CN106650812A - City water body extraction method for satellite remote sensing image - Google Patents

Abstract

The invention discloses a city water body extraction method for a satellite remote sensing image. The method comprises the steps of preprocessing remote sensing image data; obtaining two new normalized different water body indexes NNDWI1 and NNDWI2, and obtaining water body extraction results of NNDWI1 and NNDWI2; superposing a threshold segmentation result of NNDWI1 and a threshold segmentation result of NNDWI2 to obtain a threshold segmentation result of NNDWI, namely, a large-area water body object and a small-area object, performing constraint on the expanded small-area object, and performing shadow detection and removal on the constrained small-area object to obtain a small-area water body object; and superposing the large-area water body object and the small-area water body object to obtain a city water body extraction result of the satellite remote sensing image. The method provided by the invention has relatively high classification precision and relatively low missing detection and false alarm total error rate, improves subsequent water body extraction precision, and has relatively high water body edge detection precision.

Description

A kind of urban water-body extracting method of satellite remote-sensing image

Technical field

The invention belongs to water body remote sensing image technical field, relates generally to a kind of urban water-body extraction of satellite remote-sensing image Method.

Background technology

City is the embodiment of human society high development, urban water-body as factor important in urban ecological system, Maintain to be had a very important role in urban ecological system stability.Urban water-body change will opposite life birth life it is huge Change, may cause some disasters, such as urban water resource delay, shortage of water resources, or even cause some and human health to live Related disease.Therefore, understand and grasp urban water-body distribution and the change of water surface area has become people and increasingly closes The focus of note.

In recent years, with the application and development of remote sensing technology, remote sensing image is in survey of natural resources, dynamic monitoring, nature The aspects such as earth surface water source planning play more and more important role, and the monitoring for carrying out earth's surface using remote sensing technology is also obtained The concern of increasing researcher；Remote sensing image can go to observe the atural object of earth surface with a different visual angle, The change of monitoring earth's surface in real time.In Clean water withdraw technology, urban water-body information is timely and accurately obtained by remotely-sensed data Become the Clean water withdraw mode of current main flow.Up to the present, many scholars propose the Clean water withdraw side of a large amount of remote sensing images Method, but major part is all based on middle low resolution remote sensing image, because image resolution is relatively low, the less water body of area fails to have The extraction of effect；Especially for the extraction of urban area water body, urban area coverage of water size is uneven, exists many little Area artificial lake and tiny river.Therefore, high-resolution remote sensing image should be more adopted in urban water-body extraction, with As a example by No. 3 satellites of resource, No. 3 satellite remote-sensing images of resource have the wide cut of 5.8m resolution ratio and 5.2km, and it is carried for urban water-body Take there is provided preferable multispectral image data, No. 3 design parameter such as table 1 below of resource：

Table No. three satellite image parameters of 1 resource

Water body is atural object classification very common in remote sensing image, is also one of important Fundamental Geographic Information System, its dynamic letter The quick obtaining of breath, to causes such as Investigation of water resources, planning for water resources development, environmental monitoring and protections fairly obvious practical valency is suffered from Value and scientific meaning.In this regard, many scholars expand very early research to this, it is proposed that many effective Water-Body Informations are carried automatically Delivery type.4 classes can be substantially divided into：(a) single band and multiband thresholding method (single-band or multiple- Band threshold method), (b) water body index method (Water indices), (c) linear unmixed model (linear Un-mixing model), (d) supervision and non-supervised classification (supervised or unsupervised classification method).In addition, also some other method, such as：Water body based on digital elevation model Extractive technique, based on the Clean water withdraw technology of microwave remote sensing (Microwave Remote Sensing) image, object-oriented Clean water withdraw technology of (Object Oriented) technology etc..But these methods are simultaneously of little use, must for, water body index method Because its model is simple, convenience, and precision is higher, the most commonly used in practice.

