distribution, density are different difference in spectral response oftenCan cause classification difficulty, thereby a certain atural object is subdivided into some subclasses according to its spectrum and other features classifies again passableEffectively reduce classification difficulty.

Adopt the plough structure of sample set of following steps:

(1) first by visual collection arable land sample, the sample that requirement gathers is uniformly distributed and is representative, substantially containsCover the farmland types in all districts, quantity is no more than 5% of whole section object number.

(2) according to sample region, arable land, terrain slope feature, phenology feature and image feature, the sample of ploughingClassification, build multiple arable lands sample set, and form the attributive character space of each arable land sample set.

(3) arable land sample is rejected:

Although in the sample selection course of arable land, followed evenly, the rule such as homogeneity, analyzed and find, a small amount of sampleThe characteristics of objects variation on ground is too large, therefore adopt standard deviation method to reject arable land sample.

If the mean value of certain classification arable land each characteristics of objects of sample isStandard deviation isStd＝(std₁，std₂，std₃，……，std_n), for a certain arable land sample X=(x₁，x₂，x₃......，x_n), if With the plough rejecting of sample of 2 times of standard deviations.

District title, the gradient (Slope), the normalized differential vegetation index (NDVI) and bright of sample have been chosen by above methodDegree (Brightness) four indexs, to the plough structure of sample set of test block, have generated four class arable land sample sets, its sampleThe attributive character space of this collection is as shown in the table:

S4, to each arable land sample set, carries out optimum image (arable land object) feature space and selects

(1) arable land characteristics of objects space primary election

Set up storehouse, characteristics of objects space, arable land for features such as the spectrum of high resolution image, texture, shapes, as following table instituteShow:

Feature classification	Characteristic parameter
		Spectral signature	Average, brightness, normalized differential vegetation index, soil vegetative cover index
Shape facility	Length-width ratio, density, degree of compacting, the gradient
		Textural characteristics	Maximum heterogeneity, homogeneity degree, contrast, entropy

(2) the optimal characteristics space based on feature variance is selected

In arable land characteristics of objects space primary election result, the feature of selecting as far as possible to plough less and effectively forms plant extractionOptimal characteristics space. Adopt the optimal characteristics space system of selection based on feature variance, each by calculating each arable land sample setThe variance of individual feature, 2～3 features of statistical variance minimum, form the optimal characteristics space of corresponding each arable land sample set. FeatureVariance computing formula is as follows:

${S_A}^2=\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}{(C_A-\stackrel{\‾}{C_A})}^2$

Wherein, S_A ²For the variance of a certain feature, C_AFor this characteristic value of single sample,For institute in the sample set of same arable landHave this characteristic mean of sample, m is the number of sample in sample set for this reason, i=1 ... m.

Through comparative analysis to feature variance, select the optimal characteristics space of each arable land sample set, as following table:

S5, to each arable land sample set, builds the Fuzzy Classifier of plant extraction according to optimum image feature space

According to the optimal characteristics space of the difference arable land sample set obtaining, carry out Feature Combination by great many of experiments, adoptFuzzy classification carries out plant extraction experiment, has finally built multiple Fuzzy Classifiers of plant extraction.

The fuzzy membership functions of Fuzzy Classifier adopts the definite method based on Bayes criterion, is based on feature space rootJudge that according to ambiguity function imaged object is under the jurisdiction of the degree of a certain class, it is not only to differentiate classification institute according to feature spaceBelong to, simultaneously according to the subjection degree threshold value of divide of ground class, the accuracy of judgement ground class again.

$f_g\left(x\right)=\[P^{\′}\left(g\right)P_9^{*}\left(X\right)\]/\underset{i=1}{\overset{G}{\Σ}}{P}^{\′}\left(g\right)P_g^{*}\left(X\right)$

In formula: P ' is (g) prior probability that g class occurs, $P_g^{*}\left(X\right)=\frac{1}{{\left(2\mathrm{\π}\right)}^{m/2}|{\mathrm{\Σ}}_g^{*}{|}^{1/2}}\exp \[-\frac{1}{2}{(X-{\mathrm{\μ}}_g^{*})}^{\mathrm{\τ}}{\left({\mathrm{\Σ}}_i^{*}\right)}^{-1}(X-{\mathrm{\μ}}_g^{*})\],{\mathrm{\μ}}_g^{*}={({\mathrm{\μ}}_{1g}^{*},{\mathrm{\μ}}_{2g}^{*},\mathrm{...}{\mathrm{\μ}}_{mg}^{*})}^{\mathrm{\τ}}$ Be g class fuzzy vector average,Fuzzy covariance matrix, the sum that m is object, X is the object of classification, what i was object is individualNumber, i=1,2 ..., G, G is total number of g class object.

