Semantic-to-Instance Segmentation of Time-Invariant Offshore Wind Farms Using Sentinel-1 Time Series and Time-Shift Augmentation

Abstract

The rapid expansion of offshore wind energy requires effective monitoring to balance renewable energy development with environmental and marine spatial planning. This study proposes a novel offshore wind farm detection methodology integrating Sentinel-1 SAR time series, a time-shift augmentation strategy, and semantic-to-instance segmentation transformation. The methodology consists of (1) constructing a dataset with offshore wind farms labeled from Sentinel-1 SAR time series, (2) applying a time-shift augmentation strategy by randomizing image sequences during training (avoiding overfitting due to chronological ordering), (3) evaluating six deep learning architectures (U-Net, U-Net++, LinkNet, DeepLabv3+, FPN, and SegFormer) across time-series lengths of 1, 5, 10, and 15 images, and (4) converting the semantic segmentation results into instance-level detections using Geographic Information System tools. The results show that increasing the time-series length from 1 to 15 images significantly improves performance, with the Intersection over Union increasing from 63.29% to 81.65% and the F-score from 77.52% to 89.90%, using the best model (LinkNet). Also, models trained with time-shift augmentation achieved a 25% higher IoU and an 18% higher F-score than those trained without it. The semantic-to-instance transformation achieved 99.7% overall quality in per-object evaluation, highlighting the effectiveness of our approach.

1. Introduction

Offshore wind energy (OWE) has emerged as a significant contributor to the global power generation portfolio, driven by environmental sustainability, technological advancements, and policy support. Unlike onshore wind farms, offshore wind farms are strategically located in oceanic or sea regions, which are more suitable for larger-scale constructions and benefit from consistent and stronger winds [1,2]. Floating wind turbines allow OWE to expand beyond shallow waters without relying on the ocean floor and gravity-based concrete at depths [3,4]. These advances also provide a promising renewable energy solution to the inherent spatial constraints of densely populated regions or those with limited land availability.

OWE is at a significant inflection point in its evolution due to technological advances, economies of scale, and increased political support, making it a more cost-effective energy source and reducing greenhouse gas emissions [5]. According to the World Forum Offshore Wind, global offshore wind capacity surpassed 57 Gigawatts (GW) at the end of 2022, representing an increase of 9.4 GW in 2022 [6]. China has the highest installed capacity, with 26.6 GW, followed by the United Kingdom, with 13.6 GW. In 2022, China led the global offshore wind capacity under construction, with a total capacity of 3.4 GW in Chinese waters. The United Kingdom followed closely in second place with 2.8 GW, which includes notable projects like Dogger Bank A (1.2 GW) and Seagreen (1.1 GW), both with capacities in the gigawatt range [6]. The International Energy Agency presented two scenarios for global offshore wind capacity growth by 2040: the State Policy Scenario, driven by policy targets and falling technology costs, and the Sustainable Development Scenario, which includes incentives to accelerate decarbonization efforts in the electricity sector [7]. Under the Sustainable Development Scenario, global offshore wind capacity is expected to reach approximately 560 GW by 2040, representing a 65% increase compared to the State Policy Scenario.

However, exponential global growth projections for OWE infrastructure involve several interests and conflicts with legal, socioeconomic, and ecological implications [8]. Despite OWE’s clear contributions to addressing the global emergence of carbon-neutral energy, this technology has limitations and encompasses wide-ranging challenges. As OWE intensifies and expands, the tendency is to increase conflicts with nearby coastal communities with interests contrary to using marine environments, such as aquaculture, tourism, navigation, and commercial fishing [9]. OWE negatively impacts various marine life populations due to collision mortality, disruption of migration corridors, barrier effects, acoustic and electromagnetic disturbances, ecological function changes, habitat alterations, and effects on the benthic ecosystem [10,11,12,13]. In this context, studies have evaluated the establishment of institutional and participatory measures to mitigate conflicts [14,15,16], the development of decision-support mechanisms to configure maritime space while considering multiple-use projects [17,18,19], and the application of environmental protection actions [20,21,22]. Additionally, OWE growth between 2020 and 2040 will require significant quantities of raw materials, including 8.2–14.6 million tons (Mt) of iron, 129–235 Mt of steel, 3.8–25.9 Mt of concrete, 0.5–1.0 Mt of copper, 0.3–0.5 Mt of aluminum, and substantial amounts of rare earth elements [23].

Monitoring the growth of OWE and analyzing its spatial distribution helps establish marine areas that minimize conflicts with other activities, such as shipping routes, fishing areas, and environmentally sensitive areas. This spatiotemporal information helps in alerting and planning a balance between renewable energy generation and the sustainable use of marine resources. However, on-site monitoring of these large-scale installations is complex and economically costly [24]. In this context, remote sensing is a promising data source for automatically monitoring offshore wind farms.

