A wind turbine extraction method and device combined with graph attention network

Abstract

The present invention discloses a method and device for extracting a wind turbine in combination with a graph attention network, and belongs to the technical field of detection of new energy power generation facilities. It includes: constructing a posture graph of a wind turbine, taking the hub, base, and hub shadow of the wind turbine as key points in the graph model respectively, and taking the fan body connecting the hub and the base, and the fan body shadow connecting the hub shadow and the base as the edges of the graph model to form a spatial graph; modeling a wind turbine graph attention network model, using the YOLOv11 network as the target detection framework to establish a model, adding a graph attention layer to the model, using the spatial graph as an adjacency matrix for the graph attention layer, and using the graph attention layer to integrate the posture semantic information of the wind turbine into the detection head; constructing a sample data set, and constructing a remote sensing image data set of a wind turbine. The present invention makes better use of the characteristic information of the wind turbine itself, and provides a technical approach for achieving more accurate key point positioning and detection of wind turbines.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy power generation facility detection, and in particular, relates to a wind turbine extraction method and device combined with a graph attention network. Background Art

[0002] Wind energy is an environmentally friendly energy source in modern society and an important component of renewable energy. The use of wind energy helps mitigate the greenhouse effect and promote the realization of sustainable development goals. In recent years, the wind energy industry has developed. The advancement of the Sustainable Development Goals (SDGs) is expected to further expand the scale of the wind energy industry, and the location and number of wind turbines are key factors in evaluating the effectiveness of wind farms, predicting the annual power generation of wind farms, optimizing the layout of wind farms, and evaluating the development potential of wind power projects. It is essential for monitoring the development status of the wind energy industry and evaluating wind power generation production capacity.

[0003] Wind turbines consist of blades and cylindrical towers made of metal materials. In optical remote sensing images, wind turbines appear as white highlights and unique spatial structures. As a typical feature with unique morphological characteristics in remote sensing images, wind turbines have the characteristics of wide distribution range and complex distribution background. When extracting wind turbines from remote sensing images, there are problems such as poor extraction accuracy and weak generalization ability.

[0004] In the past, remote sensing extraction algorithms for wind turbines mainly included three methods: segmentation, positioning, and detection. In terms of segmentation, many scholars have used segmentation technology to extract wind turbine contours from remote sensing images. In the prior art, some people use traditional salient target detection methods to segment wind turbines in Google Earth images of the selected study area with the wind turbine body as the target, and obtain binary results. In the prior art, some people have developed a target U-Net model for segmenting wind turbines in five Gaofen-2 images, in which shadow features are used to locate wind turbines. This positioning method predicts the basic position of wind turbines in the image. In terms of positioning, in the prior art, some people have built five weakly supervised model structures based on the same backbone, namely class activation mapping (CAM), improved gradient weighted class activation mapping (Grad-Cam++), soft proposal network (SPN), weakly supervised learning of deep convolutional neural networks (WILDCAT), and peak response mapping (PRM), and compared the performance of these five structures on high-resolution satellite images. WILDCAT is the most effective weakly supervised structure for wind turbine positioning. In the prior art, some people applied adaptive threshold segmentation, morphological operations and centroid calculation to preprocessed Sentinel-1 SAR images to locate offshore wind turbines around the world. In terms of detection, some people built a fully automatic quantitative atmospheric correction (QUAAC) link for high-resolution optical images, introduced QUAAC in the GF-2 image preprocessing process, and realized accurate recognition of wind turbine targets in high-resolution remote sensing images based on the YOLOv5-CBAM model. Others proposed an improved YOLOv5 windmill extraction model based on multi-scale features of remote sensing images and small target detection.

