SAR sea wind power remote sensing detection method and system based on structural perception constraint

Abstract

The application belongs to the technical field of marine remote sensing, and discloses a SAR offshore wind power remote sensing detection method and system based on structure perception constraint. The method changes the offshore wind power detection paradigm from single target local strong scattering discrimination to a cooperative discrimination mechanism of deep learning initial detection and wind farm array spatial structure constraint; a training data set is constructed; a deep learning target detection model is used to generate a high recall rate of candidate center points; a minimum rotating rectangle candidate boundary is constructed based on density constraint clustering; adaptive threshold binaryzation and connected domain extraction are carried out in the candidate boundary; and the wind turbine detection result is output through double-scale clustering merging, dynamic intensity threshold construction and anisotropic K nearest neighbor spatial constraint screening. The application changes the detection logic from single target local strong scattering discrimination to wind farm level spatial structure consistency discrimination, effectively reduces the false detection caused by ships, platforms and strong scattering noise in complex sea conditions, and improves the cross-sensor and cross-time phase detection stability.

Description

Technical Field

[0001] This invention belongs to the field of marine remote sensing technology, specifically relating to a SAR (Synthetic Aperture Radar) remote sensing detection method and system for offshore wind power based on structure-aware constraints. In particular, it relates to a technical solution that transforms the offshore wind power detection paradigm from a single-target local strong scattering discrimination system to a deep learning-based initial detection and a collaborative discrimination mechanism based on the spatial structure constraints of the wind farm array, thereby achieving high-precision and robust automatic identification of offshore wind turbines under complex sea conditions. Background Technology

[0002] Satellite remote sensing, with its advantages of wide coverage, all-weather operation, and high efficiency, has become an important means of monitoring offshore wind power. However, optical remote sensing images are affected by clouds, lighting conditions, and imaging timing, resulting in insufficient observation continuity. Synthetic Aperture Radar (SAR) possesses active microwave imaging capabilities and can stably acquire information on marine targets under day and night conditions and complex weather conditions, making it an effective means of large-scale offshore wind power monitoring. However, offshore wind turbines, ships, offshore oil platforms, strong scattering waves, and speckle noise can all exhibit high-intensity scattering characteristics in SAR images, posing a fundamental challenge to detection methods based on local features. Simply relying on threshold detection, constant false alarm rate (CFAR) methods, or end-to-end deep learning target detection models is insufficient to distinguish wind turbines from strong scattering interference targets such as ships and platforms at the level of local scattering features, leading to a significant increase in false detection rates under real sea conditions. Even if deep learning target detection models achieve high evaluation indicators on labeled samples, they may still misclassify ships and platforms as wind turbines in the context of real sea surfaces, resulting in insufficient stability and reliability of large-scale detection results. Furthermore, different SAR sensors differ in imaging modes, polarization methods, noise levels, and spatial resolution, and rely solely on local textures or single-point scattering intensity for identification, limiting their generalization capabilities across sensors and time phases.

[0003] Since offshore wind farms typically consist of multiple wind turbines clustered together at certain intervals and in an array orientation, they possess relatively stable spatial structure characteristics. Therefore, to address the aforementioned issues, this invention proposes a SAR offshore wind power remote sensing detection method based on structure-aware constraints. This method transforms the detection paradigm from single-target local discrimination to multi-target spatial structure constraint discrimination. By recovering the wind farm-level spatial structure and using it to constrain single-unit identification, high-precision and robust offshore wind turbine detection is achieved under complex sea conditions.

[0004] Based on the above analysis, the existing problems and shortcomings of SAR offshore wind power detection methods are as follows: (1) Difficulty in distinguishing local scattering features: In SAR images, wind turbines, ships, offshore platforms, and strongly scattering waves are highly similar in local scattering features. A single wind turbine occupies only a few pixels, and the discrimination ability of deep learning target detection models at the single-unit scale has a physical upper limit. Simply increasing the confidence threshold or enhancing the model structure is difficult to maintain a high recall rate while suppressing false detections.

[0005] (2) The spatial structure of wind farm arrays is not effectively utilized: Existing methods do not make full use of the prior knowledge of the fixed spacing, uniform orientation, and clustered distribution of wind farm arrays; density clustering cannot directly handle problems such as unknown array orientation and missing edge units; directly using convex hull or axis-aligned rectangles to construct boundaries is inaccurate; conventional Euclidean distance or isotropic K-nearest neighbor cannot adapt to array structures with different unit spacing in the main axis direction and vertical direction, and normal wind turbines are easily misjudged as outliers.

[0006] (3) Insufficient generalization ability across sensors and time phases: Different SAR sensors have significant differences in imaging mode, polarization mode, noise level and spatial resolution. Existing methods rely on local texture or single-point scattering intensity and lack an adaptive mechanism, resulting in poor detection stability across sensors and time phases.

[0007] (4) Existing single or simple concatenation methods are difficult to coordinate and take into account multi-scale distortion: Deep learning initial detection, density clustering, binarization segmentation, K-nearest neighbor outlier removal and other steps are independent or simply concatenated, which cannot coordinate to handle multi-scale problems such as target splitting, fragmentation, speckle noise, and low-intensity interference; single-scale connected domain merging cannot simultaneously deal with pixel-level splitting and structural fragmentation, making it difficult to achieve high recall, low false detection rate and cross-scene robustness at the same time. Summary of the Invention

[0008] To overcome the problems existing in related technologies, the present invention discloses a SAR remote sensing detection method and system for offshore wind power based on structure-aware constraints, the technical solution of which is as follows: This invention is implemented as follows: a SAR offshore wind power remote sensing detection method based on structure-aware constraints. This method transforms the offshore wind power detection paradigm from single-target local strong scattering discrimination to a collaborative discrimination mechanism of deep learning initial detection and wind farm array spatial structure constraints. The collaborative discrimination mechanism uses the deep learning output as a candidate seed for wind farm spatial structure reconstruction, not as the final detection result, and uses the reconstructed field-level boundary, array principal axis direction, and typical turbine spacing as inverse constraints for single-unit target screening. Specifically, it includes the following steps: S1. Training dataset construction: SAR images are acquired based on offshore wind power ownership data, converted into grayscale images for annotation, and a training dataset is constructed. S2. Preliminary detection result generation: The offshore wind power detection network is trained based on the training dataset using a deep learning target detection model to detect SAR images of the target sea area and output preliminary detection boxes and center points. The preliminary detection boxes and center points exist in the form of a noisy candidate set with high recall, which is used to provide input for subsequent wind farm spatial structure restoration and is not directly used as the final wind turbine detection result. S3. Generation of candidate wind farm boundaries: Project the output preliminary detection box and center point to the UTM coordinate system, identify wind farm clusters based on density-constrained clustering method and remove isolated strong scattering interference points, construct the candidate wind farm boundary with minimum rotation circumscribed rectangle based on the spatial distribution of clusters, transform the noisy preliminary detection results into wind farm-level spatial structure priors, and strictly limit the subsequent fine identification process to the area with wind power array spatial structure; Using wind farm clusters rather than single points as the boundary generation basis is the first key step in introducing spatial structure into the detection process, ensuring the spatial targeting of subsequent processing. S4. Image binarization and candidate target extraction: Under the boundary constraints of candidate wind farms, adaptive threshold binarization and connected component extraction are performed on SAR images, and the area, centroid and maximum gray value of candidate connected components are calculated. This step is not a simple repetition of the detection in step S2, but rather a re-extraction of pixel-level candidate targets within the determined wind farm-level spatial range. This retains weakly scattering targets or those missed by step S2 while avoiding a large number of interfering candidate regions introduced by direct binarization of the full-frame image. S5. Fine-tuning of candidate targets and output of results: The candidate connected components are clustered and merged on two scales to construct a dynamic intensity threshold. Anisotropic K-nearest neighbor spatial consistency constraints with reference to the main axis direction of the wind farm array are used to remove outliers and low-intensity noise targets that do not conform to the arrangement of the wind farm array, so as to obtain the detection results of offshore wind turbines.