Alongside remote sensing images, recent advances in deep learning have achieved state-of-the-art object detection and segmentation in digital image processing [25,26,27], mainly using Convolutional Neural Networks (CNNs) [28,29,30]. These models can learn the differences between foreground and background features, such as oceans, wind farms, and other structures. However, previous studies aimed at detecting wind farms using remote sensing and deep learning have mainly focused on onshore areas. These onshore wind farm studies explored various combinations of optical images and CNN architectures: (a) Gaofen-2 images and U-Net [31]; (b) aerial photographs and comparative analysis between U-Net and LinkNet [32]; (c) aerial photographs and multiclass reconnaissance network [33]; (d) aerial photographs and Mask R-CNN [34]; (e) China–Brazil Earth Resources Satellite (CBERS) 4A scenes and CNN architecture comparative analysis (U-Net, U-Net++, LinkNet DeepLabv3+, and Feature Pyramid Network (FPN)) [35]; and (f) high-resolution Google Earth imagery and YOLOv5 [36]. Finally, Chen et al. [37] applied a combination of target detection techniques (Faster R-CNN, YOLOv5, DEtection TRansformer (DETR)—a transformer-based method—and CenterNet—an anchor-free method) using Google Earth imagery for onshore wind farm detection.

In the context of mapping offshore wind farms, the first studies used traditional techniques employing various methodological approaches, image data, and study areas: (a) a cloud-based geoprocessing algorithm using synthetic aperture radar (SAR) and the Google Earth Engine (GEE) to detect offshore infrastructure, including oil platforms and wind turbines in the Gulf of Mexico [38]; (b) a visual saliency detection algorithm based on optical satellite images (Landsat family and Sentinel-2) in the North Sea and surrounding waters [39]; (c) a percentile-based reduction algorithm using Sentinel-1 SAR time-series images, constructing a dataset containing over 6900 turbines across 14 coastal nations [40]; and (d) the Random Forest method using Sentinel-1 data on the GEE in the Yellow Sea of China and the North Sea of Europe [41].

Studies on offshore wind farm detection based on deep learning techniques have focused on the use of Sentinel-1 data and object detection methods (rather than deep segmentation), considering the following purposes: (a) the development of the DeepOWT dataset with spatiotemporal information on OWE infrastructure on a global scale [42]; (b) the proposal of SyntEO, a system that generates large datasets for deep learning, merging existing and synthetic data through the incorporation of specialized and structured knowledge [43]; (c) the estimation of the global capacity of offshore wind turbines based on the analysis of spatiotemporal patterns of offshore wind turbines, revealing the substantial growth of the sector due to the European Union, China, and the United Kingdom [44]; and (d) the identification of the temporal evolution of offshore wind distribution in China using the YOLOv5s-CR model, which integrates attention mechanisms and receptive fields to enhance detection accuracy [45]. In addition, some studies have used optical imagery to detect offshore wind farms, such as the mapping in the Shandong Sea of China using Sentinel-2 imagery and object detection models [46], or the global mapping integrating Sentinel-1 and Sentinel-2 data using deep learning and GEE [47].

Despite significant advances in offshore wind farm studies using object detection techniques, most previous approaches have relied on single-date images, which cannot capture temporal consistency, differentiate between fixed and transient objects (for example, wind farms and ships), or provide fine-grained spatial segmentation. Furthermore, object detection remains the predominant approach, while semantic segmentation models are often simpler and obtain instance-level results.

Thus, the main objective of this paper is to introduce a novel integrated methodology that combines Sentinel-1 SAR time series, time-shift augmentation, and semantic-to-instance transformation to enhance offshore wind farm monitoring. The novelty of this research lies not only in these individual components but in their combination, providing a new approach to wind farm detection in offshore environments:

• Novel time-shift augmentation for time-invariant target segmentation:
  We propose a new augmentation strategy that randomizes the temporal order of time-series images, preventing overfitting to a fixed chronological sequence and improving generalization. Unlike conventional augmentation techniques (rotation, flipping, and noise injection), this approach is designed explicitly for time-invariant target segmentation, which addresses unique challenges in offshore wind farm detection.
• Dataset for offshore wind farm segmentation using SAR time series and benchmarking: Unlike previous datasets that focus on object detection from single-date images, this dataset contains multi-temporal SAR patches, allowing models to leverage time-series information for improved accuracy. This dataset includes over 5000 labeled patches and is publicly available. We evaluate six segmentation architectures (U-Net, U-Net++, LinkNet, FPN, DeepLabv3+, and SegFormer) to establish a benchmark for this dataset.
• Semantic-to-instance segmentation using Geographic Information System tools for instance-level detection: We introduce a simple yet effective transformation from semantic to instance segmentation using a Geographic Information System (GIS) to derive wind farm instances from pixel-wise segmentations. This method enables object-based wind farm monitoring without requiring specialized instance segmentation architectures.

2. Materials and Methods

2.1. Dataset Construction

This study focused on the United Kingdom, a global pioneer in offshore wind technology [48], which has the second-largest offshore wind capacity worldwide [49]. This study selected ten regions distributed throughout the United Kingdom (Figure 1).

