A method for identifying offshore wind turbines based on multi-source remote sensing images

Abstract

The application discloses a kind of off-shore wind turbine identification methods based on multi-source remote sensing image, it is related to remote sensing image recognition technique, effectively improve the identification accuracy of off-shore wind turbine.The application is divided into tile by global coastal exclusive economic zone, and then the sentinel1 image in each tile range is generated, a plurality of connected domains are set in the sentinel1 image and the candidate plot is obtained, and then the sentinel2 image slice at candidate plot is generated, the plot classification is carried out to all sentinel2 image slices by classification algorithm, the time-back scattering signal curve is generated according to the classification result, the construction time of off-shore wind turbine is obtained by convolution calculation on time-back scattering signal curve, and then the wgs84 coordinates of each tile and plot are further generated, and then the information of off-shore wind turbine is stored.

Description

TECHNICAL FIELD

[0001] The present application relates to remote sensing image recognition technology, in particular to a method for identifying offshore wind turbines based on multi-source remote sensing images. BACKGROUND

[0002] Offshore wind farms are composed of clusters or arrays of wind turbines and are widely recognized as a renewable energy source, which is an effective way to reduce greenhouse gas emissions and promote a net-zero carbon economy. In recent years, the proportion of offshore wind farms in the global renewable energy market and the proportion of global installed capacity have increased significantly. In order to reduce construction and maintenance costs, most offshore wind farms are located near the coast.

[0003] Accurately grasping the distribution of existing offshore wind turbines is a prerequisite for spatial planning, environmental assessment, etc. of offshore wind farms. Further, long-term, real-time identification of offshore wind farms and their distribution is of great significance for construction progress detection, quality control, environmental monitoring, and safety management of wind farms under construction.

[0004] In the prior art, SAR satellite images are the main source of information for identifying offshore wind turbines. SAR satellite images can stably obtain radar backscatter data without being affected by day and night and cloud cover. Moreover, due to the dihedral structure of offshore wind turbines, their VV polarized radar backscatter signals are significantly different from the background, and have a higher signal-to-noise ratio. In general, under simple background conditions, global or local threshold methods can easily identify offshore wind turbines. However, for areas with sandbars, small islands, and significant sea clutter, confusion is likely to occur, resulting in low identification accuracy. Moreover, for masts, substations, signal towers, and other similar radar signals that often appear near offshore wind farms, neither traditional threshold segmentation identification methods nor deep learning segmentation identification methods can accurately identify them.

[0005] In the prior art, another technical route is to use optical satellite images to identify offshore wind turbines. However, the obvious disadvantages are as follows: first, optical images are often affected by clouds and mist in coastal areas, making it difficult to identify offshore wind turbines; second, freely available optical satellite images have limited resolution, and typical offshore wind turbines only occupy a limited number of pixels in the image, making it difficult to accurately identify them; if high-resolution commercial images are used to identify offshore wind turbines globally, the cost is extremely high.

[0006] Due to the above factors, existing international offshore wind farm data has obvious wind turbine number omissions and wind turbine location recording errors. So far, there is no global offshore wind turbine dataset or service with low cost, high identification accuracy, and near-real-time updates in the public domain. SUMMARY

[0007] To solve the above technical problems, the purpose of the present application is to provide an offshore wind turbine identification method based on multi-source remote sensing images.

[0008] In order to achieve the above purpose, the present application provides the following technical solutions:

[0009] An offshore wind turbine identification method based on multi-source remote sensing images, comprising the following steps:

[0010] S1, data preparation, including obtaining the vector range of global coastal exclusive economic zone and carrying out tile division;

[0011] S1.1, obtaining the vector range of global coastal exclusive economic zone;

[0012] S1.2, tile division is carried out to the global coastal exclusive economic zone, and then a plurality of tiles are generated;

[0013] S2, generating sentinel2 image slice at candidate graph spot in each tile range;

[0014] S2.1, generating preprocessed sentinel1 image in each tile range, and using minimum coverage number method to obtain the minimum value of the number of preprocessed sentinel1 images in each time range;

[0015] S2.2, performing set operation on the preprocessed sentinel1 image and the global coastal exclusive economic zone, and then generating economic zone filtering image;

[0016] S2.3, using connected domain algorithm to obtain all connected domains in the economic zone filtering image and offshore wind turbine mask, and then obtaining the candidate graph spot of each connected domain;

[0017] S2.4, generating processed sentinel2 image;

[0018] S2.5, generating sentinel2 image slice at candidate graph spot according to the candidate graph spot of each connected domain and the processed sentinel2 image;

