Mapping Wind Turbine Distribution in Forest Areas of China Using Deep Learning Methods

Abstract

Wind power plays a pivotal role in the achievement of carbon peaking and carbon neutrality. Extensive evidence has demonstrated that there are adverse impacts of wind power expansion on natural ecosystems, particularly on forests, such as forest degradation and habitat loss. However, incomplete and outdated information regarding onshore wind turbines in China hinders further systematic and in-depth studies. To address this challenge, we compiled a geospatial dataset of wind turbines located in forest areas of China as of 2022 to enhance data coverage from publicly available sources. Utilizing the YOLOv10 framework and high-resolution Jilin-1 optical satellite images, we identified the coordinates of 63,055 wind turbines, with an F1 score of 97.64%. Our analysis indicated that a total of 16,173 wind turbines were situated in forests, primarily within deciduous broadleaved forests (44.17%) and evergreen broadleaved forests (31.82%). Furthermore, our results revealed significant gaps in data completeness and balance in publicly available datasets, with 48.21% of the data missing and coverage varying spatially from 28.96% to 74.36%. The geospatial dataset offers valuable insights into the distribution characteristics of wind turbines in China and could serve as a foundation for future studies.

1. Introduction

The rapid increase in global carbon emissions driven by fossil fuel consumption poses a great challenge for climate change mitigation [1], causing extreme weather such as heat waves and floods [2,3,4]. Consequently, the reduction of carbon emissions within the energy sector has emerged as an urgent priority. In 2023, global renewables demonstrated an unprecedented development rate, with an addition of 565 GW [5]. As the world’s largest emitter and leader in the expansion of renewable energy, China has installed over 470 GW of wind power by 2024, surpassing its capacity targets for 2030 [6]. Wind power, recognized as a clean energy source, imposes a lower environmental burden compared to traditional fossil-fuel-generated power, highlighting its critical importance in sustainable electricity generation [7]. It is anticipated that wind power will continue to play a pivotal role in the future, with a projected target of 3200 GW by 2060 [8].

Although large-scale deployment of wind power is necessary, its potential adverse impacts on residents and the environment cannot be overlooked [9,10]; thus, a clear understanding of its spatial distribution is required [11]. Wind farms located in China’s forest areas are situated closer to cities with high electricity demand compared to the large wind power bases in northwestern provinces. This locational advantage, combined with abundant wind resources in mountainous regions, suggests a high theoretical potential for developing wind power in forest areas [12]. Nonetheless, the construction of wind farms inevitably causes disturbances to ecosystems, particularly forest ecosystems [13]. Increasing attention has been directed toward the impacts on vegetation and wildlife. For instance, the construction of wind turbines and associated infrastructure is likely to lead to forest loss, which subsequently results in increased soil erosion and forest fragmentation [14]. The cascading loss of natural habitats threatens wildlife and results in declines in biodiversity [15]. In addition to the construction phase, the operation of turbines may also directly cause concerning shifts in species distribution and collision mortality for mammals and birds [16]. Moreover, damage to ecosystems and biodiversity, particularly in the planet’s critical areas, may further reduce nature’s contributions to people in return [17]. Given the significant impacts highlighted by existing case studies, it is imperative to conduct a more comprehensive assessment of ecological impacts on a larger scale in the context of wind farm expansion. However, the availability of data regarding wind turbine distributions in forest areas of China remains a limitation for related research. Although OpenStreetMap (OSM) provides a wide range of valuable data on a global scale, including wind turbines, the incompleteness and spatial inequalities of these data due to the reliance on manual annotation may raise concerns regarding data quality among researchers [18,19]. Many turbine points are labelled in the OSM; however, there are cases where entire wind farms or parts of turbines within individual wind farms are missing. Consequently, the subsequent analysis associated with the data coverage may contain uncertainties. More efforts are needed to fill this gap.

With the advent of numerous high-resolution remote sensing (RS) images, there has been a notable increase in research initiatives focused on object interpretation. Compared to traditional methods that rely on manual visual interpretation or feature design, deep learning methods in computer vision show great advantages in efficiency and generalization [20,21]. Earlier object detection algorithms mainly include the You Only Look Once (YOLO) [22], YOLO 9000 [23], Single Shot MultiBox Detector (SSD) [24], and families of region-based convolutional network methods (R-CNN) [25,26,27]. In recent years, the YOLO series has undergone continuous development, leading to enhancements not only in network architecture [28,29] but also in the model’s applicability and scalability [30,31]. YOLOv8 achieved an important advancement, as it not only balanced optimal accuracy and speed, but also integrated multiple tasks such as semantic segmentation. Based on YOLOv8, YOLOv10 improved performance by eliminating non-maximum suppression (NMS), using a large-kernel convolution and lightweight self-attention module, etc. These improvements enabled it to outperform other state-of-the-art models, such as Gold-YOLO [32], Real-Time DEtection Transformer (RT-DETR) [33], and Real-Time Models for object Detection (RTMDet) [34]. Therefore, mapping distribution patterns utilizing satellite RS images and deep learning algorithms shows significant research value and potential [35,36,37]. YOLOseries-based object detection models have been extensively employed [20,38]. The spatial mapping of renewable energy facilities, including photovoltaics and offshore wind farms, has contributed to a more profound understanding of their development trends and established a foundational basis for subsequent studies [39,40].

