CN104331698A - Remote sensing type urban image extracting method - Google Patents

Abstract

The invention provides a remote sensing type urban image extracting method. The method comprises the following steps: S1, performing characteristics extracting for altitude data of an ASTER VNIR satellite remote sensing image and derived products and a sample of a PALSAR HH/HV satellite remote sensing image; S2, extracting obvious urban and non-urban points based on the spectral characteristics of the urban and non-urban part; S3, performing confidence spreading by the LLGC using the obvious urban and non-urban points as the initial information based on the characteristics distribution feature of data to be classified, so as to obtain an urban confidence map; S4, obtaining the confidence of the whole remote sensing image, weighting and randomly sampling to obtain a training sample; S5, classifying the urban based on SVM, namely, classifying by the SVM method on the basis of the characteristic vectors extracted in the step S1 and the sample data extracted in step S4, and then obtaining a urban map subjected to binarization according to a classification label. With the adoption of the method, the problems of high cost and high time consumption and the like caused by manual sampling in the prior art can be solved.

Description

A kind of remote sensing images city extracting method

Technical field

The present invention relates to Remote Sensing Image Processing Technology, more specifically, relate to a kind of remote sensing images city extracting method.

Background technology

In population in the world, the ratio sustainable growth of urban population.Have the inhabitants live exceeding half at present in city.Correlative study important in city planning, social pattern analysis, environmental change, Disaster control etc. of urbanization.The geographical relevant information in city is generally grasped by administrative unit, but has certain limitation, comprises as follows:

There are differences about the definition in " city " between-various countries;

-describe comparatively delayed to the dynamic change in city;

The information sharing of-city not easily;

-developing country accurate city acquisition of information not easily.

In the current analysis to city map, the city cartographic analysis based on remote sensing images becomes a kind of main way of urbanization correlative study.Remote sensing analysis method is analyzed based on the multispectral characteristic in city, to obtain high-resolution city map.It does not rely on administrative division definition, still can keep consistency, be conducive to carrying out Global Urbanization correlation analysis between country variant, region.

But current problems faced is, the city of Different Climatic Zones often differs greatly in multispectral characteristic, and Fig. 1-3 shows the pseudo-color data of ASTER (b1 ~ b3) spectrum of three different cities, and the region in figure in circle roughly represents city.As can be seen from the figure the color distribution characteristic in city differs greatly in three width figure, and non-city is also like this.

For this problem, supervised classification method is the main stream approach of the current urban area recognition based on remote sensing images.The method by manual city pixel in image range and the non-city pixel chosen as training sample, then by supervised classification method (as support vector machines, artificial neural network, decision tree etc.), the study spectral characteristic in city/non-city Modling model, then carry out classification acquisition city map to remotely-sensed data.Fig. 4 shows a training sample and gathers schematic diagram, this figure is from " by carrying out the automated process of earth city mapping in conjunction with ASTER satellite image and GIS data " (Miyazaki et al. of the people such as document Miyazaki, An Automated Method for Global Urban Area Mapping by Integrating ASTER Satellite Images and GIS Data), the figure shows the visual deciphering for researching and developing ground truth data collection.According to this figure, by specific human-computer interaction interface, choose sample point by hand by researchist.

Generally speaking, the multi-spectral remote sensing image city maps processing flow process based on supervised classification method is as follows:

(1) from multispectral image, eigenvector is extracted

Multispectral image is generally represented by multi-spectral remote sensing image data set, such as, be expressed as:

{ Img _b, b=1,2 ..., B, B are wave band sum

For the pixel i in image, extract correlated characteristic by various method, generating feature vector

(2) by training sample, city disaggregated model is learnt

For supervised classification method, need known training sample set:

l is the generic label (Label) of sample point

(3) by study after models applying in entire image, obtain the class label (city/non-city) of each pixel, thus obtain the city map of city binaryzation.

For city classification problem, the codomain of L is set {-1,1}.Value is 1 represent city, and-1 represents non-city.

There are the following problems for said method:

● the acquisition of training sample is wasted time and energy

It is hand-manipulated that the method needs have the researchist of abundant process experience to come to remote sensing images, and pointwise is chosen.And, due to the change of different geographical city spectral characteristic, sample point generally only applicable present image and adjacent domain.

● analyze for global city, need to sample to each place, city, manpower and materials consume huge, with high costs.

