CN107507202A - A kind of vegetation rotary island towards high-resolution remote sensing image automates extracting method - Google Patents

Abstract

The invention discloses a method for automatically extracting vegetation roundabouts from high-resolution remote sensing images, including four steps: object-oriented image segmentation and rule-based extraction, support vector machine-based road classification, road hole extraction, and inspection of roundabout blocks and Empty topological relationships. The object-oriented image segmentation of the present invention has two steps of segmentation and fusion. The segmentation has two modes: edge and intensity, and the edge mode is preferred; the fusion has two modes: full lambda mode or fast lambda mode, and full lambda mode is preferred; The road classification based on the support vector machine of the present invention has four functions as kernel function candidates: linear, polynomial, radial basis function, and S-shaped, and the radial basis function is preferred. The invention can effectively extract the roundabouts containing vegetation, and has the characteristics of high accuracy, strong adaptability, automation, no dependence on other auxiliary data, and the like.

Description

An automatic extraction method of vegetation surrounding islands for high-resolution remote sensing images

technical field

The invention belongs to the technical field of visual processing, and relates to a method for automatically extracting vegetation roundabouts from high-resolution multi-band remote sensing images.

Background technique

A roundabout is a traffic facility commonly found in areas with less traffic pressure and sufficient locations. Because the roundabout turns the meeting point of vehicles into a driving point, it can slow down the speed of vehicles and reduce the occurrence of collisions, thus improving traffic quality. The radius of roundabouts is generally 12 to 30 meters, which varies with road grades, and some roundabouts can even become landmarks or leisure squares. Vegetation is often planted in the roundabout, and these vegetation are generally evergreen plants, flowers or low shrubs, and should have obvious edges and a height that does not affect the driver's driving.

The extraction of vegetation roundabouts can help understand traffic conditions, help municipal planning, improve electronic maps that assist manual or automatic driving, or measure the prosperity of an area, which has certain significance for urban transportation, design, and humanistic research. There are few studies on round-island extraction from remote sensing images. In 2009, Ravanbakhsh proposed a geospatial database as a priori knowledge for round-island extraction, thus developing the research on round-island extraction. However, the current round-island extraction methods still have many deficiencies, such as low extraction accuracy, failure to make good use of vegetation information in remote sensing images, or having to rely on other auxiliary data, etc.

Therefore, it is a difficult problem to be solved in this field to propose a method for automatically extracting vegetation roundabouts from high-resolution multi-band remote sensing images.

Contents of the invention

In order to solve the above-mentioned technical problems, the present invention provides a method for automatic extraction of vegetation surrounding islands for high-resolution remote sensing images.

The technical solution of the present invention is: an algorithm for automatically extracting vegetation roundabouts from high-resolution multi-band remote sensing images, comprising the following steps:

Step 1: Execute process A and process B in parallel;

Process A: Carry out image segmentation on the original image, and extract the roundabout and roundabout blocks;

Process B: Carry out road classification on the original image, extract road holes and hole blocks;

Step 2: Check the topological relationship between the roundabout blocks and holes;

Step 3: Obtain the roundabout extraction result.

In the present invention, image segmentation has two watershed algorithms and two fusion modes to choose from, and support vector machine classification has four kernel functions to choose from, which is more flexible; the present invention has no higher requirements for processed images and can adapt to The high-resolution multi-band remote sensing images of different sizes, different clippings, different resolutions, and different band configurations have strong adaptability and robustness; The result of interpretation has a higher coincidence rate; the data processing time of the present invention is shorter, and a large number of roundabouts can be processed in a short time; It does not rely on other auxiliary knowledge or data, and has very strong independence; the present invention can still make a rough judgment on the topological relationship when the data situation is not ideal or the error caused by the algorithm itself, and has a certain fault tolerance rate, ensuring Reliability of extraction results. The invention improves and innovates the algorithm for extracting the surrounding island features from remote sensing images, reduces the cost for automatic identification and extraction tasks around the island, increases stability and accuracy, and has a positive role in promoting the related research on remote sensing ground feature extraction .

