CN105930772A - City impervious surface extraction method based on fusion of SAR image and optical remote sensing image - Google Patents

Abstract

Provided is a city impervious surface extraction method based on fusion of an SAR image and an optical remote sensing image. The method comprises that a general sample set formed by samples of a research area is selected in advance, and a classifier training set, a classifier test set and a precision verification set of impervious surface extraction results are generated from the general sample set in a random sampling method; the optical remote sensing image is configured with the SAR image of the research area, and features are extracted from the optical remote sensing image and the SAR image; training is carried out, the city impervious surface is extracted preliminarily on the basis of a random forest classifier, and optimal remote sensing image data, SAR image data and an impervious surface RF preliminary extraction result are obtained; decision level fusion is carried out by utilizing a D-S evidence theory synthesis rule, and a final impervious surface extraction result of the research area is obtained; and the precision of each extraction result is verified via the precision verification set. Advantages of the optical remote sensing image and SAR image data sources are utilized fully, the SAR image and optical remote sensing image fusion method based on the RF and D-S evidence theory is provided, and the impervious surface of higher precision in the city is obtained.

Description

The city impervious surface extracting method merged with optical remote sensing image based on SAR image

Technical field

The invention belongs to remote sensing image data information extraction field, relate to a kind of city impervious surface extracting method merged with optical remote sensing image based on SAR image.

Background technology

Impervious surface refers to that water can not be by the non-natural ground surface being seeped in soil under it, including various cultures, road, parking lot, airport etc..Impervious surface is as the indicator of tolerance Development of Urbanization degree, in recent years, along with the continuous quickening of urbanization process, the substantial increase of impervious surface, causes the appearance of a series of Urban Environmental Problems such as earth's surface transpiration minimizing, rainwash acceleration, and urban heat island.Therefore, extract the distribution of impervious surface and the assessment of urban ecological environment and the monitoring of urbanization process and management etc. are all had and be of great significance by level of coverage.Remote sensing is due to its feature such as low cost, convenience, it has also become the Main Means of impervious surface estimation.At present, the remotely-sensed data of impervious surface estimation is divided into following several:

(1) low resolution optical remote sensing image in, such as Landsat TM/ETM, MODIS image, ASTER image, Hyperion image etc.；

(2) high-resolution remote sensing image, such as IKONOS, Quick Bird etc.；

(3) other remote sensing image, such as SAR, LiDAR image etc.；

Usually, based on remote sensing image, impervious surface is estimated main method and have impervious surface based on various indexes extraction, spectral mixture analysis (Spectral Mixture Analysis, SMA), classification and regression tree CART algorithm (Classification And Regression Tree, CART), artificial neural network (Artificial Neural Networks, ANN), SVMs (Support Vector Machine, SVM), object-oriented (Object Oriented, OO) etc..These methods are to estimate impervious surface based on single optical remote sensing image data source mostly, due to the similitude between complexity and the construction material of city impervious surface, it is difficult to solve, at some research scene single image data source of net income that places an order, all obstacles that current impervious surface extracts, needs the extraction accuracy adding extra information to improve impervious surface.SAR image data as one of advanced earth observation technology, because of its there is round-the-clock, round-the-clock, the process time is short and the feature such as low cost and more and more applied.Different from optical image, roughness, shape, structure and the dielectricity of SAR data ground surface target are the most sensitive so that building can produce the strongest backscattering echo, can provide complementary information to optical image.

Summary of the invention

For problem above, the present invention proposes a kind of city impervious surface extracting method merging SAR image and optical remote sensing image, and is used for extracting city impervious surface distribution.Extracting method different from the past, the present invention utilizes SAR image, the respective advantage of optical remote sensing image data, using SAR image as the compensated information of optical image, by random forest (Random Forests, RF) the basic confidence level that in classifier calculated each evidence source, each pixel is of all categories, build BPA function, utilize D-S evidence theory composition rule to carry out the final extraction of city impervious surface.