[0009] The candidate connected components obtained in step S4 are clustered and merged on two scales to restore the spatial integrity of the split targets. A dynamic intensity threshold is constructed to remove low-intensity noise targets. Anisotropic K-nearest neighbor spatial consistency constraints with reference to the main axis direction of the wind farm array are used to remove outliers that do not conform to the arrangement rules of the wind farm array. Finally, the detection results of offshore wind turbines are output. The anisotropic K-nearest neighbor constraints construct a dimensionless distance index with the typical turbine spacing as the normalization scale, so that the outlier judgment criteria are transformed from a fixed physical distance threshold to a dimensionless index that is adaptive to the wind farm array structure, thereby having the ability to generalize across sensors and wind farms.

[0010] In step S1, the training dataset construction includes: radiometric calibration, geometric correction, and speckle removal of the original SAR image to generate a backscatter intensity image in dB; selecting a co-polarization channel (VV / HH) based on the sensor's polarization characteristics; and converting the backscatter intensity image into a uniform grayscale image for model training and detection. This preprocessing procedure ensures that images from different sensors and at different times have a unified radiometric reference before entering the detection network, which is a prerequisite for achieving cross-sensor detection generalization capability.

[0011] In step S2, the deep learning target detection model includes YOLO, RetinaNet, Faster R-CNN, or a Transformer-based detection model. After training, the SAR image of the target sea area is converted into a grayscale image and input into the detection network, which outputs a preliminary detection box and its corresponding center point. The output of step S2 is only used as a candidate set. The final identification is completed by the spatial structure constraint mechanism from step S3 to step S5, rather than relying on the classification confidence of the detection model itself.

[0012] In SAR imagery, wind turbines exhibit significant point-like strong scattering characteristics, but interfering targets such as ships and offshore platforms also possess similar scattering responses, making reliable differentiation impossible based solely on local features. Step S2 utilizes a deep learning target detection model to rapidly and comprehensively detect candidate targets in the target sea area, aiming to obtain a preliminary candidate set with high recall, providing sufficient input for subsequent spatial structure constraint screening. After training, the SAR image of the target sea area is converted into a grayscale image and input into the detection network, outputting preliminary detection boxes and corresponding center points. It is particularly important to emphasize that the preliminary detection results inevitably contain interference from ships, offshore platforms, and strong scattering noise.

[0013] In step S3, the generation of candidate wind farm boundaries includes: The obtained preliminary detection boxes and center points are projected onto the UTM (Universal Transverse Mercator) coordinate system, and clustering is performed using the DBSCAN density constraint algorithm. DBSCAN does not require a preset number of clusters and can naturally distinguish between core points and noise points, making it suitable for situations where the number of wind farms is unknown in real-world detection scenarios. The neighborhood and core point determination conditions satisfy the following formulas: ; ; In the formula, For a neighborhood set of points, when satisfying hour, Considered the core point; For set X, divide Any preliminary wind power testing point outside, This is the set of all preliminary wind power detection points in the UTM coordinate system. These are preliminary wind power detection points in the UTM coordinate system. The maximum neighborhood distance between target points within the same cluster. The value is 4000m; The minimum number of targets required to form an effective cluster. When the value is 10, the integrity of the wind farm can be maintained while effectively eliminating isolated interference; After obtaining the clusters, based on the point sets within the clusters Constructing a minimum rotation bounding rectangle is crucial because offshore wind farm arrays are typically not horizontal or vertically aligned. Using axis-aligned rectangles would introduce a large number of non-wind farm areas, significantly increasing the false positive rate in subsequent steps. The minimum rotation bounding rectangle can adapt to different wind farm array orientations and, compared to axis-aligned rectangles, effectively reduces the range of non-wind farm areas entering subsequent processing. The latitude and longitude extrema of the clusters satisfy the following conditions: ; ; ; ; In the formula, These are the maximum and minimum longitude values, respectively. These are the latitude maximum and minimum values, respectively. The first in the cluster Longitude of the initial wind power testing points The first in the cluster Latitude of the initial wind power testing points; An initial axis-aligned bounding rectangle is constructed using the latitude and longitude extrema of the clusters, serving as an initial reference for rotation optimization; the rotation angle is then adjusted. Perform a traversal search to determine the optimal rotated bounding rectangle based on the minimum area criterion; An expansion buffer is set on the basis of the minimum rotational bounding rectangle to obtain the candidate wind farm boundary, so as to avoid the wind farm edge units being truncated.

[0014] In step S4, image binarization and candidate target extraction include: Under the constraint of candidate wind farm boundaries, the cropped SAR image is subjected to adaptive threshold binarization. Compared with the global fixed threshold, the adaptive threshold can adapt to the background intensity changes caused by local sea conditions within the candidate boundary, reducing missed detections.

[0015] First, the SAR image is normalized to the set grayscale mapping range. and adopt Gaussian filtering is used to smooth the normalized grayscale image to reduce the impact of speckle noise and local grayscale abrupt changes on the binarization result; let the normalized and smoothed SAR grayscale image be... Then pixel The local adaptive threshold at point is defined as: ; In the formula, For pixels The corresponding local adaptive threshold, In pixels Centered Within a local window The average gray level, Let be the threshold offset constant. The local adaptive threshold is expressed as: ; Based on the aforementioned local adaptive threshold, the binarization function is defined as: ; In the formula, This is the binarized result; when When, it indicates that the pixel is determined to be a foreground pixel; when When this occurs, it indicates that the pixel is determined to be a background pixel; After obtaining the binarization result, connected component labeling is performed on the foreground pixel set, let... For the first If there are 1 connected component, then the pixel area of ​​that connected component is... The calculation formula is: ; In the formula, This represents the number of foreground pixels contained in the connected component.

[0016] In step S5, the fine-tuning of candidate targets and the output of results include: (1) Dual-scale clustering and merging: The small-scale DBSCAN with a search neighborhood radius of 150m is used to merge spatially adjacent split connected components; the connected components are divided into core targets and fragment targets according to their area size, and the fragment targets are assigned to the nearest core target. The core targets and the assignment conditions are respectively satisfied as follows: ; ; In the formula, The connected component with the largest area within the cluster is selected as the core objective. These are the centroid coordinates of the fragment and the core target, respectively. This is the threshold for large-scale search of the neighborhood. The value is 250m; This is the Euclidean distance function in the UTM plane coordinate system. The Euclidean distance between the centroids of fragment targets and the core target is represented in meters (m). Connected components after dual-scale clustering are considered as candidate wind power targets. (2) Dynamic intensity threshold construction: For each wind power candidate target after dual-scale clustering and merging, the normalized grayscale image of its coverage area is statistically analyzed. The maximum grayscale value is denoted as The set consisting of the maximum gray values ​​of all candidate targets. A dynamic intensity threshold is constructed; the dynamic threshold adaptively adjusts according to the grayscale distribution of the specific scene; compared with a fixed threshold, it has stronger robustness to different SAR sensors and different sea states. The dynamic intensity threshold satisfies: ; ; In the formula, For dynamic intensity threshold, The minimum intensity threshold, This is the strength correction factor. The 5th percentile of the set, For the first The maximum grayscale value of each candidate target The set consisting of the maximum grayscale values ​​of all candidate targets. This represents the total number of candidate targets obtained after merging the two-scale clusters. In this embodiment, , When the maximum grayscale value of the candidate target is lower than When this happens, the candidate target is discarded as a low-intensity noise target.

[0017] (3) Anisotropy Nearest Neighbor Spatial Consistency Constraint: The centroid of the candidate target is transformed from the UTM coordinate system to the main axis frame of the wind farm array, and the transformation relationship satisfies the following formula: ; In the formula, For the first Coordinates of each candidate target in the UTM coordinate system The coordinates of the rotated main shaft frame are... The rotation angle of the candidate wind farm boundary; spindle direction Aligned with the main direction of the wind farm array, Perpendicular to the main direction of the wind farm array; Within the main shaft frame, in Euclidean... Nearest Neighbor Sets Construct axial displacement samples in the principal axis and vertical directions: ; ; In the formula, It is a set of axial displacement samples consisting of the absolute coordinate differences between each candidate target and its Euclidean K-nearest neighbor along the principal axis. It is a set of axial displacement samples consisting of the absolute coordinate differences in the vertical direction between each candidate target and its Euclidean K-nearest neighbor. For the first Coordinates of the nearest neighbor candidate targets within the main axis frame; This represents the absolute value operation. and They represent the i-th Candidate targets and their first The non-negative axial distance between the nearest neighbor candidate targets in the principal axis and vertical direction, in meters.