Description of drawings

Fig. 1 is a schematic flow chart of an embodiment of the present invention;

Fig. 2 is a schematic diagram of the preprocessing flow chart of the embodiment of the present invention;

Fig. 3 is a schematic flow chart of an object-oriented image segmentation process according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a rule-based extraction process according to an embodiment of the present invention;

Fig. 5 is the schematic flow chart of the road classification based on support vector machine of the embodiment of the present invention;

FIG. 6 is a simplified flow chart of hole detection according to an embodiment of the present invention;

detailed description

In order to facilitate those of ordinary skill in the art to understand and implement the present invention, the present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described here are only used to illustrate and explain the present invention, and are not intended to limit this invention.

The invention is an algorithm for automatically extracting vegetation roundabouts from high-resolution multi-band remote sensing images, which includes the following four steps: object-oriented image segmentation and rule-based extraction, road classification based on support vector machines, road hole extraction, and inspection Topological relationship between island blocks and voids. Before the first step, there can be appropriate image preprocessing; after the fourth step, the extraction results can be compared with the results of human visual interpretation to evaluate the results of the automated algorithm. The first step can be paralleled with the second and third steps, the second step must be before the third step, and the fourth step must be carried out after the first, second and third steps are completed, as shown in Figure 1.

Preprocessing refers to geometric correction, radiometric calibration, sky-map registration, cropping and other processing on the original remote sensing image, so as to improve the quality of the original remote sensing image as much as possible, hoping to improve the accuracy of the results of the present invention from the data source . In terms of order, there is no specified order for geometric correction and radiometric calibration. Space map registration must be completed after the first two are completed, and clipping must be performed after the first three are completed, as shown in Figure 2. Geometric correction refers to eliminating or reducing the geometric deformation of the original image; radiometric calibration refers to converting the dimensionless DN value recorded by the sensor into the radiance or reflectance of the top layer of the atmosphere with actual physical meaning; sky map registration refers to the geometric correction , radiometrically calibrated images are registered with a standard, orthographic world map; cropping is the process of cutting a large remote sensing image into a small size containing one or more complete roundabouts in order to reduce processing time. The entire preprocessing is not necessary, but proper preprocessing can guarantee the accuracy of the extraction results of the present invention within a certain range, and can reduce the data processing time.

The specific implementation plan of embodiment is:

Geometric correction: In the ENVI5.3sp software, open the Gaofen-2 original image data and its metadata (other high-resolution multi-band images are also acceptable), select the RPCOrthorectification Workflow tool under Orthorectification under Geometric Correction, and select the output pixel size For example, "4 meters", select the resampling method as "cubic equation", and after completing the settings such as the output path, the geometric correction can be implemented according to the set parameters.

Radiation calibration: In ENVI5.3sp software, open the Gaofen-2 original image data or geometrically corrected data, select the Radiometric Calibration tool under Radiometric Correction, and set the calibration type to Radiance data, which is automatically set by ENVI software The data type required by the FLAASH atmospheric correction tool, the storage order is BIL or BIP, the data type is FLOAT, and the emissivity data unit adjustment factor is set to 0.1. After completing the relevant settings, the radiation calibration can be implemented according to the set parameters.

For sky-map registration, you can use ArcMap 10.2 in ArcGIS 10.2 software to call its Georeferencing function, and select a few pairs of obvious artificial features by clicking on the area captured by the image of the orthographic standard world map (standard world map http ://www.scgis.net.cn/imap/iMapServer/defaultRest/services/newtiandidom/WMS) for registration.

For cropping, you can use the resize function in the ENVI5.3sp software to specify the corner position and crop size after cropping (for example, the coordinates of the upper left corner are 114.468, 30.471, and the size is 400*400 pixels), and then the cropping can be completed.