The technical solution adopted in the present invention is a kind of city impervious surface extracting method merged with optical remote sensing image based on SAR image, comprises the following steps:

Step a, the sample formation sample that some study area is pre-selected always collects, and uses the method for stochastical sampling always to concentrate from sample and generates classifier training collection, grader test set and the precision test collection of impervious surface extraction result；

Step b, carries out the optical remote sensing image of study area and the registration of SAR image；

Step c, according to step b acquired results, carries out the feature extraction of SAR image and optical remote sensing image respectively, obtains the textural characteristics of SAR image and the spectral signature of optical remote sensing image；

Step d, the SAR image extracted for step c and optical remote sensing image feature, use step a gained classifier training collection, it is separately input in random forest grader set up disaggregated model, for step a gained grader test set, the random forest sorter model set up is tested, obtains the test error of grader；Utilize the random forest sorter model the set up optical remote sensing data and spectral signature, SAR image data and textural characteristics thereof respectively to study area to carry out city based on random forest grader impervious surface and tentatively extract, obtain optical remote sensing image data source, the impervious surface in SAR image data source tentatively extracts result；

Step e, uses what random forest grader carried out impervious surface to optical remote sensing image, SAR image tentatively to extract result according to step d, utilizes D-S evidence theory composition rule to carry out decision level fusion, obtain the final impervious surface in study area and extract result；

Step f, result is tentatively extracted for step d gained study area optically-based remote sensing image RF, SAR image RF tentatively extracts result and step e gained study area impervious surface finally extracts result, use step a gained precision test collection, use confusion matrix method respectively said extracted result to be carried out precision test.

And, in step a, select sample with reference to Google earth and high spatial resolution image.

And, in step e, it is achieved mode is as follows,

According to step d acquired results, calculate each pixel to be divided in each evidence source to basic confidence level of all categories, build BPA function；

According to Dempster composition rule, optical remote sensing image, SAR imaging results are merged, obtain each pixel and be assigned to all kinds of always trust interval；

By pixel maximum probability decision rule, each pixel to be divided is sorted out, obtain impervious surface D-S final decision result.

The technical scheme that the present invention provides has the beneficial effect that the advantage making full use of optical image with two kinds of data sources of SAR image, proposes SAR image based on RF and D-S evidence theory and optical remote sensing image fusion method, obtains the impervious surface of higher precision in city.

Accompanying drawing explanation

Fig. 1 is the flow chart of the embodiment of the present invention.

Fig. 2 is random forests algorithm flow chart.

Detailed description of the invention

When being embodied as, the provided flow process of technical solution of the present invention can be used computer software technology to realize automatically by those skilled in the art and run.In order to be more fully understood that technical scheme, the present invention is described in further detail with embodiment below in conjunction with the accompanying drawings.

Seeing Fig. 1, the embodiment of the present invention comprises the following steps:

nullStep a，With reference to Google Earth image、High spatial resolution remote sense image is (such as Quick Bird,IKONOS，WorldView series etc.)，According to study area land type、The features such as image resolution，Some sample formation sample is pre-selected always collect，The method using stochastical sampling always concentrates generation classifier training collection (setting up for random forest grader) from sample、(the test for random forest grader of grader test set，The quality of classification of assessment device is carried out with grader test error，Test error is the least，Represent the random forest classifier performance set up the best) and impervious surface extraction result precision test collection (extracting the precision test of result for impervious surface)，Sample includes optical remote sensing data and the feature thereof of correspondence respectively、SAR image data and feature thereof.When being embodied as, those skilled in the art can be according to sets itself sample sizes such as study area size, image resolutions.

The registration of step b, study area optical remote sensing image and SAR image；Optical remote sensing image is different from the imaging mechanism of SAR image, and owing to being affected by radar image geometric distortion, speckle noise, back scattering, shade etc., SAR image is the committed step that impervious surface extracts with the high registration accuracy of optical remote sensing image.

First study area optical remote sensing image data and the SAR image data of acquisition are pre-processed, generally include and optical remote sensing image is carried out the process such as atmospheric correction, ortho-rectification；SAR image carries out radiant correction, multiple look processing, topographical correction etc. and processes；After pretreatment, optical remote sensing image data and SAR image data registration are the important steps of optical remote sensing image and SAR visual fusion, when being embodied as, software can be processed as platform with existing remote sensing image, such as ENVI 5.X, utilize autoregistration instrument that optical image and SAR image are carried out autoregistration, require less than 0.5 pixel as image subject to registration, registration accuracy using optical remote sensing data as reference images, SAR data during registration.