However as the raising of remote sensing image resolution ratio, most of high-resolution remote sensing image is (such as；WorldView-2, IKONOS, RapidEye, and resource 3) do not have as the so how utilizable wave bands of Landsat TM/ETM+/OLI are used for The extraction of water body, therefore, MNDWI (improved normalization difference water body index method) and AWEI (automatic water body index) will be unable to make With because most high-resolution remote sensing image only has 4 wave bands (blue, green, red, near infrared band), lacking MNDWI/AWEI calculates required short infrared band (SWIR).Therefore, using NDWI (normalization difference water body index) to height Some problems, the especially problem that such as shade cannot be removed, city will be just produced when resolution image carries out Clean water withdraw The shade of city area high-rise, shows especially prominent on high resolution image.On high-resolution remote sensing image, city The high-rise shade in region is difficult to distinguish with water body, although there is related scholar to study this aspect at present, such as Based on Object-oriented technology, high-rise shade is carried out by calculating the texture features of high-rise shadow region Detection；Although can produce a desired effect, due to texture description with calculate it is relatively complex and time consuming longer, so from calculating Consider that the method is not a preferable shadow detection method on time.Also the shade carried out based on SVM features trainings examines Survey, to reaching the impact that water body detects middle-high building thing shade is removed.SVM is a kind of higher method of nicety of grading, but SVM training needs have spent the more time, especially when training sample number is more and sampling feature vectors dimension is higher.If By using high-resolution remote sensing image shadow detection method (morphological shadow index, MSI) and NDWI phases With reference to mode Clean water withdraw is carried out to WorldView-2 high resolution images, to improve water body accuracy of detection；Although the party Method principle is simple, but the basis due to the method in NDWI methods as Clean water withdraw, and its accuracy of detection can't be very high, especially It is the minute area waters of dense vegetation around, the spectral characteristic of water body will be polluted by serious, the performance of water spectral feature It is extremely unstable, while city water body has, Remote Sensing of Suspended Sediment Concentration is high, body eutrophication is serious, by various contaminants compared with Big the features such as so that the city water body optical characteristics different from unpolluted water body performance in nature.

The content of the invention

Clean water withdraw problem for how from high resolution ratio satellite remote-sensing image, carrying out urban area, particularly has The precision problem for distinguishing high-rise shade and water body and raising Clean water withdraw of effect, the present invention proposes a kind of satellite remote sensing The urban water-body extracting method (Automatic urban water extraction method, AUWEM) of image.

The concrete technical scheme of the present invention is as follows：

A kind of urban water-body extracting method of satellite remote-sensing image, comprises the following steps：

Step 1：The pretreatment of remote sensing image data, i.e., carry out ortho-rectification and atmospheric correction to remote sensing image data；

Step 2：Pretreated remote sensing image data, including blue band1 data, green band2 data, redness Band3 data and near-infrared band4 data, choose the blue band1 data in pretreated remote sensing image data and replace returning Green band2 data in one computing formula for changing difference water body index NDWI, obtain new normalization difference water body index NNDWI1, the computing formula of new normalization difference water body index NNDWI1 is：

This computing formula is NNDWI1 exponential models, and the threshold value of NNDWI1 is obtained by Threshold segmentation using this model Segmentation result, also as NNDWI1 Clean water withdraws result；

Step 3：It is four wave band datas that pretreated remote sensing image data is included, i.e. blueness band1 data, green Color band2 data, redness band3 data and near-infrared band4 data carry out PCA conversion, and after PCA is converted first it is main into Divide component Component1 to substitute the green band2 data in the computing formula of normalization difference water body index NDWI, obtain another One new normalization difference water body index NNDWI2, i.e.,：

Wherein, Component1 represents the first principal component component that PCA is converted, and this computing formula is NNDWI2 index moulds Type, using this model the Threshold segmentation result of NNDWI2 is obtained by Threshold segmentation, also the Clean water withdraw knot of as NNDWI2 Really；

Step 4：The threshold value of the NNDWI2 obtained in the Threshold segmentation result of the NNDWI1 that step 2 is obtained and step 3 point Cut result to be overlapped, the result for obtaining is defined as into the Threshold segmentation result of new normalization difference water body index NNDWI, i.e., The Threshold segmentation result of NNDWI1 is overlapped with the Threshold segmentation result of NNDWI2, and its computing formula is：

NNDWI=(segmentation_NNDWI1) ∪ (segmentation_NNDWI2)

Segmentation_NNDWI1 represents the Threshold segmentation result of NNDWI1, segmentation_NNDWI2 tables in formula Show the Threshold segmentation result of NNDWI2, this computing formula is NNDWI exponential models, obtains NNDWI's using this exponential model Clean water withdraw result；