Despite considerable advancements in global offshore wind turbine detection, there have been limited efforts directed toward large-scale mapping of onshore wind turbines [41,42]. Most research on onshore wind turbine interpretation consists of case studies, emphasizing methodological enhancements. For example, Zhang et al. assessed the feasibility of deep learning in the remote sensing interpretation of wind turbines by employing four object detection models, including Faster R-CNN [43]. The accuracy of each model reached 98% following post-processing. Manso-Callejo et al. used segmentation models, such as LinkNet architecture with an EfficientNet-b3 backbone, to extract onshore wind turbines [44]. Based on YOLOv5, Zhang et al. examined the impact of varying resolutions of Google Earth images on the model performance and found that adding keypoints could enhance average precision by 0.11% to 6.01% across image resolutions ranging from 1.2 m to 5.4 m [45]. However, fewer studies have applied models to large-scale detection [46]. This research gap can be attributed to the complexities inherent in the background and the resolution of publicly available satellite imagery. Offshore wind turbines are typically installed in regular patterns, making them relatively straightforward to identify based on the backscatter values of Sentinel-1 synthetic aperture radar images [47]. In contrast, the presence of significant background noise on land means that onshore wind turbines could only be accurately identified in high-resolution images [45]. Case studies have proven the feasibility of high-quality Google Earth imagery in various object detection tasks [43]. However, the temporal inconsistencies associated with Google Earth images, especially at the national level, may introduce biases in pattern analysis. The timing of image tiles in different regions of China vary by as much as 20 years, potentially affecting the accuracy of the analysis [48]. To address these limitations, time-consistent products, such as Jilin-1 satellite images, present a promising alternative. The Changguang Jilin-1 optical RS images offer global coverage, characterized by high image quality and a maximum resolution of 0.5 m, rendering them appropriate for remote sensing interpretation of relatively small objects, such as wind turbines [49].

Therefore, we aim to (1) establish a wind turbine interpretation framework and obtain task-specific deep learning models based on YOLOv10, (2) develop a comprehensive wind turbine distribution dataset within the forest areas of China to supplement existing publicly available data, and (3) analyze the spatial characteristics of wind turbine distribution. The subsequent sections of the article are organized as follows: Section 2 describes the study area, data sources, and research methods. Section 3 presents the results, which include the accuracy assessment, the visualization map of detected wind turbines, and an analysis of spatial distribution characteristics. Section 4 provides a discussion regarding the main findings, significance, and limitations of the study. Section 5 is the conclusion.

2. Materials and Methods

2.1. Study Area

We identified study areas based on existing publicly available data regarding forest regions [50], wind turbine distribution [51,52], and wind farm distribution [53] (Figure 1). First, we filtered forest areas in China, which cover approximately 2.29 million square kilometers and span from tropical to temperate regions [50]. By establishing buffer zones and conducting overlay analysis, areas of interest were narrowed down. Specifically, to account for potential inaccuracies associated with the coordinates of wind farms reported in the Global Energy Monitor (GEM) [53], we applied a buffer zone of 10 km around each wind farm site, the size of a typical wind farm. Similarly, a buffer zone of 5 km was created for each wind turbine point obtained from OpenStreetMap (OSM) to address possible missing data (Figure 1a,c) [51]. Subsequently, we established a fishnet grid comprising 0.3° × 0.3° cells that covers the entire land area of China. Finally, after selecting fishnet grids containing forest areas and interests with the designated two buffer zones, we identified 1899 grids covering an area of 1.75 million square kilometers (Figure 1b).

2.2. Data Sources

2.2.1. Satellite Imagery and Other Datasets

We used Changguang Jilin-1 satellite images for the detection task. The Jilin-1 optical remote sensing imagery provides global coverage with a maximum resolution of 0.5 m and has been utilized in environmental monitoring and disaster assessment [49]. Considering data availability and detection efficiency, this research used satellite images with a resolution ranging from 1.5 to 2.0 m for the year 2022. Additionally, for the pattern of wind resource utilization, we used wind speed data at 100 m above ground level [54]. To analyze the land cover type occupied by turbines, we adopted GLC_FCS30D, a finely classified global dataset that contains 35 land cover categories spanning the years 1985 to 2022 [55]. The land cover data for the year 2000 and 2022 were used.

