A road network extraction method based on SAR images considering angle texture features and POI

Abstract

The present invention relates to a method for extracting a road network from a SAR image considering angle texture features and POI, comprising: step 1: obtaining input data, including SAR image data, binary segmentation data, POI point data and POI connection data; step 2: constructing a road network extraction model considering angle texture features and POI; step 3: inputting the input data in step 1 into the road network extraction model constructed in step 2 to obtain a road network extraction result. Compared with the prior art, the present invention has the advantages of effectively extracting a road network from a SAR image, having a good extraction effect, and having good versatility.

Description

Technical Field

[0001] The present invention relates to the technical field of SAR image road network extraction, and in particular to a SAR image road network extraction method considering angle texture features and POI. Background Art

[0002] Urban road network information is an important part of modern spatial information infrastructure. In the era of big data, urban road network information is crucial for high-precision road navigation and autonomous driving. Field measurement, geographic information collection, remote sensing, etc. are important means to obtain urban road network information. In particular, for the supplement and update of road network information, ground and satellite remote sensing have unique advantages in expressway mapping. Satellite remote sensing is widely used in road extraction due to its numerous data sources and rich information. Among them, optical images and synthetic aperture radar images are more widely used. Of the two, SAR images are more difficult to extract road information due to their special imaging methods. In particular, road network extraction from high-resolution SAR images is affected by image noise and complex road network structures, and it is difficult to obtain good results. In recent years, driven by big data, crowd-source geographic information has been increasingly applied to surveying and mapping remote sensing. POI data is one of them. It not only contains location information such as shopping malls, restaurants, and scenic spots, but also contains a large amount of road-related information. This information undoubtedly provides the possibility for effective road identification.

[0003] At present, there are few studies on using POI data for road extraction, especially for SAR image road extraction. Therefore, for road network extraction, it is necessary to study the combination of SAR image road extraction and POI information. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a SAR image road network extraction method that effectively extracts road networks in SAR images, has good extraction effect and good versatility, and takes into account angle texture features and POI.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for extracting a road network from a SAR image considering angle texture features and POI, the method comprising:

[0007] Step 1: Obtain input data, including SAR image data, binary segmentation data, POI point data, and POI connection data;

[0008] Step 2: Construct a road network extraction model considering angle texture features and POI;

[0009] Step 3: Input the input data in step 1 into the road network extraction model constructed in step 2 to obtain the road network extraction results.

[0010] Preferably, the method for acquiring the SAR image data is:

[0011] Geometric correction of SAR images using orbit files and corresponding DEM;

[0012] The method for obtaining the binary segmentation data is:

[0013] The corrected SAR image is binarized using the threshold segmentation method. In the binary segmented image, the road, water, and shadow areas are divided into black areas with a pixel value of 0, while other areas are divided into white areas with a pixel value of 1.

[0014] Preferably, the method for acquiring the POI point data is:

[0015] The POI class of the intersection name is used. Within the POI range, a rectangular box is used for movement calculation. The number of points in the rectangular box is counted. If the number is greater than 1, the center position of these points will be calculated, and only the point closest to the center position will be retained, and the other points will be deleted; if the number of points is 2, one of them will be retained;

[0016] The calculation formula is:

[0017]

[0018]

[0019] Among them, (c x ,c y ) is the coordinate of the center of the rectangular frame; (x Pi ,y Pi ) are the coordinates of each point in the rectangular frame; d i is the distance from each point to the center position; the value range of i is 1 to n, n>2.

[0020] More preferably, the step 2 is specifically as follows:

[0021] Step 2-1: Create an initial template;

[0022] Step 2-2: Construct a road recognition sub-model;

[0023] Step 2-3: Construct intersection recognition sub-model;

[0024] Step 2-4: Set up mobile strategy.

[0025] More preferably, the initial template in step 2-1 is a rectangular template.