Step c, for optical remote sensing image in step b and SAR image autoregistration result, carries out the extraction of feature respectively to optical remote sensing image and SAR image；

For specifically including that the feature extraction of SAR image, different according to image resolution, 3x3 can be chosen, 5x5,7x7, or 9x9 window etc., gray level co-occurrence matrixes (Gray-level Co-occurrence Matrix, GLCM) is utilized to extract the textural characteristics of SAR image；Optical remote sensing image extracts its spectral signature, such as normalized differential vegetation index (Normal Differential Vegetation Index, NDVI), normalization water body index (Normalized Difference Water Index, NDWI) etc., depending on the extraction of spectral signature can be according to factors such as the optical remote sensing data source selected and study area land cover pattern classifications thereof.

Utilizing algorithm of co-matrix to carry out the extraction of texture information SAR image, gray level co-occurrence matrixes is that a statistics describes two pixel gray scales in a regional area in image or whole region adjacent pixel or a determining deviation and presents the matrix of certain relation.Element value in this matrix represents combination condition probability density P between gray level (i, j | d, θ), P (i, j | d, θ) represent when given space length d and direction θ,, with i as initial point, there is the probability (namely frequency) that gray level is j in gray scale.Conventional index includes:

(1) average

$Mean=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×i---\left(1\right)$

The regular degree of average reflection texture, texture is disorderly and unsystematic, inenarrable, is worth less；Regularity is strong, be prone to description, is worth bigger.

(2) variance

$Variance=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×{(i-Mean)}^2---\left(2\right)$

(3) standard deviation

$Std=\sqrt{\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×{(i-Mean)}^2}---\left(3\right)$

Variance, standard deviation reflection pixel value and the tolerance of mean bias, when in image, grey scale change is bigger, variance, standard deviation are bigger.

(4) homogeneity degree

$Homogeneity=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×\frac{1}{1+{(i-j)}^2}---\left(4\right)$

Also being inverse difference moment (unfavourable balance away from), be the tolerance of image local gradation uniformity, if the uniform gray level of image local, homogeneity degree value is bigger.

(5) contrast

$Contrast=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×{(i-j)}^2---\left(5\right)$

Local gray level change total amount in reflection image.In the picture, the gray scale difference such as local pixel pair is the biggest, then the contrast of image is the biggest, and the visual effect of image is the most clear.

(6) non-similarity

$Dissmilarity=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×\left|i-j\right|---\left(6\right)$

Measure similar with contrast, but be linearly increasing.If the contrast of local is the highest, the most non-similarity is the highest.

(7) entropy

$Entropy=-\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})\×\ln p(i,j)---\left(7\right)$

It is the tolerance of the information content that image is had, is the characteristic parameter measuring grey level distribution randomness, characterizes the complexity of texture in image.The texture of image is the most complicated, and entropy is the biggest；Otherwise, the gray scale in image is the most uniform, then entropy is the least.

(8) angle second moment

$ASM=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}P{(i,j|d,\mathrm{\θ})}^2---\left(8\right)$

Angle second moment is the tolerance of gradation of image distributing homogeneity.When Elemental redistribution relatively concentrates near leading diagonal in gray level co-occurrence matrixes, gradation of image distribution uniform in regional area being described, ASM value is the biggest；On the contrary, if all values of co-occurrence matrix is the most equal, then ASM (Angular Second Moment) value is less.

(9) correlation

$Correlation=\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}\frac{(i-Mean)\×(j-Mean)\×P{(i,j|d,\mathrm{\θ})}^2}{Variance}---\left(9\right)$

Describing in GLCM similarity degree between row or column element, it reflects certain gray value development length along certain direction, if extend is the longest, then correlation is the biggest, and it is the tolerance of grey level sexual intercourse.

Optical remote sensing image is carried out the extraction of spectral signature, such as the extraction of the indexes such as NDVI (Normalized Difference Vegetation Index), NDWI (Normalized Difference Water Index).