[0019] S3, offshore wind turbine information extraction, the offshore wind turbine information includes: the offshore wind turbine longitude and latitude coordinates, construction time, identification confidence;

[0020] S3.1, using classification algorithm, carrying out graph spot classification on offshore wind turbine mask;

[0021] S3.2, using clustering algorithm, carrying out clustering on the graph spot in the offshore wind turbine mask filtered after the graph spot classification;

[0022] S3.3, confirming the satellite image start providing image service time and query cut-off time, and then obtaining the processed sentinel1 image corresponding to the time range intersected with each tile range, and then generating the pre-processed sentinel1 image set, traversing all the polygons in the offshore wind turbine mask filtered by polygon clustering, and then generating the time-backscatter signal curve, and performing convolution calculation on the time-backscatter signal curve to obtain the construction time of each offshore wind turbine;

[0023] S3.4, obtaining the wgs84 coordinates corresponding to each tile and its corresponding polygon, and then storing each information of the corresponding offshore wind turbine.

[0024] Further, the generation process of the pre-processed sentinel1 image includes:

[0025] Each tile is numbered i, where i=1, 2, 3, …, k, k is a natural number greater than 0, and the latest N period IW mode sentinel1 VV polarization Ground Range Detected(GRD) image in the range of the tile i is downloaded, and the IW mode sentinel1 VV polarization Ground Range Detected(GRD) image is radiometrically calibrated, geometrically corrected, terrain corrected, and dB processed; secondly, the N period pre-processed sentinel1 image is superimposed and averaged to generate a processed sentinel1 image, where N is a natural number greater than 0.

[0026] Further, the process of calculating the processed sentinel1 image by the minimum coverage number method includes:

[0027] Determine the range of the tile i, the satellite image start providing image service time t1, and the query cut-off time t2, where t2>t1; secondly, query all the pre-processed sentinel1 images intersected with the range of the tile i in the t1 to t2 time range;

[0028] Further, from the t2 time, calculate the intermediate time t3, where t2>t3>t1, so that all the pre-processed sentinel1 images in the t3 to t2 time range can completely cover the range of the tile i; finally, the number of all the pre-processed sentinel1 images in the t3 to t2 time range is the minimum value of N.

[0029] Further, the generation process of the candidate polygon includes:

[0030] The pre-processed sentinel1 image is intersected with the global coastal exclusive economic zone to generate an intersection region, and the pixel values of the pre-processed sentinel1 image outside the intersection region are set to zero to generate an economic zone filtered image;

[0031] All connected domains in the economic zone filtered image are obtained using a connected domain algorithm, and an offshore wind turbine mask is generated, the offshore wind turbine mask being a binary image with the same length and width as the economic zone filtered image but with a channel of 1 and a gray value of 0 or 255, the connected domain being a connected pixel with a gray value of 255 in the offshore wind turbine mask, and the connected domain algorithm using an 8-neighbor connected algorithm.

[0032] The center pixel coordinates, minimum bounding rectangle width, height, and pixel area of all connected domains are obtained, and then the threshold value W of the minimum bounding rectangle width, the threshold value H of the minimum bounding rectangle height, and the connected domain pixel area threshold value A are determined. The gray value of the connected domain in the offshore wind turbine mask is set to zero if the minimum bounding rectangle width is greater than W, or the minimum bounding rectangle height is greater than H, or the connected domain pixel area is greater than A, and finally a graph shape filtered offshore wind turbine mask is generated, wherein each connected domain is a candidate graph.

[0033] Further, the generation process of the processed sentinel2 image includes:

[0034] First, in the range of tile i, the latest M period sentinel2 visible band image is downloaded, and after the pre-processing procedure, the M period corrected sentinel2 image is generated.

[0035] The pre-processing procedure includes: geometric correction and atmospheric correction of the original sentinel2 image; secondly, cloud and shadow identification is performed on the M period corrected sentinel2 image, and the cloud and shadow area is recorded as a transparent area to generate an M period cloud-removed sentinel2 image; finally, the latest image in the M period cloud-removed sentinel2 image is taken as the topmost layer, and the remaining M period cloud-removed sentinel2 images are used in descending order of time, and the transparent area is filled using a recursive method, wherein M is greater than the number of images required by the recursive method, and finally a processed sentinel2 image is generated.