[0018] To suppress the interference of local high-density structures such as offshore substations on spacing estimation, the sample is truncated at the 10% to 90% quantile, resulting in: ; ; In the formula, To The set of principal axis displacement samples obtained after truncation of 10% to 90%. To The set of vertical displacement samples obtained after truncation of 10% to 90% quantile. For any sample value in the corresponding axial displacement sample set. Main axis direction displacement sample set The 10th percentile, A set of vertical displacement samples The 10th percentile, Main axis direction displacement sample set The 90th percentile, A set of vertical displacement samples The 90th percentile; On the truncated sample, the 75th percentile is used as the estimate of typical unit spacing in both directions: ; ; In the formula, These are the typical unit spacings in the main shaft direction and the vertical direction, respectively; Anisotropic normalization of coordinates is performed using the typical turbine spacing as a scale to construct an anisotropic distance metric that reflects the structure of the wind turbine array: ; In the formula, For any two candidate targets and The relative spatial spacing under the normalized main axis framework; due to the difference in the spacing between units in the main axis direction and the vertical direction in different wind farms, the isotropic Euclidean distance will cause the false rejection of normal targets in the array direction. Anisotropic normalization can eliminate the influence of spacing difference on outlier determination, and the distance index after normalization of typical unit spacing is a dimensionless ratio, which is adaptive to different unit spacing and different sensor resolutions.

[0019] For each candidate target, calculate the anisotropic distance of its Kth nearest neighbor. Candidate targets exceeding the dimensionless threshold are identified as outliers and removed. ; ; In the formula, For the first The candidate target is sorted by its anisotropic normalized distance from other candidate targets. Small distance, i.e., the first Nearest neighbor distance; since this distance is normalized by typical unit spacing, it is a dimensionless quantity. The threshold for determining a dimensionless outlier is when... If the distance between the candidate target and its surrounding targets is too large within the normalized principal axis framework, indicating insufficient spatial neighborhood support, then it is identified as an outlier and removed. In this invention, the number of nearest neighbors K is 3, and the dimensionless outlier determination threshold is... Take 5. For all An operator that sorts the anisotropic normalized distances in ascending order. For the number of nearest neighbors, The threshold for determining a dimensionless outlier. For the first One candidate target; (4) Determination and Result Output of Effective Wind Farms: Effective wind farms are determined based on the number of wind turbines within the candidate wind farm area. ; ; In the formula, For the first One candidate wind farm area, Candidate wind farm area The final set of wind turbine targets retained after meticulous screening of candidate targets in the Inner Canon; The cardinality of a set is the number of elements contained in the set. The number of wind turbines in the area, when At that time, The offshore wind farm is deemed a valid offshore wind farm; this threshold is designed based on the economic effects and scale statistics of offshore wind farm construction; the centroid of the connected domain is used as the output of the wind turbine center location, and the wind turbine center location satisfies: ; In the formula, For the first The final retained wind turbine connectivity domain, This represents the number of pixels contained in the connected component. The coordinates of the center of latitude and longitude of the connected domain. For connected components The latitude and longitude coordinates of any pixel within the range. For the first The final set consists of all pixels within the connected components.

[0020] Furthermore, the value of the nearest neighbor number K is 3, and the dimensionless outlier determination threshold is... The value is 5.

[0021] Another objective of this invention is to provide a SAR offshore wind power remote sensing detection system based on structure-aware constraints. This system is used to execute the aforementioned SAR offshore wind power remote sensing detection method based on structure-aware constraints. The system includes: The data preparation module is used to acquire SAR images, perform radiometric processing and format conversion, and complete the annotation and training dataset construction based on ownership data. The preliminary detection module is used to train an offshore wind power detection network based on a deep learning target detection model, detect SAR images of the target sea area, and output preliminary detection boxes and center points. The candidate boundary generation module is used to identify wind farm clusters based on density-constrained clustering and generate candidate wind farm boundaries that accurately match the orientation of the wind farm array using the minimum rotation circumscribed rectangle. The candidate target extraction module is used to perform adaptive threshold binarization and connected component extraction within the candidate boundary, and to calculate the area, centroid, and maximum gray value of the candidate connected component. The fine recognition module is used to perform dual-scale clustering and merging, dynamic intensity threshold screening, and anisotropic K-nearest neighbor outlier removal, and output the detection results of offshore wind turbines.

[0022] Furthermore, the anisotropic K-nearest neighbor constraint in the fine recognition module constructs a dimensionless distance index with reference to the main axis direction of the wind farm array, which is adaptive to different sensor resolutions and different wind farm layouts.

[0023] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the SAR offshore wind power remote sensing detection method based on structure-aware constraints.

[0024] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, in SAR imagery, strong scattering noise from wind turbines, ships, offshore platforms, and the sea surface can manifest locally as point-like or patchy high-intensity scattering, making it difficult to establish stable discrimination based solely on the image features of a single target. Initial detection results typically include false positives, false negatives, and fragmented targets, and cannot be directly used as the final detection result. Furthermore, wind farm arrays exhibit characteristics such as non-fixed orientation and inconsistent turbine spacing between the main axis and the vertical direction, rendering ordinary spatial distance constraints incapable of directional adaptation. To address these technical problems, this invention provides a SAR remote sensing detection method for offshore wind power based on structure-aware constraints. This method first acquires SAR images based on offshore wind power ownership data, converts them into grayscale images, and annotates them to construct a training dataset. Then, a deep learning target detection model is used to generate high-recall preliminary detection results, which are used as candidate seeds for wind farm spatial structure reconstruction, rather than as the final wind turbine identification results. Based on this, density-constrained clustering is used to eliminate isolated strong scattering interference points, constructing candidate wind farm boundaries that match the orientation of the wind farm array. Under the constraints of the candidate boundaries, the SAR images are re-adaptively thresholded and binarized, and connected component extraction is performed to obtain candidate targets. Finally, by combining dual-scale clustering merging, dynamic intensity thresholding, and anisotropic K-nearest neighbor spatial consistency constraints based on the main axis direction of the wind farm array and the spacing between typical turbines, outliers and low-intensity noise targets that do not conform to the wind farm array arrangement are eliminated, resulting in the offshore wind turbine detection results. This invention transforms the detection logic from single-target local strong scattering discrimination to wind farm-level spatial structure consistency discrimination, which can effectively reduce false detections caused by ships, platforms and strong scattering noise under complex sea conditions, and improve the stability of cross-sensor and cross-temporal detection. It also solves the problem that offshore wind turbines, ships, offshore platforms, strong scattering sea conditions and speckle noise are difficult to distinguish in terms of local scattering characteristics in SAR images.

[0025] Secondly, this invention systematically constructs a collaborative discrimination mechanism that integrates deep learning initial detection, density-constrained clustering boundary generation, connected component feature extraction, and anisotropic K-nearest neighbor spatial structure constraints. This mechanism is not a linear concatenation of existing algorithms, but rather transforms the deep learning initial detection results into a wind farm-level spatial structure prior. This prior is then used to inversely constrain candidate target extraction and single-machine target screening. This solves the problems in existing technologies where single strong scattering feature discrimination easily misclassifies ships, offshore platforms, and strong scattering noise as offshore wind turbines, and the insufficient stability of detection results under cross-sensor conditions. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention. Figure 1 This is a flowchart of the SAR offshore wind power remote sensing detection method based on structure-aware constraints provided in the embodiments of the present invention; Figure 2 This is a schematic diagram illustrating typical characteristics of offshore wind power in SAR and optical images provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the core components of the SAR offshore wind power remote sensing detection method based on structure-aware constraints provided in the embodiments of the present invention. Among them, (a) is a schematic diagram of the SAR offshore wind power target detection network structure based on YOLO26; (b) is a schematic diagram of density-constrained clustering and minimum rotation circumscribed rectangle boundary generation that takes into account the spatial distribution of offshore wind power; and (c) is a schematic diagram of the offshore wind power fine identification strategy that integrates connected component features and spatial constraints. Figure 4 These are schematic diagrams of the detection results of five types of deep learning object detection models provided in this embodiment of the invention on Sentinel-1 images. Among them, (a) is a schematic diagram of the detection results of the RetinaNet object detection model on Sentinel-1 images; (b) is a schematic diagram of the detection results of the Faster R-CNN object detection model on Sentinel-1 images; (c) is a schematic diagram of the detection results of the Yolov5 object detection model on Sentinel-1 images; (d) is a schematic diagram of the detection results of the Yolov8 object detection model on Sentinel-1 images; (e) is a schematic diagram of the detection results of the Yolov26 object detection model on Sentinel-1 images; and (f) is a schematic diagram of the actual location of offshore wind power on Sentinel-1 images. Detailed Implementation