2.2.2. Image Processing

To obtain the best models, we generated two datasets using satellite imagery: one for model tuning during the training stage and the other for additional evaluation. The first dataset, containing 35 regions and approximately 2% of the total study area, was used for model training, validation, and testing (Figure 2a). After conducting a quality assessment, we cropped each satellite image into small tiles (640 × 640 pixels). To enhance detection and positioning accuracy, each wind turbine was labeled with a bounding box and three keypoints, facilitating the differentiation of distinctive features from the surrounding background (Figure 2b) [45]. The bounding box covers all visible wind turbine components within a tile, including the tower, blades, shadow, etc. The three keypoints represent the base, hub, and shadow of the hub of a wind turbine, respectively. We used Labelme, a tool widely used for constructing deep learning datasets, to perform labeling [56]. After filtering images without wind turbines, the training, validation, and testing sets comprised 1179, 294, and 294 images, respectively. The datasets collectively included a total of 2728 bounding boxes, 2715 base points, 2704 hub points, and 2623 shadow points.

An additional dataset was utilized to evaluate the model’s external validity (Figure 2a). Due to the similarity between the testing set and the training set, there was potential bias in the accuracy assessment during the training phase. To address the issue, we randomly selected one-tenth of the satellite images within the study area to generate an evaluation set, which is independent of the training, validation, and testing sets. A total of 8503 wind turbines, representing 26% of the publicly available data, were labeled as ground truth to evaluate the accuracy of the keypoints.

2.3. Data Analyses

2.3.1. Workflow

Our research consists of four parts: (1) processing RS images and generating datasets for training, validation, testing, and evaluation; (2) constructing and optimizing keypoint detection models based on YOLOv10; (3) applying optimal models to study areas for the identification of wind turbine distribution; and (4) uncovering the spatial distribution patterns of wind turbines (Figure 3).

2.3.2. Model Construction

YOLO detectors were initially known as having real-time inference capability rather than high accuracy [22]. After several generations of updates, the YOLO model has established itself as state-of-the-art and achieved a well-balanced combination of accuracy and speed [57]. Given the state-of-the-art performance and widespread application of YOLO series models in real-time object detection tasks, we selected YOLOv10 as the benchmark (Figure 4). The YOLO series detector comprises three components: the backbone network for feature extraction, the neck for feature fusion, and the head for prediction. The latter two components may also be regarded as one part.

We compared the performance of eight models, including two benchmarks (YOLOv10n, YOLOv10s), RT-DETRv2, two hybrid YOLO models using backbones from RT-DETRv2 [33,58] (named YOLOv10-HGNet-n and YOLOv10-HGNet-s), and three older versions of YOLO (YOLOv5, v8, v9). The seven YOLO models used keypoint heads, enabling the identification of both bounding boxes and keypoints [31]. The object detection model RT-DETRv2 was not modified. YOLOv10 incorporates an improved backbone derived from YOLOv8, featuring spatial-channel decoupled downsampling modules, Compact Inverted Blocks (CIB), and a self-attention module. For the purpose of this study, the backbones of YOLOv10-n and YOLO-v10s (Cross Stage Partial Network, CSPNet), were referred to as benchmark-n and benchmark-s, respectively (Figure 4a). In contrast to the C2f (faster implementation of Cross Stage Partial Bottleneck with 2 convolutions) module in YOLOv10-n before the pooling layer, YOLOv10-s employs a strengthened C2fCIB (convolutional block with C2f and Compact Inverted Block) module to enhance feature extraction capabilities. In addition, the other two hybrid models were based on PP-HGNetv2 (Paddle Paddle High performance GPU Net-v2), the networks employed in RT-DETRv2 [58]. Corresponding to two benchmarks, we referred to them as HGNet-n and HGNet-s (Figure 4b,c). For the HGNet, we added Spatial Pyramid Pooling-Fast layer (SPPF) [30] and the Partial Self-Attention module (PSA) after the backbone, similar to YOLOv10. SPPF was used to extract multi-scale feature information with different sizes of pooling kernels. Then, PSA, a lightweight attention block, was added to enhance computational efficiency.