Step 5：Row threshold division is entered to the near-infrared band4 data in pretreated remote sensing image data, obtains near The Threshold segmentation result of infrared band4 data；

Step 6：Water body in large object and small area object in the Clean water withdraw result of NNDWI is split, In the Clean water withdraw result of NNDWI, number of pixels is water body in large object more than given threshold, and number of pixels is less than or equal to Given threshold for small area object；

Step 7：Small area object to obtaining in step 6 carries out mathematical morphology expansion process, little after being expanded Object oriented, the Threshold segmentation result of the near-infrared band4 data that step 5 is obtained is adopted after expansion as constraints Small area object and the mode that seeks common ground of Threshold segmentation result of near-infrared band4 data the small area object after expansion is entered Row constraint, the mathematic(al) representation of constraint is：

Component2=(dilate_component) ∩ (segmentation_band4)

In formula, dilate_component represents the small area object after expansion, and segmentation_band4 represents near The Threshold segmentation result of infrared band4 data, component2 represents the small area object after constraint；

Step 8：Small area object after the constraint obtained to step 7 carries out shadow Detection and removes, and obtains facet ponding Body object；

Shadow Detection and removal, referring to carries out the description of wave spectrum relation to each pixel in each small area object, and Judge whether the pixel meets the condition of shade pixel, record and count the number of shade pixel in each small area object, when When shade pixel proportion is more than threshold value T in one small area object, the small area object is judged to building effects pair As, then it is judged to small area water body object less than or equal to small area object during threshold value T, shade pixel proportion is small area The number of shade pixel and the ratio of total pixel number in the small area object, distinguish small area object small areas water in object The function expression of body object and shadow object is：

In formula, n represents total pixel number in a certain small area object, and m is the number of shade pixel in the small area object；

The condition of shade pixel, refers to the wave spectrum magnitude relationship for meeting shade pixel, that is, meet three below inequality bar Part：

Step 9：The small area water body object obtained in the water body in large object obtained in step 6 and step 8 is carried out Superposition, the water body in large object that will be obtained in step 6 and the small area water body object obtained in step 8 seek union, obtain The urban water-body of satellite remote-sensing image extracts result.

Beneficial effects of the present invention are as follows：

(1) new normalization difference water body index NNDWI1 and NNDWI2 effectively can carry out initial extraction to water body, Improve the extraction accuracy of follow-up water body.

(2) in terms of shadow object acquisition, in order to more accurately obtain shadow object, small area object is expanded Process；Simultaneously in order to be limited in real surface shadow region, the object to expanding adopts the Threshold segmentation of near-infrared band4 data As a result row constraint is entered.

(3) to avoid shadow object detection process in time loss, to small area object using being retouched based on spectral characteristic State, reduce the time loss being described using textural characteristics in tradition, improve the computational efficiency of method.

(4) nicety of grading of AUWEM methods is higher than the nicety of grading of NDWI and the nicety of grading of maximum likelihood method.5 In the experimental result of individual test block, the average Kappa coefficients of AUWEM are about the average Kappa coefficients of 93.0%, NDWI and are about 86.2%, maximum likelihood method nicety of grading falls between, and average Kappa coefficients are about 88.6%；Simultaneously AUWEM missing inspections with False-alarm total false rate will also be less than the classification results of NDWI and the classification results of maximum likelihood method, the average missing inspections of AUWEM and false-alarm Total false rate is about 11.9%, the average missing inspection of maximum likelihood method and false-alarm total false rate be about the average missing inspections of 18.2%, NDWI with False-alarm total false rate is about 22.1%%.In addition, AUWEM has higher coastal waters accuracy of detection.

Description of the drawings

Fig. 1 is the Clean water withdraw result of the NNDWI in the specific embodiment of the invention, wherein figure (a) be from resource No. 3 defend The remote sensing image that star is obtained, figure (b) is NNDWI1 Clean water withdraw results, and figure (c) is NNDWI2 Clean water withdraw results, is schemed (d) NNDWI Clean water withdraw results；

Fig. 2 is the expansion constraint process schematic in the specific embodiment of the invention；

Fig. 3 (a)～(e) is the shadow region difference spectral characteristic curve in the specific embodiment of the invention；

Fig. 4 is the urban water-body extracting method flow chart of the satellite remote-sensing image in the specific embodiment of the invention；

Fig. 5 is experimental result of the distinct methods in the specific embodiment of the invention 5 test blocks；

Fig. 6 (a)～(f) is 6 index Nogatas of the distinct methods in the specific embodiment of the invention 5 test blocks Figure；

Fig. 7 is that region to be assessed obtains schematic diagram in the coastal waters accuracy evaluation in the specific embodiment of the invention；

Fig. 8 (a)～(c) is that the distinct methods in the specific embodiment of the invention are smart in 5 test block coastal waters detections Degree compares.