[0027] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0028] The innovation of this invention lies in: (1) Detection paradigm innovation: The detection paradigm of offshore wind power is transformed from a single-target local strong scattering discrimination system to a collaborative discrimination mechanism of deep learning initial detection + wind farm array spatial structure constraint. The deep learning output is used as a candidate seed for wind farm spatial structure recovery rather than the final detection result. The recovered field-level boundary, array main axis direction and typical unit spacing are used to inversely constrain single-unit target screening, fundamentally excluding strong scattering interference from ships, offshore platforms and other sources. (2) Innovation in generating prior knowledge of wind farm-level spatial structures: Based on density-constrained clustering, isolated strong scattering interference points are eliminated, and the candidate boundary of the minimum rotation circumscribed rectangle is constructed with the overall spatial distribution of the cluster. This transforms the noisy preliminary detection point set into a prior knowledge of wind farm-level structures that can adapt to any array orientation, overcoming the problem of inaccurate boundaries caused by missed detections or array rotation of convex hull or axis-aligned rectangles. (3) Innovation of re-detection and multi-scale restoration within candidate boundaries: Under the constraint of candidate wind farm boundaries, SAR images are re-adaptively thresholded and connected component extracted, instead of simply filtering the initial detection results, so as to retain wind turbines with weak scattering or those that were missed; and pixel-level splitting and structural fragmentation are processed simultaneously by dual-scale clustering and merging to restore the spatial integrity of split targets. (4) Innovation of anisotropic spatial consistency constraint: Taking the main axis direction of the wind farm array as a reference frame, the typical unit spacing between the main axis direction and the vertical direction is statistically estimated from the candidate target itself, and a dimensionless anisotropic distance index is constructed accordingly. This transforms the outlier judgment standard from a fixed physical distance threshold to a dimensionless index that is adaptive to the array structure, avoiding misjudging wind turbines that are normally distributed along the array direction as outliers, and possessing the generalization ability across sensor resolution and wind farm layout. (5) Innovation in dynamic intensity threshold and effective wind farm determination: The dynamic intensity threshold is constructed by the statistical quantile of the maximum gray value of the candidate target to adaptively eliminate low-intensity noise targets; and the effective wind farm unit number threshold is set based on the economic scale statistics of wind farm construction to eliminate pseudo wind farms formed by occasional strong scattering interference, thereby further improving the reliability of the detection results.

[0029] Example 1: The key to this invention's use of structure-aware constraints for offshore wind power detection lies in breaking through the paradigm limitations of existing methods' single-target local discrimination. It fully utilizes the array-like spatial distribution structure of wind farms as the basis for identification, fundamentally excluding targets with strong scattering interference such as ships and platforms. Its implementation relies on the coordinated operation of deep learning initial detection, spatial structure constraint boundary generation, connected component feature extraction, and anisotropic K-nearest neighbor spatial consistency constraints. It should be noted that there is a progressive constraint relationship between the steps in this method, rather than a simple sequential superposition of existing algorithms. Step S2 intentionally retains a noisy candidate set with high recall; Step S3 recovers the wind farm-level candidate boundary from this set, transforming the noisy point set into a structural prior; Step S4 re-extracts pixel-level candidate targets from SAR images within the candidate boundary, rather than simply filtering the results of Step S2; Step S5 then constructs spatial consistency constraints based on the array's main axis direction and typical turbine spacing. The output of the previous step is transformed into a new structural prior or screening scale in the subsequent step, thus forming a collaborative discrimination process where single-machine target identification is constrained by the field-level structure.

[0030] like Figure 1 As shown, the SAR offshore wind power remote sensing detection method based on structure-aware constraints provided in this embodiment of the invention specifically includes the following steps: S1. Training dataset construction: SAR images are acquired based on offshore wind power ownership data, converted into grayscale images for annotation, and a training dataset is constructed. The raw SAR data was radiometrically calibrated, geometrically corrected, and despeculiarized using SNAP and ENVI 5.6 combined with the SARscape module, generating backscatter intensity images in dB. Based on the sensor's polarization characteristics, a co-polarization channel (VV / HH) was selected, and the backscatter intensity images were converted into grayscale images of a uniform format for model training and detection. This preprocessing workflow ensures a unified radiometric reference for images from different sensors and time phases, a prerequisite for achieving subsequent cross-sensor detection generalization capabilities. In this embodiment, a total of 1251 sample images were obtained, containing 31058 wind power targets.

[0031] like Figure 2 As shown, Figure 2 Figure (a) and Figure 2 Figure (b) shows the overall representation of offshore wind power in SAR and optical images, respectively. Figure 2 Figure (c) in the middle and Figure 2Figure (d) shows the local characterization of a single wind turbine in SAR and optical images, respectively. In SAR images, wind turbines appear as point-like bright spots with strong scattering arranged regularly along the array direction; however, the local scattering pattern of a single bright spot is similar to that of strong scattering interference from ships, offshore platforms, etc., and it is difficult to reliably distinguish them based solely on the local features of a single target. Therefore, this invention further utilizes the array-type spatial distribution structure of the wind farm to implement constrained discrimination.

[0032] S2. Preliminary detection result generation: The offshore wind power detection network is trained based on the training dataset using a deep learning target detection model to detect SAR images of the target sea area and output preliminary detection boxes and center points. The preliminary detection boxes and center points exist in the form of a noisy candidate set with high recall, which is used to provide input for subsequent wind farm spatial structure restoration and is not directly used as the final wind turbine detection result. In SAR imagery, wind turbines exhibit significant point-like strong scattering characteristics, but interfering targets such as ships and offshore platforms also have similar scattering responses, making reliable differentiation impossible based solely on local features. Step S2 aims to utilize a deep learning target detection model to perform rapid and comprehensive candidate target detection over the target sea area, prioritizing targets with high recall rates, and outputting bounding boxes and center points to provide sufficient candidate input for subsequent spatial structure constraint screening.

[0033] It is particularly important to emphasize that the output of step S2 inevitably contains interference such as ships, offshore platforms, and strong scattering noise. These false positives are not random noise, but often originate from real strong scattering targets and cannot be eliminated without sacrificing recall by simply increasing the confidence threshold. Therefore, the output of step S2 is only considered as a candidate set, and the final identification and elimination of interfering targets is accomplished by the spatial structure constraint mechanism in subsequent steps S3 to S5, rather than relying on the classification confidence of the detection model itself.

[0034] like Figure 3 As shown in Figure (a), in this embodiment, a 640×640 pixel SAR grayscale image is input into the YOLO26 offshore wind power target detection network. Multi-scale features are extracted sequentially through the Backbone, upsampling and feature fusion are performed through the Neck, and the Head outputs the detection results at each scale to obtain the preliminary detection box and its center point. The network output is only used as a candidate seed for subsequent density-constrained clustering and spatial structure restoration, and is not directly used as the final detection result.

[0035] S3. Generation of candidate wind farm boundaries: Project the output preliminary detection box and center point to the UTM coordinate system, identify wind farm clusters based on density-constrained clustering method and remove isolated strong scattering interference points, construct the candidate wind farm boundary with minimum rotation circumscribed rectangle based on the spatial distribution of clusters, transform the noisy preliminary detection results into wind farm-level spatial structure priors, and strictly limit the subsequent fine identification process to the area with wind power array spatial structure; The initial detection results contain a large number of isolated, strongly scattering interference points, which would lead to numerous false positives if used directly for identification. Wind farms exhibit clustered array structural characteristics in their spatial distribution. Density clustering can be used to naturally distinguish wind farm clusters from isolated noise points, and candidate boundaries can be constructed accordingly, limiting subsequent fine-grained identification to areas with wind farm spatial structures. This step is the first key step in this method of introducing farm-level spatial structure into the detection process.