In contrast to the challenge of detecting onshore wind farms using SAR images—where the lack of clear distinction in the terrestrial environment introduces complexities—offshore wind farms are larger and, over the vast ocean, produce a high-backscatter pattern that facilitates their detection. Therefore, this research used SAR images from the Sentinel-1 mission, developed by the European Space Agency (ESA), which consists of a constellation of two polar-orbiting satellites: Sentinel-1A and Sentinel-1B. These sensors operate at a 5.4 GHz C-band frequency and capture SAR images in VV (vertical transmission, vertical reception) and VH (vertical transmission, horizontal reception) polarizations, with free access to the Copernicus Open Access Hub. The data specifications included Interferometric Wide-Swath (IWS) images, which provide wide-area coverage with consistent spatial resolution using burst-mode acquisition. These images were acquired in VV polarization in Level-1 Ground Range Detected (GRD) format, which consists of multi-looked, georeferenced SAR data projected to ground range with reduced speckle noise. VV polarization was chosen for its lower noise influence and its ability to better discern the target characteristics in the sea [40]. The analysis encompassed a time series of up to 15 SAR images captured in 2022 and 2023. Image pre-processing was performed using the Sentinel Application Platform (SNAP) software (version 11.0.0) and included the following steps [50]: (a) applying the orbit file, (b) removing thermal noise, (c) calibrating the digital pixel values to radiometrically calibrated SAR backscatter, (d) applying speckle filtering using the Lee Sigma filter, (e) performing range-Doppler terrain correction utilizing the Global Earth Topography and Sea Surface Elevation data at 30 arc-second resolution, and (f) converting backscattering values to decibels (dB).

2.2. Ground-Truth Generation

One difficulty in mapping wind farms is the presence of different reflections from the metal structures, which differ in each image with various sizes and positions (Figure 2). In this context, using time series allows us to reduce noise by selecting only those pixels that demonstrate constant brightness over time located in the central part of the wind farm. Therefore, we propose a method that calculates the average image of the time series and generates a mask from a high-backscattering threshold value in the VV band, where we chose values above 0. This approach resembles the approach developed by [40]. The data refinement used vector format (shapefile) in the ArcGIS Pro software (version 3.4), enabling manual noise elimination. In this annotation task, the optical images served as a support to differentiate the target from other elements, such as oil platforms or boats.

Offshore wind farm detection using time-series data can overcome difficulties in detecting wind farms, increasing the chances of obtaining clear and discernible images of wind farms and removing floating or temporarily mobile objects. Therefore, time-series data strengthen the learning process of deep architectures to mitigate problems and ensure reliable results, allowing us to capture the invariant behavior of wind farms in contrast to other moving features or noise and identify their morphological variations due to backscattering patterns.

2.3. Semantic Segmentation Dataset Generation

The generation of fixed-size patches for training used the methodology developed by [51], which allows for strategic choice of patches rather than systematic clipping through sliding windows. This study selected a patch size of 128 × 128 pixels, yielding 4947 patches and encompassing 2910 distinct offshore wind farms. The patch count outweighs the number of wind farms, aiming to capture wind farms under various scenarios.

The training, validation, and testing split used distinct sets to avoid overlap. Therefore, six of the ten chosen areas were allocated for training, two for validation, and the other two for testing.

Table 1. Detailed description of the dataset, including the data collection locations; the number of instances; the number of patches; data partitioning into training, validation, and testing subsets; and the time intervals between the first and last images in the time series.

Location	Number of Wind Farms	Number of Patches	Split	Time Interval
3°${36}^{\prime }$${9645}^{\prime \prime }$ W 54°${33}^{\prime }$${2492}^{\prime \prime }$ N	374	652	Train	14 August 2022 to 29 January 2023
3°${26}^{\prime }$${\mathrm{37,919}}^{\prime \prime }$ W 53°${29}^{\prime }$${\mathrm{28,928}}^{\prime \prime }$ N	269	565	Train	14 August 2022 to 29 January 2023
0°$9^{\prime }$${3647}^{\prime \prime }$ E 53°${48}^{\prime }$${\mathrm{23,561}}^{\prime \prime }$ N	70	131	Train	16 August 2022 to 31 January 2023
0°${54}^{\prime }$${\mathrm{38,874}}^{\prime \prime }$ E 53°${16}^{\prime }$${\mathrm{50,015}}^{\prime \prime }$ N	474	611	Val	16 August 2022 to 31 January 2023
1°${46}^{\prime }$${\mathrm{53,292}}^{\prime \prime }$ E 53°${55}^{\prime }$${\mathrm{11,708}}^{\prime \prime }$ N	308	402	Test	16 August 2022 to 31 January 2023
1°${22}^{\prime }$${3046}^{\prime \prime }$ E 51°${33}^{\prime }$${\mathrm{52,405}}^{\prime \prime }$ N	387	903	Train	16 August 2022 to 31 January 2023
1°${57}^{\prime }$${\mathrm{20,788}}^{\prime \prime }$ E 51°${51}^{\prime }$${\mathrm{40,077}}^{\prime \prime }$ N	220	557	Test	16 August 2022 to 31 January 2023
2°${31}^{\prime }$${\mathrm{22,737}}^{\prime \prime }$ E 52°${12}^{\prime }$${\mathrm{26,383}}^{\prime \prime }$ N	96	144	Train	16 August 2022 to 31 January 2023
2°${58}^{\prime }$${\mathrm{19,339}}^{\prime \prime }$ E 51°${39}^{\prime }$${\mathrm{41,574}}^{\prime \prime }$ N	592	708	Val	16 August 2022 to 31 January 2023
0°${16}^{\prime }$${\mathrm{34,541}}^{\prime \prime }$ W 50°${40}^{\prime }$${\mathrm{11,965}}^{\prime \prime }$ N	120	274	Test	09 August 2022 to 17 February 2023
Total	2910	4947	-	-

lists the characteristics of each location, including its assigned data set, the number of patches, and the count of wind farms.