By great many of experiments, obtain four Fuzzy Classifiers that arable land sample set is corresponding, be respectively:

Wherein, the Fuzzy Classifier that f (i) is i sample set.

S6, carries out plant extraction based on arable land sample set and Fuzzy Classifier

S61, according to the attributive character space of arable land sample set, the optimum segmentation layer one by one step S2 being obtained is after cutting apartObject x (x=1,2 ... n wherein n is the total number of object) carry out attributive character calculating, comprise district title, the gradient, NDVI and brightDegree;

S62, the attributive character of the arable land sample set that the attributive character result of calculation respectively S61 being obtained and step S3 buildSpatial dimension value is mated, if the attributive character value of object, within the scope of the attributive character of a certain arable land sample set, is rememberedBe 1, otherwise be designated as 0, obtain mating matrix D_ii

$D_{ij}={\left[\begin{array}{c}{l}_1,l_2,l_3,l_4,\mathrm{...}{l}_i\\ p_1,p_2,p_3,p_4,\mathrm{...},p_i\\ n_1,n_2,n_3,n_4,\mathrm{...}{n}_i\\ b_1,b_2,b_3,b_4,\mathrm{...}{b}_i\\ \mathrm{...................}\\ w_1,w_2{w}_3,w_4,\mathrm{...}{w}_i\end{array}\right]}^T$

Wherein, l, p, n, b ..., w is respectively the attributive character of arable land sample set, such as l, and p, n, b represents respectively district nameClaim, the gradient, NDVI and brightness, i is the number of arable land sample set, the number that j is attributive character, the transposition of T representing matrix;

S63, according to definition arable land sample set in the shared weighted value of each attributive character with mate matrix D_ij, calculate respectivelyThe similarity of object and each arable land sample set, obtains similarity matrix S_i，

$S_i=D_{ij}\×{\mathrm{\δ}}_j={\left[\begin{array}{c}{l}_1,l_2,l_3,l_4,\mathrm{...}{l}_i\\ p_1,p_2,p_3,p_4,\mathrm{...}{p}_i\\ n_1,n_2,n_3,n_4,\mathrm{...}{n}_i\\ b_1,b_2,b_3,b_4,\mathrm{...}{b}_i\\ \mathrm{...................}\\ w_1,w_2,w_3{w}_4,\mathrm{...}{w}_i\end{array}\right]}^T\×\left[\begin{array}{c}{\mathrm{\δ}}_1\\ {\mathrm{\δ}}_2\\ {\mathrm{\δ}}_3\\ {\mathrm{\δ}}_4\\ \mathrm{...}\\ {\mathrm{\δ}}_j\end{array}\right]=\left[\begin{array}{c}{S}_1\\ S_2\\ S_3\\ S_4\\ \mathrm{...}\\ S_i\end{array}\right]$

Wherein, δ_iFor sample set attributive character weight matrix, δ_iCarry out according to region situation self-defined, generally, δ_iBe 0.5.

In the present embodiment, j=4, the district title in sample set attributive character, the gradient, the weight of NDVI and brightness is respectivelyBe 0.3,0.2,0.3,0.2, $\mathrm{\δ}=\left[\begin{array}{c}0.3\\ 0.2\\ 0.3\\ 0.2\end{array}\right];$

S64, sets similarity matrix S_iThreshold decision condition, if the similarity of object x and certain arable land sample set is fullFoot threshold decision condition, object and certain arable land sample set match;

In the present embodiment, threshold condition is set as:

IfS_iA >=0.5 object x and j arable land sample set match.

Otherwise, restart step S61, calculate next object x+1.

Wherein, object x can match with multiple sample sets simultaneously.