$NDVI=\frac{NIR-R}{NIR+R}---\left(10\right)$

NDVI is utilized to can detect that the information such as vegetation growth state, vegetation coverage.

$NDWI=\frac{G-NIR}{G+NIR}---\left(11\right)$

NDWI is normalized ratio index based on green wave band Yu near infrared band.This NDWI is generally used to extract the Water-Body Information in image, and effect is more preferable than NDVI.

Wherein, G represents reflectance value at green wave band, R and NIR is respectively the reflectance value at red wave band and near infrared band.

Step d, the SAR image extracted for step c and optical remote sensing image characteristic information, use the classifier training collection extracted in step a, it is separately input in random forest grader set up disaggregated model, for the grader test set extracted in step a, the random forest sorter model set up is tested, obtains the test error (R_error) of grader.Utilize the random forest sorter model the set up optical remote sensing data and spectral signature, SAR image data and textural characteristics thereof respectively to study area to carry out city based on random forest grader impervious surface and tentatively extract, obtain optically-based remote sensing image data, the RF impervious surface of SAR image data tentatively extracts result.

The embodiment of the present invention, on the basis of step a, b, c, carries out Spectra feature extraction in step c to optical remote sensing image, and SAR image carries out texture feature extraction.After extracting two kinds of respective characteristic informations in remotely-sensed data source, the training set collection extracted in step a is utilized to set up random forest grader and utilize the test set in step a that it is verified, optical remote sensing image and spectral signature thereof are input in the random forest grader set up as optical remote sensing data source, obtain optical remote sensing data impervious surface based on RF grader and tentatively extract result (RF1 in corresponding diagram 1), calculate each pixel to be divided basic confidence level to each Land cover types in optical data sources, build BPA function (BPA1 in corresponding diagram 1).In like manner, the training set extracted in step a is utilized to set up random forest grader and utilize the test set in step a that it is verified, SAR image and textural characteristics thereof are input to as SAR remotely-sensed data source the random forest grader trained, obtain SAR remotely-sensed data impervious surface based on RF grader and tentatively extract result (RF2 in corresponding diagram 1), calculate each pixel to be divided basic confidence level to each Land cover types in SAR data source, build BPA function (BPA2 in corresponding diagram 1).

Random forest (Random Forests, RF) algorithm is proposed by Leo Breiman and Adele Cutler, is tree-structure network { h (x, a β_k), k=1 ... } set.Wherein meta classifier h (x, β_k) it is the post-class processing not having beta pruning built with two points of recursive subdivision (Classification And Regression Tree is called for short CART) algorithm；X is input vector；β_kIt is independent identically distributed random vector, determines the growth course of single tree；The output of forest uses the simple average (for returning) of simple majority ballot method (for classification) or single tree output result to obtain.The basic thought of RFC: first, utilize bootstrap sampling from original training set (N number of sample, N > k) k sample of extraction, and the sample size of each sample is as original training set；Secondly, k sample is set up k decision-tree model respectively, obtains k kind classification results；Finally, it is finally classified according to k kind classification results, each record to be voted decision.Fig. 2 is random forests algorithm flow chart.It is widely used for remote sensing image research: a. shows well on data set, the introducing of two randomnesss so that random forest is not easy to be absorbed in over-fitting owing to random forest has the advantage that；B., on current a lot of data sets, other algorithms relatively have the biggest advantage, the introducing of two randomnesss so that random forest has good noise resisting ability；C. it can process the data of the most high-dimensional (feature is a lot), and it goes without doing feature selecting, adaptable to data set, can process discrete data, also can process continuous data, and data set is without standardization.

Step e, PRELIMINARY RESULTS impervious surface extracted for step d, propose SAR image based on RF and D-S evidence theory and optical remote sensing image fusion method and city impervious surface is carried out final decision extraction and exports.The present invention proposes a kind of SAR image based on RF and D-S evidence theory and extracts city impervious surface with optical remote sensing image fusion method, use random forest grader that optical remote sensing image, SAR image are carried out the preliminary of impervious surface and extracted based on step d, utilize D-S evidence theory composition rule to carry out decision level fusion, obtain the final impervious surface in study area and extract result.