[0036] The technical challenge of this step lies in the fact that the initial detection point set simultaneously contains false positives, false negatives, and localized dense interference. Simply performing point-level noise removal cannot provide a stable spatial range suitable for subsequent image processing. This invention generates a minimum rotating bounding rectangle based on the overall spatial distribution of clusters, transforming noisy candidate points into wind farm-level boundary priors. Furthermore, by rotating the rectangle to adapt to different wind farm array orientations, it reduces the intrusion of non-wind farm areas into subsequent processing. Figure 3 As shown in Figure (b).

[0037] Furthermore, step S3 specifically includes: S3.1. Project the preliminary detection box and center point obtained in step S2 onto the UTM (Universal Transverse Mercator) coordinate system, and perform clustering using the DBSCAN density constraint algorithm; DBSCAN does not require a preset number of clusters and can naturally distinguish between core points and noise points, making it suitable for situations where the number of wind farms is unknown in real-world detection scenarios; the neighborhood and core point determination conditions satisfy the following formulas respectively: ; ; In the formula, For a neighborhood set of points, when satisfying hour, Considered the core point; For set X, divide Any preliminary wind power testing point outside, This is the set of all preliminary wind power detection points in the UTM coordinate system. These are preliminary wind power detection points in the UTM coordinate system. The maximum neighborhood distance between target points within the same cluster. The value is 4000 m; The minimum number of targets required to form an effective cluster. When the value is 10, the integrity of the wind farm can be maintained while effectively eliminating isolated interference; S3.2. After obtaining the clusters, based on the point sets within the clusters... Constructing a minimum rotation bounding rectangle is crucial because offshore wind farm arrays are typically not horizontal or vertically aligned. Using axis-aligned rectangles would introduce a large number of non-wind farm areas, significantly increasing the false positive rate in subsequent steps. The minimum rotation bounding rectangle can adapt to different wind farm array orientations and, compared to axis-aligned rectangles, effectively reduces the range of non-wind farm areas entering subsequent processing. The latitude and longitude extrema of the clusters satisfy the following conditions: ; ; ; ; In the formula, These are the maximum and minimum longitude values, respectively. These are the latitude maximum and minimum values, respectively. The first in the cluster Longitude of the initial wind power testing points The first in the cluster Latitude of the initial wind power testing points; An initial axis-aligned bounding rectangle is constructed using the latitude and longitude extrema of the clusters, serving as an initial reference for rotation optimization; the rotation angle is then adjusted. Perform a traversal search to determine the optimal rotated bounding rectangle based on the minimum area criterion; for example... Figure 3 As shown in Figure (b); S3.3. Based on the minimum rotational circumscribed rectangle, an expansion buffer is set to obtain the candidate wind farm boundary, so as to avoid the wind farm edge units being truncated; such as Figure 3 As shown in Figure (b).

[0038] S4. Image binarization and candidate target extraction: Under the boundary constraints of candidate wind farms, adaptive threshold binarization and connected component extraction are performed on SAR images, and the area, centroid and maximum gray value of candidate connected components are calculated. Image binarization and candidate target extraction include: Under the constraint of candidate wind farm boundaries, the cropped SAR image is subjected to adaptive threshold binarization. Compared with the global fixed threshold, the adaptive threshold can adapt to the background intensity changes caused by local sea conditions within the candidate boundary, reducing missed detections.

[0039] First, the SAR image is normalized to the set grayscale mapping range. and adopt Gaussian filtering is used to smooth the normalized grayscale image to reduce the impact of speckle noise and local grayscale abrupt changes on the binarization result; let the normalized and smoothed SAR grayscale image be... Then pixel The local adaptive threshold at point is defined as: ; In the formula, For pixels The corresponding local adaptive threshold, In pixels Centered Within a local window The average gray level, Let be the threshold offset constant. The local adaptive threshold is expressed as: ; Based on the aforementioned local adaptive threshold, the binarization function is defined as: ; In the formula, This is the binarized result; when When, it indicates that the pixel is determined to be a foreground pixel; when When this occurs, it indicates that the pixel is determined to be a background pixel; After obtaining the binarization result, connected component labeling is performed on the foreground pixel set, let... For the first If there are 1 connected component, then the pixel area of ​​that connected component is... The calculation formula is: ; In the formula, This represents the number of foreground pixels contained in the connected component.

[0040] In SAR images within the candidate boundaries, wind turbine targets exhibit high-intensity scattering, but due to speckle noise, single targets may exhibit spatial fragmentation after binarization. This step is not a simple repetition of the detection results in step S2, but rather a re-extraction of pixel-level candidate targets within the wind farm-level candidate boundaries formed in step S3: adaptive threshold binarization of the SAR image is performed under the constraints of the candidate wind farm boundaries, connected components are extracted, and the area of ​​each connected component is calculated. A i Centroid and the maximum gray value of the corresponding region I imax This provides structured feature input for the dual-scale merging and fine-tuning in step S5.

[0041] Compared to a globally fixed threshold, an adaptive threshold can adapt to changes in background intensity caused by local sea conditions within the candidate boundary, reducing missed detections due to improper threshold settings while maintaining a high recall rate. By limiting the processing range to the candidate wind farm boundary, it can retain weakly scattering wind turbines or targets missed in step S2, while effectively avoiding a large number of candidate regions with strong ship, sea surface scattering, and speckle noise introduced by direct binarization of the full-frame SAR image, thus making subsequent fine screening more spatially targeted. S5: Fine screening of candidate targets and output of results.

[0042] The candidate targets contain interference such as target fragmentation, low-intensity noise, and outlier scattering points. This step restores the integrity of fragmented targets through dual-scale clustering, removes low-intensity noise targets through dynamic intensity thresholding, and eliminates spatially isolated outliers using anisotropic K-nearest neighbor spatial consistency constraints. These three methods work synergistically to significantly reduce the false detection rate while maintaining high recall. Figure 3 As shown in Figure (c).

[0043] The technological breakthrough in this step lies in the fact that outlier removal does not use fixed physical distances or isotropic Euclidean distances. Instead, it first determines the array's main axis frame based on the candidate wind farm boundaries, then statistically estimates the typical turbine spacing in the main axis direction and vertical direction from the candidate targets themselves, and constructs dimensionless anisotropic distances accordingly. This allows it to adapt to different wind farm orientations, different longitudinal and transverse spacings, and different SAR sensor resolutions, avoiding the accidental deletion of wind turbines that are normally distributed along the array direction.

[0044] S5. Fine-tuning of candidate targets and output of results: The candidate connected components are clustered and merged on two scales to construct a dynamic intensity threshold. Anisotropic K-nearest neighbor spatial consistency constraints with reference to the main axis direction of the wind farm array are used to remove outliers and low-intensity noise targets that do not conform to the arrangement of the wind farm array, so as to obtain the detection results of offshore wind turbines.

[0045] The candidate connected components obtained in step S4 are clustered and merged on two scales to restore the spatial integrity of the split targets. A dynamic intensity threshold is constructed to remove low-intensity noise targets. Anisotropic K-nearest neighbor spatial consistency constraints with reference to the main axis direction of the wind farm array are used to remove outliers that do not conform to the arrangement rules of the wind farm array. Finally, the detection results of offshore wind turbines are output. The anisotropic K-nearest neighbor constraints construct a dimensionless distance index with the typical turbine spacing as the normalization scale, so that the outlier judgment criteria are transformed from a fixed physical distance threshold to a dimensionless index that is adaptive to the wind farm array structure, thereby having the ability to generalize across sensors and wind farms.