[0005] In recent years, wind turbine extraction technology combining positioning and detection methods has become increasingly popular. In the prior art, a method WT-YOLO based on the YOLOv5 model is proposed to detect and locate wind turbines in remote sensing images. This method uses the hub, base and hub shadow of the wind turbine as key points and incorporates them into the head regression term of the YOLOv5 basic framework to regress and locate the positions of these three key points. The base is used to determine the exact position of the wind turbine, and finally the specific position of the wind turbine is detected. Within the spatial resolution range of 0.6 m ~ 5.4 m, WT-YOLO's detection capability for wind turbines has been enhanced, and the average precision (AP) has been greatly improved compared with the existing wind turbine extraction methods. However, the model has a large error in key point positioning, especially the deviation in the positioning of the shadow hub point needs to be further improved.

[0006] Existing technologies have not fully considered the unique attitude semantic information and contextual information of wind turbines in remote sensing images. Due to the characteristics of vertical orthophotos, the attitude structural features of the wind turbine fuselage and its fuselage shadow in remote sensing images are relatively stable. Compared with simple visual features, making full use of wind turbine attitude features is conducive to improving the accuracy of wind turbine target detection under complex backgrounds. At the same time, because traditional convolutional neural networks are sensitive to changes in factors such as rotation and scale, and have a small receptive field, they are prone to losing the semantic information and contextual information of wind turbine features, resulting in the problem of difficulty in extracting wind turbine structural features. Summary of the invention

[0007] The purpose of the present invention is to overcome the above shortcomings of the prior art by introducing a spatial graph representing the connection relationship between the key points of the wind turbine, and combining the graph attention network (GAT), a deep learning technology that focuses on graph structure, to make full use of the posture information characteristics of the wind turbine, combined with target box regression and key point positioning, so as to improve the positioning and extraction accuracy of wind turbines in complex backgrounds.

[0008] The present invention uses the graph attention module to aggregate the wind turbine attitude information composed of the key points of the wind turbine, and splices it with the feature extraction results of the Backbone part of YOLOv11 to achieve feature extraction. The regression results of the target frame and key point positioning of the wind turbine are obtained through the forward propagation of the neural network. The loss is calculated based on the regression results and the real target frame and shutdown point coordinates, and the loss is back-propagated layer by layer to complete the model training. The positioning method of the key points of the wind turbine is further extended to the wind turbine attitude estimation task, while considering the key point features and the semantic features of the wind turbine, and effectively combined with the graph attention network to participate in the training process of the model, aiming to make the model fully utilize the semantic feature information of the wind turbine, so as to achieve higher positioning and extraction accuracy.

[0009] The technical solution of the present invention is: a wind turbine extraction method combined with a graph attention network, comprising the following steps:

[0010] Step 1, constructing a wind turbine posture graph, including: taking the hub, base, and hub shadow of the wind turbine as key points in the graph model respectively, taking the wind turbine body connecting the hub and the base, and the wind turbine body shadow connecting the hub shadow and the base as edges of the graph model, and setting the edge weights as learnable parameters to form a spatial graph;

[0011] Step 2, modeling a wind turbine graph attention network model, including: establishing a wind turbine graph attention network model using a YOLOv11 network as a target detection framework, adding a graph attention layer to the wind turbine graph attention network model, using the spatial graph as a spatial weight matrix for the graph attention layer, and using the graph attention layer to integrate the posture semantic information of the wind turbine into the detection head;

[0012] Step 3, sample data set construction, including: constructing a wind turbine remote sensing image data set, including six types of backgrounds, namely farmland, forest, terrain, water, desert and grassland, using annotation software to annotate the wind turbines in the image with detection boxes, when selecting the boxes, the shadows of the wind turbines are also selected, and the key points of the wind turbine hub, base, and hub shadow are respectively annotated.