Furthermore, model scaling, a highly effective and common technique in computer vision, was adopted to control model size [28,59]. We set the scale parameters of depth, width, and maximum channels at 0.33, 0.25, and 1024, respectively, for YOLOv10-HGNet-n and YOLOv10-HGNet-s, the same as YOLOv10-n [57]. The other models used default parameters: depth, width, and maximum channels set at 0.33, 0.25, and 1024 for YOLOv10-n, YOLOv8, and YOLOv5; 0.33, 0.50, and 1024 for YOLOv10-s; and 1.0, 1.0, and 1024 for RT-DETRv2. YOLOv9 does not incorporate additional scaling parameters [29].

The total loss function ($L_{total}$) is a weighted combination of five loss functions [31], including those measuring box loss ($L_{box}$), classification loss ($L_{cls}$), distribution focal loss ($L_{df}$), keypoint objectness loss ($L_{kobj}$), and keypoint loss ($L_{pose}$):

$$L_{total}={\lambda }_{box}\cdot L_{box}+{\lambda }_{cls}\cdot L_{cls}+{\lambda }_{df}\cdot L_{df}+{\lambda }_{kobj}\cdot L_{kobj}+{\lambda }_{pose}\cdot L_{pose}$$

(1)

where ${\lambda }_{box}$ = 7.5, ${\lambda }_{cls}$ = 0.5, ${\lambda }_{df}$ = 1.5, ${\lambda }_{kobj}$ = 2, and ${\lambda }_{pose}$ = 15. These weighting parameters were configured based on default settings, loss magnitude, and task priority. We adjusted the latter two parameters governing wind turbine keypoint identification while keeping the other three unchanged [31]. This decision first stemmed from observed magnitude disparities: keypoint losses (<0.5) and keypoint objectness losses (<0.1) were significantly smaller than box loss (1.0–1.5) at convergence, suggesting a need to weight them differently. To adjust, we empirically increased the keypoint-related weights, ${\lambda }_{pose}$ and ${\lambda }_{kobj}$, from 12 and 1 to 15 and 2, respectively, achieving a 0.55% AP improvement.

(1)

Box loss

The box loss function measures the error of predicted bounding box, which is represented by Complete Intersection over Union ($CIoU$) [60]:

$$L_{box}=1-CIoU$$

(2)

$$CIoU=IoU-\left({\displaystyle \frac{{\rho }^2\left(b^p,b^g\right)}{c^2}}+{\displaystyle \frac{v^2}{\left(1-IoU\right)+v}}\right)$$

(3)

where $IoU$ represents the intersection over union of predicted and ground-truth boxes; $\rho \left(b^p,b^g\right)$ represents the distance between the centers of predicted box $b^p$ and ground truth box $b^g$; $c$ represents the diagonal length of the smallest box enclosing $b^p$ and $b^g$; and $v$ represents the aspect ratio consistency term.

(2)

Classification loss

$L_{cls}$ is measured by Binary Cross-Entropy Loss:

$$L_{cls}=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N\left(y_i\log \left({\hat{y}}_i\right)+\left(1-y_i\right)\log \left(1-{\widehat{y}}_i\right)\right)}$$

(4)

where $y_i$ is the ground truth class probability, ${\widehat{y}}_i$ is the predicted class probability, and $N$ is the number of samples.

(3)

Distribution focal loss

Distribution focal loss uses probabilistic distribution for bounding box regression [61]:

$$L_{df}=-\left(\left(y_{i+1}-y\right)\log \left(S_i\right)+\left(y-y_i\right)\log \left(S_{i+1}\right)\right)$$

(5)

where $S_i=\left(y_{i+1}-y\right)/\left(y_{i+1}-y_i\right)$, $S_{i+1}=\left(y-y_i\right)/\left(y_{i+1}-y_i\right)$; $y$ is the regression label; and $y_i$ is the value between range $\left(y_0,y_n\right)$.

(4)

Keypoint objectness loss

$L_{kobj}$ measures whether keypoints are accurately identified or not:

$$L_{kobj}=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N\left(k_i\log \left({\widehat{k}}_i\right)+\left(1-k_i\right)\log \left(1-{\widehat{k}}_i\right)\right)}$$

(6)

where $k_i$ and ${\widehat{k}}_i$ represent ground truth point (1 if objectness exists, 0 otherwise) and predicted point, respectively [31].

(5)

Keypoint loss

Keypoint regression loss is used to measure the accuracy of predicted keypoint coordinates, obtained by weighting point distance $d_i$:

$$L_{pose}=1-{\displaystyle \frac{1}{{\displaystyle \sum _i\delta (v_i>0)+}\epsilon }}{\displaystyle \sum _{i=1}^N\exp \left(-\frac{{d_i}^2}{2\cdot {\left(s_i+\epsilon \right)}^2\cdot {\left(2\cdot {\sigma }_i\right)}^2}\right)}\cdot \delta \left(v_i>0\right)$$

(7)

where $d_i$ is the Euclidean distance between predicted point and ground truth, $\delta \left(v_i>0\right)$ indicates the visible points, $\sigma$ is the stand error of coordinate, $s_i$ is the area of bounding box, and $\epsilon$ equals to 1 × 10⁻⁹ [31].

