Offshore wind driven generator identification method based on multi-source remote sensing image

Abstract

The invention discloses an offshore wind driven generator identification method based on a multi-source remote sensing image, relates to a remote sensing image identification technology, and effectively improves the identification accuracy of an offshore wind driven generator. According to the method, tile division is performed on a global coastal dedicated economic area, then sentinel1 images in each tile range are generated, a plurality of connected domains are set in the sentinel1 images, candidate pattern spots are obtained, sentinel2 image slices at the candidate pattern spots are generated, pattern spot classification is performed on all the sentinel2 image slices through a classification algorithm, and the classification efficiency of the candidate pattern spots is improved. And a time-back scattering signal curve is generated according to a classification result, convolution calculation is performed on the time-back scattering signal curve to obtain the construction time of the offshore wind power generator, and various information of the offshore wind power generator is stored according to wgs84 coordinates of each tile and each pattern spot.

Description

Technical Field

[0001] The present invention relates to remote sensing image recognition technology, and specifically to a method for identifying offshore wind turbines based on multi-source remote sensing images. Background Art

[0002] An offshore wind farm consists of a cluster or array of wind turbines and is widely regarded as a renewable energy source and 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. 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 and real-time identification of the distribution of offshore wind farms and offshore wind turbines is of great significance for the construction progress detection, quality control, environmental monitoring, safety management, etc. of under-construction wind farms.

[0004] In the prior art, SAR satellite images are the main information source for identifying offshore wind turbines. SAR satellite images can obtain radar backscattering data stably without being affected by day and night and cloud cover. And due to the dihedral structure of offshore wind turbines, their VV polarization radar backscattering signals are obvious, with a higher signal-to-noise ratio compared to the background. In general, it is easy to identify using global or local threshold methods when the background is simple. However, in areas with sandbanks, small islands, and obvious sea clutter, it is easy to be confused, resulting in a low recognition accuracy. Moreover, for radar signals similar to masts, substations, signal towers, etc. that often appear near offshore wind farms, neither traditional threshold segmentation recognition methods nor deep learning segmentation recognition methods can achieve accurate recognition.

[0005] In the prior art, another technical route is to use optical satellite images to identify offshore wind turbines. However, the disadvantages are obvious: First, optical images are often affected by clouds and haze in coastal areas, which makes it difficult to identify offshore wind turbines; Second, the resolution of freely available optical satellite images is limited, and typical offshore wind turbines only occupy a limited number of pixels in the images, resulting in difficult accurate identification; If high-resolution commercial images are used for global identification of offshore wind turbines, the cost is extremely high.

[0006] Due to the above factors, there are obvious omissions in the fan numbers and errors in the fan position records in existing international offshore wind farm data. So far, there is no global offshore wind turbine dataset or service in the public domain with low cost, high recognition accuracy, and near-real-time update. Summary of the Invention

[0007] To solve the above technical problems, the purpose of the present invention is to provide a method for identifying offshore wind turbines based on multi-source remote sensing images.

[0008] To achieve the above purpose, the present invention provides the following technical solutions: A method for identifying offshore wind turbines based on multi-source remote sensing images, comprising the following steps: S1. Data preparation, including obtaining the vector range of the global coastal exclusive economic zone and performing tile division; S1.1. Obtain the vector range of the global coastal exclusive economic zone; S1.2. Perform tile division on the global coastal exclusive economic zone to generate a number of tiles; S2. Generate sentinel2 image slices at candidate patches within each tile range; S2.1. Generate preprocessed sentinel1 images within each tile range, and use the minimum covering number method to obtain the minimum value of the number of preprocessed sentinel1 images within each time range; S2.2. Perform a union operation on the preprocessed sentinel1 image and the global coastal exclusive economic zone to generate an economic zone filtered image; S2.3. Use the connected component algorithm to obtain all connected components and the offshore wind turbine mask in the economic zone filtered image, and then obtain candidate patches for each connected component; S2.4. Generate processed sentinel2 images; S2.5. Generate sentinel2 image slices at candidate patches according to the candidate patches of each connected component and the processed sentinel2 image; S3. Offshore wind turbine information extraction, where the offshore wind turbine information includes: the longitude and latitude coordinates of the offshore wind turbine, the construction time, and the recognition confidence; S3.1. Use a classification algorithm to classify the patches of the offshore wind turbine mask; S3.2. Use a clustering algorithm to cluster the patches in the offshore wind turbine mask after patch classification filtering; S3.3. Confirm the start time of satellite image service provision and the query cut-off time, and then obtain the processed sentinel1 images that intersect with each tile range within the corresponding time range, and then generate a preprocessed sentinel1 image set. Traverse all patches in the offshore wind turbine mask after patch clustering filtering to generate a time-backscatter signal curve, and perform convolution calculation on the time-backscatter signal curve to obtain the construction time of each offshore wind turbine; S3.4. Obtain the WGS84 coordinates of each tile and its corresponding patch, and then store the information of the corresponding offshore wind turbines.