[0013] A wind turbine extraction device combined with a graph attention network, comprising the following modules:

[0014] A spatial graph forming module is used to take the hub, base, and hub shadow of the wind turbine as key points in the graph model, take the wind turbine body connecting the hub and the base, and the wind turbine body shadow connecting the hub shadow and the base as edges of the graph model, and set the weight of the edge as a learnable parameter to form a spatial graph;

[0015] Wind turbine graph attention network model, using YOLOv11 network as the target detection framework to establish a wind turbine graph attention network model, adding a graph attention layer to the wind turbine graph attention network model, using the spatial graph as the adjacency matrix for the graph attention layer, and using the graph attention layer to incorporate the posture semantic information of the wind turbine into the detection head;

[0016] The sample dataset construction module constructs a wind turbine remote sensing image dataset, which includes six types of backgrounds, namely farmland, forest, terrain, water, desert and grassland. The wind turbines in the image are annotated with detection boxes using annotation software. When selecting the box, the shadow of the wind turbine is also selected, and the key points of the wind turbine hub, base and hub shadow are marked separately.

[0017] An electronic device comprises: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned wind turbine extraction method combined with a graph attention network.

[0018] A computer-readable storage medium stores executable instructions thereon, which, when executed by a processor, enable the processor to implement the above-mentioned wind turbine extraction method combined with a graph attention network.

[0019] The present invention has the following beneficial effects:

[0020] Based on the unique posture information of wind turbines in remote sensing images, the present invention studies the component posture structure diagram of modeling wind turbines, and uses the graph attention network to capture the long-distance dependency characteristics of wind turbines in remote sensing images. Based on the wind turbine extraction model WT-YOLOGAT fused with YOLOv11 and graph attention network, the present invention models the topological relationship between the key points of the wind turbine fuselage, and uses the graph attention network to effectively utilize the semantic information of the wind turbine in the remote sensing image. The present invention establishes a spatial graph representing the spatial relationship between the three key points of the wind turbine, and combines the graph attention network (GAT), a deep learning technology focusing on graph structure, to comprehensively consider the influence of the posture semantic information of the wind turbine on the detection effect of the wind turbine, forming a new method that is different from only considering the prediction of the wind turbine fuselage, only considering the wind turbine fuselage and shadow, only considering the wind turbine fuselage and shadow, and key point positioning, so as to better utilize the characteristic information of the wind turbine itself, and provide a technical approach for achieving more accurate wind turbine key point positioning and detection.

[0021] The present invention introduces attitude estimation and graph attention methods into the wind turbine extraction task. The present invention believes that when extracting wind turbines, it is necessary to consider the unique attitude information of wind turbines in remote sensing images, while the existing extraction technology does not take this information into account. Because traditional convolutional neural networks are sensitive to changes in factors such as rotation and scale, and have a small receptive field, they are prone to losing semantic information and contextual information of wind turbine features, resulting in the problem of difficulty in effectively extracting wind turbine structural features.

[0022] Based on the YOLOv11 model, the present invention constructs a spatial graph according to the connection relationship between the key points of the wind turbine, aggregates the posture semantic information of the wind turbine with the help of the graph attention module, and splices the key point feature information with the feature information extracted by Backbone to realize feature extraction, aiming to effectively use the wind turbine feature information to improve the extraction accuracy of the wind turbine, and at the same time improves the corresponding loss function. The total loss function (Loss) includes the target box loss (Bbox.Loss), the key point loss (Kpts.Loss) and the target information loss composed of the key points (Kpts.obj.Loss). BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is the spatial adjacency matrix of the key points of the fan Get the schematic diagram;

[0024] Figure 2 is the spatial weight matrix Calculation diagram;

[0025] Figure 3Schematic diagram of the attention network model for a wind turbine graph;

[0026] Figure 4 Schematic diagram of the detection head structure. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned purpose, the present invention adopts the following technical scheme.