2.3.3. Strategies for Training and Post-Processing

In the training phase, each model ran 150 epochs to optimize the weights while ensuring a smooth convergence of the validation loss function. For a comprehensive fitness assessment, we used a weighted combination of $mA{P}_{50}$:$mA{P}_{50:95}$ in a ratio of 0.1:0.9 to identify the optimal weights. Models were trained using an NVIDIA Quadro RTX 6000 GPU (TSMC, Hsinchu, Taiwan, China), with the PyTorch (version 2.3.1+cu121) framework.

The main objective of our study is to obtain a more complete geospatial dataset, so we harmonized the detection results to enhance data quality in the post-processing stage (Figure 2). The disparities in model backbone structure would impact the features learned by the models, subsequently leading to differences in distributions of false positives and false negatives. To take advantage of different algorithms, we implemented a straightforward yet effective approach—overlay detection results from two models. Specifically, we aimed to minimize false negatives for each result by first using a relatively low confidence threshold of 0.5, followed by balancing precision with recall based on the intersection. If the keypoints inferred from the two models were in close spatial proximity (less than 50 m apart), we kept those with higher scores as correct labels. Otherwise, those isolated points were classified as false positives and were excluded from the harmonized results.

2.3.4. Accuracy Assessment

Several standard metrics were used during both the training and post-processing phases, including Precision, Recall, and F1 score:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(8)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(9)

$$F1={\displaystyle \frac{2\times Precision\times Recall}{Precision+Recall}}$$

(10)

where TP stands for true positive, FP stands for false positive, and FN stands for false negative. For object detection, the results were based on the IOU between the inferred bounding box and the ground truth. For keypoint detection, the accuracy was calculated based on Object Keypoint Similarity (OKS) [62]:

$$OKS={\displaystyle \frac{{\displaystyle {\sum }_i\exp \left(-\frac{d_i^2}{2{\left(s_j+\epsilon \right)}^2\times k_i^2}\right)}\times \delta \left(v_i>0\right)}{{\displaystyle {\sum }_i\delta \left(v_i>0\right)+\epsilon }}}$$

(11)

where $d_i$ represents the Euclidean distance between the predicted and ground-truth point; $s_j$ represents the scale of the bounding box; $k_i$ is the keypoint-specific constant which represents uncertainty and equals twice the stand error; $v_i$ represents whether the keypoint is visible or not; $\delta \left(v_i>0\right)$ counts the number of visible keypoints; and $\epsilon$ is a small value, set to 1 × 10⁻⁷ to avoid division by zero.

In further assessment using the evaluation set, we identified the accuracy of the keypoint (wind turbine base) by measuring its proximity to the ground truth. For simplicity, we classified points as TPs if the proximity to the ground truth was less than 30 m, the radius of the areas affected by wind turbine construction. Otherwise, the points were classified as FPs.

3. Results

3.1. Accuracy Assessment Results

We first assessed the performance of each model on the test and evaluation sets and then investigated the effectiveness of post-processing on accuracy enhancement. After that, we selected three regions for case studies to compare the detection results of the benchmark model with YOLOv10-HGNet.

Using test sets, we found that the YOLOv10-s and HGNet-based YOLOv10-s exhibited better performance in object detection and keypoint detection tasks, respectively (

Table 1. Performance comparison of different models. P: Precision, R: Recall. The highest value for each metric is shown in bold.

Metrics	Object Detection				Keypoint Detection
Model	P (%)	R (%)	$\mathit{A}{\mathit{P}}_{50}^{\mathit{v}\mathit{a}\mathit{l}}$(%)	$\mathit{A}{\mathit{P}}_{50:95}^{\mathit{v}\mathit{a}\mathit{l}}$(%)	P (%)	R (%)	$\mathit{A}{\mathit{P}}_{50}^{\mathit{v}\mathit{a}\mathit{l}}$(%)	$\mathit{A}{\mathit{P}}_{50:95}^{\mathit{v}\mathit{a}\mathit{l}}$
YOLOv10-n	96.59	94.54	98.60	63.11	97.96	95.99	99.07	97.99
YOLOv10-s	97.77	95.60	98.51	64.79	97.25	97.63	99.07	98.44
YOLOv10-HGNet-n	97.91	95.08	99.06	63.32	96.36	97.27	99.18	98.45
YOLOv10-HGNet-s	97.94	95.07	98.42	64.78	98.69	95.80	99.19	98.57
YOLOv9	95.80	94.50	97.47	59.42	96.20	95.60	98.71	97.69
YOLOv8	96.01	94.06	97.55	61.38	95.39	95.59	98.63	97.55
YOLOv5	94.84	94.54	97.53	61.04	93.99	94.92	97.94	97.13
RT-DETRv2	94.61	94.35	96.30	60.42	-	-	-	-