Specific embodiment

With reference to the accompanying drawings and detailed description the present invention is described in detail.

A kind of urban water-body extracting method of satellite remote-sensing image, with No. 3 satellite remote-sensing image data instances of resource, including Following steps：

Step 1：The pretreatment of remote sensing image data, i.e., carry out ortho-rectification and atmospheric correction to remote sensing image data；It is right The image of Experimental Area is carried out without control point ortho-rectification using RPC+30DEM, using the FLAASH of Feyisa G L et al. (Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes) atmospheric correction models The atmospheric correction of image is carried out, is completed in ENVI5.2 softwares.Wherein, the calibration that No. 3 FLAASH atmospheric corrections of resource need Coefficient can pass through website http://www.cresda.com/CN/Downloads/dbcs/index.shtml is downloaded, light Spectrum receptance function can pass through website http://www.cresda.com/CN/Downloads/gpxyhs/index.shtml is carried out Download.

Step 2：Pretreated remote sensing image data, including blue band1 data, green band2 data, redness Band3 data and near-infrared band4 data, choose the blue band1 data in pretreated remote sensing image data and replace returning Green in one computing formula for changing difference water body index NDWI (Normalized Difference Water Index) Band2 data, obtain new normalization difference water body index NNDWI1 (New Normalized Difference Water Index 1), the computing formula of new normalization difference water body index NNDWI1 is：

This expression formula is NNDWI1 exponential models.Pass through threshold value in programming software Matlab 2014a using this model Segmentation obtains NNDWI1 Threshold segmentation results, also as NNDWI1 Clean water withdraws result.

Step 3：It is four wave band datas that pretreated remote sensing image data is included, i.e. blueness band1 data, green Color band2 data, redness band3 data and near-infrared band4 data carry out PCA conversion, and after PCA is converted first it is main into Divide component Component1 to substitute the green band2 data in the computing formula of normalization difference water body index NDWI, obtain another One new normalization difference water body index NNDWI2 (New Normalized Difference Water Index 2), i.e.,：

Wherein, Component1 represents the first principal component component that PCA is converted.

This expression formula is NNDWI2 exponential models.Pass through threshold value in programming software Matlab 2014a using this model Segmentation obtains NNDWI2 Threshold segmentation results, also as NNDWI2 Clean water withdraws result.

Step 4：The threshold value of the NNDWI2 obtained in the Threshold segmentation result of the NNDWI1 that step 2 is obtained and step 3 point Cut result to be overlapped, the result for obtaining is defined as into new normalization difference water body index NNDWI (New Normalized Difference Water Index) Threshold segmentation result, i.e. the Threshold segmentation result of NNDWI1 and the threshold value of NNDWI2 point Cut result to be overlapped, its computing formula is：

NNDWI=(segmentation_NNDWI1) ∪ (segmentation_NNDWI2)

Segmentation_NNDWI1 represents the Threshold segmentation result of NNDWI1 index images, segmentation_ in formula NNDWI2 represents the Threshold segmentation result of NNDWI2 index images.This computing formula is NNDWI exponential models, using this model NNDWI Clean water withdraw results are obtained in programming software Matlab 2014a.

By the Threshold segmentation of experimental analysis, the threshold value that the Threshold segmentation result of NNDWI1 is adopted in experiment and NNDWI2 As a result the threshold value for being adopted all is set to 0, can obtain ideal experimental result.The Clean water withdraw result of NNDWI is combined The Clean water withdraw result of NNDWI1 and the Clean water withdraw result of NNDWI2, it is to avoid single index is sent out the detection leakage phenomenon of water body It is raw.As shown in Fig. 1 (a)～(d).NNDWI1 is more obvious to the high water body Detection results containing silt, and NNDWI2 is to by vegetation The water body of spectral information serious interference is more sensitive, therefore, in actual Clean water withdraw with reference to NNDWI1 Clean water withdraw result and The Clean water withdraw result of NNDWI2, integrates and generates new Clean water withdraw result, improves follow-up Clean water withdraw precision.