[0026] More preferably, the road recognition sub-model is specifically:

[0027] Define the rectangle template as:

[0028] Template=(α,M,V,S)

[0029] Among them, α is the rotation angle of the rectangular template; M is the grayscale mean corresponding to α; V is the grayscale variance corresponding to α; S is the road information in the binary image corresponding to α;

[0030] When the rectangular template is rotated, the corresponding ones are:

[0031]

[0032] Sort M in ascending order, record the angles corresponding to the first several values, extract V and S corresponding to these angles and sort them in ascending order respectively; score several values ​​of M, V and S, with the highest score being 1 and the remaining values ​​decreasing by 0.1; after obtaining three sets of scores, combine the three using the following formula and sort the new array in descending order:

[0033]

[0034] Among them, δ, γ and ω are coefficients, ranging from 0 to 1; A is the total score after combining the three types of information, and A′ is the result of sorting A in descending order;

[0035] After sorting, select several directions with the highest total scores as candidate directions, calculate the difference in the grayscale mean between the current rectangular template and the next rectangular template, and record the corresponding angle;

[0036] The scoring method is also used to select the direction with smaller difference and smaller rotation angle as the advancing direction of the rectangular template.

[0037] More preferably, the intersection recognition sub-model is specifically:

[0038] The intersection is identified based on POI points and POI lines. When the rectangular template moves, the extension direction of the intersection is determined after the POI point is found. When the direction of the POI line is consistent with the road direction, the POI line is used to sort the road information. The corresponding grayscale mean, grayscale variance and road information of the binary segmentation image are obtained respectively, and then the scores are sorted by scoring. The direction with smaller grayscale mean difference and rotation angle will be used as the candidate direction for intersection extension.

[0039] When the POI connection direction is inconsistent with the road direction, the road recognition sub-model is directly used for judgment.

[0040] More preferably, the mobile strategy is specifically:

[0041] Move the rectangular template forward according to the road recognition sub-model, advancing half the template length each time.

[0042] More preferably, the steps 2-4 further include:

[0043] Set the stop movement policy. The movement will be stopped in any of the following situations:

[0044] Reach the edge of the image;

[0045] Arrived at the extracted road.

[0046] Preferably, the road network extraction model is specifically:

[0047] The initial template continues to advance and makes the following judgments:

[0048] If the extracted road is reached, the candidate intersection is determined;

[0049] When there is a candidate intersection and the corresponding road has not been extracted, the road recognition sub-model is used to extract the road;

[0050] Otherwise, the next intersection will be judged until the extraction requirements are met;

[0051] If there are no candidate intersections, or all intersections have been determined and do not meet the extraction requirements, the program is terminated;

[0052] If the extracted road has not been reached, it will continue to determine whether it has reached a road intersection; if it is determined to be a road intersection, it will be determined and stored according to the intersection recognition sub-model; otherwise, it will continue to determine whether it has reached the edge of the image; if it has reached the edge of the image, it will determine whether there is a candidate intersection waiting, if not, it will continue to determine the template direction;

[0053] The entire extraction process is the moving process of the rectangular template, and the judgment will be repeated each time until the stop movement strategy is met.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. Effectively extracting road networks from SAR images: The SAR image road network extraction method of the present invention fully combines the SAR image angle texture features and POI information, effectively extracting roads while ensuring the integrity of the road network structure.

[0056] 2. Good extraction effect: The SAR image road network extraction method of the present invention combines the grayscale mean, grayscale variance and binary segmentation information of the SAR image in a scoring manner, which solves the problem of inconsistent measurement scales between different information and effectively improves the extraction effect of the road network.

[0057] 3. Good versatility: Experiments have shown that the SAR image road network extraction method of the present invention can be used to extract different types of road networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the process of the SAR image road network extraction method in the present invention;

[0059] Figure 2 Schematic diagram of POI processing in an embodiment of the present invention;

[0060] in, Figure 2 (a) is a schematic diagram of the process of deleting duplicate POI data. Figure 2 (b) Schematic diagram of the process of determining the direction for connecting POI points;

[0061] Figure 3 is a schematic diagram of an initial rectangular template in an embodiment of the present invention;

[0062] in, Figure 3 (a) is a schematic diagram of the initial rectangular template created. Figure 3 (b) and Figure 3 (c) Two cases after adjusting the initial rectangular template;

[0063] Figure 4 Schematic diagram of a road recognition sub-model in an embodiment of the present invention;

[0064] Figure 5 A schematic diagram of an intersection recognition sub-model in an embodiment of the present invention;