). Specifically, for object detection, YOLOv10-s achieved the highest recall (95.60%) and $A{P}_{50:95}^{val}$ (64.79%) while YOLOv10-HGNet-s had the fewest false positives. In keypoint identification, the latter achieved the highest $A{P}_{50:95}^{val}$ at 98.57% across all three metrics.

Utilizing 8503 ground truths to assess model generalization, we demonstrated the effectiveness of post-processing (

Table 2. Accuracy assessment of keypoints of base models and harmonized results. TPs: True Positives, FPs: False Positives, FNs: False Negatives, P: Precision, R: Recall, F1: F1 score. The highest value for P, R, and F1 are shown in bold.

Results	Models	TPs	FPs	FNs	P (%)	R (%)	F1 (%)
Base models	YOLOv10-n (1)	8239	887	262	90.28	96.92	93.48
	YOLOv10-s (2)	8294	695	207	92.27	97.57	94.84
	YOLOv10-HGNet-n (3)	8329	574	172	93.55	97.98	95.71
	YOLOv10-HGNet-s (4)	8346	798	155	91.27	98.18	94.60
	YOLOv9 (5)	8235	1613	268	83.62	96.85	89.75
	YOLOv8 (6)	8096	1326	407	85.93	95.21	90.33
	YOLOv5 (7)	8148	1671	355	82.98	95.83	88.94
Harmonized results	YOLOv10-Benchmark (1 and 2)	8155	149	346	98.21	95.93	97.05
	YOLOv10-HGNet (3 and 4)	8267	165	234	98.04	97.25	97.64

). Four base models (1–4) performed well on the evaluation set, with YOLOv10-HGNet-s and YOLOv10-HGNet-n achieving the highest recall and precision at 98.18% and 93.55%, respectively. Furthermore, their harmonized results showed an improved overall performance. The precision and F1 score of YOLOv10-HGNet increased by 4.49% and 1.93%, respectively, while the recall only decreased by 0.93%.

We selected several wind farms in three provinces as cases to illustrate the performance of two harmonized results (Figure 5). Generally, YOLOv10-HGNet exhibited fewer false positives and false negatives compared to YOLOv10-Benchmark. Complex backgrounds may lead to false detections. Some objects that share similar color and texture characteristics with wind turbine towers, such as snow-covered narrow roads, may be misidentified (case 1 in Figure 5c). Roads with significant dark-colored features may be detected as the shadow of a wind turbine, thereby increasing false positive rates (case 2 in Figure 5c). Additionally, the turbine shadow serves as a crucial characteristic for identification. When it is intermingled with dark-colored vegetation, fewer distinguishable features are available, leading to potential omissions (case 3 in Figure 5c).

3.2. Spatial Distribution of Wind Turbines

Using YOLOv10-HGNet, we identified a total of 63,055 wind turbines, which is 1.93 times the quantity recorded in OSM (32,658) as of 2022 (Figure 6). Compared with the YOLOv10-Benchmark (62,304), the modified model obtained more points with a higher recall rate, as evidenced by the previous accuracy assessment (Table 2). Within the designated study areas, northern regions had a higher concentration of turbines distributed around forests. For example, Shaanxi province had the most installations (7935), followed by Hebei (5714) and Liaoning (4654). Additionally, a significant gap was identified in the completeness of the data from open sources compared to our results at the provincial level. Among provinces with more than 1000 turbines, the average completeness of public data was 52.09%, with explicit spatial variance even between neighboring regions. Notably, Anhui and Guangdong exhibited the highest data coverage, at 74.36% and 72.15%, respectively. In contrast, much lower data coverage was observed in Hunan, at only 28.96%, and Guangxi, at 29.40%.

3.3. Analysis of Spatial Distribution Characteristics

Within our study area, wind turbines situated in the northern, eastern coastal, and southwestern regions exhibited a greater utilization of available wind resources (Figure 7). By ranking the wind speeds corresponding to the turbine sites across five levels, we found that 68.85% and 66.56% of the turbines in Hebei and Inner Mongolia, respectively, utilized the most wind. In addition, the coastal province of Fujian and the southwestern province of Yunnan each had over 60% of the wind turbine installations located at the top two wind speed levels. In contrast, wind turbines in the central region were positioned in areas where wind resources were comparatively limited.

