AI-Based Mapping of Offshore Wind Energy Around the Korean Peninsula Using Sentinel-1 SAR and Numerical Weather Prediction Data

Abstract

Offshore wind farm projects are being promoted in the seas surrounding the Korean Peninsula to secure renewable energy. To support site selection, offshore wind resource maps were generated using deep neural networks trained on Sentinel-1 SAR imagery, numerical weather prediction data, offshore wind observations, sea surface temperature, and bathymetry. The deep neural network (DNN) framework consisted of six sub-models targeting eastward and northward wind components across three regions—the Yellow Sea, Korea Strait, and East Sea—to account for spatial heterogeneity. The proposed models outperformed existing approaches, achieving mean absolute errors (MAE) ranging from 1.31 to 1.69 m/s and correlation coefficients (CC) between 0.827 and 0.913. These DNN models were then applied to produce offshore wind energy maps at a 150 m resolution, effectively capturing seasonal and regional variability. The resulting high-resolution maps provide valuable insights for evaluating the suitability of existing wind farm sites and identifying potential new candidates.

1. Introduction

Offshore wind power is an effective source of renewable energy that contributes to sustainable development. Since wind farms must be located in areas with high-quality wind resources, assessing wind potential through direct or indirect measurements is essential for identifying suitable sites. The most accurate method for evaluating wind resources is the installation of offshore observation towers equipped with anemometers and lidar systems. However, this approach is costly and limited to point-based measurements, making it primarily suitable only for preselected candidate sites. To identify broader areas with abundant wind resources before final site selection, large-scale spatial information is required.

In this context, the United States developed a comprehensive wind resource map covering North America [1], while the European Union produced the New European Wind Atlas, which includes the North Sea, the Mediterranean, parts of the Atlantic, and the Baltic Sea [2,3]. Similarly, the Technical University of Denmark, in collaboration with the World Bank Group, created a global wind atlas based on the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) dataset [4]. Australia has also developed a national wind resource map, including its exclusive economic zone (EEZ), using ERA5 data [5]. South Korea, as one of the countries actively promoting offshore wind energy, has developed wind resource maps based on regional weather models [6]. However, these maps, with spatial resolutions on the order of several kilometers, are insufficient to capture detailed spatial distributions in coastal areas, where complex shorelines contribute to high uncertainty in wind conditions. Consequently, there is an increasing demand for higher-resolution wind resource maps around the Korean Peninsula.

Synthetic Aperture Radar (SAR) remote sensing has been proposed as a complementary approach to enhance the spatial resolution of offshore wind resource assessments [7]. Although SAR data have relatively low temporal resolution due to infrequent satellite overpasses, they provide spatial resolutions below 100 m—significantly higher than those of regional weather models. With resolutions ranging from a few meters to several tens of meters, along with the ability to capture wide swaths under all weather conditions, SAR is particularly well-suited for monitoring complex coastal regions. SAR estimates ocean wind speeds by detecting variations in surface roughness. As surface roughness increases due to stronger ocean winds, the SAR backscatter intensity from the ocean surface also increases, forming the basis of wind speed retrieval using SAR [8]. Geophysical Model Functions (GMFs) have been widely employed to retrieve coastal wind fields from SAR observations [9,10]. More recently, artificial intelligence (AI) models have demonstrated comparable or even superior accuracy relative to GMFs in offshore wind estimation [11,12].

A well-known limitation of SAR satellites is their low temporal resolution, resulting from relatively long revisit intervals. However, in the context of wind resource mapping, the primary focus is on long-term average wind conditions—such as annual and monthly means—rather than instantaneous measurements. This limitation can therefore be mitigated by compiling multi-year SAR datasets to derive long-term averages. Sentinel-1A and Sentinel-1B, the two satellites employed in this study, each have a revisit cycle of 12 days, providing a combined revisit interval of 6 days for a given location. By collecting data over more than six years, it becomes possible to achieve an average temporal resolution of approximately one observation per day.