[0036] Further, the generation process of the candidate graph sentinel2 image slice includes:

[0037] Querying the center pixel coordinates A of all candidate polygons, calculating the corresponding coordinates B in the processed Sentinel2 image for each coordinate A, and using the cropping method to crop a square image slice from the processed Sentinel2 image with the coordinates B as the center;

[0038] If the square image slice is completely contained in the processed Sentinel2 image, directly crop the square region;

[0039] If the square image slice is not completely contained in the processed Sentinel2 image, crop the containing part from the processed Sentinel2 image, and set the remaining pixels to zero; finally, generate a corresponding Sentinel2 image slice for each candidate polygon.

[0040] Further, the process of polygon classification and clustering for all Sentinel2 image slices includes:

[0041] Using a classification algorithm to classify all Sentinel2 image slices, the classification algorithm used in the classification process is a binary classification algorithm or a multi-classification algorithm, the output of the classification algorithm is a one-dimensional vector, the length of the vector is the number of target classes, and the value of the vector is the probability value of the ground object contained in the Sentinel2 image slice belonging to the target class, and the target class corresponding to the maximum value of the vector is the classification result of the Sentinel2 image slice;

[0042] In the offshore wind turbine mask after the polygon shape filtering, querying the candidate polygon C, the classification result of the Sentinel2 image slice corresponding to the position of the candidate polygon C is not an offshore wind turbine, and removing the polygon C on the basis of the polygon shape filtering offshore wind turbine mask in S2.3 to obtain the polygon classification filtering offshore wind turbine mask;

[0043] Using a clustering algorithm to cluster the polygons in the polygon classification filtering offshore wind turbine mask into a cluster with a number not less than 1, and marking the cluster D, the cluster D only contains one polygon, and finally removing all polygons contained in the cluster D from the polygon classification filtering offshore wind turbine mask to obtain the polygon clustering filtering offshore wind turbine mask.

[0044] Further, the generation process of the time-backscatter signal curve includes:

[0045] Determine the satellite image starting time t1 of providing image service and the query cut-off time t2 in the tile i range, which is generally the data analysis day, query and download the pre-processed sentinel1 image intersecting with the tile range in the t1-t2 time range, and generate the corresponding pre-processed sentinel1 image set, and the pre-processed sentinel1 images in the pre-processed sentinel1 image set are arranged in descending order of shooting time;

[0046] Traverse all the polygons in the offshore wind turbine mask after the polygon clustering filtering, and set a number j for each polygon, wherein j=1, 2, …, a, and a is a natural number greater than 0;

[0047] Generate a time-backscattering signal curve for the polygon j, and the time-backscattering signal curve is obtained by the following steps: first, determine the pixel range PAj of the polygon j, and second, sequentially obtain the maximum value Vj in the gray scale range of PAj in the pre-processed sentinel1 image set, and the shooting time of the pre-processed sentinel1 image is tj, to obtain the tj-Vj curve belonging to the polygon j, and the convolution calculation is performed on the tj-Vj curve, and the convolution kernel satisfies the following conditions:

[0048]

[0049] wherein, is the convolution kernel window length, ; in the embodiment, the value of k is taken as 1 , , and the time corresponding to the maximum value k in the result is taken as the construction time of the offshore wind turbine.

[0050] Further, the wgs84 coordinates TLi corresponding to the upper left corner of the tile i are obtained, then the wgs84 coordinate offset TLij of the center pixel coordinates of the polygon j relative to TLi is obtained, then the absolute wgs84 coordinates of the polygon j are calculated, and finally, the information of the offshore wind turbine is stored after traversing i and j.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] (1) The present application generates a processed sentinel1 image by superimposing and averaging N pre-processed sentinel1 images, and determines the most suitable N value by using the minimum coverage number method, which can ensure that the near real-time offshore wind turbine signal with strong timeliness is captured, and the timeliness of offshore wind turbine identification is greatly improved compared with the prior art. ​

[0053] (2) In the present application, in the case of focusing on real-time recognition effect, the signal-to-noise ratio of the processed sentinel1 image is low, a multi-stage filtering step is introduced, including economic zone mask filtering, patch shape filtering, etc., which effectively filters noise signals, and further, multiple source remote sensing images are used comprehensively, first, the sentinel1 image is used to filter most of the confusing ground objects, and the candidate results are obtained, and then combined with the sentinel2 optical image in the candidate area, the classification algorithm and clustering algorithm are used to efficiently filter small islands, masts, substations, signal towers, ships and other ground objects that cannot be recognized by sentinel1 image, thereby greatly improving the recognition accuracy compared with the prior art;

[0054] (3) In the present application, the sentinel1 image and sentinel2 image used are open source satellite images, and have high temporal resolution and spatial resolution, and the use of the method for identifying offshore wind turbines has low cost and high practicability. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application.