By analyzing land use patterns at both the national and provincial levels, we observed that wind turbines were predominantly located on cropland (43.70%), followed by forest (25.65%) and grassland (24.80%) (Figure 8). Although our primary focus was on forests, our analysis revealed that among the 26 provinces, 15 exhibited the highest proportion of wind turbines located on cropland. Provinces with developed agriculture, such as Henan in central China and Shandong in eastern China, exhibited a significant concentration of wind turbines installed on cropland, particularly in areas adjacent to forests. In contrast, forest types, which account for a quarter of the total, were more concentrated in the southern and northeastern regions, exemplified by provinces such as Guangxi and Heilongjiang. In addition, a quarter of the installations were distributed in grassland, with notable concentrations in the northern provinces of Shanxi and Hebei.

We further analyzed the spatial patterns of forest occupation. We found that the Heilongjiang, Hunan, and Yunnan provinces exhibited the highest number of wind turbine installations in forests, with totals of 2174, 1921, and 1789 installations, respectively. A final categorization of forest types revealed that wind turbines primarily occupied deciduous broadleaved forests (44.17%) and evergreen broadleaved forests (31.82%), with significant spatial differences between northern and southern regions (Figure 9). Consistent with distribution patterns of tree species, wind turbines in southwestern China were mainly located within evergreen needle-leaved forests. This trend is particularly evident in the provinces of Yunnan and Sichuan, where evergreen needle-leaved forests compose 85.64% and 75.33% of the total occupied forests, respectively. To understand the potential impacts due to wind power development, we calculated land use changes from 2000 to 2022. The transformation suggested noteworthy forest degradation at the wind turbine sites, with a total of 63.49% of the area converted from forest to non-forest types. This conversion includes rainfed cropland (14.80%), irrigated cropland (14.80%), grassland (13.30%), etc.

4. Discussion

In summary, we mapped the wind turbine distribution in the forest regions of China utilizing a modified YOLOv10 keypoint detector and revealed the characteristics of the layout. This study represents an important effort to map onshore wind turbine distributions on such a scale by employing time-consistent satellite images. We used keypoint detection models based on YOLOv10 and identified 63,055 wind turbines in the forest areas of China. Our dataset significantly enhanced the turbine numbers as of 2022, about doubling that of publicly available data. This discrepancy highlights the necessity of assessing dataset quality, particularly when the true distribution of the data may influence the conclusions drawn from such analyses. Below, we discuss the advantages, key findings, and limitations of our research.

The advantages of high-quality Jilin-1 satellite images and the effective detection framework ensured that our results were reliable (Table 1 and Table 2 and Figure 5). Considering the irregular distribution of wind turbines within forest areas and the presence of heterogeneous backgrounds on land, identifying onshore wind turbines through conventional offshore wind turbine detection methods presented significant challenges [47]. Thus, satellite images with 2 m or even sub-meter resolutions were commonly used to extract small objects [63]. In this study, we used Jilin-1 satellite images with a resolution ranging from 1.5 to 2.0 m. The results corroborated the research design, with an $A{P}_{50}^{val}$ metric achieving 99% (Table 1). To enhance performance on this task, we modified the YOLOv10 model by replacing the backbone and heads. Instead of merely obtaining bounding boxes typical in common object detection, we adopted the keypoint detection method to accurately ascertain the geographical location of each turbine, which yielded favorable performance in further evaluations (Table 2). Finally, we filtered out the majority of the false positives with a distance-based post-processing strategy. This process also facilitated a balance between precision and recall and improved overall performance (Table 2).

By statistically analyzing the detection results, we found significant spatial differences in the coverage of wind turbine data in OSM (Figure 6). The unequal pattern mainly arises from the way OSM data are generated, which are primarily labeled by global volunteers, communities, and corporations [64]. The spatial distribution of data contributors is uneven, with developed regions having more mappers and, consequently, a higher quality of labeled data [18]. Although China has the highest installed wind power capacity globally, the distribution of turbines is relatively dispersed due to its vast land area. The manual labeling of these turbines necessitates considerable effort, which may lead to regional-variant omissions.