[0065] in, Figure 5 (a) is a schematic diagram of the rotation process of the rectangular template at the intersection. Figure 5 (b) is a schematic diagram of template judgment at the intersection;

[0066] Figure 6 Schematic diagram of the experimental area selected in the embodiment of the present invention;

[0067] Figure 7 This is a schematic diagram of the experimental area-I after POI processing in an embodiment of the present invention;

[0068] in, Figure 7 (a) is the superposition display of POI lines and POI points. Figure 7 (b) is the POI connection diagram of experimental area-I. Figure 7 (c) is the POI distribution map of experimental area-I;

[0069] Figure 8 Schematic diagram of road network extraction results in an embodiment of the present invention;

[0070] in, Figure 8 (a) corresponds to an intersection experiment, Figure 8 (b) Corresponding to the two intersection experiments, Figure 8 (c) corresponds to the three intersection experiments; Figure 8 (d) corresponds to the experiment at four intersections;

[0071] Fig. 9 A schematic diagram of an abnormal region of an extraction result in an embodiment of the present invention;

[0072] Fig.10 This is a schematic diagram of the extraction results of other experimental areas in the embodiment of the present invention;

[0073] in, Fig.10 (a) is the extraction result of experimental area-II, Fig.10 (b) is the extraction result of experimental area-III; DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0075] A road network extraction method for SAR images considering angle texture features and POIs, the process is as follows Figure 1 As shown, including:

[0076] Step 1: Obtain input data, including SAR image data, binary segmentation data, POI point data, and POI connection data;

[0077] Step 2: Construct a road network extraction model considering angle texture features and POI;

[0078] Step 3: Input the input data in step 1 into the road network extraction model constructed in step 2 to obtain the road network extraction results.

[0079] The following is a detailed description of each step:

[0080] 1. Data Processing

[0081] (1) SAR image data:

[0082] Geometric correction of SAR images using orbit files and corresponding DEM;

[0083] (2) Binary segmentation data

[0084] The corrected SAR image is binarized using the threshold segmentation method. In the binary segmented image, the road, water, and shadow areas are divided into black areas with a pixel value of 0, while other areas are divided into white areas with a pixel value of 1. In the road extraction process, the black area will be used as auxiliary information.

[0085] (3) POI data

[0086] This embodiment uses the POI class of intersection names. POI data is collected spontaneously by people, so there is a lot of duplicate information. The purpose of applying POI data in this embodiment is to identify road intersections, so duplicate data needs to be removed. To achieve this goal, this embodiment adopts a filter-like method. Within the POI range, a 100×100 meter rectangular box is used for motion calculation. Count the number of points in the rectangular box, and if the number is greater than 1, calculate the center position of these points. Only keep the point closest to the center position and delete the other points. If the number of points is exactly 2, keep one of them. The calculation formula is as follows:

[0087]

[0088]

[0089] Among them, (c x ,c y ) is the coordinate of the center of the rectangular frame; (x Pi ,y Pi ) are the coordinates of each point in the rectangular frame; d i is the distance from each point to the center position; the value range of i is 1 to n, n>2.

[0090] Figure 2 (a) shows the POI processing process. The five-pointed star represents the center position of all POIs in the rectangular box, and the point closest to the center position will be retained. The processed POI points need to be buffered, because if the POI point position is inaccurate, a small template rotation angle may miss the POI point. The buffer will increase the target range of the POI point, making it easier to identify.

[0091] (4) POI connection data

[0092] The processed POI data only retains one point at each intersection, so these points can be used to identify road intersections. In actual roads, intersections are connected to each other. If the POI data can be effectively connected, it can be used to determine the direction of the road. For roads, they can generally be considered as continuous linear features with small curvature. In dense road networks, in most cases, the extension directions of road intersections (except for overpasses or special roads) do not exceed four. Based on these characteristics, this embodiment connects the POI data.