The first formal step of the present invention: object-oriented image segmentation, using the watershed algorithm to segment the entire high-resolution multi-band remote sensing image, the result of the segmentation is an image block, and for many image blocks, it will enter the rule-based Extract for processing. The watershed algorithm originated from the concept of hydrology, that is, as the water level rises, water bodies in different elevation areas will converge, and the watershed is built at the confluence of water bodies from different elevation areas. In the field of digital image processing, the watershed algorithm is a commonly used image segmentation algorithm. The ideal segmentation result should be consistent with the objects in the real world, that is, ideally, a segmented block should correspond to a real entity. This is an object-oriented idea. embodiment. Watershed segmentation includes two steps of segmentation and fusion, as shown in Figure 3.

There are two fusion modes for segmentation: edge or intensity. The edge mode refers to the edge detection of the original image by the sobel operator, and then forms a gradient image to perform segmentation by the watershed algorithm. The intensity mode refers to the original image and selects parts from it. Band, take the mean value within a certain range to generate an intensity map, and then perform segmentation by the watershed algorithm.

Fusion is to avoid the problem of over-segmentation. After judging the degree of correlation between the divided image blocks, the fusion is performed. There are two fusion modes available for fusion: full λ mode or fast λ mode; the formula of full λ mode is as follows, if the merging cost t _i,j of two pixels is less than the threshold, then merge:

Among them: O _i is the area i in the image;

|O _i | is the area of region i;

μ _i is the mean value of region i;

||μ _i -μ _j || is the Euclidean distance between the spectral values of region i and region j;

is the length of the common boundary of O _i and O _j ;

The formula of the fast lambda mode is as follows. The smaller the lambda value, the smaller the Euclidean color distance, and the larger the common boundary length, the more they should be merged. If the lambda value is smaller than the threshold, they should be merged;

Where: N1 is the number of pixels in area 1;

E is the Euclidean color distance between area 1 and area 2;

L is the common boundary length of area 1 and area 2.

The first step of the rule-based extraction is to consider the obvious characteristics of the vegetation roundabout: size, roundness and vegetation index for the segmentation results, and set the comprehensive consideration of the three as a rule, that is, the size is within the range of the country and region to which the image belongs. The radius of the roundabout specified by the transportation department is within the range of the upper and lower limits; the roundness (the formula is as follows) is within a certain range close to 1; the vegetation index uses NDVI (normalized differential vegetation index, the formula is as follows ) and the NDVI value of the image block where the vegetation roundabout is located should not be lower than the first 40% of the NDVI value of the entire image, as shown in Figure 4. After the three attributes of the rules are set, filter all image blocks with reference to the rules; there should be only one image block left after filtering, that is, the roundabout block.

Where: R is the roundness of the region;

S is the area of the region;

C is the perimeter of the area;

Where: NDVI is the normalized difference vegetation index;

eps is a constant small enough to prevent the denominator from being 0;

b1 is the infrared band;

b2 is the near-infrared band;

The specific implementation plan of embodiment is:

For the preprocessed image, select the Rule Based Feature Extraction Workflow tool in ENVI, check the Normalized difference option, and set the third band (red band) of the image to band 1, and the fourth band (infrared band) to band 2. Take this as the setting to participate in the calculation of NDVI. Next, when entering image segmentation, you need to set the watershed algorithm mode and parameters for image segmentation, as well as the fusion mode and parameters, which can be set to: "edge mode, degree 50, full lambda mode, 0".

Next set the rules. Since the rules in the present invention include three items of size, roundness and vegetation index, then first create the rules, then click on the created rules to create three attributes, and the attributes can be set as follows respectively: "Spatial attribute-size: 452-2826 ; Spatial property - circularity: 0.5-1.2; Band property - Normalized Normalized Index (NDVI): 60% - 100%". Run according to the rules, perform screening, and view the filtered results, mainly focusing on two aspects: whether the roundabout is left after the rule-based screening; whether there are redundant non-roundabout blocks after the screening. If the former is yes and the latter is no, object-oriented image segmentation and rule-based extraction are completed. The extraction result can be converted to evf and then to shp file for subsequent operation. It is worth noting that the shapefile data format (.shp file) of ESRI is not the only feasible data format of the present invention, depending on the specific operating software and production requirements.