Building on previous advances, recent studies have employed a fusion of SAR and numerical weather data using AI techniques to generate offshore wind energy maps around the Korean Peninsula [12,13]. However, the spatial resolution of these maps, the robustness of the models with optimal variables, and the interpretation of regional and seasonal characteristics remain insufficient for the Korean Peninsula. To address these limitations, we developed multiple AI models to produce high-accuracy offshore wind resource maps at a 150 m resolution on a monthly basis for the three different waters surrounding the Korean Peninsula. We utilized nine years (2015–2023) of Sentinel-1A/B SAR imagery, regional weather model data, sea surface temperature (SST), and bathymetry data to optimize the combination of input variables, unlike previous research. The resulting monthly 150 m wind energy maps also reveal distinct regional and seasonal wind patterns around the Korean Peninsula.

2. Materials and Methods

The study area, shown in Figure 1, covers the Exclusive Economic Zone (EEZ) of South Korea, extending from 32°30′ N to 38°27′ N latitude and from 122°52′ E to 131°30′ E longitude. This region includes three major seas: the Yellow Sea, the Korea Strait, and the East Sea (Sea of Japan). These waters are bordered by several neighboring countries, including China, Japan, Russia, and North Korea.

The Yellow Sea lies on the continental shelf between China and the Korean Peninsula, with average and maximum depths of approximately 44 m and 140 m, respectively. It is characterized by a large tidal range and extensive mudflats [14]. The coastal zones of the Yellow Sea are heavily industrialized, with major industrial complexes located in both China and Korea. In addition, numerous offshore oil fields operate in the northwestern part of the sea. The Korea Strait, also situated on the continental shelf, connects the East China Sea to the East Sea and serves as a passage for the northward-flowing warm Kuroshio waters. It has a relatively shallow depth of about 120 m and receives substantial freshwater discharge from the lower Yangtze River, particularly during the rainy season [15]. The East Sea is a semi-enclosed marginal sea bordered by South Korea, Japan, Russia, and North Korea. With depths reaching up to 3500 m, it functions as a deep basin. The Kuroshio Current enters the East Sea through the Korea Strait, with outflow occurring through two narrow northern straits [16]. Together, the Yellow Sea, the Korea Strait, and the East Sea exhibit distinct oceanographic characteristics that are critical for assessing offshore wind resources around the Korean Peninsula.

To generate offshore wind energy maps for the study area, 7477 Sentinel-1 SAR images acquired between 2015 and 2023 were obtained from the NASA Alaska Satellite Facility Distributed Active Archive Center (ASF DAAC) [17]. From the downloaded Ground Range Detected (GRD) products, data collected in Interferometric Wide (IW) swath mode with Vertical–Vertical (VV) and Vertical–Horizontal (VH) polarizations were extracted. Orbit correction, radiometric calibration, and geometric correction were then performed using the Sentinel Application Platform (SNAP) software (ver. 2.72) provided by the European Space Agency (ESA) [18]. In addition, numerical weather prediction (NWP) data from the Local Data Assimilation and Prediction System (LDAPS), operated by the Korea Meteorological Administration (KMA), were collected for the same period [19].

Observed offshore wind data from marine buoys (2015–2023) were obtained from both the KMA [20] and the Korea Hydrographic and Oceanographic Agency (KHOA) [21]. Daily sea surface temperature (SST) data were acquired from the Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) dataset [22], while bathymetry data for the study area were obtained from KHOA [23]. Unlike previous studies, we incorporated SST and bathymetry as new explanatory variables, since seawater viscosity is physically related to temperature—warmer water, having lower viscosity, tends to produce rougher surface conditions under the same wind speed. The correlation between SST and offshore wind has also been demonstrated in previous statistical studies of the coastal seas around Korea [24]. In addition, wave development can be influenced by ocean depth [25]. Given that the study area is located in the mid-latitudes, seasonality is also expected to play an important role. Although not considered in earlier research [13], seasonality was incorporated by encoding the month as a cosine-transformed variable, allowing the model to capture its circular annual pattern. For generating the training dataset, all variables in the data-matching procedure were aligned to the nearest timestamp relative to the Sentinel-1 SAR acquisition time. Consequently, the temporal mismatch was constrained to less than 30 min for buoy wind observations, less than 3 h for LDAPS fields, and less than 1 day for SST data. Furthermore, although the datasets possess different spatial resolutions, the values assigned were taken from the nearest pixels relative to the buoy points.