[0028] The present invention proposes a wind turbine extraction method combined with a graph attention network, comprising the following steps:

[0029] Step 1, constructing a posture graph of a wind turbine, including: using the hub, base, and hub shadow of the wind turbine as key points in the graph model, respectively, using the fan body connecting the hub and the base, and the fan body shadow connecting the hub shadow and the base as the edges of the graph model, and setting the weights of the edges as learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for a subsequent graph attention module, and can effectively aggregate the semantic feature information of the wind turbine according to the weights to enhance the head component of YOLO's ability to extract wind turbines in remote sensing images.

[0030] Step 2, modeling a wind turbine graph attention network model, including: establishing a wind turbine graph attention network model using YOLOv11 as a target detection framework, using a spatial graph as a spatial weight matrix for the wind turbine graph attention network model, and using the wind turbine graph attention network model to integrate the posture semantic information of the wind turbine into the detection head, extracting the posture information of the wind turbine in the remote sensing image, thereby reducing the positioning error of the key points of the wind turbine and improving the extraction accuracy of the wind turbine in a complex background.

[0031] Step 3: Sample dataset construction, including: constructing a high-resolution wind turbine remote sensing image dataset, including six types of backgrounds, namely farmland, forest, terrain, water, desert and grassland. The data source is Google Earth, with a spatial resolution of 0.6m. Labelme annotation software is used to annotate the wind turbines in the image with detection boxes, a total of 353 images, for model training, verification and testing. It should be noted that the shadow of the wind turbine needs to be selected when selecting the box, and the three key points of the wind turbine hub, base and hub shadow are marked separately. These datasets are key resources in the research process.

[0032] Step 1 specifically includes: taking the hub, base, and hub shadow of the wind turbine as key points in the graph model, and defining a matrix representing the spatial adjacency relationship of the three key points with the wind turbine body connecting the hub and the base, and the wind turbine body shadow connecting the hub shadow and the base as edges respectively. ,like Figure 1 Matrix 1 indicates that the key points are connected by edges, and 0 indicates that the key points are not connected, such as the matrix The 1 in the first row and first column indicates the key point Connected to itself, the 1 in the first row and second column indicates a key point With key points Connected, the 0 in the first row and third column indicates the key point With key points Not connected.

[0033] ,

[0034] like Figure 2 As shown, the edge weights are then set as learnable parameters , defines a parameterized matrix Represents the weight relationship between key points:

[0035] ,

[0036] Matrix representing spatial adjacency The parameterized matrix of the relationship between and key points Hadamard

[0037] The product result is the final spatial weight matrix ,in Represents the Hadamard product operation.

[0038] ,

[0039] Step 2 includes: adding the graph attention layer GAT to the detection head of the YOLOv11 model to build a wind turbine graph attention network model to achieve higher accuracy wind turbine extraction.