Summary of the invention

For problems of the prior art, the present invention proposes a kind of new remote sensing images city extracting method, comprise: S1), for the sample of ASTER VNIR satellite remote sensing images and derived product altitude figures and PALSAR satellite remote sensing images, carry out feature extraction; S2), based on the spectral characteristic priori in city and non-city, carry out remarkable city and non-city point extracts; S3), using this part remarkable city and non-city point as initial information, in conjunction with the feature distribution character of data to be sorted, by LLGC method, carry out degree of confidence diffusion, obtain city confidence map; S4), after obtaining the degree of confidence of whole remote sensing images, carry out stochastic sampling with this weighting, generate training sample; S5), carry out city classification based on SVM, comprising: with the eigenvector extracted in step S1, and based on the sample data extracted in step S4, classified by traditional SVM method, obtain the city map of binaryzation according to tag along sort.

Method of the present invention is a kind of adaptive remote sensing images city extracting method.It is based on the spectrum essential information in city in remote sensing images, in conjunction with the spectral distribution property of data to be sorted, by the mode that degree of confidence spreads, realize automatically choosing of training sample, then based on the training sample generated, by the support vector machines method in supervised classification method, remotely-sensed data is classified, generate city map.

Accompanying drawing explanation

Fig. 1-3 shows the pseudo-color data of ASTER (b1 ~ b3) spectrum of three different cities;

Fig. 4 shows a training sample and gathers schematic diagram;

Fig. 5-6 is the process flow diagram of method of the present invention;

Fig. 7 is the experimental result picture of method of the present invention.

Embodiment

The treatment scheme of method of the present invention as seen in figs. 5-6.Method of the present invention comprehensively uses the remotely-sensed data of two types:

ASTER (advanced spaceborne heat radiation and reflection measuring set, Advanced Spaceborne Thermal Emission and Reflection Radiometer) middle VNIR (visible ray and near-infrared radiometer, visible and near-infrared radiometer) obtain 4 wave bands of satellite remote sensing images, be designated as Aster respectively _b1~ Aster _b4spectral range be respectively 0.52-0.60 μm, 0.63-0.69 μm, 0.76-0.86 μm (nadir view), 0.76-0.86 μm (backward scan), and related derivative product altitude figures (slope data, is designated as slope); And

PALSAR (L-band phased array synthetic-aperture radar, Phased Array L-band Synthetic Aperture Radar) HH, HV satellite remote sensing images (2 wave bands, be designated as hh, hv), and the HH wave band data after carrying out correction process to incident angle (is designated as hh _cor).

With reference to figure 6, in step S1, for the sample of ASTER VNIR satellite remote sensing images and derived product altitude figures and PALSAR HH/HV satellite remote sensing images, carry out feature extraction.

In the present invention, calculate altogether and employ 12 features, 8 wherein original class input data (Aster _b1~ Aster _b4, slope, hh, hv, hh _cor) be also the part of feature, this is 8 features.Except original input data, extra 4 kinds of features are also used: NDVI (normalized differential vegetation index in the present invention, Normalized Difference Vegetation Index), NDWI (normalization aqua index, Normalized Difference Water Index), hh _suband hh _ent.Wherein, NDVI and NDWI is two kinds of general features, is applicable to distinguish vegetation and water body, all according to ASTER VNIR satellite remote sensing images data Aster _b1~ Aster _b4calculate.Hh _subaccording to PALSAR HH satellite remote sensing images data hh and hh _corcalculate, hh _subcertain effect is had to differentiation mountain range (non-city).Hh _entcalculate according to PALSAR HH satellite remote sensing images data hh, hh _entenrich degree for what describe texture information, the texture information of general remarkable non-city part is all less.The computing method of these 4 kinds of features are as formula (1)-(4):

$\mathrm{NDVI}=\frac{{\mathrm{Aster}}_{b3}-{\mathrm{Aster}}_{b2}}{{\mathrm{Aster}}_{b3}+{\mathrm{Aster}}_{b2}}---\left(1\right)$

$\mathrm{NDWI}=\frac{{\mathrm{Aster}}_{b1}-{\mathrm{Aster}}_{b3}}{{\mathrm{Aster}}_{b1}+{\mathrm{Aster}}_{b3}}---\left(2\right)$

hh _sub＝|hh-hh _cor| (3)

hh _ent＝EntropyFilt(hh) (4)

Wherein, EntropyFilt () represents that entropy filtering is (see Eddins, S.L.; Gonzalez, R.; Woods, R.Digital image processing using Matlab.Princeton Hall Pearson Education Inc., New Jersey 2004), in the present invention, EntropyFilt () acts on (analysis window is of a size of 15x15 pixel) on hh image.