To develop AI models for wind speed estimation, we conducted a preliminary performance test comparing a random forest (RF) as a bagging model, a gradient boosting machine (GBM) as a boosting model, and a deep neural network (DNN) as a neural network model. Based on the results, we selected the DNN for this study to estimate offshore wind speeds around the Korean Peninsula. DNNs are well-suited for capturing nonlinear relationships through multilayer architectures optimized intensively across the network. The use of rectified linear unit (ReLU) activation functions mitigates the vanishing gradient problem during loss minimization. Regularization techniques help reduce overfitting by preventing the model from assigning excessive weight to specific neurons. Furthermore, dropout methods enhance generalization by randomly deactivating neurons during training, thereby improving robustness to unseen data. The specifications of the training dataset used for the DNN model are summarized in

Table 1. Summary of the dataset comprising 9117 records used to train the DNN models for offshore wind estimation, including mean, standard deviation (SD), spatial resolution (SR), temporal resolution (TR), and data source for each variable.

Type	Name	Mean	SD	Unit	SR	TR	Source
In situ observation with marine buoys	Wind speed	5.38	3.39	m/s	Point	1 to 30 min	KMA and KHOA
	Wind direction	188.9	115.4	°
SAR intensity of Sentinel-1A/1B	VV intensity	−18.80	4.99	dB	10 m	12 days	ESA
	VH intensity	−29.30	3.61	dB
LDAPS numerical weather data	Eastward wind speed	0.55	4.29	m/s	1.5 km	3 h	KMA
	Northward wind speed	−1.62	4.46	m/s
Synthetic product with in situ and satellite data	Sea surface temperature	291.05	5.59	°K	1 km	1 day	OSTIA
Echo sounding	Bathymetry	183.4	420.2	meter	150 m	N.A.	KHOA

.

To train the DNN models, the datasets were spatially and temporally aligned to create a unified matchup dataset. Using the acquisition date and pixel location of each Sentinel-1 SAR image as reference points, all datasets were merged into a two-dimensional table comprising 9117 rows. Each row included the following variables: buoy-observed wind speed and direction, longitude, latitude, timestamp, SAR VV and VH backscatter intensities, LDAPS-derived eastward and northward wind components (at pixel resolution), SST, bathymetry, and calendar month (to account for seasonality).

Based on this training dataset, two separate DNN models were constructed, as illustrated in Figure 2: one for estimating eastward wind speed and the other for northward wind speed. Given the substantial oceanographic differences among the Yellow Sea, Korea Strait, and East Sea—three distinct subregions within the study area—separate DNN models were developed for each subregion and wind component. In total, six DNN models were constructed. Model training and hyperparameter optimization were conducted in RStudio (ver. 2024.12.1+563) using the H2O library [26,27]. Figure 3 shows a section of the code, including the input features, the target variable, and the selected hyperparameters for the h2o.deeplearning function. We utilized the default optimizer, AdaDelta, and adopted an early stopping option when the optimization began to diverge. The remaining hyperparameters—such as the activation function, number of hidden layer neurons, training epochs, learning rate decay, and dropout ratios—were manually tuned through empirical experiments to identify the best performance, as summarized in

Table 2. Hyperparameter settings of the DNN models for eastward (EW) and northward (NW) winds across the Yellow Sea, Korea Strait, and East Sea.

Hyperparameters	Yellow Sea		Korea Strait		East Sea
	EW	NW	EW	NW	EW	NW
Activation function	ReLU	ReLU	ReLU	ReLU	ReLU	ReLU
Hidden layer neurons	100	100	100	100	100	100
Training epochs	5000	5000	2000	2500	5000	2000
Learning rate decay	0.95	0.95	0.99	0.9	0.99	0.95
Hidden dropout ratio	0.4	0.4	0.4	0.5	0.1	0.1
Input dropout ratio	0.05	0.05	0.05	0.0	0.05	0.1
Optimizer	AdaDelta	AdaDelta	AdaDelta	AdaDelta	AdaDelta	AdaDelta

. Figure 4 further presents a diagram summarizing the implementation of the DNN models with the optimal hyperparameters detailed in Table 2. Furthermore, 5-fold cross-validation was employed during the training and validation process to test model performance.