[0093] Assume that point a (at Figure 2 (b) is the first point to be judged. Calculate the distance between point a and other points based on the position. Sort the distances from small to large and take values ​​in order. Calculate whether there is a point on the extension line of the connecting line between point a and other points that falls within an angle of -30° to 30°. If so, the direction of the straight line connecting point a and the point is the initial direction, otherwise, the direction of the shortest distance is used as the initial direction. Then determine whether the angle between other directions and the initial direction is greater than a threshold (such as 80°). Keep directions with larger angles until all directions are judged. After removing point a, repeat the previous process to judge the remaining POI points in turn. Connect each pair of points and export the connection data for buffer processing.

[0094] exist Figure 2 In (b), α is the angle between ab and ac, and β is the angle corresponding to -30° to 30°. The point indicated by the arrow falls within β, and this direction is the initial direction. Points b, c, d and e are the direction points retained after judgment. POI lines cannot represent the real road network. Since roads are complex, there may be wrong connections. This embodiment is only to use POI data to assist in determining the direction, so a reasonable number of connections will be retained as much as possible.

[0095] 2. Model Construction

[0096] The model building process includes:

[0097] Step 2-1: Create an initial template;

[0098] The initial template in this embodiment is rectangular, and its advantage is that the width of the road can be kept unchanged during the extraction process. The shape and position of the rectangular template are determined by the coordinates of three points using geometric relationships. The first point determines the direction of the road and is selected in the middle of the road. The second and third points are used to determine the bottom edge (width) of the template and are selected on both sides of the road. Generally speaking, the length of the template should be greater than 2 times the width. The position of the second point should fall on the edge of the road as accurately as possible. In order to make the rectangular template better match the road, after the three points are determined, the template will be adjusted with the second point as the center. When creating a rectangular template, there are many situations due to different initial directions. Figure 3 (a) is one of them, and the corresponding formula is as follows:

[0099]

[0100]

[0101] X n =X C +S·sinα

[0102] Y n =Y C -S·cosα

[0103] X m =X n -BC·cosα

[0104] Y m =Y n -BC·sinα

[0105] Where (X A ,Y A ),(X B ,Y B ),(X C ,Y C ),(X m ,Y m )and(X n ,Y n )yes Figure 3 (a) The coordinates of the points, α is the ∠BCP. Figure 3 (b) and (c) show two possible situations that may occur during the template adjustment process.

[0106] Step 2-2: Construct a road recognition sub-model;

[0107] In SAR images, the grayscale of roads is slightly lower than that of surrounding objects, and in binary segmentation images, the value of roads is generally 0. Based on this information, the calculation content of angle texture features is improved. It mainly combines grayscale mean, grayscale variance and binary segmentation information to achieve road extraction. The model assumes that the rectangular template is defined as:

[0108] Template=(α,M,V,S)

[0109] Among them, α is the rotation angle of the rectangular template; M is the grayscale mean corresponding to α; V is the grayscale variance corresponding to α; S is the road information in the binary image corresponding to α;

[0110] When the rectangular template is rotated, the corresponding ones are:

[0111]

[0112] Sort M in ascending order, record the angles corresponding to the first 10 values, extract the V and S corresponding to these 10 angles and sort them in ascending order respectively; score the 10 values ​​of M, V and S, with the highest score being 1, and the remaining values ​​decreasing by 0.1; taking M as an example, the scores of the first 10 smallest values ​​are (1, 0.9, ..., 0.1). Similarly, V and S also get a set of scores, such as (0.9, 1, ..., 0.2) and (1, 0.9, ..., 0.3). It is normal for the scores of V and S to be different from those of M because the sorting is different.

[0113] After obtaining the three sets of scores, combine them using the following formula and sort the new array in descending order:

[0114]

[0115] Among them, δ, γ and ω are coefficients, ranging from 0 to 1; A is the total score after combining the three types of information, and A′ is the result of sorting A in descending order; in the formula, δ, γ and ω are used to determine the weights of the three types of information, which can be modified according to the specific situation of the image. The coefficients used in this embodiment are all 1. After sorting, the 5 directions with the highest total scores are selected as candidate directions. After sorting, the 5 directions with the highest total scores are selected as candidate directions, and the difference in the grayscale mean between the current rectangular template and the next rectangular template is calculated, and the corresponding angle is recorded;

[0116] The scoring method is also used to select the direction with smaller difference and smaller rotation angle as the advancing direction of the rectangular template.