Owing to two kinds of images are from different sensors, entirely different on imaging mechanism, the most separate.Therefore, the RF of optical remote sensing image and SAR image is tentatively extracted result as two kinds of independent evidence sources, propose a kind of SAR image based on RF and D-S evidence theory and optical remote sensing image fusion method, carry out the extraction of city impervious surface.

Described SAR image based on RF and D-S evidence theory and optical remote sensing image fusion method, be optical remote sensing image, SAR image utilize RF grader obtain the probability that each pixel belongs to each ground class after, by optical remote sensing image, the RF of SAR image tentatively extracts result as independent evidence source, utilize D-S evidence theory method according to Dempster composition rule, obtain the final result extracted with optical remote sensing image impervious surface based on the SAR image that RF and D-S evidence theory merge.

First D-S evidence theory is proposed in 1967 by Dempster, a kind of inexact reasoning grown up further in 1976 by Shafer is theoretical, also referred to as Dempster/Shafer evidence theory (abbreviation D-S evidence theory), can preferably process uncertain information, be mainly characterized by: meet the condition more weak than Bayesian probability opinion；There is " uncertain " and the ability of " unknown " directly expressed.D-S evidence theory principle is as follows:

If Θ is an identification framework, or claim to assume space.

(1) basic probability assignment (Basic Probability Assignment is called for short BPA)

BPA on identification framework Θ is one 2^ΘThe function m of → [0,1], referred to as mass function.And meet

And

$\underset{A\⊆\mathrm{\Θ}}{\mathrm{\Σ}}m\left(A\right)=1$

Wherein so that the A of m (A) > 0 is referred to as burnt unit (Focal elements), m (A) is referred to as the basic probability function of A,For empty set.

(2) Dempster composition rule (Dempster ' s combinational rule) also referred to as combining evidences formula, it is defined as follows:

ForTwo mass function m on identification framework Θ₁,m₂Dempster composition rule be:

$m_1\⊕m_2\left(A\right)=\frac{1}{K}\underset{BIC=A}{\mathrm{\Σ}}{m}_1\left(B\right)\·m_2\left(C\right)$

Wherein, A, B, C are burnt unit.

Optical remote sensing image and spectral signature thereof are the most separate with SAR image and textural characteristics thereof, using optical remote sensing image and spectral signature, SAR image and textural characteristics thereof as two independent evidence sources.RF from different sensors different characteristic being identified, information carries out D-S theory fusion, and algorithm steps is as follows:

1. RF grader identification: extract in optical remote sensing image spectral signature, SAR image texture feature base in step c, use the training sample obtained in step a, to optical remote sensing image and spectral signature thereof in step d, SAR image and textural characteristics thereof carry out the preliminary extraction of impervious surface.

2. BPA construction of function: on the basis of step d, calculates each pixel to be divided in each evidence source to basic confidence level of all categories, builds BPA function, see above (1) basic probability assignment；

3. Decision fusion: according to Dempster composition rule, extracts result by the RF of optical remote sensing image, SAR image and merges, and obtains each pixel and is assigned to all kinds of always trust interval, sees above (2) Dempster composition rule；

4. by pixel maximum probability decision rule (i.e. differentiating that pixel belongs to the maximum probability of a certain classification, using the category as this pixel generic), each pixel to be divided is sorted out, obtain impervious surface based on RF and D-S and finally extract result.

Step f, for step d, the step e estimation result to impervious surface, using gained precision test collection in step a, the impervious surface RF to remotely-sensed data optically-based in step d tentatively extracts result respectively, impervious surface RF based on SAR image data tentatively extracts impervious surface based on RF and D-S in result and step e and finally extracts result and carry out precision test.

Conventional Accuracy Assessment is error matrix or confusion matrix (Error Matrix) method, error matrix is a n × n matrix (n is classification number), various precision statistics value can be calculated from error matrix, as producer's precision (Producer ' s Accuracy, PA), user's precision (User ' s Accuracy, UA), overall accuracy (Overall Accuracy, OA), Kappa coefficient etc., be used for simply comparing reference point and classification point.