Road classification based on support vector machine refers to the use of support vector machine algorithm (Support Vector Machine, SVM) for image classification after selecting an appropriate number of road sample points. SVM is a commonly used algorithm for pattern recognition. Its idea is: for the case of linear inseparability, the linear inseparable samples of low-dimensional input space are transformed into high-dimensional feature space (hyperplane) by using nonlinear mapping algorithm to make them linearly separable. Therefore, it is possible to linearly analyze the nonlinear characteristics of samples using linear algorithms in high-dimensional feature spaces.

Road sample points can be manually selected in ENVI software to form ROI (Regions of interest) to obtain road pixel samples by selecting polygons with the mouse, or you can specify classification samples in other ways. There are four options for the kernel function of the classification, namely linear, polynomial, radial basis function, and S-shape (the formula is as follows), and the output is the classification result including the banded entity. It is recommended that if there is no special requirement, you can choose the radial basis function as the kernel function, because it works better in most cases. In addition, three parameters required by the SVM algorithm must be set: the number of pyramid layers, the penalty parameter, and the classification probability threshold (as shown in Figure 5). The number of pyramid levels is the number of levels of hierarchical processing applied in a support vector machine to avoid classification or overclassification. The multiplication parameter allows a certain degree of misjudgment, and theoretically allows some training points to be on the wrong side of the hyperplane, so that the support vector machine model has strong fault tolerance and flexibility. The classification probability threshold represents the closest part point in the hyperplane that represents the confidence of that class, so a higher threshold results in fewer closest points being classified and an increase in the unclassified part of the image. Simply put, the classification probability threshold controls the degree of classification.

In the present invention, a satisfactory SVM-based road classification result is a continuous strip-shaped entity, that is, a strip-shaped road, in which there should be a large hole close to a circle. Since this big hole should be caused by the roundabout, the spatial positions of this hole and the roundabout should basically coincide. Subtle errors are unavoidable, such as the fine holes and uneven edges of the road extraction results caused by the shadows of vehicles or street trees, and some gaps or overlaps at the edges of holes and roundabouts. The present invention can tolerate these unavoidable errors, which are reflected in the next two steps.

Linear: K(x _i ,x _j )=x _i ^T x _j

Polynomial: K(x _i ,x _j )=(gx _i ^T x _j +r) ^d

Radial basis function: K(x _i ,x _j )=exp(-g||x _i -x _j || ² )

S shape: K(x _i ,x _j )=tanh(gx _i ^T x _j +r)

Among them: g is the gamma parameter required for functions other than the linear kernel function;

d is the parameter required by the polynomial kernel function;

r is the partial term parameter required by polynomial and sigmoid kernel functions.

The specific implementation plan of embodiment is:

Open the image to be classified, and select the SupportVector Machine tool of Supervised under Classification in ENVI to perform SVM classification. The following settings can be made: select the radial basis function, set gamma to 0.25, penalty parameter to 100, pyramid level to 0, and classification probability threshold to 0.7. It should be noted that the classification effect of these settings varies with different images, and it may sometimes require multiple appropriate adjustments to achieve the desired effect.

The road hole extraction only extracts the holes in the strip road, and regards the largest hole as the hole caused by the roundabout. This is because under normal circumstances, the SVM-based classification effect of roads is difficult to produce an extraction error with a radius exceeding 12 meters, that is, an error larger than the roundabout in area.

Hole extraction includes four steps, each step is completed with GIS (Geographic Information System) tools, three of which are conversion; the first step is to convert polygons into points, that is, convert strip road polygons into center points, and the center point attributes include number FID , shape Shape, type_name Class_Name, type_number Class_ID, block number Parts, length Length, area Area, source number ORIG_FID; the second step is to convert polygons into lines, that is, convert strip road polygons into boundary lines, and these boundaries The line has only a few attributes, and the boundary line attributes include number FID, shape Shape, type_name Class_Name, type_number Class_ID, block number Parts, length Length, area Area, source element ID; the third step is to convert the line to a polygon, that is Regenerate polygons according to the boundary line, polygon attributes include number FID, shape Shape, source element ID, type_number Class_ID, block number Parts, length Length, area Area, source number ORIG_FID; according to the ID or other signs of each polygon, Add the attributes of the point generated in the first step to the polygon generated in the third step. If the source polygon does not exist, the additional attribute is empty; the fourth step, type_number Class_ID, block number Parts, length Length, area Area The polygons whose source element IDs are all 0 are holes; finally, the holes are sorted by area, and the largest one is selected, which is the road hole caused by the roundabout, called a hole block (the process is shown in Figure 6).