With reference to figure 6, in step S2, for ASTER VNIR satellite remote sensing images and PALSAR HH satellite remote sensing images, based on the spectral characteristic in city/non-city, carry out remarkable city/non-city point and extract.

According to priori, so-called remarkable city/non-city point refers to, it has spectral characteristic comparatively clearly, by the clear and definite discrimination in addition of simple many assembled classifiers.The spectral characteristic of remarkable city/non-city point has good accuracy, and is applicable to the data of different geographical.

But in general, the significant point negligible amounts extracted, can not represent the real features distribution of city/non-city data.Therefore, significant point can not directly as training sample, and also needing distributes with the feature of input data further combines and perfect.

The present invention to the principle that significant point extracts is, based on the binaryzation mask image calculated in satellite remote sensing images, then two to the form opening and closing operation with binaryzation mask image determines remarkable city and non-city point, wherein, the comparison threshold value of binaryzation mask image is according to NDVI, DNWI, hh of this satellite remote sensing images _sub, hh and/or hh _entcalculate.

The concrete grammar that significant point of the present invention extracts is see following formula (5) ~ (18).For non-city, produce 6 mask (binaryzation mask image, for representing the non-city significant point meeting specified conditions).For mask ₁~ mask ₆if condition meets, then the value of mask is 1, otherwise is 0.Mask ₁and mask ₂set according to NDVI and NDWI respectively.If the value of pixel is higher than mean value, then mark this pixel, represent vegetation region and pool.Mask ₃in order to identify that hh wave band incident angle corrects the non-city indicated by the difference of front and back, mountain range (non-city) part effectively can be identified.In addition, because the buildings in city has higher reflectivity, but not city partial reflectance is relatively low, and therefore in PALSAR HH image, the value in non-city will lower than average, mask ₄setting namely based on this.Degree is enriched, mask according to what analyze slope data and texture ₅and mask ₆for compared with given threshold value.Based on experience, threshold value thresh ₅and thresh ₆be set to 15 and 4.5 respectively.

Then, by specific morphology operations template (10x10 size, complete 1) to mask ₁~ mask ₆the two-value form opening and closing operation (being designated as MorphFilt ()) of carrying out, for refining mask.Its objective is and remove isolated non-city (that includes a small amount of pixel), and obtain mask by the mask after merging these refinements _nonurban, as shown in formula (15).

To the prediction in city based on mask ₄similar mode, with mask _nonurbanmerge and obtain mask after carrying out identical form opening and closing operation _urban, as shown in formula (16)-(18).

mask ₁＝NDVI＞thresh ₁(5)

thresh ₁＝mean(NDVI)+std(NDVI) (6)

mask ₂＝NDWI＞thresh ₂(7)

thresh ₂＝mean(NDWI)+std(NDWI) (8)

mask ₃＝hh _sub＞thresh ₃(9)

thresh ₃＝mean(hh _sub)+std(hh _sub) (10)

mask ₄＝hh＜thresh ₄(11)

thresh ₄＝mean(hh)-std(hh) (12)

mask ₅＝slope＞thresh ₅(13)

mask ₆＝hh _ent＜thresh ₆(14)

${\mathrm{mask}}_{\mathrm{nonurban}}=\underset{i=1}{\overset{6}{\∪}}\mathrm{MorphFilt}\left({\mathrm{mask}}_i\right)---\left(15\right)$

mask ₇＝hh＞thresh ₇(16)

thresh ₇＝mean(hh)+std(hh) (17)

mask _urban＝MorphFilt(mask ₇∩Not(mask _nonurban)) (18)

Here, mean () represents the mean value of input picture, and std () represents standard variance, and Not () represents the reciprocal value of binary value.Wherein threshold value thresh ₅and thresh ₆for empirical value setting, other thresh are the adaptive threshold calculated according to present analysis image.

By above-mentioned steps, remarkable city/non-city point can be obtained.

With reference to figure 6, in step S3, using this part remarkable city/non-city point is as initial information, in conjunction with the feature distribution character of data to be sorted, by LLGC (Learning with Local and Global Consistency) method (see the learning method based on local and global coherency, Zhou, D.; Bousquet, O.; Lal, T.N.; Weston, J.; b.Learning with local and global consistency.Advances in neural information processing systems 2004,16,321 – 328.), carry out degree of confidence diffusion, obtain city confidence map.So-called unfiled data refer to also does not determine it is the data of city or non-city point.