Table 2. Summary of the number of unique instances and the number of patches for each data subset.

Set	Number of Wind Farms	Number of Patches
Train	1196 (41.1%)	2395 (48.41%)
Val	1066 (36.63%)	1319 (26.27%)
Test	648 (22.27%)	1233 (24.92%)

lists the synthesized characteristics of the dataset.

2.4. Deep Learning Approach

2.4.1. Deep Learning Models

This study compared six deep learning architectures: Deeplabv3+ [52], U-Net [53], U-Net++ [54], FPN [55], LinkNet [56], and SegFormer [57]. EfficientNet-B7 [58] served as the backbone for all models.

The hyperparameters were the same for all models to ensure a consistent comparison: (1) the Adam optimizer, (2) a learning rate of 0.001, (3) a batch size of 20, and (4) 150 epochs. The model selection considered the cross-entropy loss and saved the model with the lowest validation loss for each architecture for further statistical analysis. All models were tested on the same computer, equipped with an NVIDIA RTX 4090 with 24 GB of Random Access Memory.

2.4.2. Novel Data Augmentation Strategy

While time series typically depend on the data frames’ sequential order, as in agriculture and phenology studies, our research focuses on a time-invariant target. In this context, the sequential order of images within the time series is irrelevant and could introduce bias. If a model is invariably trained on an identical sequence of images, there is a risk of learning specific patterns from that sequence.

This study proposes a new data augmentation strategy for time series and semantic segmentation that randomly shuffles the temporal order of images during the training process, preventing the model from becoming familiar with any specific sequence. A similar approach has been taken for instance segmentation of center pivots using optical imagery, in which the main objective was to avoid cloud interference in the model [59].

2.4.3. Experiments

The experiment analyzed datasets with 1, 5, 10, and 15 images. The selection of these time-series lengths was designed to assess how increasing temporal information impacts segmentation performance. The single-date image case (1 image) served as a baseline to assess how well models performed without any temporal context. The 5-image time series captures short-term variations and helps filter out transient maritime objects while maintaining high temporal resolution. The 10-image sequence balances noise reduction and computational feasibility, allowing the model to better differentiate between temporary and persistent high-backscatter objects. The 15-image sequence represents a longer-term analysis to ensure that the detected structures are truly permanent and to reduce the impact of SAR noise and momentary reflections.

Even while preserving the training, validation, and testing samples, the analysis established different time-series periods. To maintain consistency when comparing models across different time-series lengths, we performed a single shuffle in the validation and test samples, preserving their ordering across all evaluated models. Furthermore, to assess the impact of the proposed augmentation strategy, we compared the best-performing model with and without its application in different time-series configurations.

2.4.4. Sliding-Window Approach

The sliding-window technique is a moving window (with the size used in the training) in dimensions x and y using a predetermined stride value. The largest possible stride value corresponds to the window’s dimension, and stride values less than this result in overlapping pixels. These overlaps improve the resulting classification by mitigating discontinuities between adjacent frames, creating a smoother and more coherent segmentation output [60]. The sliding-window approach is necessary to classify large images, which is the case for most remote sensing data, in which the patch size used for training is much smaller than the data to be analyzed.

2.4.5. Accuracy Metrics

This study’s accuracy assessment employed two types of metrics: per-pixel and per-object [61]. The per-pixel metrics evaluate the accuracy of semantic segmentation at the pixel level by comparing the predicted and reference masks, which makes them useful for fine-grained analysis [62,63]. The per-object (per-polygon) metrics assess segmentation performance at the instance level, focusing on correctly identifying and counting individual wind farms. This analysis reduces the impact of minor contour misalignments [64].

Table 3. Mathematical formulations for the per-pixel (overall accuracy, precision, recall, F-score, and Intersection over Union (IoU)) and per-object (overall quality, completeness, and correctness) accuracy metrics. For the per-pixel metrics, TP, FP, FN, and TN refer to individual pixels, whereas for the per-object metrics, they refer to wind farm instances based on an IoU threshold of 50%.

Metric	Formula
Per-Pixel Metrics
Overall Accuracy	$(T{P}_{pixel}+T{N}_{pixel})/(T{P}_{pixel}+T{N}_{pixel}+F{P}_{pixel}+F{N}_{pixel})$
Precision	$T{P}_{pixel}/(T{P}_{pixel}+F{P}_{pixel})$
Recall	$T{P}_{pixel}/(T{P}_{pixel}+F{N}_{pixel})$
F-Score	$2\times \left({\displaystyle \frac{\mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}\right)$
IoU (Intersection over Union)	$T{P}_{pixel}/(T{P}_{pixel}+F{P}_{pixel}+F{N}_{pixel})$
Per-Object (Per-Polygon) Metrics
Overall Quality	$T{P}_{object}/(T{P}_{object}+F{P}_{object}+F{N}_{object})$
Correctness	$T{P}_{object}/(T{P}_{object}+F{P}_{object})$
Completeness	$T{P}_{object}/(T{P}_{object}+F{N}_{object})$

presents the mathematical formulations for the per-pixel and per-object metrics. The per-pixel metrics were computed on 128 × 128-pixel test patches, and the per-object (per-polygon) evaluation method analyzed complete images reconstructed using the sliding-window approach. A true positive object was defined based on an IoU threshold of 50% between the predicted and reference polygons.