That is to say, the hole extraction of the road takes advantage of the loss and retention of attributes in various transformations of GIS, and filters out holes; since the hole is not a single polygonal entity at the beginning, after the three-step conversion, A polygon consistent with the shape of the hole is generated from the boundary of the hole. Compared with all the polygons before conversion, there are naturally differences in the completeness and the number of items in the attributes.

The specific implementation plan of embodiment is:

In ArcMap10.2, "feature to points", "polygon to lines", and "feature to polygons" can be used successively to convert polygons to points, polygons to lines, and lines to polygons. It should be emphasized that in the third step of the conversion operation, use the points converted in the first step as label features, and select the preserveattributes option, in order to add the attributes obtained in the first step to the polygon generated in the third step, and then find all hollow. Finally, select all the polygon shapefiles generated in the third step, use attribute-based queries to find those polygons that are missing attributes such as classname, area and original fid, and click the sort button to sort by size. Select the polygon with the largest area and save it in shp format, which is called a hole block.

The purpose of testing the topological relationship between the roundabout block and the hole is to tolerate the incomplete coincidence of the roundabout block and the hole due to many errors, and to allow the gap or coincidence of the roundabout block and the boundary of the hole, that is, to accommodate the relationship between the road and the roundabout block It is not a phenomenon of inclusion relationship in an ideal state, but at the same time, it still roughly checks the topological relationship between the roundabout block and the road, and provides a basis for judging whether the roundabout block is a roundabout. The specific method is to calculate the center point of the roundabout block and the hole separately, then measure the distance between the two center points, and compare the distance with a certain threshold. If the distance is less than the threshold, it is determined that the extracted roundabout block is a roundabout. If the distance is greater than the threshold , the extraction is invalid.

The specific implementation plan of embodiment is:

In ArcMap10.2, use "feature to points" on the roundabout block and hollow block shapefile extracted in the previous steps to obtain the center of gravity, or center, of the two. The center of gravity is calculated by taking the mean value of the two-dimensional coordinates of all boundary points of the entity. Display the two centers differently, such as red triangles and black dots. Then use ArcMap's Measure tool to measure the distance between two points. A rough threshold can be set with reference to the size and resolution of remote sensing images, combined with the general width of the road. For example, the original remote sensing image is 7300*6908 pixels, and the cropped remote sensing image is 400*400 pixels, each pixel represents the actual ground area of 4*4 meters (the ground resolution is 4m*4m), the area where the roundabout to be judged is located For Huangshi, Hubei Province, China, the road width is about 8-30 meters. Taking these factors into consideration, the threshold can be set to 5 meters. Comparing the distance with the threshold (5m), make a corresponding judgment, and roughly judge whether the approximate topological relationship between the roundabout block and the road hole is coincident, that is, whether the extracted roundabout block can be regarded as a real roundabout. The specific value of the threshold is not too limited, as long as it can reflect the feature of judging the extraction effect of the roundabout from the distance between two points in the present invention, and can roughly judge the extraction effect of the roundabout.