After building the DNN models, wind resource maps at a spatial resolution of 150 m were generated by inputting Sentinel-1 VV and VH backscatter intensities, LDAPS-derived eastward and northward wind components, OSTIA SST, and bathymetry data for each pixel within the target sea areas. A total of 7477 Sentinel-1 SAR images acquired between 2015 and 2023 were used in this process. The estimated wind speed maps were subsequently aggregated on monthly and annual timescales.

3. Results

3.1. Accuracy of the DNN Models for Wind Speed

To evaluate the accuracy of the DNN models within a 5-fold cross-validation framework, several statistical metrics were computed, including mean bias error (MBE), mean absolute error (MAE), root-mean-square error (RMSE), and the correlation coefficient (CC). In addition, scatter plots were generated to compare buoy-observed wind speeds with model estimates, as shown in Figure 5. Key statistical indicators reflecting model performance were also calculated in

Table 3. Validation results of the six DNN models for eastward (EW) and northward (NW) wind components in the Yellow Sea, Korea Strait, and East Sea.

Metrics	Yellow Sea		Korea Strait		East Sea
	EW	NW	EW	NW	EW	NW
MBE (m/s)	0.029	−0.020	−0.058	0.014	0.079	−0.100
MAE (m/s)	1.311	1.404	1.514	1.488	1.543	1.687
CC	0.827	0.912	0.913	0.878	0.848	0.861

.

The CC values of the DNN models for eastward and northward wind components in the Yellow Sea were 0.827 and 0.912, respectively. In the Korea Strait, the CC values were 0.913 and 0.878, while in the East Sea they were 0.848 and 0.861. The corresponding MAE values were 1.311 and 1.404 m/s in the Yellow Sea, 1.514 and 1.488 m/s in the Korea Strait, and 1.543 and 1.687 m/s in the East Sea for eastward and northward winds, respectively. All models exhibited negligible bias, with MBE values close to zero. Too low wind speeds are not critical because they are not suitable for wind energy development. Conversely, excessively high wind speeds are also undesirable due to safety concerns for wind turbines.

These results indicate that both CC and MAE varied across the three sea regions, as well as between the eastward and northward wind models within each region. Notably, in the Korea Strait the eastward wind model achieved a higher accuracy than the northward model, whereas in the Yellow Sea the northward model outperformed the eastward one. The smallest discrepancy in correlation between wind components was observed in the East Sea. Overall, based on both CC and MAE, the DNN models demonstrated the highest accuracy in the Korea Strait, followed by the Yellow Sea and the East Sea.

3.2. Feature Importance in the DNN Models for Wind Speed

To assess the relative importance of the input variables in the DNN models, a Permutation Feature Importance (PFI) analysis was conducted. As summarized in

Table 4. Permutation feature importance (%) for the DNN models used in wind speed estimation.

	Regions	Yellow Sea		Korea Strait		East Sea
Features		East	North	East	North	East	North
SAR VV Intensity		16.6	13.8	16.7	14.2	12.0	12.5
SAR VH Intensity		11.6	9.8	6.9	9.7	8.1	7.9
LDAPS Eastward wind speed		19.3	14.0	33.9	9.5	17.8	13.4
LDAPS Northward wind speed		12.6	29.3	12.9	33.1	13.7	20.0
Sea surface temperature		11.8	12.6	9.5	11.1	14.0	12.3
Bathymetry		10.9	8.8	13.0	13.0	21.9	19.7
Month		17.2	12.4	7.2	9.4	12.4	14.1

, the eastward and northward wind models across the three sea regions showed varying contributions from each input feature. The eastward wind components from the regional NWP model contributed between 9.5% and 33.9% of the total feature importance, while the northward components contributed between 12.6% and 33.1%. Sentinel-1 SAR VV polarization accounted for 12.0–16.7%, and VH polarization for 6.9–9.7%. SST contributed 9.5–14.0%, bathymetry 8.0–21.9%, and the seasonal variable Month (included as a categorical input) accounted for 7.2–17.2%.