[0040] The wind turbine graph attention network model structure is as follows Figure 3As shown in the figure. It consists of three basic components: the backbone network, the neck component, and the head component. First, the backbone component acts as a feature extractor and uses a convolutional neural network to convert the original image data into a multi-scale feature map. The backbone network of YOLOv11 consists of a convolutional neural network (CNN), which can extract multi-scale feature maps of the image. These feature maps contain spatial information at different levels. CNN extracts local features in the image layer by layer by stacking multiple convolutional layers and nonlinear activation functions, and generates feature maps of different resolutions at different levels. For the specific task of wind turbine recognition, the convolutional neural network can recognize the rough shape and position in the image and generate feature maps containing spatial information at different levels. Secondly, the neck component of YOLOv11 fuses feature maps from different levels so that the model can capture local details and global information in the image at the same time. This process is achieved through the feature pyramid network FPN or the path aggregation network PANet, in which the feature pyramid network FPN amplifies the low-resolution feature map through upsampling operations and cascades it with the high-resolution feature map to enhance the model's multi-scale feature learning ability. Finally, the features of different scales are combined and passed to the detection head component for prediction. PANet is an improvement on FPN. On the basis of FPN, a bottom-up path is added to further enhance the feature fusion capability. In the wind turbine detection task, the role of the neck component is particularly important, because the wind turbine is composed of multiple key parts (such as blades, shafts, etc.), and the model needs to be able to capture these different scales of information. The neck component of this model is the path aggregation network PANet. Finally, the head component is responsible for generating the final prediction in terms of object detection and classification as a prediction mechanism. The head component consists of three parallel detection heads. Each detection head processes the feature map passed from the neck component and finally outputs the bounding box and category label of the object in the image, showing the location and classification of the object. The output of this part includes information such as bounding box and key points, which are used to detect targets in the image. The wind turbine graph attention network model adds two layers of graph attention layers GAT for the key point location of the wind turbine and the target information composed of the key points in the detection head. The graph attention layer GAT is placed at the front end of the detection head to receive the multi-scale feature maps from the neck component and further process these feature maps. The graph attention layer models the adjacency relationship of each key point using the spatial weight matrix to calculate the attention weights between key points. Then, the information of adjacent key points is propagated to the target key point by weighted averaging, and finally the enhanced key point features are output. This process enables the model to model and detect different parts of the wind turbine more accurately. The role of the graph attention layer is not limited to the weighted adjustment of the feature map. More importantly, it models the relationship between key points through the graph structure to form an effective dependency network. In the detection process of wind turbine targets, the graph attention layer can identify and strengthen the relationship between key parts such as the wind turbine hub, wind turbine base, and wind turbine hub shadow. Specifically, the graph attention mechanism adjusts the attention weights according to the similarity between key points, and dynamically updates the representation of key points through these weights, so that the key point information is more prominent. In this way, the graph attention layer GAT can help the network model focus on the positioning of key parts of the wind turbine and reduce attention to irrelevant background. During the training process, the spatial weight matrix Add to the graph attention layer GAT, such as Figure 4 The calculation formula of the graph attention (GAT) in the graph attention layer is shown in formulas (1)-(4) to calculate the key points Its adjacent key points Take the attention coefficient and weighted sum as an example.

[0041] (1)

[0042] (2)

[0043] (3)

[0044] (4)

[0045] in, Indicates key points In the The feature vector of the layer, Indicates that key point j is in The feature vector of the layer, It means that the key point after convolution is The feature vector of the layer, represents the activation function, represents the attention score between keypoint M and keypoint j. Indicates The shared parameters of the layer convolution, Indicates the use of shared parameters right Perform a linear transformation, Indicates the key points and ( ) The transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. It is a kind of activation function.

[0046] Figure 4 In the example, Bbox.Loss stands for bounding box regression loss, which is used to evaluate the difference between the predicted bounding box and the true bounding box. Kpts.Loss stands for keypoint regression loss, which is used to evaluate the difference between the predicted keypoints and the true keypoints. Kpts.obj.Loss stands for keypoint object existence loss, which is used to evaluate the confidence of the object constituted by the predicted keypoints. GAT stands for graph attention layer, which helps the model better understand the feature dependencies between the keypoints of the wind turbine by establishing the spatial relationship between the keypoints of the wind turbine. Conv2d layer stands for standard convolution layer, which is used to perform convolution operations on the output from GAT to extract higher-level features and optimize spatial understanding. Conv2d layer further refines features based on the relationship information extracted by GAT and passes these features to the final object detection task, i.e. bounding box regression, keypoint localization and keypoint object existence evaluation. The total loss function (Loss) includes the target box loss (Bbox.Loss), the key point loss (Kpts.Loss) and the key point target existence loss (Kpts.obj.Loss), as shown in formula (5):

[0047] (5)

[0048] in, , , are predefined weight parameters.

[0049] Step 3 includes:

[0050] A total of 353 wind turbine images with a spatial resolution of 0.6m were collected in different backgrounds such as farmland, forest, terrain, water, desert and grassland. There is at least one wind turbine in each image.

[0051] According to the above method, 353 samples were constructed and divided into training set, validation set and test set in the ratio of 8:1:1. The training set is used to train the model; the validation set is used to adjust the model hyperparameters; and the test set is used to test the final accuracy of the model.