The best way to test the roundabout extraction effect of the present invention is to conduct statistical and comprehensive analysis of the distance between multiple extracted roundabout blocks and hollow blocks, and to compare the extracted roundabout blocks with the results of human visual interpretation. The effect is compared. Table 1 shows the statistical results of the distance and the comparison results with visual interpretation in a total of 10 island roundabout experiments in Wuhan, Huangshi, and Ezhou, Hubei Province, China. Table 1 shows that when the ground resolution is 4m*4m or 3.7m*3.7m, the distance between the two center points is less than 4 meters, that is, the side length of about one pixel, and the rough topological judgment can play a better role. It shows robustness to different data sources, and the extraction effect can be considered to be relatively accurate. Table 1 also shows that the producer accuracy (producer accuracy, the percentage of pixels marked as roundabouts by both automatic and visual interpretation to the number of pixels marked as roundabouts by visual interpretation) exceeded 85% in all 10 experiments. %, has a fairly good extraction accuracy, and has obvious accuracy advantages compared with existing algorithms. Table 2 also shows the location, size, roundness and NDVI of the 10 roundabouts. The area of the roundabouts is basically within the expected range and meets the requirements of the transportation department. The roundness is basically between 0.4-1.0, which is in line with the geometric laws and the present invention According to the idea, NDVI basically has a higher value in the whole or cropped image where the roundabout is located, so it can be seen that the idea of screening based on rules is very correct and effective. These results can prove that the algorithm for automatically extracting vegetation roundabouts from high-resolution multi-band remote sensing images is available and accurate, indicating that the algorithm for automatically extracting vegetation roundabouts from high-resolution multi-band remote sensing images is feasible and has high accuracy , Adaptability, automation, and the advantages of not relying on other auxiliary data.

GTR(m) Distance(m) Producer Accuracy 1 3.7 3.036 91.67% 2 3.7 3.271 86.02% 3 3.7 3.409 85.71% 4 3.7 2.308 92.63% 5 3.7 3.821 87.37% 6 3.7 3.286 90.32% 7 4 0.858 95.47% 8 4 2.461 94.06% 9 4 2.604 96.67% 10 4 2.486 94.74%

Table 1 The ground resolution (GTR), center point distance (Distance) and production accuracy (Producer Accuracy) of ten experiments

ID Coordinates Rd.Name Size(m ² ) Roundness NDVI 1 114.368,30.571 Oriole Road 555.4803 0.584201 0.044186 2 114.377,30.569 Yanhu Road 1353.093 0.473994 0.122496 3 114.255,30.626 Evergreen Road 897.3143 0.710775 0.072343 4 115.059,30.255 Yellowstone Road 1623.712 0.737449 0.094100 5 114.378,30.580 Liyuan Road 1285.934 0.816796 0.117352 6 115.097,30.211 Yiyang Road 1373.633 0.764884 0.035440 7 114.825,30.417 Sihai Road 3926.975 0.908119 0.229562 8 114.920,30.398 Situ Lu 1600.000 0.810996 0.239780 9 114.882,30.391 Soochow Road 2432.000 0.855590 0.207698 10 114.882,30.367 Guanliu Road 1104.000 0.820611 0.166888

Table 2 Coordinates (Coordinates), name (Rd.Name), size (Size), roundness (Roundness) and vegetation index (NDVI) of the ten experiments

It should be understood that the parts not described in detail in this specification belong to the prior art.

It should be understood that the above-mentioned descriptions for the preferred embodiments are relatively detailed, and should not therefore be considered as limiting the scope of the patent protection of the present invention. Within the scope of protection, replacements or modifications can also be made, all of which fall within the protection scope of the present invention, and the scope of protection of the present invention should be based on the appended claims.

Claims (8)

1. A vegetation rotary island automatic extraction method facing a high-resolution remote sensing image is characterized by comprising the following steps:

step 1: executing the flow A and the flow B in parallel;

scheme A: carrying out image segmentation on an original image, and extracting a roundabout and a roundabout block;

and (B) a process: carrying out road classification on the original image, and extracting road cavities and cavity blocks;

step 2: checking the topological relation between the ring island blocks and the holes;

and step 3: and acquiring a roundabout extraction result.