Among all input features, the LDAPS-derived eastward and northward wind components exhibited the greatest importance across all six DNN models. For eastward wind estimation, the LDAPS eastward component was dominant, while in the northward models, the LDAPS northward component held greater importance. SAR-derived features ranked next in importance. Combined, the LDAPS wind components explained 32–47% of the total feature importance, whereas the Sentinel-1 VV and VH polarizations together accounted for 20–28%, highlighting the balanced integration of numerical model outputs and satellite observations in the DNN architecture. Despite their relatively coarse spatial resolution of 1.5 km, the LDAPS wind components proved effective for estimating wind fields at a 150 m resolution. The SAR backscatter intensities also served as appropriate variables, as they capture variations in ocean surface roughness.

In the East Sea, bathymetry emerged as a particularly significant factor, contributing up to 22%. SST demonstrated a moderate level of importance across all regions and models. The Month variable had relatively low influence (below 10%) in the Korea Strait but became more critical in the Yellow Sea and East Sea, reaching up to 17%.

3.3. Offshore Wind Energy Maps Using DNN Models

3.3.1. Spatial Distribution of Surface Wind Speed over the Korean Seas

The annual mean offshore wind speeds estimated using the DNN models and associated datasets are shown in Figure 6 at a spatial resolution of 150 m. Surface wind speeds ranged from 3 to 9 m/s, generally increasing with distance from the coastline across the study region. In the Yellow Sea, nearshore areas exhibited relatively weak winds of 3–4 m/s, rising westward toward the open sea to about 6 m/s. In the Korea Strait, coastal waters dotted with islands showed wind speeds of 3–4 m/s, increasing to 5–7 m/s south of Jeju Island. Similarly, in the East Sea, nearshore regions experienced weaker winds of about 3 m/s, increasing to roughly 6 m/s farther offshore.

Regions with annual mean wind speeds below 4 m/s—shown in light and dark green in Figure 6—covered extensive areas of the Yellow Sea. In contrast, such low-wind zones in the Korea Strait and East Sea were confined to narrow coastal bands. Overall, the strongest wind resources were observed in the southern offshore areas of the Korea Strait and the waters between Jeju Island and Tsushima Island, while the weakest were found along the western coast of the Korean Peninsula in the Yellow Sea. The East Sea exhibited intermediate wind conditions, with wind speeds between those of the Yellow Sea and the Korea Strait.

3.3.2. Monthly Variation in Surface Wind Speed over the Korean Seas

Seasonal variations in offshore wind speeds, estimated using the DNN models and associated datasets, are shown in Figure 7 at a spatial resolution of 150 m. The results reveal a clear seasonal cycle typical of the Northern Hemisphere, with stronger winds in winter (e.g., January) and weaker winds in summer (e.g., July).

In January, wind speeds across the study area—including the Yellow Sea, Korea Strait, and East Sea—ranged from 5 to 9 m/s (Figure 7a). As the region transitioned into spring, wind speeds declined. In April, they ranged from 3.0 to 5.2 m/s in the Yellow Sea, 4.0–7.0 m/s in the Korea Strait, and 4.0–6.0 m/s in the East Sea (Figure 7b). Wind speeds reached their seasonal minimum in July, with values of 2.5–5.0 m/s in the Yellow Sea, 3.0–5.2 m/s in the East Sea, and 3.0–6.0 m/s in the Korea Strait (Figure 7c).

Following summer, wind speeds increased again in autumn. In October, they ranged from 3.0 to 6.0 m/s in the Yellow Sea, 4.0–9.0 m/s in the Korea Strait, and 3.0–7.5 m/s in the East Sea (Figure 7d). This cycle culminates in January, when wind speeds reach their annual maximum across all regions (Figure 7a).