[0056] Figure 1 is a flowchart of the present application;

[0057] Figure 2 is a processed sentinel1 image with N value of 5;

[0058] Figure 3 is a processed sentinel1 image with N value of 30;

[0059] Figure 4 is an experimental result graph of the classification recognition algorithm;

[0060] Figure 5 is an experimental result graph of the classification recognition algorithm superimposed on the sentinel2 base map;

[0061] Figure 6 is a mask graph of offshore wind turbines after patch classification filtering;

[0062] Figure 7 is a mask graph of offshore wind turbines after patch clustering filtering;

[0063] Figure 8 is a time-maximum gray value curve corresponding to the patch;

[0064] Figure 9 is Figure 8The result map after convolution calculation. DETAILED DESCRIPTION

[0065] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described in detail below. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the scope of protection of the present application.

[0066] As shown in the drawings, a method for identifying offshore wind turbines based on multi-source remote sensing images comprises the following steps: Figure 1

[0067] S1, data preparation, including obtaining the vector range of global coastal exclusive economic zone and performing tile division;

[0068] S1.1, obtaining the vector range of global coastal exclusive economic zone, offshore wind turbines only exist in the range of the coastal exclusive economic zone, and the global coastal exclusive economic zone adopts wgs84 coordinates and uses direct projection of longitude and latitude;

[0069] S1.2, dividing the coastal exclusive economic zone into tiles, in the embodiment, the tile size is 50KM*50KM, the coastal exclusive economic zone is contained in the set of all the tiles, and the tile division of part of the coastal exclusive economic zone is shown, and each tile is numbered i, wherein i=1, 2, 3, …, k, k is a natural number greater than 0.

[0070] S2, in the range of each tile i in S1.2, a pre-processed sentinel1 image is generated, and a sentinel2 image slice is selected according to the pre-processed sentinel1 image candidate plot, in the embodiment, the sentinel2 image resolution is 10 meters, and the slice size is 64 pixels*64 pixels;

[0071] S2.1, generating a pre-processed sentinel1 image, the specific steps are as follows: first, in the range of the tile i, downloading the latest N period IW mode sentinel1 VV polarization Ground Range Detected (GRD) image in time, and generating an N period pre-processed sentinel1 image after pre-processing;

[0072] ​The preprocessing procedure comprises: radiometric calibration, geometric correction, terrain correction, and decibel processing on an IW mode sentinel1 VV polarization Ground Range Detected (GRD) image; secondly, a preprocessed sentinel1 image is generated by stacking and averaging N period preprocessed sentinel1 images, wherein N is a natural number greater than 0;

[0073] Wherein, the greater the value of N, the more the level of radar image inherent random noise and moving target interference signals can be suppressed, and the higher the signal-to-noise ratio of the preprocessed sentinel1 image, but the smaller the value of N, the more conducive to identifying newly built offshore wind turbines in the short term, so an optimal N value needs to be determined for each tile, and in the case of focusing on real-time identification effect, the signal-to-noise ratio of the preprocessed sentinel1 image is improved as much as possible;

[0074] The minimum coverage number method is used to determine the most suitable N value, and the minimum coverage number method comprises the following steps: first, determining the range of the tile i, the satellite image service start time t1, and the query end time t2, wherein t2>t1; secondly, querying all the preprocessed sentinel1 images that have an intersection with the range of the tile i within the time range from t1 to t2;

[0075] Further, the intermediate time t3 is calculated from the time t2, wherein t2>t3>t1, so that all the preprocessed sentinel1 images within the time range from t3 to t2 can completely cover the range of the tile i; finally, the number of all the preprocessed sentinel1 images within the time range from t3 to t2 is the minimum value of N;

[0076] It should be noted that, Figure 3 The tile range marked by X contains a newly built offshore wind farm, which is located in the upper left part of the tile. After manual interpretation using cloud-free sentinel2 images, it is determined that the construction start time of the offshore wind turbine in this area is about December 1, 2023, and the end time is about March 10, 2024. The query end time is set to March 20, 2024. The N value determined by the minimum coverage number method for this place is 5, i.e. the number of all preprocessed sentinel1 images for about 2 months. The corresponding processed sentinel1 image is shown in Figure 2 It can be seen that the newly built offshore wind turbine in the upper left corner is relatively clear, but the level of random noise in the image background is high, and the interference signal is obvious, which makes the subsequent offshore wind turbine identification prone to errors and omissions; the specified N value is 30, i.e. the number of all preprocessed sentinel1 images for nearly 1 year. The corresponding processed sentinel1 image is shown in Figure 3As shown, the background random noise level is low, but the original information of the newly built offshore wind turbine in the upper left corner has been basically lost, and it is difficult to make up for it later using technical means;