[0009] Furthermore, the generation process of the preprocessed Sentinel-1 image includes: Set a number i for each tile, where i = 1, 2, 3, ……, k, and k is a natural number greater than 0. Within the range of tile i, download the latest N IW-mode Sentinel-1 VV polarization Ground Range Detected (GRD) images in terms of time, and perform radiometric calibration, geometric correction, terrain correction, and decibelization on the IW-mode Sentinel-1 VV polarization Ground Range Detected (GRD) images; Secondly, stack and average the N preprocessed Sentinel-1 images to generate a processed Sentinel-1 image, where N is a natural number greater than 0.

[0010] Furthermore, the process of calculating the minimum covering number for the processed Sentinel-1 image includes: Determine the range of tile i, the start time t1 when the satellite image starts providing image services, and the query cut-off time t2, where t2 > t1; Secondly, query all the preprocessed Sentinel-1 images that intersect with the range of tile i within the time range from t1 to t2; Furthermore, starting from time t2, calculate the intermediate time t3, where t2 > t3 > t1, so that all the preprocessed Sentinel-1 images within the time range from t3 to t2 can completely cover the range of tile i; Finally, the number of all the preprocessed Sentinel-1 images within the time range from t3 to t2 is the minimum value that N can take.

[0011] Furthermore, the generation process of the candidate patches includes: Perform a union operation on the preprocessed Sentinel-1 image and the global exclusive economic zone of the coast to generate an intersection area, set the pixel values of the intervals outside the intersection area in the preprocessed Sentinel-1 image to zero, and generate an economic zone filtered image; Use the connected component algorithm to obtain all the connected components in the economic zone filtered image, and generate an offshore wind turbine mask. The offshore wind turbine mask is 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 component is the connected pixels with a gray value of 255 in the offshore wind turbine mask. The connected component algorithm uses the 8-neighborhood connectivity algorithm; Obtain the central pixel coordinates, minimum bounding rectangle width, height, and pixel area of all connected components; then, determine the threshold W for the minimum bounding rectangle width, the threshold H for the minimum bounding rectangle height, and the threshold A for the connected component pixel area. Set the gray value of the connected component in the off-shore wind turbine mask whose minimum bounding rectangle width is greater than W, or minimum bounding rectangle height is greater than H, or connected component pixel area is greater than A to zero. Finally, generate the off-shore wind turbine mask after filtering the patch shape, where each connected component is a candidate patch.

[0012] Furthermore, the generation process of the processed sentinel2 image includes: First, within the range of the tile i described in S2.1, download the M latest sentinel2 visible light band images in terms of time. After the preprocessing process, generate M corrected sentinel2 images. The preprocessing process includes: performing geometric correction and atmospheric correction on the original sentinel2 image; secondly, identifying clouds and shadows in the M corrected sentinel2 images, marking the cloud and shadow areas as transparent areas, and then generating M sentinel2 images after cloud removal; finally, taking the image with the latest time in the M sentinel2 images after cloud removal as the topmost layer, using the remaining M sentinel2 images after cloud removal, and using the recursive method to fill the transparent areas in the descending order of time, where M is greater than the number of image periods required by the recursive method. Finally, generate a processed sentinel2 image.

[0013] Furthermore, the generation process of the sentinel2 image slice at the candidate patch includes: Query the central pixel coordinates A of all candidate patches, calculate the corresponding coordinates B in the processed sentinel2 image for each coordinate A, and use the 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, directly crop it according to the square area. If the square image slice is not completely contained in the processed sentinel2 image, crop the included part from the processed sentinel2 image and set the pixel values of the remaining non-included parts to zero; finally, generate a corresponding sentinel2 image slice for each candidate patch.