The embodiment of the present invention, for the extraction result of impervious surface in step d, step e, uses the checking sample obtained in step a, utilizes confusion matrix respectively the extraction result of impervious surface in step d, step e to be carried out precision test.

Kappa coefficient is the multivariate statistical method of classification of assessment precision, and the estimation to Kappa is referred to as KHAT statistics, and Kappa coefficient represents and being evaluated the ratio reduced than completely random classification generation mistake of classifying, and computing formula is as follows:

$^K=\frac{N.\underset{i}{\overset{r}{\mathrm{\Σ}}}{x}_{ii}-\mathrm{\Σ}(x_{i+}.x_{+i})}{N^2-\mathrm{\Σ}(x_{i+}.x_{+i})}---\left(14\right)$

In formula: ^K is Kappa coefficient, r is the line number of error matrix, x_iiIt is the value on i row i row (leading diagonal), x_i+And x_+iBe respectively the i-th row and with i-th row and, N is total sample.

Table 0 kappa statistical value and nicety of grading corresponding relation

In sum, the combination SAR image of present invention proposition and optical remote sensing image feature, the method that SAR image based on RF and D-S evidence theory and optical remote sensing image fusion method carry out city impervious surface extraction, at aspects such as the degree of accuracy and stability that impervious surface extracts, Billy is with data mapping more effectively, owing to optical remote sensing image, SAR image data source are easy to obtain, operating procedure is simple, and the method more conforms to the application demand of reality.

Practicality and effect for the explanation present invention, it is provided that experiment below:

(1) experimental data: optical remote sensing data decimation 2 scape GF-1 (high score one) multispectral data on April 14th, 2015, data derive from geographical spatial data cloud website (http://www.gscloud.cn/).Totally four wave bands, wherein 3 visible light wave ranges, 1 near infrared band, spatial resolution is 16 meters.ENVI 5.2 carries out atmospheric correction, ortho-rectification etc. for platform and processes.

nullSAR image data source is chosen 2 scapes on February 17th, 2015 and is interfered wide cut (IW，Interferometric Wide swath)) the sentinel-1A GRD radar image of pattern，Image is downloaded from Sentinels Scientific Data Hub website (https: //scihub.esa.int/dhus/)，Image incidence angle is 29～46 °，Resolution ratio 5m × 20m，Fabric width 250km，Polarization mode HH+HV (Horizontal Vertical+Vertical Vertical)，Sentinel-1A image is carried out radiant correction、Multiple look processing、Topographical corrections etc. process.

(2) feature extraction: optical remote sensing image, SAR image are carried out feature extraction.

(3) RF grader tentatively identifies: according to the GF-1 image spectral signature extracted, Sentinel 1-A image texture feature, use RF algorithm in step d, utilize the classifier training collection that optical remote sensing image in step a, SAR image are the most corresponding, set up random forest grader；According to optical remote sensing image in step a, grader test set that SAR image is corresponding, the RF grader being set up GF-1 image, Sentinel 1-A image is tested, RF grader test error (R_error) obtaining GF-1 optical remote sensing image and Sentinel 1-ASAR image is respectively 0.007,0.548.GF-1 image and Sentinel 1-A image are added separately in respective RF grader carry out the preliminary extraction of city impervious surface.

(4) result is tentatively extracted based on RF grader to GF-1 optical remote sensing image, Sentinel 1-A SAR image in (3), obtain each pixel to probability of all categories, structure BPA function, using optical remote sensing image and SAR image as independent evidence source, utilize D-S evidence theory composition rule, according to pixel maximum probability decision rule, carry out final decision fusion, obtain impervious surface fusion results based on SAR image and optical remote sensing image.