[0077] S2.2, economic zone mask filtering, the specific process is: the pre-processed sentinel1 image is intersected with the global coastal special economic zone to generate an intersection region, the pixel values of the pre-processed sentinel1 image outside the intersection region are set to zero, and an economic zone filtered image is generated;

[0078] S2.3, shape filtering, specifically: first, using a connected domain algorithm, all connected domains in the economic zone filtered image in S2.2 are obtained, and an offshore wind turbine mask is generated, the offshore wind turbine mask is a binary image with the same length and width as the economic zone filtered image, but the channel is 1 and the gray value is 0 or 255, the connected domain is a connected pixel with a gray value of 255 in the offshore wind turbine mask, and the connected domain algorithm uses an 8-neighbor connected algorithm;

[0079] Secondly, the center pixel coordinates, the minimum bounding rectangle width, the minimum bounding rectangle height and the pixel area of all the connected domains are calculated; then, the threshold value W of the minimum bounding rectangle width, the threshold value H of the minimum bounding rectangle height and the connected domain pixel area threshold value A are determined, the gray value of the connected domain in the offshore wind turbine mask is set to zero if the minimum bounding rectangle width is greater than W, or the minimum bounding rectangle height is greater than H, or the connected domain pixel area is greater than A, and finally a shape filtered offshore wind turbine mask is generated, wherein each connected domain is a candidate patch in S2;

[0080] S2.4, generating a processed sentinel2 image, the specific steps are: first, in the range of tile i in S2.1, download the latest M period sentinel2 visible light band image, and generate an M period corrected sentinel2 image after preprocessing;

[0081] The preprocessing process includes: geometric correction and atmospheric correction of the original sentinel2 image, in this embodiment, the sentinel2 visible light band image is a 3-channel color image with a resolution of 10 meters; secondly, the M period corrected sentinel2 image is subjected to cloud and shadow identification, and the cloud and shadow area is recorded as a transparent area to generate an M period cloud-removed sentinel2 image; finally, the latest image in the M period cloud-removed sentinel2 image is taken as the topmost layer, and the remaining M period cloud-removed sentinel2 images are used in descending order of time, and the transparent area is filled using a recursive method, wherein M is greater than the number of images required by the recursive method, and finally a processed sentinel2 image is generated;

[0082] S2.5, generating a sentinel2 image slice at the candidate graph spot, the specific steps are: first, querying the center pixel coordinates A of all candidate graph spots in S2.3, calculating the coordinates B corresponding to the sentinel2 image in S2.4 after processing for each coordinate A, using the clipping method to clip the square image slice from the processed sentinel2 image with the coordinates B as the center;

[0083] The clipping method is: if the square image slice is completely contained in the processed sentinel2 image, then directly clip the square region;

[0084] If the square image slice is not completely contained in the processed sentinel2 image, then clip the contained part from the processed sentinel2 image, and set the pixel values of the remaining non-contained part to zero; finally, a corresponding sentinel2 image slice is generated at each candidate graph spot in S2.3.

[0085] S3, offshore wind turbine information extraction, obtaining the offshore wind turbine information including the offshore wind turbine latitude and longitude coordinates, construction time and recognition confidence;

[0086] S3.1, using a classification algorithm to classify all the sentinel2 image slices in S2.5, the classification algorithm used in the classification process is a binary classification algorithm or a multi-classification algorithm, the target classes of the binary classification algorithm are offshore wind turbines and background respectively, and the target classes of the multi-classification algorithm include but are not limited to small islands, masts, substations, signal towers, ships, and high-frequency interference ground objects such as broken clouds; the output of the classification algorithm is a one-dimensional vector, the length of the vector is the number of target classes, and the value of the vector is the probability value of the ground object contained in the sentinel2 image slice belonging to the target class corresponding to the maximum value of the vector, which is the classification result of the sentinel2 image slice;

[0087] In the offshore wind turbine mask after the graph shape filtering in S2.3, querying the candidate graph spot C, the classification result of the sentinel2 image slice corresponding to the position of the candidate graph spot C is not an offshore wind turbine, removing the graph spot C based on the offshore wind turbine mask after the graph shape filtering in S2.3 to obtain the offshore wind turbine mask after the graph classification filtering;