[0117] The main reason for adopting this method is that the reference quantities have different value ranges, and if they are simply added, the meaning of the combination is lost. The method in this embodiment not only retains the order of each reference quantity in the grouping, but also takes into account the importance of each value in the extraction process. Figure 4 Taking a template as an example, we calculated the values ​​of its rotation from -30° to 30°. We selected 10 angles with smaller grayscale mean values, obtained the corresponding grayscale variance and binary segmentation image information, and then calculated their combined scores. We selected the 5 directions with the largest scores as candidates, calculated the difference between the grayscale mean values ​​of each template, and selected the smaller difference and rotation angle as the forward direction.

[0118] Step 2-3: Construct intersection recognition sub-model;

[0119] The intersection is identified based on POI points and POI lines. When the rectangular template moves, the extension direction of the intersection is determined after the POI point is found. When the direction of the POI line is consistent with the road direction, the POI line is used to sort the road information. The corresponding grayscale mean, grayscale variance and road information of the binary segmentation image are obtained respectively, and then the scores are sorted by scoring. The direction with smaller grayscale mean difference and rotation angle will be used as the candidate direction for intersection extension.

[0120] Since most of the intersections in this embodiment are four-way intersections, for the rectangular template, it is only necessary to determine the left and right sides of the traveling direction, that is, -50° to -130° and 50° to 130° of the template's traveling direction, such as Figure 5 (a) is shown. Although the POI lines are buffered, there are connection errors and inaccurate directions, so they need to be judged before application. When the line direction is consistent with the road direction, the POI lines are sorted according to the amount of road information covered, and the corresponding grayscale mean, grayscale variance and road information of the binary segmentation image are obtained, and then the scores are sorted. Similar to the road recognition sub-model, the direction with a smaller grayscale mean difference and rotation angle will be used as an alternative direction for intersection extension. When the POI line direction is inconsistent with the road direction, the road recognition method is used directly for judgment.

[0121] Step 2-4: Set up mobile strategy.

[0122] The mobile strategy is as follows:

[0123] The rectangular template is advanced according to the road recognition sub-model, each time by half the length of the template. For roads, there are generally no large turns or sudden angles. Taking advantage of this feature, this embodiment adds a criterion that the moving direction remains unchanged when the grayscale mean changes slightly. That is, if the difference between the grayscale mean of the current rectangular template and the grayscale mean after two advances is less than the threshold, the direction remains unchanged; otherwise, the road recognition method is used to determine the forward direction.

[0124] The stop movement strategy in this paper is divided into two cases: one is reaching the edge of the image; the other is reaching the road that has been extracted. Both cases require the judgment of the remaining intersections. If there are still intersections that have not been extracted, enter the next intersection for extraction; otherwise, terminate the program. In this process, it is necessary to determine whether the road corresponding to the intersection has been extracted. If it has been extracted, judge the next intersection until all intersections are identified; if it has not been extracted, directly extract the road.

[0125] The road network extraction model in this embodiment is specifically:

[0126] The initial template continues to advance and makes the following judgments:

[0127] If the extracted road is reached, the candidate intersection is determined;

[0128] When there is a candidate intersection and the corresponding road has not been extracted, the road recognition sub-model is used to extract the road;

[0129] Otherwise, the next intersection will be judged until the extraction requirements are met;

[0130] If there are no candidate intersections, or all intersections have been determined and do not meet the extraction requirements, the program is terminated;

[0131] If the extracted road has not been reached, it will continue to determine whether it has reached a road intersection; if it is determined to be a road intersection, it will be determined and stored according to the intersection recognition sub-model; otherwise, it will continue to determine whether it has reached the edge of the image; if it has reached the edge of the image, it will determine whether there is a candidate intersection waiting, if not, it will continue to determine the template direction;

[0132] The entire extraction process is the moving process of the rectangular template, and the judgment will be repeated each time until the stop movement strategy is met.