[0014] Furthermore, the process of classifying and clustering patches for all sentinel2 image slices includes: Using a classification algorithm, classify all Sentinel-2 image slices. The classification algorithm used in the classification process is a binary classification algorithm or a multi-class classification algorithm. The output of the classification algorithm is a one-dimensional vector. The length of the vector is the number of target classes. The values of the vector are the probability values that the ground features contained in the Sentinel-2 image slice belong to the target class. The target class corresponding to the maximum value of the vector is the classification result corresponding to the Sentinel-2 image slice; In the filtered offshore wind turbine mask by patch shape, query candidate patch C. The classification result of the Sentinel-2 image slice corresponding to the location where the candidate patch C is located is not an offshore wind turbine. Based on the filtered offshore wind turbine mask by patch shape in S2.3, remove the patch C to obtain a filtered offshore wind turbine mask by patch classification; Using a clustering algorithm, cluster the patches in the filtered offshore wind turbine mask by patch classification into clusters D with a quantity of not less than 1, and label the clusters D. Each cluster D contains only one patch. Finally, remove all patches included in the clusters D from the filtered offshore wind turbine mask by patch classification to obtain a filtered offshore wind turbine mask by patch clustering.

[0015] Furthermore, the generation process of the time-backscatter signal curve includes: Determine the satellite image start service time t1 and query cut-off time t2 within the tile i. This time is generally the day of data analysis. Query and download the preprocessed Sentinel-1 images that intersect with the tile range within the time range from t1 to t2, and generate a corresponding set of preprocessed Sentinel-1 images. The preprocessed Sentinel-1 images in the set of preprocessed Sentinel-1 images are arranged in reverse order of the shooting time; Traverse all patches in the filtered offshore wind turbine mask by patch clustering, and set a number j for each patch, where j = 1, 2,..., a, and a is a natural number greater than 0; Generate a time-backscatter signal curve for patch j. The time-backscatter signal curve is obtained according to the following steps: First, determine the pixel range PAj covered by the patch j. Secondly, sequentially obtain the maximum gray value Vj within the range PAj in the set of preprocessed Sentinel-1 images. The shooting time corresponding to the preprocessed Sentinel-1 image is tj, and a tj-Vj curve belonging to the patch j is obtained. Perform convolution calculation on the tj-Vj curve. The convolution calculation uses a discrete convolution kernel which satisfies the following conditions: where T is the convolution kernel window length, t ∈ R; in this embodiment, take , when T = 5, the time tjk corresponding to the maximum value k in the result is the construction time of the offshore wind turbine corresponding to the patch j.

[0016] Furthermore, obtain the wgs84 coordinate TLi corresponding to the upper left corner of the tile i. Secondly, obtain the wgs84 coordinate offset TLij of the center pixel coordinate of the patch j relative to TLi. Then, calculate the absolute wgs84 coordinate of the patch j. Finally, after traversing i and j, store the information of the offshore wind turbine.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In the present invention, N - period pre - processed sentinel1 images are superimposed and averaged to generate a processed sentinel1 image, and the minimum covering number method is used to determine the most appropriate value of N, which can ensure capturing near - real - time offshore wind turbine signals with strong timeliness, and greatly improves the timeliness of offshore wind turbine recognition compared with the prior art; (2) In the present invention, when focusing on the real - time recognition effect and the signal - to - noise ratio of the processed sentinel1 image is low, multi - stage filtering steps are introduced, including economic zone mask filtering, patch shape filtering, etc., which effectively filter noise signals. Further, multi - source remote sensing images are comprehensively used. First, most of the easily confused ground objects are filtered using the sentinel1 image to obtain candidate results, and then combined with the sentinel2 optical image at the candidate area, classification algorithms and clustering algorithms are used to efficiently filter out ground objects such as small islands, masts, substations, signal towers, ships, etc. that cannot be recognized by the sentinel1 image, thus greatly improving the recognition accuracy rate compared with the prior art; (3) In the present invention, the used sentinel1 image and sentinel2 image are both open - source satellite images, and have high time resolution and spatial resolution. Using this method for the recognition of offshore wind turbines has low cost and high practicability. Description of the Drawings

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments recorded in the present invention.