2. The automatic extraction method of the vegetation rotary island facing to the high-resolution remote sensing image according to claim 1, characterized in that: in the step 1, carrying out image segmentation on the whole high-resolution multiband remote sensing image by using a watershed algorithm, wherein the segmentation result is an image block;

the watershed algorithm comprises two steps of segmentation and fusion;

there are two fusion modes of segmentation: edge or strength; the edge mode is to perform edge detection of sobel operator on the original image, then form a gradient image and perform segmentation of watershed algorithm; the intensity mode is that aiming at the original image, part of wave bands are selected from the original image, and an average value in a certain range is taken to generate an intensity map;

there are two fusion modes of fusion: full λ mode or fast λ mode;

the full λ mode formula is as follows:

<mrow> <msub> <mi>t</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mfrac> <mrow> <mfrac> <mrow> <mo>|</mo> <msub> <mi>O</mi> <mi>i</mi> </msub> <mo>|</mo> <mo>&amp;CenterDot;</mo> <mo>|</mo> <msub> <mi>O</mi> <mi>j</mi> </msub> <mo>|</mo> </mrow> <mrow> <mo>|</mo> <msub> <mi>O</mi> <mi>i</mi> </msub> <mo>|</mo> <mo>+</mo> <mo>|</mo> <msub> <mi>O</mi> <mi>j</mi> </msub> <mo>|</mo> </mrow> </mfrac> <mo>&amp;CenterDot;</mo> <mo>|</mo> <mo>|</mo> <msub> <mi>&amp;mu;</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>&amp;mu;</mi> <mi>j</mi> </msub> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> </mrow> <mrow> <mi>l</mi> <mi>e</mi> <mi>n</mi> <mi>g</mi> <mi>t</mi> <mi>h</mi> <mrow> <mo>(</mo> <mo>&amp;part;</mo> <mo>(</mo> <mrow> <msub> <mi>O</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>O</mi> <mi>j</mi> </msub> </mrow> <mo>)</mo> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>;</mo> </mrow>

wherein: o is_iIs the area i in the image; i O_iI is the area of region i; mu.s_iIs the mean of region i; | mu |_i-μ_jI is the Euclidean distance of the spectral values of the region i and the region j;is O_iAnd O_jThe length of the common boundary; if the merging cost t of two pixels_i，jIf the sum is less than the threshold value, merging;

the fast λ mode formula is as follows:

<mrow> <mi>L</mi> <mi>a</mi> <mi>m</mi> <mi>b</mi> <mi>d</mi> <mi>a</mi> <mo>=</mo> <mo>&amp;lsqb;</mo> <mfrac> <mrow> <msub> <mi>N</mi> <mn>1</mn> </msub> <mo>&amp;CenterDot;</mo> <msub> <mi>N</mi> <mn>2</mn> </msub> </mrow> <mrow> <msub> <mi>N</mi> <mn>1</mn> </msub> <mo>+</mo> <msub> <mi>N</mi> <mn>2</mn> </msub> </mrow> </mfrac> <mo>&amp;rsqb;</mo> <mfrac> <mi>E</mi> <mi>L</mi> </mfrac> <mo>;</mo> </mrow>

wherein: n1 is the number of pixels in region 1; e is the euclidean distance of the region 1 and the region 2; l is the common boundary length of region 1 and region 2; the smaller the lambda value is, the smaller the Euclidean color distance is, the larger the common boundary length is, the more should be merged, and if the lambda value is smaller than the threshold value, the merging is performed.

3. The automatic extraction method of the vegetation rotary island facing to the high-resolution remote sensing image according to claim 1, characterized in that: in the step 1, based on the image segmentation result, the size, the roundness and the vegetation index of the image block are considered at the same time, and the setting rule of the three are considered comprehensively: the size of the circle is within the range of the upper limit and the lower limit of the circle area with the radius of the roundabout specified by the traffic department of the country and the region to which the image belongs; the roundness R is within a predetermined range close to 1; the vegetation index is normalized vegetation index NDVI and the NDVI value of the image block where the vegetation rotary island is located is not lower than the first 40 percent of the NDVI value of the whole image; screening all image blocks according to a rule; the screening result is a roundabout block;

wherein,r is the roundness of the region, S is the area of the region, C is the perimeter of the region;NDVI is the normalized vegetation index, eps is a constant small enough to avoid denominators other than 0, b1 is the infrared band, b2 is the near infrared band.