3. Results

3.1. Per-Pixel Metrics

This study evaluated the performance of six models in different time-series configurations.

Table 4. Per-pixel results for the overall accuracy (OA), precision, recall, F-score, Intersection over Union (IoU), training time per epoch (TT), and inference time (IT) for time series with 1, 5, 10, and 15 images.

Model	OA	Precision	Recall	F-Score	IoU	TT (s)	IT (s)
Time Series with 15 images
LinkNet	99.96	87.62	92.29	89.90	81.65	32.44	0.04
U-Net	99.96	87.75	91.59	89.63	81.21	27.78	0.04
U-Net++	99.96	87.43	92.10	89.71	81.33	28.40	0.04
DeepLabv3+	99.93	79.82	86.72	83.13	71.13	28.15	0.04
FPN	99.92	79.81	85.01	82.33	69.96	27.97	0.04
SegFormer	99.92	79.56	85.97	82.64	70.42	28.15	0.04
Time Series with 10 images
LinkNet	99.95	86.65	89.11	87.86	78.35	32.24	0.04
U-Net	99.95	86.04	88.55	87.28	77.42	26.65	0.04
U-Net++	99.94	84.83	89.44	87.07	77.11	28.08	0.04
DeepLabv3+	99.92	78.82	83.60	81.14	68.26	28.20	0.04
FPN	99.91	75.76	85.47	80.32	67.11	27.54	0.04
SegFormer	99.92	76.57	85.78	80.91	67.94	27.89	0.04
Time Series with 5 images
LinkNet	99.93	82.94	86.03	84.45	73.09	32.89	0.04
U-Net	99.93	83.29	85.54	84.40	73.01	25.65	0.04
U-Net++	99.93	83.06	86.47	84.73	73.51	27.89	0.04
DeepLabv3+	99.90	73.76	82.88	78.06	64.01	29.37	0.04
FPN	99.90	72.94	83.00	77.64	63.46	27.45	0.04
SegFormer	99.90	74.40	79.93	77.07	62.69	28.14	0.04
Single Image
LinkNet	99.90	76.63	78.42	77.52	63.29	33.77	0.04
U-Net	99.90	71.29	85.51	77.76	63.61	20.22	0.03
U-Net++	99.90	75.34	79.08	77.16	62.82	23.70	0.03
DeepLabv3+	99.88	67.70	80.38	73.50	58.10	29.93	0.03
FPN	99.88	69.76	77.96	73.63	58.27	23.24	0.02
SegFormer	99.89	70.65	79.15	74.66	59.56	26.87	0.03

presents the segmentation accuracy results, showing how the IoU and F-score improved as the number of images in the time series increased. Models trained with longer time series showed better generalization, as the increased temporal context helped mitigate the impact of transient objects, such as ships. Overall accuracy remained high across all models due to class imbalance. Since most pixels corresponded to the ocean background, models that misclassified wind farms but correctly classified the background still achieved high overall accuracy.

LinkNet achieved the best results, closely followed by U-Net++ and U-Net. These models performed consistently across different time-series lengths, with IoU values above 80% when trained with 15 images. In contrast, FPN, DeepLabv3+, and SegFormer yielded lower scores, with IoU values of around 70%. These models had difficulty capturing the fine details of wind farm structures, especially when trained with shorter time-series sequences.

Training times varied significantly, with LinkNet and DeepLabV3+ requiring the longest times, while U-Net trained more efficiently. Inference times remained consistent across models, indicating that once trained, all architectures achieved near-real-time performance. The increase in time-series length had only a minor effect on inference time but significantly impacted training duration, highlighting the computational trade-offs of using longer time sequences.

Table 5. LinkNet results without the augmentation strategy for the overall accuracy (OA), precision, recall, F-score, and Intersection over Union (IoU) for time series with 5, 10, and 15 images.

Time Series	OA	Precision	Recall	F-Score	IoU
15	99.90	90.89	60.67	72.77	57.20
10	99.86	63.87	73.96	68.55	52.15
5	99.89	87.18	53.39	66.22	49.50

compares the performance of the LinkNet model with and without the proposed time-shift augmentation strategy. The results indicate that removing this strategy decreased performance, particularly in the IoU and F-score. The 15-image time series model showed a 24.45% reduction in the IoU (from 81.65 to 57.20) and a 17.13% drop in the F-score (from 89.90 to 72.77). A similar trend was observed for the 10-image and 5-image time-series models. The absence of augmentation primarily affected recall, as seen in the substantial increase in false negatives (FN). The model struggled to recognize some wind farms that were previously detected when augmentation was applied. This effect was more pronounced in shorter time-series models, which already had lower generalization ability.