[0019] Figure 1 It is the flow chart of the present invention; Figure 2 It is a schematic diagram of the vector range for obtaining the global coastal exclusive economic zone; Figure 3 It is a schematic diagram of the tile division of a part of the coastal exclusive economic zone; Figure 4 The processed Sentinel-1 image with N = 5; Figure 5 The processed Sentinel-1 image with N = 30; Figure 6 The experimental result graph of the classification and recognition algorithm; Figure 7 The experimental result graph of the classification and recognition algorithm superimposed on the Sentinel-2 base map; Figure 8 The offshore wind turbine mask map after polygon classification and filtering; Figure 9 The offshore wind turbine mask map after polygon clustering and filtering; Figure 10 The time-maximum gray value curve corresponding to the polygon; Figure 11 For Figure 10 The result graph after convolution calculation. Detailed implementation manners

[0020] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other implementation manners obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope protected by the present invention.

[0021] As Figure 1 shown, a method for identifying offshore wind turbines based on multi-source remote sensing images includes the following steps: S1. Data preparation, including obtaining the vector range of the global coastal exclusive economic zone and performing tile division; S1.1. Obtain the vector range of the global coastal exclusive economic zone. The obtained vector range of the global coastal exclusive economic zone is as Figure 2 shown by the shaded part. Most of the land parts are excluded from the area, and offshore wind turbines only exist within the scope of this coastal exclusive economic zone. The global coastal exclusive economic zone uses the WGS84 coordinate and is directly projected using longitude and latitude; S1.2. Perform tile division on the coastal exclusive economic zone. In this embodiment, the tile size is 50KM * 50KM, and the coastal exclusive economic zone is included in the set of all the tiles. As Figure 3 shows the tile division of a part of the coastal exclusive economic zone. Each tile is numbered i, where i = 1, 2, 3,..., k, and k is a natural number greater than 0.