Step 5：Row threshold division is entered to the near-infrared band4 data in pretreated remote sensing image data, obtains near The Threshold segmentation result of infrared band4 data.

Found by analyzing the image data of a large amount of NNDWI Clean water withdraws results, except Some City area it is less artificial Outside pond and lake, in general, the area of urban skyscraper thing shadow object is typically less than water body object Area.So in actual analysis, only the atural object relatively small to area is detected, because in this section area is relatively Little object includes almost all of shadow object, while also including small area water body object.

Step 6：Water body in large object and small area object in the Clean water withdraw result of NNDWI is split, In the Clean water withdraw result of NNDWI, number of pixels is water body in large object more than given threshold, and number of pixels is less than or equal to Given threshold for small area object；

Small area object acquisition model, is represented by：

Wherein, t is the threshold value of setting, and its value takes the number of pixels of maximum shadow object, and component represents NNDWI water Discrete objects in body exponent extracting result, it is discrete right in result that area (component) represents that NNDWI water body indexes are extracted The area of elephant, area (component)>T be water body in large object, area (component)≤t be small area water body or Shadow object.

Step 7：Small area object to obtaining in step 6 carries out mathematical morphology expansion process, little after being expanded Object oriented, the Threshold segmentation result of the near-infrared band4 data that step 5 is obtained is adopted after expansion as constraints Small area object and the mode that seeks common ground of Threshold segmentation result of near-infrared band4 data the small area object after expansion is entered Row constraint, the mathematic(al) representation of constraint is：

Component2=(dilate_component) ∩ (segmentation_band4)

In formula, dilate_component represents the small area object after expansion, and segmentation_band4 represents near The Threshold segmentation result of infrared band4 data, component2 represents the small area object after constraint；

The more complete shadows pixels for including small area water body object and shade of small area object after the expansion are right Small area object after expansion enters row constraint and also maintains the expansive working of small area object to be limited in earth's surface life shadow region In the range of.

By small area object that step 6 is obtained for the expansion constraint process schematic such as Fig. 2 as a example by building effects object It is shown, from No. 3 pseudo color coding hologram images of resource, building effects object and Band4 image data Threshold segmentation results are obtained respectively, The building effects object for obtaining is carried out into mathematical morphology expansion process, the expansion results of building effects object are obtained, will The expansion results of building effects object carry out the process that seeks common ground with Band4 image data Threshold segmentation results, that is, obtain The building effects object after expansion under the constraint of Band4 image data Threshold segmentations result.

Step 8：Small area object after the constraint obtained to step 7 carries out shadow Detection and removes, and obtains facet ponding Body object；

In NNDWI Clean water withdraw results, substantially only comprising small area water body object and shade, so only needing to little Area water body object carries out research and analysis with the feature of shade, finds the spy for being adapted to distinguish small area water body object and shade Levy.Find in an experiment, although textural characteristics can be very good to describe small area water body object and shade, due to the line of atural object Reason feature, such as：Gray level co-occurrence matrixes, its calculating is relative complex, time-consuming longer, is not particularly suitable for for small area water body object and the moon The differentiation of shadow, so being described the pixel of small area water body object and shade using the spectral signature of atural object in an experiment, is made with this It is according to differentiation small area water body object and shade.

By the spectral characteristic curve for analyzing substantial amounts of water body and shade, the wave spectrum relation for obtaining water body pixel meets Formula：

band2>band4

And the pixel spectral profile of shade is complex, analysis and summary has gone out following 5 kinds of spectral characteristic curves in experiment, such as Shown in Fig. 3 (a)～(e).

According to the corresponding magnitude relationship of each wave band of above-mentioned spectral profile, the wave spectrum relation of shade pixel is summed up in experiment to expire Sufficient three below inequality condition：

It follows that the condition of shade pixel, the wave spectrum relation for referring to shade pixel meets three above inequality condition.