In addition, we also observed false positives and false negatives in our results that need to be avoided (Figure 5). As previously indicated, objects with similar features might be subjected to misclassifications, including narrow roads, buildings, and high-voltage transmission towers [65]. It is therefore necessary to investigate strategies for reducing misdetections. Apart from expanding datasets to retrain the models, better trade-offs between precision and recall could also be achieved during post-processing. By comparing confidence scores, we validated our hypothesis that significant differences exist between the distributions of TPs and FPs, as evidenced by p-values < 0.0001 (Figure 10). Across four models, higher mean scores were observed for TPs (ranging from 0.9918 to 0.9933) compared to FPs (ranging from 0.9448 to 0.9619). These disparities revealed the feasibility of post-processing measures, as a lower confidence score is associated with a decreased likelihood of misdetection by the dual models, which was reflected in the absence of overlapping points.

While we identified 63,055 wind turbines in our study areas related to forest distribution, it is important to note that most of them were not constructed in forests. This discrepancy arises from the criteria employed in selecting the study area and differences in the land cover classification system. As we previously illustrated, nearly 70% of the wind turbines were observed to be situated within cropland (43.70%) and grasslands (24.80%), particularly in central and northern provinces. On the one hand, our expanded interest areas encompassed many non-forest regions. The rapid expansion of agricultural land has resulted in the reclamation of numerous flat or gently sloping regions adjacent to forests for cultivation, which could also serve as potential sites for decentralized wind farms. Aiming to reduce omissions, we included these potential areas using buffer zones. On the other hand, the fine resolution of land cover classification products contributed to distinguishing non-forest areas in forests. The forests in central and northern China exhibit significantly lower tree coverage on their ridges than in their valleys, leading to the potential classification of these areas as grasslands or cultivated land. However, to attain higher wind speeds, wind turbines are typically situated in mountainous ridges, thereby contributing to discrepancies in the classification outcomes.

The rapid expansion of wind power has inevitably increased and may continue to increase the pressure on conservation efforts [66]. Many regions with abundant wind resources are also essential zones for maintaining biodiversity and providing key ecosystem services to human societies. Forest degradation resulting from the construction of wind farms can have significant adverse impacts. In addition to the reduction of vegetation coverage and the exacerbation of soil erosion [13,67], it also leads to the loss of wildlife habitats [68], the disruption of migration corridors [69], and an increase in accidental mortality [16], all of which pose a considerable threat to biodiversity. To a certain extent, the conflicts between development and conservation are inevitable. Although the large-scale development of renewable energy is an established trend for mitigation [70], national and local planning should be conducted in a more thoughtful, responsible, and long-term manner, taking into account the trade-offs between multiple targets. At the same time, the adoption of environmentally friendly technologies also shows a promising solution to mitigate potential disruptions to the physical environment and wildlife, particularly in instances where spatial conflicts exist.

While this study presents valuable insights, further research is necessary. First, the current models still have much room for improvement. Apart from expanding the training set, it is also advisable to test alternative modules [71,72]. Second, the detection area could be further expanded to enhance the coverage and completeness of the results. Due to cost considerations, we did not survey entire forest areas; instead, we focused on regions with a high likelihood of wind turbine installation and their associated buffer zones. In addition to forests, mapping wind turbines in grasslands or barren areas also holds great value for related studies. Third, it is essential to further investigate the spatiotemporal dynamics of wind turbines, not just the spatial patterns. Lastly, future research could focus more on the cross-scale ecological impacts associated with climate change. For instance, the historical and projected impacts of wind power expansion on shifts in species distribution could be quantified and modeled through the integration of site observations [73].

5. Conclusions

Driven by the rapid expansion of renewable power, China is steadily pacing towards carbon neutrality. Wind power provides clean energy; however, its development can lead to deforestation, habitat loss, and a decline in biodiversity. Given the various adverse impacts that have emerged from the construction and operation of wind farms on forest ecosystems, a clear understanding of wind turbine distribution patterns is necessary for a comprehensive ecological impact assessment at the national level. However, the coverage and completeness of existing data remain a research gap. Here, we identified wind turbines located in forest areas of China and analyzed the characteristics of associated distribution patterns. The main findings are as follows: (1) The HGNet-based YOLOv10 keypoint detection model exhibited good performance in wind turbine detection, achieving a F1 score of 97.64%. (2) A total of 63,055 wind turbines were identified as of 2022, representing 1.93 times the amount of publicly available data. (3) Among these, 16,173 wind turbines were located in forests, primarily within deciduous broadleaved forests (44.17%) in the northern region and evergreen broadleaved forests (31.82%) in the southern region. Our findings suggest that a more comprehensive understanding of wind turbine distribution patterns is fundamental for future studies. The research framework and methods employed in this study may serve as a reference for other researchers. Future research is necessary to further refine the dataset, elucidate the spatiotemporal dynamics of wind power distribution, and quantify the impacts on biodiversity and ecosystem services.