S65, according to the matching result of S64, carries out Fuzzy Classifier corresponding to arable land sample set matching to this object,Obtain class categories function E (i) and classification degree of membership, its principle is that multiple Fuzzy Classifiers execution classification results are set to 1With 0, wherein 1 for ploughing, and 0 is bare place, and the degree of membership value of then Fuzzy Classifier being ploughed, as weighted value, adopts weighting to throwTicket rule judges, meets the final differentiation of certain threshold range condition for ploughing.

S66, adopts the Nearest Neighbor with Weighted Voting method weighted calculation of ploughing, and computing formula is as follows:

$E\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}\left(E\right(i)\×\mathrm{\ω}(i\left)\right)$

Wherein, E (x) is arable land weighting discriminant function, and n is the Fuzzy Classifier sum matching with object x, and i is fuzzyThe number of grader, i=1 ..., n;

Whether S67, be to plough to judge to object, and Rule of judgment is as follows:

Repetitive cycling, until all calculation and object completes.

Claims (8)

1. the region adaptivity plant extraction method based on high resolution image, is characterized in that, described extracting method bagDraw together:

S1, carries out pretreatment to data;

S2, carries out multi-scale division to high resolution image, determines the optimum segmentation yardstick of ploughing, and obtains the optimum of ploughing and classifyingDividing layer;

S3 builds multiple arable lands sample set in optimum segmentation layer, forms the attributive character space of each arable land sample set;

S4, to each arable land sample set, carries out optimum image feature space and selects;

S5, to each arable land sample set, builds the Fuzzy Classifier of plant extraction according to optimum image feature space;

S6, carries out plant extraction based on arable land sample set and Fuzzy Classifier:

S61, according to the attributive character space of arable land sample set, carries out attributive character to the object after step S2 optimum segmentation one by oneCalculate;

S62, the attributive character space of the arable land sample set that the attributive character result of calculation respectively S61 being obtained and step S3 buildValue range mates, if the attributive character value of object is within the scope of the attributive character of a certain arable land sample set, is designated as 1,Otherwise be designated as 0, obtain mating matrix D_ij

$D_{ij}={\left[\begin{array}{c}{l}_1,l_2,l_3,l_4,...l_i\\ p_1,p_2,p_3,p_4,...p_i\\ n_1,n_2,n_3,n_4,...n_i\\ b_1,b_2,b_3,b_4,...b_i\\ \mathrm{...................}\\ w_1,w_2,w_3,w_4,...w_i\end{array}\right]}^T$

Wherein, l, p, n, b ..., w is respectively the attributive character of arable land sample set, and i is the number of arable land sample set, and j is attribute spyThe number of levying, the transposition of T representing matrix;

S63, according to definition arable land sample set in the shared weighted value of each attributive character with mate matrix D_ij, calculating object respectivelyWith the similarity of each arable land sample set, obtain similarity matrix S_i，

$S_i=D_{ij}\×{\mathrm{\δ}}_j={\left[\begin{array}{c}{l}_1,l_2,l_3,l_4,...l_i\\ p_1,p_2,p_3,p_4,...p_i\\ n_1,n_2,n_3,n_4,...n_i\\ b_1,b_2,b_3,b_4,...b_i\\ ..................\·\\ w_1,w_2,w_3,w_4,...w_i\end{array}\right]}^T\×\left[\begin{array}{c}{\mathrm{\δ}}_1\\ {\mathrm{\δ}}_2\\ {\mathrm{\δ}}_3\\ {\mathrm{\δ}}_4\\ ...\\ {\mathrm{\δ}}_j\end{array}\right]=\left[\begin{array}{c}{S}_1\\ S_2\\ S_3\\ S_4\\ ...\\ S_i\end{array}\right]$

Wherein, δ_iFor sample set attributive character weight matrix, δ_iCarry out self-defined according to region situation;

S64, sets similarity matrix S_iThreshold decision condition, if the similarity of object and certain arable land sample set meets threshold valueRule of judgment, object and certain arable land sample set match;

S65, according to the matching result of S64, carries out Fuzzy Classifier corresponding to arable land sample set matching to this object, obtainClass categories function E (i) and classification degree of membership ω (i),

S66, adopts the Nearest Neighbor with Weighted Voting method weighted calculation of ploughing, and computing formula is as follows:

$E\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}\left(E\right(i)\×\mathrm{\ω}(i\left)\right)$

Wherein, E (x) is arable land weighting discriminant function, and n is the Fuzzy Classifier sum matching with object x, and i is fuzzy classificationThe number of device, i=1 ..., n;

Whether S67, be to plough to judge to object, and Rule of judgment is as follows:

2. the region adaptivity plant extraction method based on high resolution image as claimed in claim 1, is characterized in that, numberData preprocess comprises to be processed high resolution image and digital elevation model, and wherein, image processing comprises to be carried out imageRadiation and geometric correction, digital elevation model processing comprises the generation gradient.