[0133] Here is a specific experimental case:

[0134] 1. Experimental Area and Dataset Selection

[0135] The experimental area was selected in City A because the urban expansion driven by the road network construction there is relatively fast. In the city center, the characteristics of the roads are: relatively narrow, covered with trees, and there are many high-rise buildings on both sides of the road. In the urban fringe areas, the roads are relatively wide and there are fewer trees. In general, the road structure of City A is relatively curved and complex, and it is difficult to extract road network information. This experiment selected three areas to test the road network extraction method, of which Experimental Area-I and Experimental Area-III are located in Area a, and Experimental Area-II is located in Area b. Figure 6As shown. GF-3 SAR image data, as shown in Table 1, is used as experimental data to extract road network information. In experimental area-I, four roads form a "tic-tac-toe" road network, which is a typical road network structure and is used to test the road network extraction model proposed in the present invention, namely the ATP-ROAD model. Experimental area-II and experimental area-III are used to verify the versatility of the ATP-ROAD model, among which experimental area-II has a main road with a large curvature and two branch roads on the same side; the road network structure of experimental area-III is irregular. Since the backscatter information of the SAR image road contains information about the surrounding objects, the difficulty of road extraction is increased. Testing and verification in these complex areas will promote better application of the model in other areas.

[0136] Table 1 Conditions of the three experimental areas (GF-3 SAR image with 1m resolution)

[0137] Famous City Location Imaging Mode Width (km) Polarization Range (pixels) Experimental Area-I Area A Slider bunching 10×10 Unipolarization 1643×1637 Experimental Area-Ⅱ Zone B Slider bunching 10×10 Unipolarization 911×1517 Experimental Area-III Area A Slider bunching 10×10 Unipolarization 875×1342

[0138] For POIs, they are mostly distributed on roads or on both sides of roads, and have obvious road relevance, which can provide assistance for road extraction. Generally speaking, the categories of POIs are relatively rich, and POI points of intersection names are selected as auxiliary information for road extraction. The POI data used in the experiment comes from the data released online by Amap in 2021. There are 23 POI classes, including road facilities, transportation services, place names, and addresses. The POI data of intersection names is a subclass of the place name and address information class. Only a few classes related to roads are shown in Table 2, which are used to illustrate the subordination of intersection name POI data, and refer to the POI classification code table of Amap.

[0139] Table 2 Subordinate relations of POI data related to roads

[0140]

[0141] 2. POI Data Processing

[0142] Corresponding to the experimental area-I, the POI range should be larger than the range of this area. This embodiment processes the POI data according to the POI processing and POI connection methods, and the buffer zone is 5 meters. Figure 7 The points in (a) are POI points after buffer processing and deduplication calculation, and the lines are POI connections after buffer processing. The connection lines shown in the figure are inaccurate, which also explains why the extension direction of the road cannot be determined entirely by the connection lines at the intersection. Figure 7 (b) shows the POI lines corresponding to experimental area-Ⅰ. Figure 7 (c) shows the POI points, both of which are used as input data in the experiment.

[0143] 3. Different types of road network extraction results

[0144] In order to check the implementation process of the method in this embodiment, experiments were also conducted at one intersection, two intersections, and three intersections, such as Figure 8 According to this method, the road network of the SAR image can be effectively extracted, and in each case, the roads extracted at the intersection are relatively complete.

[0145] 4. Experimental Results Analysis

[0146] Depend on Figure 8 It can be seen that the results extracted in this embodiment are basically consistent with the real road. Due to the different widths of the four roads, the results are represented as a single line. The shape of the line is relatively complete, and the intersection also conforms to the actual situation. In the experiment, 228 road points were automatically generated, and the points that deviated from the road were calculated as error points, with an accuracy rate of more than 70%, as shown in Table 3.

[0147] Table 3 Accuracy evaluation table of experimental area-I

[0148] Extraction method Manual extraction of points Automatically withdraw points Error Points Accuracy Manual extraction 159 - - - The present invention 3 228 65 71.49％

[0149] In the extraction results, there are three abnormal areas, corresponding to Fig. 9 In the region-A, region-B and region-C, this embodiment analyzes the three abnormal regions respectively.