An important observation from Table 5 is that, despite the performance degradation caused by the removal of the time-shift augmentation strategy, the precision values for 5 and 15 images were comparable to their augmented counterparts (Table 4). This phenomenon occurred because precision is primarily influenced by the number of false positives (FP), whereas the removal of the time-shift augmentation strategy primarily increased the number of false negatives (FN), which instead impacted recall and IoU. Since precision measures the proportion of correctly predicted wind farms relative to the total number of positive predictions, it remained relatively stable when the number of FPs did not increase significantly.

This behavior was particularly evident in the 15-image case, where the time series already provided substantial redundancy, making the network robust enough to avoid generating additional false positives. Similarly, for the 5-image case, while removing the augmentation strategy reduced recall (due to missing some true wind farms), it did not cause a surge in FPs, allowing precision to remain stable. Thus, the drop in performance observed in Table 5 was primarily due to a decline in recall rather than an increase in false detections, explaining why precision did not decrease as expected.

Figure 3 shows the results of the LinkNet model with various time-series configurations. Despite the significant IoU difference, all models identified the presence of the target. The main reason for this performance variation was the small size of the wind farms, where even slight errors had a considerable impact on the performance metrics. Using more images in the time series led to more consistent predictions and resulted in fewer misclassifications of other high-backscatter targets. When the augmentation strategy was not applied, the model struggled with certain features, indicating a tendency to memorize patterns within the time series.

3.2. Per-Object Metrics

The per-object metrics considered the region with the highest concentration of offshore wind farms and time series consisting of 1, 5, 10, and 15 images (

Table 6. Per-object metrics, including true positives, false positives, false negatives, overall quality, correctness, and completeness for time series with 1, 5, 10, and 15 images.

	Number of Images Used in the Time Series
	1	5	10	15
TP	296	297	297	297
FP	14	2	1	1
FN	1	0	0	0
Overall Quality (%)	95.18	99.33	99.67	99.67
Correctness (%)	95.48	99.33	99.67	99.67
Completeness (%)	99.67	100.00	100.00	100.00

). The instance-based evaluation showed that increasing the number of images improved correctness and completeness. The 15-image model successfully identified 297 wind farms, with only one false positive and no false negatives, achieving an overall quality of 99.7%. In contrast, the single-date image model had a higher false positive rate, which reduced its overall quality to 95.4%. The number of false positives (FP) dropped from 14 in the single-date image model to just 1 in the 15-image case. This result highlights the benefit of temporal analysis in filtering out transient maritime objects, which can cause misclassification in single-frame models.

Figure 4 illustrates the final instance segmentation results for a region with a high density of offshore wind farms, while Figure 5 highlights misclassified regions when using a single image. The errors in Figure 5 confirm the benefits of incorporating temporal sequences to eliminate transient high-backscatter targets.

4. Discussion

This research developed a new approach using time-series data with different deep learning models to detect offshore wind farms. The developed strategy involved generating a dataset with different numbers of temporal images and implementing an innovative augmentation procedure that promotes the shuffling of the chronological sequence. Previous studies have applied deep learning to detect offshore wind farms using remote sensing data, but the methodologies vary depending on the data sources, model architectures, and processing approaches. Our approach advances offshore wind farm detection by integrating time-series analysis, novel augmentation techniques, and GIS-based post-processing, distinguishing it from other studies.

4.1. Onshore vs. Offshore Wind Farms

Mapping strategies for onshore and offshore wind farms using optical and radar remote sensing imagery differ mainly due to variations in target size, background complexity, and segmentation particularities. In onshore regions, wind farms are smaller in size, with rotor diameters ranging from 40 to 120 m and tower heights from 60 to 150 m, and they have a heterogeneous background, which makes the use of SAR imagery challenging. Therefore, most remote sensing studies of onshore wind farms use high-resolution optical imagery [31,32,33,34,36,37]. For example, a comparison between optical images from Sentinel-2 (10 m resolution) and CBERS 4A (2 m resolution) showed that onshore wind farms in Sentinel-2 appeared with only a few pixels, making visual detection of targets complex [35]. Therefore, onshore wind farm detection using radar sensors is limited due to the following factors: (1) terrain complexity and signal interference—in terrestrial areas, the presence of elements such as buildings, dense vegetation, varied relief, and other infrastructure (poles, power towers, and other metal structures) results in a highly heterogeneous background, which causes multiple reflections and scattering of the radar signal, making it difficult to clearly distinguish wind farms from surrounding objects and generating false positives. (2) limited spatial resolution—radar sensors, especially freely accessible and continuous imagery such as Sentinel-1, have insufficient spatial resolution to identify the specific details of wind turbines, and access to high-resolution SAR data (such as TerraSAR-X or RADARSAT-2) can be expensive, limiting their large-scale application.