[0022] S2. Generate a preprocessed Sentinel-1 image within the range of each tile i as described in S1.2, and slice the Sentinel-2 image at the candidate patches based on the preprocessed Sentinel-1 image. In this embodiment, the resolution of the Sentinel-2 image is 10 meters, and the slice size is 64 pixels * 64 pixels; S2.1. Generate the preprocessed Sentinel-1 image. The specific steps are as follows: First, within the range of the tile i, download the N latest IW-mode Sentinel-1 VV-polarized Ground Range Detected (GRD) images in terms of time, and generate N preprocessed Sentinel-1 images after the preprocessing process; The preprocessing process includes: performing radiometric calibration, geometric correction, terrain correction, and decibelization on the IW-mode Sentinel-1 VV-polarized Ground Range Detected (GRD) image; Secondly, stack and average the N preprocessed Sentinel-1 images to generate a processed Sentinel-1 image, where N is a natural number greater than 0; Among them, the larger the value of N, the more the inherent random noise and moving target interference signals in the radar image can be suppressed, and the higher the signal-to-noise ratio of the preprocessed Sentinel-1 image. However, the smaller the value of N, the more conducive to identifying newly built offshore wind turbines in the near future. Therefore, an optimal N value needs to be determined for each tile. In the case of emphasizing the real-time recognition effect, the signal-to-noise ratio of the preprocessed Sentinel-1 image should be improved as much as possible; Use the minimum coverage number method to determine the most suitable value of N. The minimum coverage number method includes the following steps: First, determine the range of the tile i, the start time t1 of the satellite image providing image services, and the query cut-off time t2, where t2 > t1; Secondly, query all the preprocessed Sentinel-1 images that intersect with the range of the tile i within the time range from t1 to t2; Furthermore, starting from the time t2, calculate the intermediate time t3, where t2 > t3 > t1, so that all the preprocessed Sentinel-1 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 Sentinel-1 images within the time range from t3 to t2 is the minimum value that N can take; It should be noted that Figure 3The area within the tile marked with X contains a newly built offshore wind farm, located in the upper left part of the tile. After manual interpretation using cloudless Sentinel-2 imagery, it is determined that the construction start time of the offshore wind turbines in this area was approximately December 1, 2023, and the end time was approximately March 10, 2024. The query cut-off time was set to March 20, 2024. The N value determined using the minimum coverage number method for this location is 5, which is approximately the number of all preprocessed Sentinel-1 images for about 2 months. The corresponding processed Sentinel-1 images are as shown in Figure 4 shown. It can be seen that the newly built offshore wind turbines in the upper left corner are relatively clear, but the random noise level of the image background is high and the interference signals are obvious, making it easy to have misidentifications and omissions in the subsequent identification of offshore wind turbines; Specify the N value as 30, which is approximately the number of all preprocessed Sentinel-1 images for nearly 1 year. The corresponding processed Sentinel-1 images are as shown in Figure 5 shown. The random noise level of the background is low, but the original information of the newly built offshore wind turbines in the upper left corner has been basically lost, and it is difficult to make up for it using technical means in the later stage; S2.2. Economic zone mask filtering. The specific process is as follows: Perform a union operation on the preprocessed Sentinel-1 image and the global coastal exclusive economic zone to generate an intersection area. Set the pixel values of the intervals of the preprocessed Sentinel-1 image outside the intersection area to zero to generate an economic zone filtered image; S2.3. Patch shape filtering. Specifically: First, use the connected component algorithm to obtain all connected components in the economic zone filtered image of S2.2 and generate an offshore wind turbine mask. The offshore wind turbine mask is 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 component is the connected pixels with a gray value of 255 in the offshore wind turbine mask. The connected component algorithm uses the 8-neighborhood connectivity algorithm; Secondly, calculate the central pixel coordinates, minimum bounding rectangle width, height, and pixel area of all the connected components; Then, determine the threshold W for the minimum bounding rectangle width, the threshold H for the minimum bounding rectangle height, and the threshold A for the connected component pixel area. Set the gray values of the connected components in the offshore wind turbine mask with a minimum bounding rectangle width greater than W, or a minimum bounding rectangle height greater than H, or a connected component pixel area greater than A to zero. Finally, generate an offshore wind turbine mask after patch shape filtering, and each connected component in it is the candidate patch of S2; S2.4. Generate the processed Sentinel-2 image. The specific steps are as follows: First, within the range of tile i in S2.1, download the M latest Sentinel-2 visible light band images in terms of time, and generate M corrected Sentinel-2 images after the preprocessing process; The preprocessing process includes: geometric correction and atmospheric correction of the original Sentinel-2 image. In this embodiment, the Sentinel-2 visible light band image is a 3-channel color image with a resolution of 10 meters. Secondly, cloud and shadow recognition is performed on the Sentinel-2 image after M-phase correction, and the cloud and shadow areas are marked as transparent areas, and then the Sentinel-2 image after cloud removal in the M-phase is generated. Finally, the image with the latest time in the Sentinel-2 image after cloud removal in the M-phase is used as the topmost layer, and the remaining Sentinel-2 images after cloud removal in the M-phase are used. In descending order of time, the recursive method is used to fill the transparent areas, where M is greater than the number of image periods required by the recursive method. Finally, a processed Sentinel-2 image is generated. S2.5. Generate Sentinel-2 image slices at the candidate patches. The specific steps are as follows: First, query the central pixel coordinates A of all the candidate patches in S2.3, calculate the corresponding coordinates B in the processed Sentinel-2 image in S2.4 for each coordinate A, and use the cropping method to crop a square image slice from the processed Sentinel-2 image with the coordinate B as the center. The cropping method is as follows: If the square image slice is completely contained in the processed Sentinel-2 image, it is directly cropped according to the square area. If the square image slice is not completely contained in the processed Sentinel-2 image, the included part is cropped from the processed Sentinel-2 image, and the pixel values of the remaining non-included parts are set to zero. Finally, a corresponding Sentinel-2 image slice is generated at each candidate patch in S2.3.