[0046] S5.1. Dual-scale clustering and merging: A small-scale DBSCAN with a search neighborhood radius of 150m is used to merge spatially adjacent split connected components; then, core targets and fragmented targets are divided according to their area size, and a large-scale search neighborhood threshold is used. Fragmented targets are assigned to the nearest core target; small-scale merging addresses target fragmentation caused by speckle, while large-scale assignment addresses structural fragmentation. The two scales work together to ensure the spatial integrity of candidate targets; the core target and assignment conditions are respectively satisfied as follows: ; ; In the formula, The connected component with the largest area within the cluster is selected as the core objective. These are the centroid coordinates of the fragment and the core target, respectively. This is the threshold for large-scale search of the neighborhood. The value is 250m; This is the Euclidean distance function in the UTM plane coordinate system. The Euclidean distance between the centroids of fragment targets and the core target is represented in meters (m). Connected components after dual-scale clustering are considered as candidate wind power targets. S5.2. Dynamic Intensity Threshold Construction: Construct a dynamic intensity threshold based on the statistical distribution of the maximum grayscale value of candidate targets. When the maximum gray value of the candidate target is lower than When this happens, the candidate target is eliminated as a low-intensity noise target; the dynamic threshold is adaptively adjusted according to the intensity distribution of the specific scene, and has stronger robustness to different SAR sensors and different sea states compared with the fixed threshold. For each wind power candidate target after dual-scale clustering and merging, the normalized grayscale image of its coverage area is statistically analyzed. The maximum grayscale value is denoted as The set consisting of the maximum gray values ​​of all candidate targets. A dynamic intensity threshold is constructed; the dynamic threshold adaptively adjusts according to the grayscale distribution of the specific scene; compared with a fixed threshold, it has stronger robustness to different SAR sensors and different sea states. The dynamic intensity threshold satisfies: ; ; In the formula, For dynamic intensity threshold, The minimum intensity threshold, This is the strength correction factor. The 5th percentile of the set, For the first The maximum grayscale value of each candidate target The set consisting of the maximum grayscale values ​​of all candidate targets. This represents the total number of candidate targets obtained after merging the two-scale clusters. In this embodiment, , When the maximum grayscale value of the candidate target is lower than When this happens, the candidate target is discarded as a low-intensity noise target.

[0047] S5.3. Introducing anisotropic K-nearest neighbor spatial consistency constraints: The centroid of candidate targets is transformed to the main axis frame of the wind farm array; axial displacement samples in the main axis direction and vertical direction are constructed within the main axis frame, and interference from local high-density structures such as offshore substations on spacing estimation is suppressed by 10%~90% quantile truncation; the typical turbine spacing in the main axis direction and vertical direction is estimated based on the truncated samples at the 75th quantile; the coordinates are anisotropically normalized using the typical turbine spacing as the scale to construct a dimensionless anisotropic distance metric; since the turbine spacing in the main axis direction and vertical direction varies among different wind farms, isotropic Euclidean distance may cause false rejection of targets normally distributed along the array direction, and anisotropic normalization can eliminate the influence of spacing differences on outlier determination; this distance index is a dimensionless ratio, which is adaptive to different turbine spacings and different sensor resolutions; K is set to 3, a dimensionless threshold. If the value is 5, candidate targets that meet the outlier determination criteria are eliminated. The centroid of the candidate target is transformed from the UTM coordinate system to the main axis frame of the wind farm array. The transformation relationship satisfies the following formula: ; In the formula, For the first Coordinates of each candidate target in the UTM coordinate system The coordinates of the rotated main shaft frame are... The rotation angle of the candidate wind farm boundary; spindle direction Aligned with the main direction of the wind farm array, Perpendicular to the main direction of the wind farm array; Within the main shaft frame, in Euclidean... Nearest Neighbor Sets Construct axial displacement samples in the principal axis and vertical directions: ; ; In the formula, It is a set of axial displacement samples consisting of the absolute coordinate differences between each candidate target and its Euclidean K-nearest neighbor along the principal axis. It is a set of axial displacement samples consisting of the absolute coordinate differences in the vertical direction between each candidate target and its Euclidean K-nearest neighbor. For the first Coordinates of the nearest neighbor candidate targets within the main axis frame; This represents the absolute value operation. and They represent the i-th Candidate targets and their first The non-negative axial distance between the nearest neighbor candidate targets in the principal axis and vertical direction, in meters.

[0048] To suppress the interference of local high-density structures such as offshore substations on spacing estimation, the sample is truncated at the 10% to 90% quantile, resulting in: ; ; In the formula, To The set of principal axis displacement samples obtained after truncation of 10% to 90%. To The set of vertical displacement samples obtained after truncation of 10% to 90% quantile. For any sample value in the corresponding axial displacement sample set. Main axis direction displacement sample set The 10th percentile, A set of vertical displacement samples The 10th percentile, Main axis direction displacement sample set The 90th percentile, A set of vertical displacement samples The 90th percentile; On the truncated sample, the 75th percentile is used as the estimate of typical unit spacing in both directions: ; ; In the formula, These are the typical unit spacings in the main shaft direction and the vertical direction, respectively; Anisotropic normalization of coordinates is performed using the typical turbine spacing as a scale to construct an anisotropic distance metric that reflects the structure of the wind turbine array: ; In the formula, For any two candidate targets and The relative spatial spacing under the normalized main axis framework; due to the difference in the spacing between units in the main axis direction and the vertical direction in different wind farms, the isotropic Euclidean distance will cause the false rejection of normal targets in the array direction. Anisotropic normalization can eliminate the influence of spacing difference on outlier determination, and the distance index after normalization of typical unit spacing is a dimensionless ratio, which is adaptive to different unit spacing and different sensor resolutions.

[0049] For each candidate target, calculate the anisotropic distance of its Kth nearest neighbor. Candidate targets exceeding the dimensionless threshold are identified as outliers and removed. ; ; In the formula, For the first The candidate target is sorted by its anisotropic normalized distance from other candidate targets. Small distance, i.e., the first Nearest neighbor distance; since this distance is normalized by typical unit spacing, it is a dimensionless quantity. The threshold for determining a dimensionless outlier is when... If the distance between the candidate target and its surrounding targets is too large within the normalized principal axis framework, indicating insufficient spatial neighborhood support, then it is identified as an outlier and removed. In this invention, the number of nearest neighbors K is 3, and the dimensionless outlier determination threshold is... Take 5. For all An operator that sorts the anisotropic normalized distances in ascending order. For the number of nearest neighbors, The threshold for determining a dimensionless outlier. For the first One candidate target; S5.4. Valid Wind Farm Judgment and Result Output: Set the effective scale threshold of candidate wind farms to 25 units. When the number of wind turbines in the candidate area is not less than 25, it is judged as a valid wind farm; output the latitude and longitude center coordinates of each wind turbine using the centroid of the connected domain.

[0050] Determining effective wind farms based on the number of wind turbines within the candidate wind farm area: ; ; In the formula, For the first One candidate wind farm area, The number of wind turbines in the area. Candidate wind farm area The final set of wind turbine targets retained after meticulous screening of candidate targets in the Inner Canon; The cardinality of a set is the number of elements contained in the set. At that time, The offshore wind farm is deemed a valid offshore wind farm; this threshold is designed based on the economic effects and scale statistics of offshore wind farm construction; the centroid of the connected domain is used as the output of the wind turbine center location, and the wind turbine center location satisfies: ; In the formula, For the first The final retained wind turbine connectivity domain, This represents the number of pixels contained in the connected component. The coordinates of the center of latitude and longitude of the connected domain. For connected components The latitude and longitude coordinates of any pixel within the range. For the first The final set consists of all pixels within the connected components.

[0051] The values ​​of the above key parameters are all related to the physical characteristics and SAR imaging characteristics of offshore wind farms, rather than being randomly selected. The basis for their values ​​is as follows: (1) DBSCAN neighborhood distance The 4000m setting is based on the fact that the overall scale of an existing offshore wind farm is usually several kilometers and the distance between adjacent units is generally in the range of 500 to 2000m. The 4000m setting ensures that the units in the same wind farm are clustered together, while excluding strong scattering points such as isolated ships and platforms that are farther away from the cluster; (2) Minimum number of targets at the core point Taking 10 corresponds to the fact that a single effective wind farm usually consists of dozens of units. Taking 10 can suppress pseudo-clusters formed by occasional strong scattering points while ensuring that the real wind farm is effectively clustered; (3) The radius of the small-scale neighborhood of the dual-scale merging is 150m, which is equivalent to the spatial span of the split connected domain generated after a single wind turbine is affected by speckle, and is used to merge the fragments split from the same unit; the large-scale settling threshold Take 250m, slightly larger than the lower limit of typical unit spacing, to attach structural debris to the nearest core target without crossing adjacent units; (4) K is 3 in the anisotropic K nearest neighbors, dimensionless threshold The value of 5 is chosen because each unit in the wind power array usually has several equally spaced nearest neighbors under the normalized main axis frame. Choosing K=3 can robustly reflect the local array spacing. After normalization with the typical unit spacing, the distance of the Kth nearest neighbor of the normal unit is close to 1. Choosing the threshold of 5 can remove outliers that are significantly deviated from the array while retaining the normal array units. (5) The effective wind farm scale threshold is 25 units. This is based on the economic scale and grid connection capacity statistics of offshore wind farm construction. The number of units in a single commercial offshore wind farm is usually not less than this level. Based on this, pseudo wind farms formed by the accidental aggregation of a small number of strong scattering interferences are removed. The above parameters can be appropriately adjusted within the given range according to the sensor resolution, sea conditions and regional wind farm layout characteristics.