In the present invention, 3 wave band Aster of ASTER VNIR satellite remote sensing images _b1~ Aster _b3image be merged into a coloured image Aster as cluster feature _rgb.

In one embodiment, based on LLGC method, using remarkable city/non-city point as initial value, feature based vector similarity between any two builds diffusion strength criterion, the confidence information in city/non-city is diffused in the feature space of unfiled data.Specifically comprise:

1) LLGC method is input as image Aster _rgbsample set X={x ₁, x ₂..., x _n, wherein x _ifor vector, the feature of representative sample.N is sample size.In addition setting label matrix F is Nx2 matrix, and 1,2 elements of the i-th row represent the degree of confidence that i-th sample belongs to city, non-city respectively.

2) the remarkable city/non-city point setting of initial value by extracting of label matrix F.City point is expert at and is set to [10], and non-city is set to [01], and other are set to [00].

3) dispersion relation between two between sample is provided by diffusion matrix W and normalization diffusion matrix S, as shown in formula (19)-(21):

$\mathrm{dist}(x_i,x_j)=\underset{s\∈\{b1,b2,b3\}}{\mathrm{\Σ}}{\left|\right|{\mathrm{Aster}}_s\left(x_i\right)-{\mathrm{Aster}}_s\left(x_i\right)\left|\right|}^2---\left(19\right)$

$W_{i,j}=\left\{\begin{array}{cc}\exp [-\mathrm{dist}(x_i,x_j)/{2\mathrm{\σ}}^2],& \mathrm{if\; i}\≠j\\ 0,& \mathrm{if\; i}=j\end{array}\right.---\left(20\right)$

S＝D ^-1/2WD ^-1/2(21)

Wherein, dist () is scalar function, representation feature vector x _iand x _jdifference. constant σ represents diffusion kernel size.D is the diagonal matrix of NxN, and i-th diagonal element is the i-th row element sum of W.

4) by LLGC method, the final label matrix F after diffusion ^*be expressed as shown in formula (22):

F ^*＝(1-α)(I-αS) ^-1F(0) (22)

But, original LLGC method due to the cause actual Use Limitation rate of operand lower.The pixel number of the remote sensing images full figure analyzed generally is not less than 10,000,000, i.e. (N>10 ⁷), W, S are the non-sparse matrix of NxN size, are difficult to directly realize (generating 10 of a double type in theory by conventional method ⁷x10 ⁷matrix needs about 700TB storage space).

Therefore, the present invention proposes a kind of LLGC method of improvement, and input picture is carried out quantization index process, revises corresponding computing and increases relevant mapping process.Significantly can reduce operand and keep good diffusion effect.

In a preferred embodiment, above-mentioned LLGC method is improved, specifically comprises:

1) by image Aster _rgbcarry out quantization transform, become thumbnail, maximum color index number M is lower than N, and be generally far below N, in one example, such as M is set to 300.

2) sample set becomes X={ (x _i, n _i), i=1,2 ..., M}, wherein n _ibe the quantity of i-th sample.The pixel with same color index is considered to same sample, n _ipixel sum then corresponding to this index.

3) to the sample set application LLGC algorithm after conversion, carry out degree of confidence diffusion, obtain city confidence map.

Particularly, each calculation procedure in former LLGC algorithm is transformed to formula (23)-(25):

W _i，j＝exp[-dist(x _i，x _j)/2σ ²] (23)

$d_j=\left(\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{n}_i{W}_{i,j}\right)-1---\left(24\right)$

$S_{i,j}=\left\{\begin{array}{cc}{d}_i^{-\frac{1}{2}}{d}_j^{-\frac{1}{2}}{W}_{i,j}*n_j,& \mathrm{if\; i}\≠j\\ d_i^{-\frac{1}{2}}{d}_j^{-\frac{1}{2}}{W}_{i,j}*(n_j-1),& \mathrm{if\; i}=j\end{array}\right.---\left(25\right)$

Now incidence matrix W, S is the matrix of MxM size, and be namely no more than 300x300 size, therefore operand does not reconstruct problem.Meanwhile, set up index and only slight impact caused on diffusion process, therefore performance and former algorithm close.