Shadow Detection and removal, referring to carries out the description of wave spectrum relation to each pixel in each small area object, and Judge whether the pixel meets the condition of shade pixel, record and count the number of shade pixel in each small area object, when When shade pixel proportion is more than threshold value T in one small area object, the small area object is judged to building effects pair As, then it is judged to small area water body object less than or equal to small area object during threshold value T, shade pixel proportion is small area The number of shade pixel and the ratio of total pixel number in the small area object, distinguish small area object small areas water in object The function expression of body object and shadow object is：

In formula, n represents total pixel number in a certain small area object, and m is the number of shade pixel in the small area object. Threshold value T is got by experiment statisticses, statistics discovery is carried out to the shade pixel of No. 3 remote sensing image datas of resource, when T takes Can be very good to distinguish water body and shadow object when 0.5.

Step 9：The small area water body object obtained in the water body in large object obtained in step 6 and step 8 is carried out Superposition, the water body in large object that will be obtained in step 6 and the small area water body object obtained in step 8 seek union, obtain The urban water-body of satellite remote-sensing image extracts result.

A kind of overview flow chart of the urban water-body extracting method of satellite remote-sensing image is as shown in Figure 4.

For the validity of verification method, NDWI method Clean water withdraw results and maximum likelihood method (MaxLike) is respectively adopted Clean water withdraw result carries out contrast experiment.Different regions are being have chosen at 5 within Chinese territory and with the image of different surrounding enviroment For testing, they include lake and river, Beijing respectively positioned at CHINESE REGION, Wuhan City, Suzhou City, Guangzhou, its Middle Wuhan City has selected the image of two distinct coverage regions.No. 3 image details descriptions of resource are as shown in table 2, in experiment The facilities of detailed parameter as shown in table 3, wherein the numerical value of band4 first as the following formula naturalization to 0-255 spans, so Afterwards again selected threshold is split.

The experimental data details of table 2

The different experiments of table 3 area threshold value setting (wherein T1, T2, T3, the respectively segmentation of NNDWI1, NNDWI2, band4 Threshold value)

For ease of the visual interpretation and analysis of distinct methods classification results, grey is given to the water body pixel of correct classification, The non-aqueous body image unit of correct classification gives black, and the pixel of mistake classification gives white, and experimental result is as shown in Figure 5.From Fig. 5's Classification results visual interpretation is it is found that the Clean water withdraw nicety of grading of AUWEM proposed by the present invention is better than the water body of NDWI Extract the Clean water withdraw nicety of grading of nicety of grading and maximum likelihood method.AUWEM can be very good to coastal waters mixed pixel Classify (with reference to the water body classification results in Beijing, Wuhan_1 and Wuhan_2 area), performance is detected to tiny pond waters It is better than NDWI and maximum likelihood method (with reference to Suzhou areas water body classification results), the shade to building construction can be very good Remove (with reference to Suzhou and Wuhan_2 areas water body classification results).

Three kinds of different experiments region method Clean water withdraw nicety of grading compares statistics as shown in table 4, from the statistics of table 4 As a result middle discovery, AUWEM Clean water withdraw niceties of grading are higher than NDWI and maximum likelihood method.AUWEM is in this 5 Experimental Areas Nicety of grading is highest, and average Kappa coefficients are up to 93.0%；And the nicety of grading of NDWI is minimum, average Kappa coefficients are about 86.2%；Maximum likelihood method nicety of grading falls between, and average Kappa coefficients are about 88.6%.

The precision statisticses of the three kinds of methods in different experiments region of table 4

For extraction accuracies of the more detailed assessment AUWEM to water body, using producer's precision, user's precision, Kappa Coefficient, loss, false alarm rate, 6 indexs of total false rate are describing the Clean water withdraw precision of method.Three kinds of distinct methods are at 5 Shown in 6 index histogram such as Fig. 6 (a)～(f) in experiment area, from the histogram of each index it is found that the water of AUWEM Body extracts nicety of grading and is higher than the Clean water withdraw nicety of grading of NDWI and the Clean water withdraw nicety of grading of maximum likelihood method. In terms of the false alarm rate of Clean water withdraw, except reaching 9.1% or so in Suzhou areas, other Experimental Areas are all low to AUWEM In 5%；In terms of water body loss, 5 regional loss will be considerably less than NDWI and maximum likelihood method.When Clean water withdraw When false alarm rate and all low water body loss, total false rate is naturally namely minimum.It can be seen that AUWEM from Fig. 6 (a)～(f) Total false rate is minimum, next to that maximum likelihood method, total false rate highest is NDWI, and the average total false rates of wherein AUWEM are about 11.9%, the average total false rate of maximum likelihood method is about the average total false rates of 18.2%, NDWI and is about 22.1%%.