In contrast, offshore wind farm detection predominantly uses freely accessible SAR images due to the larger size of turbines, with rotor diameters exceeding 200 m and tower heights between 150 and 300 m, which are positioned over open waters [38,40,41,42,43,44,45]. SAR imagery captures wind turbine structures over the ocean due to strong radar backscatter and is unaffected by cloud cover or highly heterogeneous background interference. However, offshore environments present additional targeting difficulties: (a) strong radar reflections from metallic turbine structures vary over time due to radar incidence angle and turbine blade orientation, making the targeting process less stable; and (b) floating objects, such as ships and other offshore infrastructure, often exhibit similar backscatter signatures, increasing the risk of false positives. Therefore, the distinct characteristics of the targets and environments make SAR imagery the predominant choice for offshore monitoring, while high-resolution optical imagery is used in onshore areas.

4.2. Semantic Segmentation vs. Object Detection

Object detection has been the central methodology for offshore wind turbines. Hoeser et al. [42] developed an object detection pipeline using a cascading CNN approach trained on synthetic data. Chen et al. [37] and Liu et al. [46] used high-resolution optical images with object detection models such as YOLO and Faster R-CNN. Ding et al. [45] applied a modified YOLOv5s-CR model with attention mechanisms to detect offshore wind turbines in China using Sentinel-1 SAR images.

While object detection provides bounding-box-level localization, it does not offer pixel-level information. Also, object detection models typically require specialized design choices, such as selecting appropriate anchor boxes, aspect ratio priors, and fine-tuning hyperparameters specific to the dataset’s object sizes. Another complexity with object detection models is applying sliding-window approaches to process large regions since they classify individual turbines within bounding boxes. Overlapping windows can cause duplicate detections when turbines appear on the edges of multiple image patches, requiring post-processing techniques to remove redundant detections, thereby increasing computational complexity.

However, a significant advantage of object detection over semantic segmentation is individualized recognition, enabling the counting and precise localization of different instances of the same object. However, our GIS semantic-to-instance conversion facilitates instance-level mapping. Thus, using semantic segmentation for this target is more straightforward regarding model training and post-processing, resulting in a more efficient workflow. Having only isolated targets avoids the need for additional procedures when the targets present are clustered, such as using edge detection to isolate the interior of objects and multitask learning [65] or multiclass learning [66,67]. We adopted GIS-based semantic-to-instance transformation instead of using instance segmentation architectures such as Mask R-CNN, providing a simpler workflow. This aligns with geospatial applications in which segmentation masks can be converted into vectorized geospatial features for wind farm monitoring. While Chen et al. [37] used clustering to group wind turbine locations, our method directly extracts instance-level boundaries from segmentation outputs, offering a simpler workflow.

4.3. Time-Series vs. Single-Date Image Detection

SAR data have been widely used in oceanographic studies, as demonstrated in several review articles [68,69,70,71,72], and they have been specifically applied for the detection of ships [73,74,75,76], oil spills [77,78,79], and ice and icebergs [80,81,82]. However, SAR signal returns in marine environments are influenced by several factors, such as ocean surface roughness, ocean currents, reflections from other objects, and speckle noise. Therefore, within a single SAR scene, the amplitude and complexity of the background noise can vary due to environmental factors, such as internal waves, variations in wind speed, and biogenic films. These variations are more pronounced in multi-temporal data, as climatic and environmental differences affect the backscatter signals from the same marine areas over time.

In the case of wind turbines, which are made up of large metal structures (towers and blades), the interaction with radar waves presents a variable geometry due to the angle of incidence of the radar beam and the orientation of the rotating blades, providing different reflection patterns. The scattering generates high-intensity returns in multiple directions, forming different reflective patterns in the form of beams. Even when using deep learning techniques, the problem of misdetection still exists, considering single-date images. In this context, the use of time series enables the compilation of multiple images over time, reduces unwanted artifacts, and emphasizes the stable patterns of wind farms. In addition, the use of time series enables distinguishing wind farms, which are fixed objects, from other artifacts and transient objects, such as ships or other elements that also generate high reflection.

Most previous studies on offshore wind farm detection have relied on single-date imagery using SAR (Sentinel-1) data [42,43,44,45]. Zhang et al. [47] introduced a multi-temporal pre-processing step using GEE for optical and radar imagery to refine input imagery. However, their final detection model still relied on single-date imagery, meaning the deep learning model did not explicitly leverage temporal sequences for decision making. Instead, the pre-processing stage reduced noise and cloud clutter before feeding individual images into the detection pipeline.

Our approach differs by incorporating Sentinel-1 SAR time series with sequences of 1, 5, 10, and 15 images as direct inputs to deep learning models. By analyzing multiple images over time, our method enables the differentiation of transient objects, such as stationary wind farm vessels, by tracking their presence at various time intervals. This strategy exploits the time-invariant characteristics of offshore wind farms, overcoming the challenges of FPs from single-date images. The proposed method results highlight this benefit in the per-object metrics, in which the numbers of FPs were 14, 2, 1, and 1 for time series with 1, 5, 10, and 15 images, showing that larger time series are effective at increasing performance.

4.4. Dataset, Time-Series Augmentation, and Benchmark Evaluation

One of this study’s main contributions is the construction of a comprehensive offshore wind farm segmentation dataset using Sentinel-1 SAR time series. Unlike previous datasets, which predominantly focus on object detection or single-date image segmentation, our dataset incorporates multi-temporal sequences, enabling models to better differentiate between time-variant and time-invariant targets.