3. the region adaptivity plant extraction method based on high resolution image as claimed in claim 1, is characterized in that instituteState in step S2, the optimum segmentation yardstick in arable land is that to meet mean variance be that peak value, shape index variance are valley point simultaneouslyCut yardstick;

Wherein, mean variance S_C ²Computing formula be

Wherein, C_LFor single object is in the brightness average of L-band,For all objects in image equal in the brightness of L-bandValue, m is the summation of object number in image, i is object number, i=1 ... m;

Shape index variance S_SI ²Computing formula be ${S_{SI}}^2=\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m{(SI-\stackrel{\‾}{S}I)}^2,$ $SI=\frac{b}{V}$

Wherein, the shape index value that SI is single object,For the shape index average of all objects in image, m is in imageThe summation of object number, i is object number, i=1 ... m, the object girth that b is single object, the object that V is single objectVolume.

4. the region adaptivity plant extraction method based on high resolution image as claimed in claim 1, is characterized in that: instituteState in step S3, the step that builds arable land sample set is as follows:

S31, by visual collection arable land sample, the sample that requirement gathers is uniformly distributed and is representative, contains all districtsFarmland types, quantity is no more than 5% of whole section object number;

S32, according to sample region, arable land, terrain slope feature, phenology feature and image feature, the returning of the sample of ploughingClass, builds multiple arable lands sample set, and forms the attributive character space of each arable land sample set;

S33, rejects arable land sample:

If the mean value of certain classification arable land each characteristics of objects of sample isStandard deviation is Std=(std₁，std₂，std₃，……，std_n), for a certain arable land sample X=(x₁，x₂，x₃......，x_n), if With the plough rejecting of sample of 2 times of standard deviations.

5. the region adaptivity plant extraction method based on high resolution image as claimed in claim 4, is characterized in that: instituteState district title that sample region, arable land, terrain slope feature, phenology feature and image feature be respectively sample, the gradient,NDVI and brightness, the attributive character space of each arable land sample set comprises district title, the gradient, NDVI and the brightness of sample.

6. the region adaptivity plant extraction method based on high resolution image as claimed in claim 1, is characterized in that: instituteState in step S4, the step that optimum image feature space is selected is as follows:

S41, arable land characteristics of objects space primary election:

Set up storehouse, characteristics of objects space, arable land for spectrum, texture, the shape facility of high resolution image:

S42, carries out the selection of optimal characteristics space based on feature variance:

Calculate the variance of each arable land each feature of sample set, 2～3 features of statistical variance minimum, form corresponding each arable landThe optimal characteristics space of sample set;

Feature variance S_A ²Computing formula is as follows:

${S_A}^2=\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}{(C_A-\stackrel{\‾}{C_A})}^2$

Wherein, C_AFor this characteristic value of single sample,For this characteristic mean of all samples in the sample set of same arable land, m isThe number of sample in this sample set, i=1 ... m;

Feature variance is analyzed, selects the optimal characteristics space of each arable land sample set.

7. the region adaptivity plant extraction method based on high resolution image as claimed in claim 1, is characterized in that: instituteWhile stating the Fuzzy Classifier that builds plant extraction in step S5, according to the optimal characteristics space of the difference arable land sample set obtaining,Carry out Feature Combination by great many of experiments, adopt Fuzzy classification to carry out plant extraction experiment, build the multiple of plant extractionFuzzy Classifier.

8. the region adaptivity plant extraction method based on high resolution image as claimed in claim 7, is characterized in that: theThe Fuzzy Classifier of i sample set f (i) is as follows:

Wherein, Max.diff is maximum heterogeneous, and NDVI is normalized differential vegetation index, and Brightness is brightness,GLCM__ContrastFor texture contrast, Blue_Mean is blue wave band average, and Slope is the gradient.