For Clean water withdraw classification producer's precision, AUWEM Clean water withdraws classification producer's precision highest is average to be about 91.6%；Maximum likelihood method Clean water withdraw classification producer's precision is taken second place, and averagely about 84.8%；The Clean water withdraw classification of NDWI Producer's precision is minimum, and averagely about 81.6%.In Clean water withdraw user's precision aspect, the Clean water withdraw user of maximum likelihood method AUWEM and NDWI that precision is greater than, mean accuracy user is up to 96.6%；AUWEM methods are taken second place, and mean accuracy user be about 96.4%；Worst still falls within NDWI methods, and mean accuracy user is about 95.7%.

For the rim detection precision of the water body that three kinds of methods of more objective evaluation are extracted, design following methods are carrying out essence Degree is evaluated, and method is embodied as being described as follows：

First with the boundary line for obtaining the water body that three kinds of methods are extracted using Canny operators with reference to image；

Boundary line to obtaining carries out the expansion process of mathematical morphology, and it is 4 pictures to set up the radius centered on boundary line The buffer area of unit；

Then the pixel of buffer area is judged, it is assumed that the total pixel value in buffer area is N, the pixel number of correct classification Purpose N_R, missing inspection pixel number is No, and false-alarm number is Nc, then：

Wherein, A+Eo+Ec=100%.A represents that edge pixel correctly divides the ratio of classification, and here it is called edge The correct nicety of grading of pixel；Eo represents the ratio of edge pixel missing inspection, and it is called edge pixel loss；Ec represents edge pixel The ratio of false-alarm, it is called edge pixel false alarm rate.By through the boundary line of the water body of mathematical morphology expansion process and from money The remote sensing image that the satellite of source 3 is obtained is overlapped, and obtains region to be assessed in edge definition assessment.Treat in edge definition assessment It is as shown in Figure 7 that assessment area obtains schematic diagram.

The coastal waters accuracy of detection of Experimental Area is counted according to said method, edge pixel false alarm rate has been counted respectively (Commission Error), edge pixel loss (Omission Error) and the correct nicety of grading of edge pixel (Accuracy of edge detection), and experimental result is compared, such as shown in Fig. 8 (a)～(c).From the ratio of Fig. 8 The discovery that relatively result can be apparent from：Pixel correct nicety of grading in AUWEM methods edge is better than NDWI and maximum likelihood method.Its The correct nicety of grading of middle AUWEM methods coastal waters pixel is up to 93.7691% (Guangzhou is regional), and minimum precision is 79.5798% (Wuhan_2 is regional)；The correct nicety of grading of NDWI coastal waters pixels is up to 84.0917% (Suzhou ground Area), minimum precision is 69.8310% (Beijing is regional)；The correct nicety of grading of maximum likelihood method coastal waters pixel is up to 85.8149% (Guangzhou is regional), minimum precision is 69.7974% (Wuhan_2 is regional).

Claims (3)

1. the urban water-body extracting method of a kind of satellite remote-sensing image, it is characterised in that comprise the following steps：

Step 1：The pretreatment of remote sensing image data, i.e., carry out ortho-rectification and atmospheric correction to remote sensing image data；

Step 2：Pretreated remote sensing image data, including blue band1 data, green band2 data, redness band3 numbers According to near-infrared band4 data, choose the blue band1 data in pretreated remote sensing image data and replace normalization difference Green band2 data in the computing formula of water body index NDWI, obtain new normalization difference water body index NNDWI1, new Normalization difference water body index NNDWI1 computing formula be：

$NNDWI1=\frac{(Band1-Band4)}{(Band1+Band4)}$

This computing formula is NNDWI1 exponential models, and the Threshold segmentation of NNDWI1 is obtained by Threshold segmentation using this model As a result, also as NNDWI1 Clean water withdraws result；