[0088] In this embodiment, the classification algorithm is a YOLO binary classification algorithm, and the experimental results of classifying and identifying Figure 3 Figure 4 ​The white bright spot area is the area identified by the YOLO two-classification algorithm as an offshore wind turbine; as Figure 5 Further, the display area is focused on the newly built offshore wind farm area described in S2.1, and the area is zoomed in to more clearly show the offshore wind turbine. The bottom map is the sentinel2 optical image on March 10, when the wind farm was completed. The area circled by the red ring corresponds to the white bright spot area. It can be seen that the offshore wind turbine area identified by the YOLO two-classification algorithm is completely consistent with the true value, proving that the YOLO two-classification algorithm can perform high-quality offshore wind turbine identification in a near-real-time situation that maintains a 2-month time efficiency;

[0089] Further, the embodiment uses the YOLO two-classification algorithm to identify offshore wind turbines in all nearshore areas in China. After combining with manual interpretation, the experimental results of some typical nearshore areas are shown in Table 1. The accuracy is the number of correct identifications divided by the sum of correct and incorrect identifications, and the recall rate is the number of correct identifications divided by the total number of true values. As can be seen from the table, the accuracy and recall rate of the two-classification algorithm used in the embodiment are both better than 0.98, which is significantly better than the data set in the prior art. The significant improvement in this indicator is mainly because the two-classification algorithm based on sentinel2 visible light images can well distinguish offshore wind turbines from small islands, masts, substations, signal towers, and ships, while the prior art basically only uses sentinel1 image data and uses traditional threshold methods for segmentation, which cannot distinguish offshore wind turbines from small islands, masts, substations, signal towers, and small ships, resulting in many incorrect identifications;

[0090] Table 1:

[0091] Region Total number of true values Number of correct recognitions Number of incorrect recognitions Number of missed recognitions Accuracy rate Recall rate Northern Yellow Sea nearshore region 203 200 1 3 0.995 0.985 Hangzhou Bay nearshore region 487 480 0 7 1.000 0.986 Northern South China Sea nearshore region 905 887 4 18 0.996 0.980 Eastern Taiwan Strait nearshore region 325 323 2 2 0.994 0.994

[0092] S3.2 Using a clustering algorithm, the clusters of the graph patches in the offshore wind turbine mask after the graph patch classification filtering described in S3.1 are clustered into clusters with a number not less than 1, and the clusters D are marked. The cluster D only contains one graph patch. Finally, from the graph patch classification filtering offshore wind turbine mask described in S3.1, all graph patches contained in the cluster D are removed to obtain the graph patch clustering filtering offshore wind turbine mask.

[0093] In this embodiment, the clustering algorithm is DBSCAN clustering algorithm, the "neighborhood range" in the algorithm parameters is 7D pixel number, the 7D pixel number is the pixel number corresponding to 7 times the length of the offshore wind turbine rotor diameter in the offshore wind turbine mask after the patch clustering filtering; the "minimum sample size" in the algorithm parameters is 1, that is, the patch is allowed to form the cluster not less than 1; taking a wind farm near Guangzhou as an example, the offshore wind turbine mask after the patch classification filtering is as shown in Figure 6 , and the offshore wind turbine mask after the patch clustering filtering is as shown in Figure 7 , it can be seen that the DBSCAN clustering algorithm can further improve the accuracy of offshore wind turbine identification;

[0094] S3.3, construction time extraction, specifically including:

[0095] S3.3.1, determining the satellite image start providing image service time t1 and the query cutoff time t2 within the tile i range in S2.1, which is generally the data analysis day;

[0096] S3.3.2, querying and downloading the pre-processed sentinel1 image in S1.2 which has intersection with the tile range within the t1 to t2 time range, and generating the corresponding pre-processed sentinel1 image set, the pre-processed sentinel1 image in the pre-processed sentinel1 image set is arranged in descending order according to the shooting time;

[0097] S3.3.3, traversing all patches in the offshore wind turbine mask after patch clustering filtering in S3.2, and setting a number j for each patch, where j=1, 2, …, a, a is a natural number greater than 0;

[0098] generating a time-backscattering signal curve for patch j, the time-backscattering signal curve is obtained according to the following steps: first, determining the pixel range PAj of the patch j, second, sequentially obtaining the maximum value Vj in the PAj range in the pre-processed sentinel1 image set, the corresponding shooting time of the pre-processed sentinel1 image is tj, and the tj-Vj curve belonging to the patch j is obtained, as shown in Figure 8 ;

[0099] S3.3.4, performing convolution calculation on the tj-Vj curve, the convolution calculation uses a discrete convolution kernel , which satisfies the following conditions:

[0100]

[0101] wherein, is the convolution kernel window length, In this embodiment, take , , the convolution calculation results as shown in Figure 9 , the corresponding time , that is, the construction time of the offshore wind turbine corresponding to the image spot j;

[0102] S3.4, coordinate extraction, specifically: first, calculate the wgs84 coordinates TLi corresponding to the upper left corner of the tile i in S2.1, second, calculate the wgs84 coordinate offset TLij of the center pixel coordinates of the image spot j relative to TLi in S3.3, then, the absolute wgs84 coordinates of the image spot j are calculated, finally, the position information, construction time and other information corresponding to the offshore wind turbine are stored after traversing i, j.