[0052] Model input:

[0053] (1) Wind turbine sample dataset.

[0054] (2) Spatial weight matrix based on the connection relationship between key points of wind turbines .

[0055] Model Tags:

[0056] (1) Use the Labelme tool to label the bounding box coordinates and three key point coordinates of the wind turbines in the sample dataset in YOLO format.

[0057] When the wind turbine graph attention network model is applied, the data set to be predicted is input into the trained model for forward propagation to obtain the wind turbine extraction result of the data set.

Claims

1. A wind turbine extraction method combined with a graph attention network, characterized in that: The following steps are involved: Step 1, constructing a wind turbine posture graph, including: taking the hub, base, and hub shadow of the wind turbine as key points in the graph model respectively, taking the wind turbine body connecting the hub and the base, and the wind turbine body shadow connecting the hub shadow and the base as edges of the graph model, and setting the edge weights as learnable parameters to form a spatial graph; Step 2, modeling a wind turbine graph attention network model, including: establishing a wind turbine graph attention network model using a YOLOv11 network as a target detection framework, adding a graph attention layer to the wind turbine graph attention network model, using the spatial graph as a spatial weight matrix for the graph attention layer, and using the graph attention layer to integrate the posture semantic information of the wind turbine into the detection head; Step 3, sample data set construction, including: constructing a wind turbine remote sensing image data set, including six types of backgrounds, namely farmland, forest, terrain, water, desert and grassland, using annotation software to annotate the wind turbines in the image with detection boxes, when selecting the boxes, the shadows of the wind turbines are also selected, and the key points of the wind turbine hub, base and hub shadow are respectively annotated; Step 4: When the wind turbine graph attention network model is applied, the data set to be predicted is input into the trained model for forward propagation to obtain the wind turbine extraction result of the data set.

2. The wind turbine extraction method combined with a graph attention network according to claim 1, characterized in that: Step 1 includes: taking the hub, base, and hub shadow of the wind turbine as key points in the graph model, and defining a matrix representing the spatial adjacency relationship with the wind turbine body connecting the hub and base, and the wind turbine body shadow connecting the hub shadow and base as edges. : ， Then set the edge weights as learnable parameters , define a parameterized matrix Represents the weight relationship between key points: ， Matrix representing spatial adjacency The parameterized matrix of the relationship between and key points The Hadamard product result is the final spatial weight matrix : 。 3. The wind turbine extraction method combined with a graph attention network according to claim 2, characterized in that: In step 2, the wind turbine image attention network model includes a backbone network, a neck component, and a head component; the backbone network acts as a feature extractor and uses a convolutional neural network to convert the original image data into multi-scale feature maps at different levels; the neck component fuses feature maps at different levels to capture local details and global information in the image; the head component acts as a prediction mechanism to generate final predictions in terms of detection and classification.

4. The wind turbine extraction method combined with a graph attention network according to claim 3, characterized in that: The backbone network consists of a convolutional neural network (CNN). The convolutional neural network (CNN) extracts local features from the image layer by layer by stacking multiple convolutional layers and nonlinear activation functions, and generates feature maps of different resolutions at different levels.

5. The method for extracting wind turbines in combination with a graph attention network according to claim 3, characterized in that: The neck component is implemented by the path aggregation network PANet. The path aggregation network PANet adds a bottom-up path based on the feature pyramid network FPN. The feature pyramid network FPN enlarges the low-resolution feature map through upsampling operation, cascades it with the high-resolution feature map, and finally combines the feature maps of different scales and passes them to the head component for prediction.

6. The wind turbine extraction method combined with a graph attention network according to claim 3, characterized in that: The head component consists of multiple parallel detection heads; each detection head processes the feature map passed from the neck component and ultimately outputs the bounding box and category label of the object in the image, showing the location and classification of the object.