[0023] S3. Extract information of offshore wind turbines. Obtaining the information of offshore wind turbines includes the longitude and latitude coordinates, construction time, and recognition confidence of offshore wind turbines. S3.1. Use a classification algorithm to classify all the Sentinel-2 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 categories of the binary classification algorithm are: offshore wind turbines and background. The target categories of the multi-classification algorithm include, but are not limited to, small islands, masts, substations, signal towers, ships, broken clouds, and other frequently occurring interfering features. The output of the classification algorithm is a one-dimensional vector, the length of the vector is the number of target categories, and the value of the vector is the probability value that the feature contained in the Sentinel-2 image slice belongs to the target category. The target category corresponding to the maximum value of the vector is the classification result corresponding to the Sentinel-2 image slice. In the offshore wind turbine mask after the patch shape filtering described in S2.3, query the candidate patch C. The classification result of the sentinel2 image slice corresponding to the location where the candidate patch C is located is not an offshore wind turbine. Based on the offshore wind turbine mask after the patch shape filtering described in S2.3, remove the patch C to obtain an offshore wind turbine mask after patch classification filtering; In this embodiment, the classification algorithm is the YOLO binary classification algorithm. For Figure 5 The experimental results of classification and recognition are as Figure 6 shown. Among them, the white bright spot area is the area recognized as an offshore wind turbine by this YOLO binary classification algorithm; as Figure 7 Further, focus the display area on the newly built offshore wind farm area described in S2.1 and zoom in on this area to more clearly display the offshore wind turbines. Among them, the base map is the sentinel2 optical image when the wind power plant was completed on March 10th. The area circled by the red ring corresponds to the white bright spot area. It can be seen that the area of the offshore wind turbines recognized by this YOLO binary classification algorithm is exactly the same as the ground truth, proving that this YOLO binary classification algorithm can perform high-quality offshore wind turbine recognition in a near-real-time situation while maintaining a timeliness of 2 months; Furthermore, in this embodiment, this YOLO binary classification algorithm is used to identify offshore wind turbines in all the nearshore areas of China. After combining manual interpretation, the experimental results of some typical nearshore sea areas are shown in Table 1. Among them, the accuracy rate is the number of correct identifications / (the number of correct identifications + the number of incorrect identifications), and the recall rate is the number of correct identifications / (the total number of ground truths); it can be seen from the table that both the accuracy rate and the recall rate of the binary classification algorithm used in this embodiment can be better than 0.98, which is significantly better than the data sets in the prior art; the significant improvement of this indicator is mainly because this binary classification algorithm is based on sentinel2 visible light images and can well distinguish offshore wind turbines from interfering ground objects such as 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, unable to distinguish offshore wind turbines from ground objects such as small islands, masts, substations, signal towers, and small boats, resulting in more misidentification cases;

[0024] S3.2 Use a clustering algorithm to cluster the patches in the offshore wind turbine mask after patch classification filtering described in S3.1 into clusters with a quantity of not less than 1, and mark the cluster D. Only one patch is included in the cluster D. Finally, remove all the patches included in the cluster D from the offshore wind turbine mask after patch classification filtering described in S3.1 to obtain an offshore wind turbine mask after patch clustering filtering; In this embodiment, the clustering algorithm is the DBSCAN clustering algorithm. The "neighborhood range" in the algorithm parameters is 7D pixel numbers, and the 7D pixel numbers are the pixel numbers corresponding to 7 times the length of the wind turbine rotor diameter in the offshore wind turbine mask after the patch clustering filtering; the "minimum number of samples" in the algorithm parameters is 1, that is, it is allowed that no less than 1 of the said patches form the cluster; taking a wind farm near Guangzhou as an example, the offshore wind turbine mask after the patch classification filtering is as Figure 8 shown, and the offshore wind turbine mask after the patch clustering filtering is as Figure 9 shown. It can be seen that this DBSCAN clustering algorithm can further improve the accuracy of offshore wind turbine identification; S3.3. Extraction of construction time, specifically including: S3.3.1. Determine the start time t1 and the query cut-off time t2 of the satellite image providing image services within the range of the tile i in S2.1. This time is generally the day of data analysis; S3.3.2. Query and download the preprocessed sentinel1 images in S1.2 that intersect with the tile range within the time range from t1 to t2, and generate a corresponding set of preprocessed sentinel1 images. The preprocessed sentinel1 images in the set of preprocessed sentinel1 images are arranged in reverse order of the shooting time; S3.3.3. Traverse all the patches in the offshore wind turbine mask after the patch clustering filtering in step S3.2, and set a number j for each patch, where j = 1, 2,..., a, and a is a natural number greater than 0; Generate a time-backscatter signal curve for the patch j. The time-backscatter signal curve is obtained according to the following steps: First, determine the pixel range PAj covered by the patch j. Secondly, sequentially obtain the maximum gray value Vj within the range of PAj in the set of preprocessed sentinel1 images. The shooting time corresponding to the preprocessed sentinel1 image is tj, and a tj-Vj curve belonging to the patch j is obtained, as Figure 10 shown; S3.3.4. Perform convolution calculation on the tj-Vj curve. The convolution calculation uses a discrete convolution kernel , which satisfies the following conditions:

[0025] where T is the convolution kernel window length and t ∈ R; in this embodiment, take T = 5, and the result of the convolution calculation is as Figure 11As shown, the time tjk corresponding to a at the maximum value k in the result is the construction time of the offshore wind turbine corresponding to the patch j; S3.4. Coordinate extraction, specifically: First, calculate the wgs84 coordinate 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 coordinate of the patch j in S3.3 relative to TLi. Then, find the absolute wgs84 coordinate of the patch j. Finally, after traversing i and j, store the position information, construction time, and other information corresponding to the offshore wind turbine.