[0103] The above embodiments are only used to illustrate the technical method of the present application and are not limited. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A method for identifying offshore wind turbines based on multi-source remote sensing images, characterized in that, The method comprises the following steps: S1, data preparation, including obtaining the vector range of global coastal exclusive economic zone and tile division; S1.1, obtaining the vector range of global coastal exclusive economic zone; S1.2, tile division is carried out on the global coastal exclusive economic zone, thereby generating a plurality of tiles; S2, generating a candidate plot sentinel2 image slice at each tile range; S2.1, generating a preprocessed sentinel1 image in each tile range, and obtaining the minimum value of the number of preprocessed sentinel1 images in each time range by using the minimum coverage number method; S2.2, the preprocessed sentinel1 image is operated with the global coastal exclusive economic zone, thereby generating an economic zone filtered image; S2.3, using a connected domain algorithm to obtain all connected domains in the economic zone filtered image and an offshore wind turbine mask, thereby obtaining a candidate plot for each connected domain; S2.4, generating a processed sentinel2 image; S2.5, generating a candidate plot sentinel2 image slice according to the candidate plot of each connected domain and the processed sentinel2 image; S3, offshore wind turbine information extraction, obtaining the offshore wind turbine information including offshore wind turbine latitude and longitude coordinates, construction time and recognition confidence; S3.1, using a classification algorithm, performing plot classification on all offshore wind turbine masks; S3.2, using a clustering algorithm, clustering the plots in the offshore wind turbine mask filtered by plot classification; S3.3, confirming the satellite image service start time and query cutoff time, and obtaining the processed sentinel1 image in the corresponding time range intersecting with each tile range, thereby generating a preprocessed sentinel1 image set, traversing all plots in the offshore wind turbine mask filtered by plot clustering, thereby generating a time-backscatter signal curve, and performing convolution calculation on the time-backscatter signal curve to obtain the construction time of each offshore wind turbine; S3.4, obtaining the wgs84 coordinates corresponding to each tile and its corresponding plot, thereby storing each information of the corresponding offshore wind turbine. 2.The method of claim 1, wherein, The generation process of the preprocessed sentinel1 image comprises: Each tile is numbered i, wherein i=1, 2, 3, …, k, k is a natural number greater than 0, the latest N period IW mode sentinel1 VV polarization Ground Range Detected image in the range of the tile i is downloaded, the IW mode sentinel1 VV polarization Ground Range Detected image is radiometrically calibrated, geometrically corrected, terrain corrected and decibel processed; secondly, a processed sentinel1 image is generated by superimposing and averaging N period preprocessed sentinel1 images, wherein N is a natural number greater than 0. 3.The method of claim 2, wherein, The process of minimum coverage number method calculation on the processed sentinel1 image comprises: determining the range of the tile i, the satellite image starting image service time t1 and the query cut-off time t2, where t2>t1; secondly, querying all the pre-processed sentinel1 images that have an intersection with the range of the tile i within the time range from t1 to t2; further obtaining an intermediate time t3 from the time t2, where t2>t3>t1, so that all the pre-processed sentinel1 images within the time range from t3 to t2 can completely cover the range of the tile i; finally, the number of all the pre-processed sentinel1 images within the time range from t3 to t2 is the minimum value of N. 4.The method of claim 3, wherein, The generation process of the candidate graph patches includes: performing a union operation on the pre-processed sentinel1 image and the global coastal exclusive economic zone to generate an intersection region, setting the pixel values of the pre-processed sentinel1 image outside the intersection region to zero to generate an economic zone filtered image; using a connected domain algorithm to obtain all the connected domains in the economic zone filtered image, generating an offshore wind turbine mask, obtaining the center pixel coordinates, the minimum circumscribed rectangle width, the minimum circumscribed rectangle height and the pixel area of all the connected domains; then, determining the threshold value W of the minimum circumscribed rectangle width, the threshold value H of the minimum circumscribed rectangle height and the connected domain pixel area threshold value A, setting the gray value of the connected domain in the offshore wind turbine mask to zero if the minimum circumscribed rectangle width is greater than W, or the minimum circumscribed rectangle height is greater than H, or the connected domain pixel area is greater than A, and finally generating a graph shape filtered offshore wind turbine mask, wherein each connected domain is a candidate graph patch.