Based on mask _urban, find the maximum city be connected and select subgraph according to its bounding rectangles.The quantity of the city in this subgraph/non-city pixel by average, and at this area applications LLGC algorithm.Also there is according to the pixel with same index color the rule of identical degree of confidence, the city confidence map of subgraph is marked again and gets back on whole figure.In this way, the LLGC algorithm of improvement efficiently and effectively can generate city confidence map.

With reference to figure 6, in step S4, after obtaining the degree of confidence of all data of full figure of whole remote sensing images, carry out stochastic sampling with this weighting, generate training sample.

In step S4, obtain the training data (such as city/non-city pixel) for classifying further by weight sampling, wherein the degree of confidence of each pixel is used as weight.Concrete method of weighting comprises step S401-S403:

S401) to the degree of confidence C in the city/non-city of each pixel _u, C _nunormalization, such as, make these two sums be 1.

S402) for city, C is chosen _uthe pixel being greater than first threshold (such as 0.5) alternatively collects, then stochastic sampling method (such as document " a kind of sigma-t method of diffusion followed the tracks of for vision " is adopted, Isard M, Blake A.Condensation-conditional density propagation for visual tracking [J] .International journal of computer vision, 1998,29 (1): 5-28.), with C _uas weight, choose city training sample (sample points is such as 500).

S403) for non-city, C is chosen _nuthe pixel being greater than Second Threshold (such as 0.5) alternatively collects, then with C _nuas weight, identical step is adopted to obtain non-city training sample (sample points is such as 300).

With reference to figure 6, in step S5, based on SVM (support vector machine, Support Vector Machine) method (Cristianini, N.; Shawe-Taylor, J.An introduction to support vector machines and other kernel-based learning methods; Cambridge university press, 2000) city classification is carried out, specifically comprise: 10 category features extracted in step S1 are as eigenvector, based on the training sample data extracted in step S4, classified by traditional SVM method, obtain the city map of binaryzation according to tag along sort.10 used category features are Aster _b1~ Aster _b4, slope, NDVI, NDWI, hh, hv, hh _ent.

For method of the present invention, also test, experimental design is as follows:

● for verifying the validity of this method, the city of 75 Different Climatic Zones in global range is analyzed;

● city/non-city point that each city is manual chooses total about 100 as true value for checking;

● map compares;

● compare with city, Remote Sensing Products MCD12Q1 whole world map;

● using half true value point as training sample, then obtain city map by the sorting technique in SVM and compare.

Evaluation criterion adopts general quantitatively evaluating standard, and formula (26)-(29) represent user's precision, producer's precision, resultnat accuracy and kappa coefficient respectively:

${\mathrm{User}}^{\′}\mathrm{s\; accuracy}=\frac{n_{\mathrm{uu}}}{n_{u+}}---\left(26\right)$

${\mathrm{Producer}}^{\′}\mathrm{s\; accuracy}=\frac{n_{\mathrm{uu}}}{n_{+u}}---\left(27\right)$

$\mathrm{Overall\; accuracy}=\frac{n_{\mathrm{uu}}+n_{\mathrm{nn}}}{n}---\left(28\right)$

$\mathrm{Kappa}=\frac{(n_{\mathrm{uu}}+n_{\mathrm{nn}})/n-(n_{u+}{n}_{+u}+n_{n+}{n}_{+n})/n^2}{1-(n_{u+}{n}_{+u}+n_{n+}{n}_{+n})/n^2}---\left(29\right)$

In formula, each symbol implication is as shown in table 1 below, and table 1 illustrates the structure for city/non-city classification results matrix.

Table 1 city/non-city classification results matrix

Experimental result is as shown in table 2 and Fig. 7.Table 2 shows the estimation accuracy of city map.

The estimation accuracy of table 2 city map

Fig. 7 shows the comparative result of the city map in four places, (a) in Fig. 7 is ASTER/VNIR pseudo color coding hologram figure, b () is PALSAR pseudo color coding hologram figure, c city map that () obtains for method of the present invention, d () is MCD city map, (e) is SVM city map.Result shows, the city Map quality generated is better than MCD, close with supervised classification SVM method performance.

Methods combining priori of the present invention and degree of confidence method of diffusion, can the spectrum change in city between good conformity different geographical.

The Map quality that this method generates is better than MCD, very close with supervised classification SVM method performance.

The problems such as the high cost of artificial sample needed for present method solves in supervised classification method, height are consuming time, automatically carry out processing and excellent performance, are expected to be widely used in the related application such as the drawing of global city and Dynamic profiling description thereof.