[0052] Example 2: The SAR offshore wind power remote sensing detection system based on structure-aware constraints provided in this embodiment of the invention includes: The data preparation module is used to acquire SAR images, perform radiometric processing and format conversion, and complete the annotation and training dataset construction based on ownership data. The preliminary detection module is used to train an offshore wind power detection network based on a deep learning target detection model, detect SAR images of the target sea area, and output preliminary detection boxes and center points. The candidate boundary generation module is used to identify wind farm clusters based on density-constrained clustering and generate candidate wind farm boundaries that accurately match the orientation of the wind farm array using the minimum rotation circumscribed rectangle. The candidate target extraction module is used to perform adaptive threshold binarization and connected component extraction within the candidate boundary, and to calculate the area, centroid, and maximum gray value of the candidate connected component. The fine recognition module is used to perform dual-scale clustering and merging, dynamic intensity threshold screening, and anisotropic K-nearest neighbor outlier removal, and output the detection results of offshore wind turbines. The anisotropic K-nearest neighbor constraint constructs a dimensionless distance index with the main axis direction of the wind farm array as a reference, which is adaptive to different sensor resolutions and different wind farm layouts.

[0053] like Figure 4 As shown, Figure 4 Figure (a) to Figure 4 Figure (e) in the figure shows the preliminary detection results of RetinaNet, Faster R-CNN, YOLOv5, YOLOv8, and YOLO26 on Sentinel-1 images. Figure 4Figure (f) shows the actual locations corresponding to the offshore wind power ownership data, with red markers indicating the wind turbine detection locations or actual locations. A comparison reveals that all five models can identify the main wind farm arrays, but there are still varying numbers of discrete false detection points outside the actual wind farm arrays. This indicates that relying solely on the local target discrimination of deep learning models is insufficient to eliminate strong scattering interference from ships, offshore platforms, etc., and subsequent wind farm spatial structure constraints need to be introduced for further filtering.

[0054] To demonstrate the reliability of the method, comparative experiments were conducted in the nearshore area of ​​the South China Sea, testing the detection results of five detectors (RetinaNet, Faster R-CNN, YOLOv5, YOLOv8, and YOLO26) before and after the introduction of structure-aware constraints. For ease of comparison and presentation of experimental results, the method of this invention is abbreviated as SA-OWTDet (Structure-Aware Offshore WindTurbine Detector), and Table 1 shows the detection results after incorporating the method of this invention in the form of model name-SA-OWTDet.

[0055] Without the method of this invention, the false positives (FP) of each model were high. After introducing the method of this invention, the FP of each model decreased to below 100, the precision increased to above 0.969, the recall stabilized at around 0.980, and the F1 score exceeded 0.973. Specifically, YOLO26 combined with the method of this invention achieved a TP of 2124, an FP of 60, an FN of 44, a precision of 0.973, a recall of 0.980, and an F1 score of 0.976. These results consistently demonstrate that the spatial structure constraint mechanism proposed in this invention can effectively suppress false detections without sacrificing recall, exhibiting higher precision and stronger robustness.

[0056] Table 1. Accuracy metrics of deep learning object detection models before and after introducing structure-aware constraints.

[0057] In summary, this invention systematically transforms the detection paradigm from single-target local discrimination to multi-target spatial structure constraint discrimination through a collaborative mechanism of deep learning initial detection, density-constrained clustering, minimum rotation circumscribed rectangle boundary generation, connected component feature filtering, and anisotropic K-nearest neighbor spatial constraints. A progressive constraint relationship is formed between the deep learning initial detection results, candidate wind farm boundaries, connected component features, and anisotropic spatial distances. The structural information generated in the previous step becomes the filtering basis for the next step, thus overcoming the limitations of existing single methods or simple concatenation methods that cannot simultaneously achieve high recall, false positive suppression, and cross-scenario generalization ability. This effectively solves the problem of high-precision detection of offshore wind turbines under complex sea conditions.

[0058] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A SAR remote sensing method for offshore wind power detection based on structure-aware constraints, characterized in that, This method is a collaborative discrimination mechanism combining deep learning initial detection and wind farm array spatial structure constraints. The collaborative discrimination mechanism uses the deep learning output as candidate seeds for wind farm spatial structure reconstruction, and utilizes the reconstructed farm-level boundary, array main axis direction, and typical turbine spacing as inverse constraints for single-unit target selection. Specifically, it includes the following steps: S1. Training dataset construction: SAR images are acquired based on offshore wind power ownership data, converted into grayscale images for annotation, and a training dataset is constructed. S2. Preliminary detection result generation: The offshore wind power detection network is trained based on the training dataset using a deep learning target detection model to detect SAR images of the target sea area and output preliminary detection boxes and center points. The preliminary detection boxes and center points exist in the form of a noisy candidate set with high recall, which is used to provide input for subsequent wind farm spatial structure restoration and is not directly used as the final wind turbine detection result. S3. Generation of candidate wind farm boundaries: Project the output preliminary detection box and center point to the UTM coordinate system, identify wind farm clusters based on density-constrained clustering method and remove isolated strong scattering interference points, construct the candidate wind farm boundary with minimum rotation circumscribed rectangle based on the spatial distribution of clusters, transform the noisy preliminary detection results into wind farm-level spatial structure priors, and strictly limit the subsequent fine identification process to the area with wind power array spatial structure; S4. Image binarization and candidate target extraction: Under the boundary constraints of candidate wind farms, adaptive threshold binarization and connected component extraction are performed on SAR images, and the area, centroid and maximum gray value of candidate connected components are calculated. S5. Fine-tuning of candidate targets and output of results: The candidate connected components are clustered and merged on two scales to construct a dynamic intensity threshold. Anisotropic K-nearest neighbor spatial consistency constraints with reference to the main axis direction of the wind farm array are used to remove outliers and low-intensity noise targets that do not conform to the arrangement of the wind farm array, so as to obtain the detection results of offshore wind turbines.

2. The SAR offshore wind power remote sensing detection method based on structure-aware constraints according to claim 1, characterized in that, In step S1, the construction of the training dataset includes: Radiometric calibration, geometric correction, and speckle removal are performed on the original SAR image to generate a backscatter intensity image in dB. Based on the sensor's polarization characteristics, a co-polarization channel is selected, and the backscattering intensity image is converted into a grayscale image of a unified format for model training and detection.

3. The SAR offshore wind power remote sensing detection method based on structure-aware constraints according to claim 1, characterized in that, In step S2, the deep learning target detection model includes YOLO, RetinaNet, Faster R-CNN, or a Transformer-based detection model. After training, the target sea area SAR image is converted into a grayscale image and input into the detection network, which outputs a preliminary detection box and its corresponding center point. The output of step S2 is only used as a candidate set, and the final identification is completed by the spatial structure constraint mechanism from step S3 to step S5.

4. The SAR offshore wind power remote sensing detection method based on structure-aware constraints according to claim 1, characterized in that, In step S3, the generation of candidate wind farm boundaries includes: The obtained preliminary detection boxes and center points are projected onto the UTM coordinate system, and clustering is performed using the DBSCAN density constraint algorithm; the neighborhood and core point determination conditions satisfy the following formulas respectively: ； ； In the formula, For a neighborhood set of points, when satisfying hour, Considered the core point; For set X, divide Any preliminary wind power testing point outside, This is the set of all preliminary wind power detection points in the UTM coordinate system. These are preliminary wind power detection points in the UTM coordinate system. The maximum neighborhood distance between target points within the same cluster. The minimum number of targets required to form an effective cluster; After obtaining the clusters, based on the point sets within the clusters Construct the minimum rotational bounding rectangle such that the latitude and longitude extrema of the clusters satisfy the following conditions: ； ； ； ； In the formula, These are the maximum and minimum longitude values, respectively. These are the latitude maximum and minimum values, respectively. The first in the cluster Longitude of the initial wind power testing points The first in the cluster Latitude of the initial wind power testing points; An initial axis-aligned bounding rectangle is constructed using the latitude and longitude extrema of the clusters, serving as an initial reference for rotation optimization; the rotation angle is then adjusted. Perform a traversal search to determine the optimal rotated bounding rectangle based on the minimum area criterion; The candidate wind farm boundary is obtained by setting an expansion buffer based on the minimum rotational circumscribed rectangle.