To improve model generalization, we proposed a time-shift augmentation strategy, which randomizes the sequence of images within a time series during training. This augmentation strategy prevents the model from overfitting to a specific chronological order. Our experiments demonstrated that removing this augmentation process resulted in a significant performance drop, with the IoU decreasing by 25% and the F-score dropping by 18%. The augmentation strategy was particularly beneficial in handling variations in radar backscatter caused by environmental factors, ensuring that the model learned more robust and time-invariant features. However, even though this technique works for offshore wind farms, it does not apply to tasks in which the temporal order carries significant meaning, such as phenology-based studies.

To assess model performance, we conducted a benchmark evaluation using six deep learning architectures: U-Net, U-Net++, LinkNet, DeepLabv3+, FPN, and SegFormer. Models trained on longer sequences outperformed those trained on shorter sequences. LinkNet achieved the highest accuracy, followed closely by U-Net++ and U-Net. These models share a similar encoder–decoder structure that facilitates image reconstruction while preserving fine spatial details, which is essential for segmenting small and uniformly shaped objects like offshore wind farms.

In contrast, models such as FPN, DeepLabv3+, and SegFormer exhibited lower performance. DeepLabv3+, which relies on atrous spatial pyramid pooling to capture multi-scale context, is effective for datasets containing objects of varying sizes but may struggle when applied to uniformly small targets. FPN, designed to enhance feature representation at different pyramid levels, struggles in scenarios where high-resolution feature retention is crucial. SegFormer follows a different paradigm, relying on transformers instead of CNNs. While transformers emphasize global context, they often sacrifice the fine-grained spatial precision necessary to detect small targets.

Regarding computational efficiency, the inference times were very low across all models, typically a few milliseconds per image, which is particularly advantageous for remote sensing applications, where the revisit time of satellites is on the scale of days, meaning real-time inference is not a requirement, as it would be in video processing. The training time per epoch varied depending on the model and time-series length, but it did not present a significant limitation. Even with 15 images per time series, the training times remained within reasonable bounds for practical deployment, ensuring that model retraining can be performed efficiently when new data become available.

4.5. Limitations and Future Work

One primary limitation is the data requirement for constructing time series. Unlike single-date image models, our method relies on multiple SAR acquisitions over time to differentiate between stationary wind farms and transient maritime objects. This requirement increases data storage and processing demands, particularly when working with longer time-series sequences. The availability of historical SAR data can vary depending on the region, which may limit its applicability in certain areas. Future studies could explore adaptive time-series selection strategies in which models dynamically determine the optimal number of images based on regional data availability.

Another limitation is handling newly constructed wind turbines within the time series. Since our approach assumes that wind farms remain stationary over time, any turbines that become operational partway through the time series may be misinterpreted as transient objects, similar to ships or other maritime structures. This limitation could affect monitoring projects requiring continuous updates on offshore wind expansion.

Even though the dataset used in this study focuses on offshore wind farms in the United Kingdom, the methodology is designed to be generalizable to other regions with different environmental conditions, wind farm layouts, and SAR imaging characteristics. The main principles of our approach—including time-series segmentation, the time-shift augmentation strategy, and semantic-to-instance transformation—are not specific to any single geographic area and can be applied globally. Future studies aim to validate the method in diverse offshore environments across different geographic locations with varying wind turbine densities, oceanic conditions, and alternative SAR sensor configurations. Our study focuses on Sentinel-1 SAR data, ensuring consistent monitoring under all weather conditions. However, the method has not been tested with other SAR missions (e.g., RADARSAT or TerraSAR-X) or multi-sensor fusion approaches incorporating both SAR and optical imagery.

5. Conclusions

This research presented a novel approach for detecting offshore wind farms using Sentinel-1 time series, introduced a novel augmentation strategy, compared six deep learning architectures, and used GIS tools to obtain instance-level segmentation of these features. Sentinel-1 radar images revealed a valuable contrast for detecting offshore wind farms. The proposed offshore wind farm dataset extends beyond a single-date (single-channel) image by using a time series (multichannel), where the augmentation strategy introduces chronological sequence scrambling to enhance the detection of invariant targets in unknown images. Using a time series instead of a single-date image helps eliminate mobile high-backscatter features and reinforces distinctions between fixed objects. Across all semantic segmentation models, the IoU and F-score consistently increased as the number of images in the time series increased. The LinkNet model achieved the best results, slightly surpassing U-Net++ and U-Net, while the FPN, DeepLabv3+, and SegFormer models yielded the worst results. Using the novel augmentation strategy significantly improved accuracy, increasing the F-score by around 18% and the IoU by 25% in time series with 5, 10, and 15 images. These results demonstrate that the proposed augmentation strategy is a promising method for minimizing biases when detecting invariant targets. Per-object accuracy metrics help determine the optimal number of images in the time series for balancing performance, with this study identifying 5 or 10 images as the most effective. The semantic-to-instance segmentation conversion is highly effective, as offshore wind farms are spatially isolated, offering advantages such as simplified training data and improved semantic segmentation accuracy.