[0150] Area-A is a road intersection. Normally, the lines connecting the left and right sides of this area should intersect at one point, but in the experimental results, they do not intersect at the same point. Tracing back the extraction process, it is found that when the rectangular template moves upward along R1 (the initial position of the template is on R1, Fig. 9 When the vehicle reaches this intersection (marked by a five-pointed star), only the extension direction of the road on the left is retained, and the extension direction on the right is discarded because the grayscale difference is greater than the threshold. The road on the right is extracted from the template starting from the road intersection in the upper right corner. Therefore, it is judged that this extracted road does not finally merge with the initially extracted left road. The template map of area-A also proves this situation. In fact, this problem can be connected by modifying the extraction strategy. However, after checking the optical image, it is found that the actual situation is that R4 does not strictly merge at this intersection. Moreover, such intersections are not uncommon in real life, so it is believed that this situation is not an extraction error, so the extraction strategy is not modified.

[0151] Region-B is a section of road in R4. Compared with other extracted regions, the road extraction results of this section are relatively curved. After checking the optical image, it is found that there are high-rise buildings in this section to form shadows. In the SAR image, this section is basically mixed with the road. When the rectangular template enters this section, it will be affected by the shadow area, so the phenomenon of region-B appears. This phenomenon is difficult to avoid. From the template image, it can be seen that the rectangular template is adjusted when passing through this area, so that the subsequent roads are not affected by this area.

[0152] Area-C is near the road intersection in the upper right corner of the image. The extraction result of this area has a protruding point, which makes the line look not smooth enough. The cause of the problem was traced back to when the template moved from bottom to top along R3, the direction judged at the intersection here was the protruding direction in the figure and was stored. When this direction entered the road extraction as a candidate direction, according to the method in this embodiment, the template was quickly adjusted (the initial direction was changed), so the protruding point problem of area-C appeared. There are two reasons for this problem. One is that the rotation of the rectangular template itself is not flexible enough; the other is that the position of the POI point is inaccurate, which affects the timing of judging the extension direction of the intersection. Since this phenomenon is difficult to avoid, the extraction method will continue to be improved in the future to improve this situation.

[0153] V. Experimental results of other experimental areas

[0154] In addition to the experiment in Experimental Area-I, this embodiment also applies the ATP-ROAD method in two other areas, such as Fig.10 Compared with experimental area-I, the road structures of these two areas are relatively complex, but the image range is smaller and the road surface is clearer, so the extraction accuracy is higher, as shown in Tables 4 and 5.

[0155] Table 4 Accuracy evaluation table of experimental area-II

[0156] Extraction method Manual extraction of points Automatically withdraw points Error Points Accuracy Manual extraction 78 - - - The present invention 3 95 10 89.47％

[0157] Table 5 Accuracy evaluation table of experimental area-III

[0158] Extraction method Manual extraction of points Automatically withdraw points Error Points Accuracy Manual extraction 110 - - - The present invention 3 141 25 82.27％

[0159] The application of POI data is a new attempt at a road extraction method, and it will encounter many problems. For example, the problem of POI data redundancy. This embodiment uses a filter-like method to remove duplicate data. POI lines can not only identify road intersections, but also assist in determining the extension direction of the road to a certain extent. By using these POI data, the accuracy of the method in this embodiment for identifying intersections is guaranteed.

[0160] In the process of road extraction, it is difficult to maintain the integrity of the road network structure. The present invention fully combines the SAR image angle texture features and POI information, effectively extracting roads while ensuring the integrity of the road network structure. The present invention combines the grayscale mean, grayscale variance and binary segmentation information of the SAR image in a scoring manner, solving the problem of inconsistent measurement scales between different information. The experimental results of three regions show that this method can be used to extract different types of road networks.