7. The method for extracting wind turbines in combination with a graph attention network according to claim 6, characterized in that: The output of the detection head includes bounding boxes and key points, which are used to detect objects in the image. Two layers of graph attention layers GAT are added to the detection head for the location of wind turbine key points and the target information composed of key points. The graph attention layer GAT is placed at the front end of the detection head to receive multi-scale feature maps from the neck component and further process these feature maps. The graph attention layer GAT models the adjacency relationship of each key point and uses the spatial weight matrix To calculate the attention weights between key points; then, the information of adjacent key points is propagated to the target key point by weighted averaging, and finally the enhanced key point features are output.

8. The method for extracting wind turbines in combination with a graph attention network according to claim 7, characterized in that: The spatial weight matrix Add it to the graph attention layer GAT. The calculation formula of the graph attention in the graph attention layer is shown in formulas (1)-(4). Calculate the spatial weight matrix Key points in Its adjacent key points The attention coefficient and weighted sum: (1) (2) (3) (4) in, Indicates key points In the The feature vector of the layer, Indicates that key point j is in The feature vector of the layer, It means that the key point after convolution is The feature vector of the layer, represents the activation function, represents the attention score between key point M and key point j, Indicates The shared parameters of the layer convolution, Indicates the use of shared parameters right Perform a linear transformation, Indicates the key points and ( ) The transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. It is a kind of activation function; The corresponding loss function is shown in formula (5): (5) in, , , is a predefined weight parameter; represents the bounding box regression loss, which is used to evaluate the difference between the predicted bounding box and the true bounding box; represents the keypoint regression loss, which is used to evaluate the difference between the predicted keypoints and the true keypoints; Represents the key point target existence loss, which is used to evaluate the confidence of the target constituted by the predicted key point.

9. The method for extracting wind turbines in combination with a graph attention network according to claim 1, characterized in that: Step 3 includes: A number of wind turbine images with a spatial resolution of 0.6m were collected against the backgrounds of farmland, forest, terrain, water, desert and grassland to form a data set. There is at least one wind turbine in each image. The samples were divided into training set, validation set and test set in an 8:1:1 manner. The training set was used to train the model; the validation set was used to adjust the model hyperparameters; and the test set was used to test the final accuracy of the model.

10. The method for extracting wind turbines in combination with a graph attention network according to claim 1, characterized in that: The input of the wind turbine graph attention network model is: a sample wind turbine dataset; a spatial weight matrix based on the connection relationship between key points of wind turbines .

11. A wind turbine extraction device combined with a graph attention network, characterized in that: Includes the following modules: A spatial graph forming module is used to take the hub, base, and hub shadow of the wind turbine as key points in the graph model, take the wind turbine body connecting the hub and the base, and the wind turbine body shadow connecting the hub shadow and the base as edges of the graph model, and set the weight of the edge as a learnable parameter to form a spatial graph; Wind turbine graph attention network model, using YOLOv11 network as the target detection framework to establish a wind turbine graph attention network model, adding a graph attention layer to the wind turbine graph attention network model, using the spatial graph as the spatial weight matrix for the graph attention layer, and using the graph attention layer to integrate the posture semantic information of the wind turbine into the head component; The sample dataset construction module constructs a wind turbine remote sensing image dataset, which includes six types of backgrounds: farmland, forest, terrain, water, desert, and grassland. The wind turbines in the image are annotated with a detection box using annotation software. When selecting the box, the shadow of the wind turbine is also selected, and the key points of the wind turbine hub, base, and hub shadow are annotated separately. The result extraction module, when the wind turbine graph attention network model is applied, inputs the data set to be predicted into the trained model for forward propagation to obtain the wind turbine extraction result of the data set.

12. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement a wind turbine extraction method combined with a graph attention network as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement a wind turbine extraction method combined with a graph attention network as described in any one of claims 1 to 10.

Images