5. The method of claim 4, wherein, The generation process of the processed sentinel2 image includes: firstly, downloading the latest M period sentinel2 visible light band image within the range of the tile i in S2.1, and generating an M period corrected sentinel2 image after a pre-processing procedure; the pre-processing procedure includes: performing geometric correction and atmospheric correction on the original sentinel2 image, identifying clouds and shadows from the M period corrected sentinel2 image, recording the cloud and shadow area as a transparent area to generate an M period cloud-removed sentinel2 image; using the latest image in the M period cloud-removed sentinel2 image as the topmost layer, using the remaining M period cloud-removed sentinel2 images in descending order of time, using a recursive method to fill the transparent area, wherein M is greater than the number of images required by the recursive method, and further generating a processed sentinel2 image.

6. The method of claim 5, wherein the method further comprises: The generation process of the sentinel2 image slice at the candidate graph patch includes: querying the center pixel coordinates A of all the candidate graph patches, calculating the corresponding coordinate B in the processed sentinel2 image in S2.4 for each coordinate A, using a cropping method to crop a square image slice from the processed sentinel2 image with the coordinate B as the center; if the square image slice is completely contained in the processed sentinel2 image, then directly crop the square region; If the square image slice is not completely contained in the processed sentinel2 image, then crop the contained part from the processed sentinel2 image, and set the rest of the pixel values to zero; finally, generate a corresponding sentinel2 image slice for each candidate patch.

7. The method of claim 6, wherein, The patch classification and clustering process for all sentinel2 image slices includes: Using a classification algorithm to classify all sentinel2 image slices, the classification algorithm used in the classification process is a binary classification algorithm or a multi-classification algorithm, the output of the classification algorithm is a one-dimensional vector, the length of the vector is the number of target classes, and the value of the vector is the probability value of the contained feature in the sentinel2 image belonging to the target class, and the target class corresponding to the maximum value of the vector is the classification result corresponding to the sentinel2 image slice; Query the candidate patch C in the offshore wind turbine mask after shape filtering, and remove the patch C based on the offshore wind turbine mask after shape filtering to obtain the offshore wind turbine mask after patch classification filtering; Using a clustering algorithm, the patches in the offshore wind turbine mask after patch classification filtering are clustered into a cluster with a number not less than 1, and the cluster D is marked, the cluster D only contains one patch, and finally all patches contained in the cluster D are removed from the offshore wind turbine mask after patch classification filtering to obtain the offshore wind turbine mask after patch clustering filtering.

8. The method of claim 7, wherein the method further comprises: The generation process of the time-backscatter signal curve includes: Determine the satellite image start image service time t1 and the query end time t2 within the tile i range, which is generally the data analysis day, query and download the preprocessed sentinel1 image in the t1-t2 time range with intersection with the tile range, and generate the corresponding preprocessed sentinel1 image set, the preprocessed sentinel1 images in the preprocessed sentinel1 image set are arranged in descending order of shooting time; Iterate through all patches in the offshore wind turbine mask after patch clustering filtering, and set the number j for each patch, where j=1, 2, …, a, a is a natural number greater than 0; Generate a time-backscatter signal curve for patch j, the time-backscatter signal curve is obtained as follows: first, determine the pixel range PAj of the patch j, second, sequentially obtain the maximum grayscale Vj in the PAj range in the preprocessed sentinel1 image set, and the corresponding shooting time of the preprocessed sentinel1 image is tj, to obtain the tj-Vj curve belonging to the patch j, and perform convolution calculation on the tj-Vj curve, the convolution calculation uses a discrete convolution kernel f_t that satisfies the following conditions: Wherein, T is the convolution kernel window length, t∈R; take f_t=e^(t-1), 1≤t≤T, T=5, take the time t_jk corresponding to the maximum value k in the result as the construction time of the offshore wind turbine corresponding to the plot j.

9. The method of claim 8, wherein, TLi is obtained, secondly, the wgs84 coordinate offset TLij of the center pixel coordinate of the plot j relative to TLi is obtained, then, the absolute wgs84 coordinate of the plot j is obtained, finally, after traversing i and j, the information of the offshore wind turbine is stored.

Images