[0161] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for extracting road networks from SAR images considering angle texture features and POI, characterized in that: The road network extraction method comprises: Step 1: Obtain input data, including SAR image data, binary segmentation data, POI point data, and POI connection data; Step 2: Construct a road network extraction model considering angle texture features and POI; Step 3: Input the input data in step 1 into the road network extraction model constructed in step 2 to obtain the road network extraction result; The method for obtaining the POI point data is as follows: The POI class of the intersection name is used. Within the POI range, a rectangular box is used for movement calculation. The number of points in the rectangular box is counted. If the number is greater than 1, the center position of these points will be calculated, and only the point closest to the center position will be retained, and the other points will be deleted; if the number of points is 2, one of them will be retained; The calculation formula is: Among them, (c x ,c y ) is the coordinate of the center of the rectangular frame; (x Pi ,y Pi ) are the coordinates of each point in the rectangular frame; d i is the distance from each point to the center; the value range of i is 1 to n, n>2; The step 2 is specifically as follows: Step 2-1: Create an initial template; Step 2-2: Construct a road recognition sub-model; Step 2-3: Construct intersection recognition sub-model; Step 2-4: Set up mobile strategy; The intersection recognition sub-model is specifically: The intersection is identified based on POI points and POI lines. When the rectangular template moves, the extension direction of the intersection is determined after the POI point is found. When the direction of the POI line is consistent with the road direction, the POI line is used to sort the road information. The corresponding grayscale mean, grayscale variance and road information of the binary segmentation image are obtained respectively, and then the scores are sorted by scoring. The direction with smaller grayscale mean difference and rotation angle will be used as the candidate direction for intersection extension. When the POI connection direction is inconsistent with the road direction, the road recognition sub-model is directly used for judgment.

2. The method for extracting road network from SAR images considering angle texture features and POI according to claim 1, characterized in that: The method for acquiring the SAR image data is as follows: Geometric correction of SAR images using orbit files and corresponding DEM; The method for obtaining the binary segmentation data is: The corrected SAR image is binarized using the threshold segmentation method. In the binary segmented image, the road, water, and shadow areas are divided into black areas with a pixel value of 0, while other areas are divided into white areas with a pixel value of 1.

3. The method for extracting road network from SAR images considering angle texture features and POI according to claim 1, characterized in that: The initial template in step 2-1 is a rectangular template.

4. The method for extracting road network from SAR images considering angle texture features and POI according to claim 1, characterized in that: The road recognition sub-model is specifically: Define the rectangle template as: Template=(α,M,V,S) Among them, α is the rotation angle of the rectangular template; M is the grayscale mean corresponding to α; V is the grayscale variance corresponding to α; S is the road information in the binary image corresponding to α; When the rectangular template is rotated, the corresponding ones are: Sort M in ascending order, record the angles corresponding to the first several values, extract V and S corresponding to these angles and sort them in ascending order respectively; score several values ​​of M, V and S, with the highest score being 1 and the remaining values ​​decreasing by 0.1; after obtaining three sets of scores, combine the three using the following formula and sort the new array in descending order: Among them, δ, γ and ω are coefficients, ranging from 0 to 1; A is the total score after combining the three types of information, and A′ is the result of sorting A in descending order; After sorting, select several directions with the highest total scores as candidate directions, calculate the difference in the grayscale mean between the current rectangular template and the next rectangular template, and record the corresponding angle; The scoring method is also used to select the direction with smaller difference and smaller rotation angle as the advancing direction of the rectangular template.

5. The method for extracting road network from SAR images considering angle texture features and POI according to claim 1, characterized in that: The mobile strategy is specifically as follows: Move the rectangular template forward according to the road recognition sub-model, advancing half the template length each time.

6. The method for extracting road network from SAR images considering angle texture features and POI according to claim 1, characterized in that: The steps 2-4 also include: Set the stop movement policy. The movement will be stopped in any of the following situations: Reach the edge of the image; Arrived at the extracted road.

7. The method for extracting road network from SAR images considering angle texture features and POI according to claim 1, characterized in that: The road network extraction model is specifically as follows: The initial template continues to advance and makes the following judgments: If the extracted road is reached, the candidate intersection is determined; When there is a candidate intersection and the corresponding road has not been extracted, the road recognition sub-model is used to extract the road; Otherwise, the next intersection will be judged until the extraction requirements are met; If there are no candidate intersections, or all intersections have been determined and do not meet the extraction requirements, the program is terminated; If the extracted road has not been reached, it will continue to determine whether it has reached a road intersection; if it is determined to be a road intersection, it will be determined and stored according to the intersection recognition sub-model; otherwise, it will continue to determine whether it has reached the edge of the image; If it reaches the edge of the image, it determines whether there is a candidate intersection waiting. If not, it continues to determine the template direction; The entire extraction process is the moving process of the rectangular template, and the judgment will be repeated each time until the stop movement strategy is met.

Images