5. The SAR offshore wind power remote sensing detection method based on structure-aware constraints according to claim 1, characterized in that, In step S4, image binarization and candidate target extraction include: Under the boundary constraints of candidate wind farms, adaptive threshold binarization is performed on the cropped SAR image. First, the SAR image is normalized to a set grayscale mapping range. and adopt Gaussian filtering is used to smooth the normalized grayscale image to reduce the impact of speckle noise and local grayscale abrupt changes on the binarization result; let the normalized and smoothed SAR grayscale image be... Then the pixel The local adaptive threshold at point is defined as: ； In the formula, For pixels The corresponding local adaptive threshold, In pixels Centered Within a local window The average gray level, Let be the threshold offset constant. The local adaptive threshold is expressed as: ； Based on the aforementioned local adaptive threshold, the binarization function is defined as: ； In the formula, This is the binarized result; when When, it indicates that the pixel is determined to be a foreground pixel; when When this occurs, it indicates that the pixel is determined to be a background pixel; After obtaining the binarization result, connected component labeling is performed on the foreground pixel set, let... For the first If there are 1 connected component, then the pixel area of ​​that connected component is... The calculation formula is: ； In the formula, This represents the number of foreground pixels contained in the connected component.

6. The SAR offshore wind power remote sensing detection method based on structure-aware constraints according to claim 5, characterized in that, In step S5, the fine-tuning of candidate targets and the output of results include: (1) Dual-scale clustering and merging: The small-scale DBSCAN with a search neighborhood radius of 150m is used to merge spatially adjacent split connected components; the connected components are divided into core targets and fragment targets according to their area size, and the fragment targets are assigned to the nearest core target. The core targets and the assignment conditions are respectively satisfied as follows: ； ； In the formula, It is the connected component with the largest area within the cluster. These are the centroid coordinates of the fragment and the core target, respectively. This is the threshold for large-scale search neighborhood; This is the Euclidean distance function in the UTM plane coordinate system. This represents the Euclidean distance between the centroids of the fragment target and the centroid of the core target, in meters (m). (2) Dynamic intensity threshold construction: The dynamic intensity threshold is constructed based on the statistical distribution of the maximum gray value of the candidate target. ； ； In the formula, For dynamic intensity threshold, The minimum intensity threshold, This is the strength correction factor. The 5th percentile of the set, For the first The maximum grayscale value of each candidate target The set consisting of the maximum grayscale values ​​of all candidate targets. represents the total number of candidate targets obtained after dual-scale clustering and merging; where, , When the maximum grayscale value of the candidate target is lower than When this happens, the candidate target is discarded as a low-intensity noise target; (3) Anisotropy Nearest Neighbor Spatial Consistency Constraint: The centroid of the candidate target is transformed from the UTM coordinate system to the main axis frame of the wind farm array, and the transformation relationship satisfies the following formula: ； In the formula, For the first Coordinates of each candidate target in the UTM coordinate system The coordinates of the rotated main shaft frame are... The rotation angle of the candidate wind farm boundary; spindle direction Aligned with the main direction of the wind farm array, Perpendicular to the main direction of the wind farm array; Within the main shaft frame, in Euclidean... Nearest Neighbor Sets Construct axial displacement samples in the principal axis and vertical directions: ； ； In the formula, It is a set of axial displacement samples consisting of the absolute coordinate differences between each candidate target and its Euclidean K-nearest neighbor along the principal axis. It is a set of axial displacement samples consisting of the absolute coordinate differences in the vertical direction between each candidate target and its Euclidean K-nearest neighbor. For the first Coordinates of the nearest neighbor candidate targets within the main axis frame; This represents the absolute value operation. and They represent the i-th Candidate targets and their first The non-negative axial spacing of the nearest neighbor candidate targets in the principal axis and vertical direction, in meters; If the sample is truncated at the 10% to 90% quantile, then: ； ； In the formula, To The set of principal axis displacement samples obtained after truncation of 10% to 90%. To The set of vertical displacement samples obtained after truncation of 10% to 90% quantile. For any sample value in the corresponding axial displacement sample set, Main axis direction displacement sample set The 10th percentile, A set of vertical displacement samples The 10th percentile, Main axis direction displacement sample set The 90th percentile, A set of vertical displacement samples 90th percentile; On the truncated sample, the 75th percentile is used as the estimate of typical unit spacing in both directions: ； ； In the formula, These are the typical unit spacings in the main shaft direction and the vertical direction, respectively; Anisotropic normalization of coordinates is performed using the typical turbine spacing as a scale to construct an anisotropic distance metric that reflects the structure of the wind turbine array: ； In the formula, For any two candidate targets and Relative spatial intervals within a normalized principal axis framework; For each candidate target, calculate the anisotropic distance of its Kth nearest neighbor. Candidate targets exceeding the dimensionless threshold are identified as outliers and removed. ； ； In the formula, For the first The candidate target is sorted by its anisotropic normalized distance from other candidate targets. Small distance, i.e., the first Nearest neighbor distance; For all An operator that sorts the anisotropic normalized distances in ascending order. For the number of nearest neighbors, The threshold for determining a dimensionless outlier. For the first One candidate target; (4) Determination and Result Output of Effective Wind Farms: Effective wind farms are determined based on the number of wind turbines within the candidate wind farm area. ； ； In the formula, For the first One candidate wind farm area, Candidate wind farm area The final set of wind turbine targets retained after meticulous screening of candidate targets in the Inner Canon; The cardinality of a set is the number of elements contained in the set. Indicates the first The final number of wind turbines retained within each candidate wind farm area; when At that time, The offshore wind farm is deemed a valid offshore wind farm; this threshold is designed based on the economic effects and scale statistics of offshore wind farm construction; the centroid of the connected domain is used as the output of the wind turbine center location, and the wind turbine center location satisfies: ； In the formula, For the first The final retained wind turbine connectivity domain, This represents the number of pixels contained in the connected component. The coordinates of the center of latitude and longitude of the connected domain. For connected components The latitude and longitude coordinates of any pixel within the range. For the first The final set consists of all pixels within the connected components.

7. The SAR offshore wind power remote sensing detection method based on structure-aware constraints according to claim 6, characterized in that, The nearest neighbor number K is 3, and the dimensionless outlier determination threshold is... The value is 5.

8. A SAR offshore wind power remote sensing detection system based on structure-aware constraints, characterized in that, The system is used to execute the SAR offshore wind power remote sensing detection method based on structure-aware constraints as described in any one of claims 1 to 7, the system comprising: The data preparation module is used to acquire SAR images, perform radiometric processing and format conversion, and complete the annotation and training dataset construction based on ownership data. The preliminary detection module is used to train an offshore wind power detection network based on a deep learning target detection model, detect SAR images of the target sea area, and output preliminary detection boxes and center points. The candidate boundary generation module is used to identify wind farm clusters based on density-constrained clustering and generate candidate wind farm boundaries that accurately match the orientation of the wind farm array using the minimum rotation circumscribed rectangle. The candidate target extraction module is used to perform adaptive threshold binarization and connected component extraction within the candidate boundary, and to calculate the area, centroid, and maximum gray value of the candidate connected component. The fine recognition module is used to perform dual-scale clustering and merging, dynamic intensity threshold screening, and anisotropic K-nearest neighbor outlier removal, and output the detection results of offshore wind turbines.

9. The SAR offshore wind power remote sensing detection system based on structure-aware constraints according to claim 8, characterized in that, The anisotropic K-nearest neighbor constraint in the fine recognition module constructs a dimensionless distance index with reference to the main axis direction of the wind farm array, which is adaptive to different sensor resolutions and different wind farm layouts.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the SAR offshore wind power remote sensing detection method based on structure-aware constraints as described in any one of claims 1 to 7.

Images