4. The automatic extraction method of the vegetation rotary island facing to the high-resolution remote sensing image according to claim 1, characterized in that: in step 1, after selecting a proper number of road sample points, using a support vector machine algorithm to classify images, wherein classified kernel functions comprise linear kernel functions, polynomial kernel functions, radial basis kernel functions and S-shaped kernel functions, and output a classification result containing a banded entity, wherein the banded entity is a banded road and should contain a cavity which accords with the size of a roundabout;

linear kernel function: k (x)_i，x_j)＝x_i ^Tx_j，

Polynomial kernel function: k (x)_i，x_j)＝(gx_i ^Tx_j+r)^d，

Radial basis kernel function: k (x)_i，x_j)＝exp(-g||x_i-x_j||²)，

Sigmoid kernel function: k (x)_i，x_j)＝tan h(gx_i ^Tx_j+r)，

Wherein: g is a parameter required in functions other than the linear kernel function; d is a parameter required for the polynomial kernel; r is the bias term parameter required for the polynomial and sigmoid kernel functions.

5. The automatic extraction method of the vegetation rotary island facing to the high-resolution remote sensing image according to claim 1, characterized in that: in the step 1, only extracting the holes in the strip road, and regarding the largest hole as a hole caused by a roundabout;

the cavity extraction comprises four steps:

the first step is as follows: converting the polygon into points;

converting a strip-shaped road polygon into a central point, wherein the central point attribute comprises a number FID, a Shape, a type _ Name, a type _ number Class _ ID, a block number part, a Length, an Area and a source number ORIG _ FID;

the second step is that: converting the polygon into a line;

converting a strip-shaped road polygon into a boundary line, wherein the boundary line attribute comprises a number FID, a Shape, a type _ Name, a type _ number, a Class _ ID, a block number part, a Length, an Area and a source element ID;

the third step: the conversion line is a polygon;

generating a polygon according to the boundary line, wherein the polygon attribute comprises a number FID, a Shape, a number file FID _ shp, a type _ number Class _ ID, a block number part, a Length, an Area and a source element ID; adding the attributes of the points generated in the first step into the polygon generated in the third step according to the ID or other marks of each polygon, wherein if the source polygon does not exist, the additional attributes are null;

the fourth step: a polygon with type _ number Class _ ID, block number Parts, Length, Area and source element ID all being 0 is a hole; the holes are sorted according to area, and the largest hole block is selected.

6. The automatic extraction method of the vegetation rotary island facing to the high-resolution remote sensing image according to claim 1, characterized in that: in step 2, the central points of the roundabout block and the hollow block are respectively calculated, the distance between the two central points is measured, the distance is compared with a certain threshold, if the distance is smaller than the threshold, the extracted roundabout block is judged to be the roundabout, and if the distance is larger than the threshold, the extraction is invalid.

7. The method for automatically extracting the vegetation rotary island facing the high-resolution remote sensing image according to any one of claims 1 to 6, wherein the method comprises the following steps: in the step 1, firstly, preprocessing is carried out on an original image, and then image segmentation or road classification is carried out;

the pretreatment specifically comprises the following substeps:

step A1: judging whether to carry out geometric correction on the original image or not;

if so, performing geometric correction on the original image, and then performing radiometric calibration;

if not, firstly carrying out radiometric calibration and then carrying out geometric correction on the original image;

step A2: registering the geometrically corrected and radiometric images with a standard, orthoscopic world map;

step A3: and cutting the processed remote sensing image into a plurality of small remote sensing images containing one or more complete rotary islands.

8. The method for automatically extracting the vegetation rotary island facing the high-resolution remote sensing image according to any one of claims 1 to 6, wherein the method comprises the following steps: and (3) counting and comprehensively analyzing the distances between the plurality of extracted roundabout blocks and the cavity block for a plurality of times according to the roundabout extraction result obtained in the step (3), and comparing the plurality of extracted roundabout blocks with the effect of artificial visual interpretation.