(5) the precision test collection obtained in step a is utilized to carry out extracting the precision test of result

Table 1 producer's accuracy table

Table 2 user's accuracy table

Table 3 is classified overall accuracy and Kappa coefficient

nullFor study area and imaging characteristic，Study area is divided into following a few kind and carries out the extraction of impervious surface: high reflectance impervious surface (High reflectance of impervious surface,IS_H)，Antiradar reflectivity impervious surface (Low reflectance of impervious surface,IS_L)，Water body (Water body,WB)，Vegetation (Vegetation,VE)，Highlighted bare area (High reflectance of bare land,BL_H)，Dark-coloured bare area (Low reflectance of Bare land,BL_L).Table 1, table 2 represents the RF utilizing the RF of GF-1 remote sensing image and spectral signature (optical remote sensing image data source) thereof tentatively to extract result, sentinel 1-A image and textural characteristics thereof respectively and tentatively extracts the extractions result of result (SAR image data source) and optical remote sensing image based on RF-DS and SAR visual fusion to the production precision (Producer's Accuracy) of various places class and user's precision (User's Accuracy), by table 1, table 2 can be seen that different classes of between there is the leakage of a stable condition and divide and mistake point phenomenon.After optical remote sensing image and SAR visual fusion, improve the leakage between dark-coloured impervious surface, water body, dark-coloured bare area divide, wrong point phenomenon.nullClassification results is merged into permeable and the big class of impervious surface two，It is carried out precision test，Table 3 represents three kinds of (optical remote sensing data sources、SAR data source and optical remote sensing data merge with the RF-DS of SAR image) impervious surface extracts overall accuracy and Kappa coefficient，It can be seen that，The overall accuracy carrying out impervious surface based on RF grader extraction individually with GF-1 (optical remote sensing image) is 91.15%，Kappa coefficient is 0.819，Individually with sentinel1-A image (SAR image) carry out impervious surface based on RF grader extract overall accuracy be poor be 72.73%，Kappa coefficient is 0.446，It is 95.33% that the impervious surface utilizing evidence theory (DS) to carry out SAR image and optical remote sensing image extracts overall accuracy，Kappa coefficient is 0.905，Therefore，It may be concluded that introduce SAR data higher with the city impervious surface extraction accuracy that the RF-DS of optical remote sensing image merges based on SAR image.

Claims (3)

1. the city impervious surface extracting method merged with optical remote sensing image based on SAR image, comprises the following steps:

Step a, the sample formation sample that some study area is pre-selected always collects, and uses the method for stochastical sampling always to concentrate from sample Generate classifier training collection, grader test set and city impervious surface and extract the precision test collection of result；

Step b, carries out the optical remote sensing image of study area and the registration of SAR image；

Step c, according to step b acquired results, carries out the feature extraction of SAR image and optical remote sensing image respectively, obtains SAR The textural characteristics of image and the spectral signature of optical remote sensing image；

Step d, the SAR image extracted for step c and optical remote sensing image feature, use step a gained classifier training Collection, is separately input in random forest grader set up disaggregated model, for step a gained grader test set, to set up Random forest sorter model is tested, and obtains the test error of grader；Utilize the random forest sorter model set up Optical remote sensing data and spectral signature, SAR image data and textural characteristics thereof to study area are carried out based on random forest respectively The city impervious surface of grader tentatively extracts, and obtains at the beginning of the impervious surface in optical remote sensing image data source, SAR image data source Step extracts result；

Step e, uses random forest grader that optical remote sensing image, SAR image are carried out the preliminary of impervious surface according to step d Extract result, utilize D-S evidence theory composition rule to carry out decision level fusion, obtain the final impervious surface in study area and extract knot Really；

Step f, tentatively extracts result for step d gained study area optically-based remote sensing image RF, and SAR image RF tentatively carries Take result and step e gained study area impervious surface extracts result, use step a gained precision test collection, use and obscure square Battle array method carries out precision test to said extracted result respectively.

The city impervious surface extracting method merged with optical remote sensing image based on SAR the most according to claim 1, it is characterised in that: In step a, visually select sample set with reference to Google earth and high spatial resolution image.

The city impervious surface extracting method merged with optical remote sensing image based on SAR the most according to claim 1 or claim 2, its feature It is: in step e, it is achieved mode is as follows,

According to step d acquired results, calculate each pixel to be divided in each evidence source to basic confidence level of all categories, build BPA Function；

According to Dempster composition rule, optical remote sensing image, SAR image impervious surface RF PRELIMINARY RESULTS are merged, obtains What each pixel was assigned to various places class always trusts interval；

By pixel maximum probability decision rule, each pixel to be divided is sorted out, obtain city impervious surface D-S final decision result.