Step 3：Four wave band datas that pretreated remote sensing image data is included, i.e. blueness band1 data, green Band2 data, redness band3 data and near-infrared band4 data carry out PCA conversion, and the first principal component after PCA is converted Component Component1 substitutes the green band2 data in the computing formula of normalization difference water body index NDWI, obtains another Individual new normalization difference water body index NNDWI2, i.e.,：

$NNDWI2=\frac{(Component1-Band4)}{(Component1+Band4)}$

Wherein, Component1 represents the first principal component component that PCA is converted, and this computing formula is NNDWI2 exponential models, The Threshold segmentation result of NNDWI2 is obtained by Threshold segmentation using this model, also the Clean water withdraw result of as NNDWI2；

Step 4：The Threshold segmentation knot of the NNDWI2 obtained in the Threshold segmentation result of the NNDWI1 that step 2 is obtained and step 3 Fruit is overlapped, and the result for obtaining is defined as into the Threshold segmentation result of new normalization difference water body index NNDWI, i.e., The Threshold segmentation result of NNDWI1 is overlapped with the Threshold segmentation result of NNDWI2, and its computing formula is：

NNDWI=(segmentation_NNDWI1) ∪ (segmentation_NNDWI2)

Segmentation_NNDWI1 represents the Threshold segmentation result of NNDWI1 in formula, and segmentation_NNDWI2 is represented The Threshold segmentation result of NNDWI2, this computing formula is NNDWI exponential models, and using this exponential model the water of NNDWI is obtained Body extracts result；

Step 5：Row threshold division is entered to the near-infrared band4 data in pretreated remote sensing image data, near-infrared is obtained The Threshold segmentation result of band4 data；

Step 6：Water body in large object and small area object in the Clean water withdraw result of NNDWI is split, NNDWI's In Clean water withdraw result, number of pixels is water body in large object more than given threshold, and number of pixels is less than or equal to setting threshold Value for small area object；

Step 7：Small area object to obtaining in step 6 carries out mathematical morphology expansion process, the small area after being expanded Object, the Threshold segmentation result of the near-infrared band4 data that step 5 is obtained adopts little after expansion as constraints The mode that the Threshold segmentation result of object oriented and near-infrared band4 data seeks common ground is carried out about to the small area object after expansion Beam, the mathematic(al) representation of constraint is：

Component2=(dilate_component) ∩ (segmentation_band4)

In formula, dilate_component represents the small area object after expansion, and segmentation_band4 represents near-infrared The Threshold segmentation result of band4 data, component2 represents the small area object after constraint；

Step 8：Small area object after the constraint obtained to step 7 carries out shadow Detection and removes, and obtains small area water body pair As；

Step 9：The small area water body object obtained in the water body in large object obtained in step 6 and step 8 is overlapped, The water body in large object that will be obtained in step 6 and the small area water body object obtained in step 8 seek union, obtain satellite distant The urban water-body of sense image extracts result.

2. the urban water-body extracting method of a kind of satellite remote-sensing image according to claim 1, it is characterised in that described Shadow Detection in step 8 and removal, referring to carries out the description of wave spectrum relation to each pixel in each small area object, and Judge whether the pixel meets the condition of shade pixel, record and count the number of shade pixel in each small area object, when When shade pixel proportion is more than threshold value T in one small area object, the small area object is judged to building effects pair As, then it is judged to small area water body object less than or equal to small area object during threshold value T, shade pixel proportion is small area The number of shade pixel and the ratio of total pixel number in the small area object, distinguish small area object small areas water in object The function expression of body object and shadow object is：

$\left\{\begin{array}{ccc}component2=water& if& \frac{m}{n}\≤T\\ component2=shadow& if& \frac{m}{n}>T\end{array}\right.$

In formula, n represents total pixel number in a certain small area object, and m is the number of shade pixel in the small area object.

3. the urban water-body extracting method of a kind of satellite remote-sensing image according to claim 2, it is characterised in that described The condition of shade pixel, refers to the wave spectrum magnitude relationship for meeting shade pixel, that is, meet three below inequality condition：

$\left\{\begin{array}{c}band2>band1\\ band3>band2\\ band4>band3\end{array}\right.$

$\left\{\begin{array}{c}band1>band2\\ band4>band2\\ band4>band3\end{array}\right.$

$\left\{\begin{array}{c}band3>band2\\ band3>band4\\ band4>band2\end{array}\right..$