Although combined and be considered to feasible illustrative embodiments at present and describe the present invention, but will understand, the invention is not restricted to disclosed illustrative embodiments, but on the contrary, the present invention is intended to cover and is included in various distortion in the spirit and scope of claims and equivalent arrangements.

Claims (8)

1. a remote sensing images city extracting method, is characterized in that, comprising:

S1), for the sample of ASTER VNIR satellite remote sensing images and derived product altitude figures and PALSAR HH, HV satellite remote sensing images, feature extraction is carried out;

S2), for ASTER VNIR satellite remote sensing images and PALSAR HH satellite remote sensing images, based on the spectral characteristic in city and non-city, carry out remarkable city and non-city point extracts;

S3), using this part remarkable city and non-city point as initial information, in conjunction with the feature distribution character of data to be sorted, by LLGC method, carry out degree of confidence diffusion, obtain city confidence map;

S4), after obtaining the degree of confidence of whole remote sensing images, carry out stochastic sampling with this weighting, generate training sample;

S5), carry out city classification based on SVM, comprising: with the eigenvector extracted in step S1, and based on the sample data extracted in step S4, classified by traditional SVM method, obtain the city map of binaryzation according to tag along sort.

2. remote sensing images city according to claim 1 extracting method, is characterized in that, in step sl, the feature of extraction comprises: input data Aster _b1~ Aster _b4, slope, hh, hv, hh _cor, NDVI, NDWI, hh _suband hh _ent, wherein, Aster _b1~ Aster _b4represent the data of 4 wave bands of ASTER VNIR satellite remote sensing images, slope represents the derived product altitude figures of ASTER satellite remote sensing images, hh and hv represents the data of HH and the HV wave band of PALSAR satellite remote sensing images respectively, hh _correpresent the incident angle of PALSAR satellite remote sensing images carry out correction process after HH wave band data, NDVI represents normalized differential vegetation index, and NDWI represents normalization aqua index, hh _subaccording to PALSAR HH satellite remote sensing images data hh and hh _corcalculate, for distinguishing the data in non-city, hh _entcalculate, for describing texture information according to PALSAR HH satellite remote sensing images data hh.

3. remote sensing images city according to claim 2 extracting method, it is characterized in that, in step s 2, based on the binaryzation mask image calculated in satellite remote sensing images, then the two-value form opening and closing operation with binaryzation mask image determines remarkable city and non-city point, wherein, the comparison threshold value of binaryzation mask image is according to NDVI, DNWI, hh of this satellite remote sensing images _sub, hh and/or hh _entcalculate.

4. remote sensing images city according to claim 1 extracting method, it is characterized in that, step S3 also comprises:

1) by 3 wave band Aster of ASTER VNIR satellite remote sensing images _b1~ Aster _b3image be merged into a coloured image Aster as cluster feature _rgb;

2) based on LLGC method, using remarkable city/non-city point as initial value, feature based vector similarity between any two builds diffusion strength criterion, the confidence information in city/non-city is diffused in the feature space of unfiled data.

5. remote sensing images city according to claim 1 extracting method, it is characterized in that, step S3 also comprises:

1) by image Aster _rgbcarry out quantization transform, become thumbnail, setting maximum color index quantity M is lower than former sample size N;

2) for the new samples collection with new samples quantity after conversion, mark has the pixel of same color index for being same sample;

3) to new samples collection, feature based vector similarity between any two builds diffusion strength criterion, based on LLGC method, the confidence information in city/non-city is diffused in the feature space of unfiled data.

6. remote sensing images city according to claim 1 extracting method, it is characterized in that, step S4 also comprises:

S401), to the city of each pixel and the degree of confidence C in non-city _u, C _nunormalization;

S402), for city, choose C _uthe pixel being greater than first threshold alternatively collects, and then adopts stochastic sampling method, with C _uas weight, choose city training sample;

S403), for non-city, C is chosen _nuthe pixel being greater than the second roe alternatively collects, then with C _nuas weight, obtain non-city training sample.

7. remote sensing images city according to claim 6 extracting method, is characterized in that, in step S401, make degree of confidence C _u, C _nutwo sums are 1, and in step S402 and S403, described first threshold and Second Threshold are set to 0.5.

8. remote sensing images city according to claim 1 extracting method, is characterized in that, in step s 5: 10 used category features are Aster _b1~ Aster _b4, slope, NDVI, NDWI, hh, hv and hh _ent.