[0026] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for identifying offshore wind turbines based on multi-source remote sensing images, characterized in that: The following steps are involved: S1. Data preparation, including obtaining the vector extent of global coastal exclusive economic zones and performing tile division; S1.

1. Obtain the vector range of the global coastal exclusive economic zone; S1.2, divide the global coastal exclusive economic zones into tiles, and then generate a number of tiles; S2, generate sentinel2 image slices at the candidate spots within each tile range; S2.1, generate the pre-processed sentinel1 image within each tile range, and use the minimum coverage number method to obtain the minimum value of the number of pre-processed sentinel1 images within each time range; S2.2, performing a union operation on the pre-processed sentinel1 image and the global coastal exclusive economic zone, thereby generating an economic zone filtering image; S2.3, using the connected domain algorithm to obtain all connected domains in the economic zone filtered image and the offshore wind turbine mask, and then obtain candidate patches for each connected domain; S2.4, generate processed sentinel2 image; S2.5, generating sentinel2 image slices at the candidate spots according to the candidate spots of each connected domain and the processed sentinel2 images; S3, extracting offshore wind turbine information, obtaining the offshore wind turbine information including the longitude and latitude coordinates, construction time and identification confidence of the offshore wind turbine; S3.

1. Classify all offshore wind turbine masks using a classification algorithm; S3.2, using a clustering algorithm, clustering the spots in the offshore wind turbine mask after the spot classification filtering; S3.

3. Confirm the time when the satellite image starts to provide image service and the query deadline, and obtain the processed sentinel1 images that intersect with each tile range within the corresponding time range, and then generate a pre-processed sentinel1 image set, traverse all the spots in the offshore wind turbine mask after spot clustering and filtering, and then generate a time-backscatter signal curve, and perform convolution calculation on the time-backscatter signal curve to obtain the construction time of each offshore wind turbine; S3.

4. Obtain the wgs84 coordinates of each tile and its corresponding image patch, and then store various information of the corresponding offshore wind turbine.

2. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 1 is characterized in that: The generation process of the sentinel1 image after preprocessing includes: A number i is set for each tile, where i=1, 2, 3, ..., k, and k is a natural number greater than 0. Within the range of the tile i, the latest N IW mode sentinel1 VV polarization Ground Range Detected images are downloaded, and radiation calibration, geometric correction, terrain correction and decibel conversion are performed on the IW mode sentinel1 VV polarization Ground Range Detected images; secondly, a processed sentinel1 image is generated by superimposing and averaging the N pre-processed sentinel1 images, where N is a natural number greater than 0.

3. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 2 is characterized in that: The process of calculating the minimum coverage number method for the processed sentinel1 image includes: Determine the range of the tile i, the time t1 when the satellite image starts to provide image service, and the query deadline t2, where t2>t1; secondly, query all the pre-processed sentinel1 images that intersect with the range of the tile i within the time range from t1 to t2; Then, starting from time t2, the intermediate time t3 is obtained, where t2>t3>t1, so that all the pre-processed sentinel1 images within the time range from t3 to t2 can completely cover the tile i range; finally, the number of all the pre-processed sentinel1 images within the time range from t3 to t2 is the minimum value that N can take.

4. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 3 is characterized in that: The generation process of the candidate spots includes: Performing a union operation on the preprocessed sentinel1 image and the global coastal exclusive economic zone to generate an intersection area, setting the pixel values ​​of the interval of the preprocessed sentinel1 image outside the intersection area to zero, and generating an economic zone filtering image; A connected domain algorithm is used to obtain all connected domains in the economic zone filtered image, and an offshore wind turbine mask is generated to obtain the central pixel coordinates, minimum bounding rectangle width, height and pixel area of ​​all connected domains; then, the threshold W of the minimum bounding rectangle width, the threshold H of the minimum bounding rectangle height and the threshold A of the connected domain pixel area are determined, and the grayscale values ​​of the connected domains in the offshore wind turbine mask whose minimum bounding rectangle width is greater than W, or whose minimum bounding rectangle height is greater than H, or whose connected domain pixel area is greater than A are set to zero, and finally an offshore wind turbine mask after spot shape filtering is generated, in which each connected domain is a candidate spot.

5. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 4 is characterized in that: The generation process of the processed sentinel2 image includes: First, within the range of tile i described in S2.1, the latest M-period sentinel2 visible light band image is downloaded, and the M-period corrected sentinel2 image is generated after the preprocessing process; The preprocessing process includes: performing geometric correction and atmospheric correction on the original sentinel2 image, identifying clouds and shadows on the M-period corrected sentinel2 image, recording the cloud and shadow areas as transparent areas, and then generating the M-period cloud-removed sentinel2 image; taking the latest image in time among the M-period cloud-removed sentinel2 images as the top layer, using the remaining M-period cloud-removed sentinel2 images in descending time order, using a recursive method to fill the transparent area, wherein M is greater than the number of image periods required by the recursive method, and then generating a processed sentinel2 image.

6. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 5 is characterized in that: The generation process of sentinel2 image slices at the candidate spots includes: Query the central pixel coordinates A of all candidate spots, calculate the corresponding coordinates B in the sentinel2 image after processing described in S2.4 for each coordinate A, and use the cropping method to crop a square image slice from the sentinel2 image after processing with the coordinates B as the center; If the square image slice is completely contained in the processed sentinel2 image, the image is directly cropped according to the square area; If the square image slice is not completely contained in the processed sentinel2 image, the contained part is cropped from the processed sentinel2 image, and the pixel values ​​of the remaining uncontained parts are set to zero; finally, a corresponding sentinel2 image slice is generated for each candidate spot.

7. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 6 is characterized in that: The process of spot classification and clustering of all sentinel2 image slices includes: Use 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 categories. The value of the vector is the probability value that the object contained in the sentinel2 image slice belongs to the target category. The target category corresponding to the maximum value of the vector is the classification result corresponding to the sentinel2 image slice. Searching for a candidate spot C in the offshore wind turbine mask after spot shape filtering, removing the spot C based on the offshore wind turbine mask after spot shape filtering, and obtaining an offshore wind turbine mask after spot classification filtering; Using a clustering algorithm, the spots in the offshore wind turbine mask after spot classification and filtering are clustered into clusters of no less than 1, and cluster D is marked, wherein cluster D contains only one spot. Finally, from the offshore wind turbine mask after spot classification and filtering, all spots contained in cluster D are removed to obtain the offshore wind turbine mask after spot clustering and filtering.

8. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 7 is characterized in that: The generation process of the time-backscattered signal curve includes: Determine the satellite image service start time t1 and query deadline t2 within the tile i range, which is generally the day of data analysis, query and download the pre-processed sentinel1 images that intersect with the tile range within the time range of t1 to t2, and generate a corresponding pre-processed sentinel1 image set, in which the pre-processed sentinel1 images are arranged in reverse order of shooting time; Traversing all the spots in the offshore wind turbine mask after the spot clustering filtering, and setting a number j for each spot, where j=1, 2, ..., a, and a is a natural number greater than 0; Generate a time-backscatter signal curve for the spot j, and the time-backscatter signal curve is obtained according to the following steps: first, determine the pixel range PAj included in the spot j, and then obtain the grayscale maximum value Vj within the PAj range in the pre-processed sentinel1 image set in turn, corresponding to the shooting time of the pre-processed sentinel1 image as tj, and obtain the tj-Vj curve belonging to the spot j, and perform convolution calculation on the tj-Vj curve. The convolution calculation uses a discrete convolution kernel f_t, which satisfies the following conditions: Wherein, T is the convolution kernel window length, t∈R; in this embodiment, f_t=e^(t-1), 1≤t≤T, T=5, and the time t_jk corresponding to the maximum value k in the result is taken as the construction time of the offshore wind turbine corresponding to the spot j.

9. The offshore wind turbine identification method based on multi-source remote sensing images according to claim 8, characterized in that: Obtain the wgs84 coordinate TLi corresponding to the upper left corner of the tile i, and then obtain the wgs84 coordinate offset TLij of the central pixel coordinate of the spot j relative to TLi, and then calculate the absolute wgs84 coordinate of the spot j. Finally, after traversing i and j, store the various information of the offshore wind turbine.

Images