4. Discussions

4.1. Comparisons with Previous Studies

In previous studies, Horstmann and Koch [28] reported RMSE values ranging from 2.11 to 2.85 m/s and biases between −0.16 and 0.85 m/s for SAR-based wind speed retrievals. Wei et al. [29] sought to improve accuracy using Geophysical Model Function versions 5 and 7 (CMOD5 and CMOD7) applied to C-band SAR images in the eastern coastal waters of the United States, achieving RMSE values of 1.9 m/s and a bias of 2.1 m/s. The European Space Agency (ESA) has developed Ocean (OCN) products by integrating Sentinel-1 SAR data with European Centre for Medium-Range Weather Forecasts (ECMWF) model outputs at 1 km resolution, with reported RMSE values generally ranging from 1.2 to 2.0 m/s depending on the region [30]. A study conducted in South Korea focusing on the East Sea reported RMSE values of 1.30–1.72 m/s for SAR-derived offshore wind speeds [31].

Our model achieved RMSE values of 1.75–2.17 m/s and MBE values between −0.10 and 0.08 m/s, which are comparable to the performance reported in previous SAR-based studies (

Table 5. Accuracy comparison between the proposed DNN models on SAR and existing studies on SAR wind retrieval.

	Accuracy	RMSE	MAE	MBE	Method	Region	Reference
Study
Horstmann and Koch [28]		2.11–2.85	NA	−0.16–0.85	GMF	Spitsbergen, Norway	NWP forecast
Wei et al. [29]		1.31–5.50	1.00–3.64	−0.06–1.88	GMF	Miami, U.S.	Marine buoy
OCN Products [30]		1.2–2.0	NA	NA	GMF	European Seas	Marine buoy
Kim et al. [31]		1.30–1.72	NA	NA	GMF	East Sea, South Korea	Reanalysis
Ours		1.75–2.17	1.31–1.69	−0.10–0.08	DNN	Offshore, South Korea	Marine buoy

). It should be noted, however, that performance comparisons are not entirely equivalent: Wei et al. [29] for the Miami region and the OCN products [30] for the European seas used in situ buoy data as reference, whereas the accuracy of SAR-derived winds can vary considerably across regions. In the case of South Korea, Kim et al. [31] did not use in situ buoy observations, and Yun et al. [6] compared their estimated wind maps to only a single buoy location. In this context, our results can be regarded as sufficiently accurate when compared with previous studies conducted both in South Korea and in other regions.

4.2. Regional Characteristics

The wind speed values and model errors varied by region (west, south, and east) and by wind component (eastward and northward). In the Yellow Sea (West Sea), most eastward wind speeds ranged between −10 and 10 m/s, whereas the northward component showed many cases below −10 m/s, indicating strong southward winds. This is partly due to the influence of the Siberian High-Pressure System, whose center lies over Mongolia and northern China, driving cold air southward toward the Korean Peninsula [6]. In contrast, eastward winds were slightly stronger than northward winds in the Korea Strait (South Sea), reflecting the influence of the westerlies, which generally blow from west to east throughout the year. In the Korea Strait, southward winds (below −5 m/s) were much more dominant than northward winds (above 5 m/s), because the presence of landmasses to the north of the strait tends to suppress the development of winds blowing from south to north. Positive correlations were generally observed between wind speed and wave height; however, such relationships were slightly weaker in the Korea Strait than in the Yellow Sea or East Sea [32,33]. Nevertheless, the accuracy statistics—such as MAE and CC—were similar for both eastward and northward wind components across the three regions, suggesting that the model’s performance was overall stable.

The PFI analysis in Table 4 showed that bathymetry contributed approximately 10% to model importance in the Yellow Sea and Korea Strait, but increased to about 20% in the East Sea. The Yellow Sea and Korea Strait are shallow continental shelf seas, with maximum depths of roughly 100 m and 130 m, respectively, whereas the East Sea is a deep oceanic basin reaching depths of around 3500 m. The greater importance of bathymetry in the East Sea model may reflect the underlying oceanographic dynamics. The East Sea spans approximately 1,000,000 km² and features steep bathymetric gradients, in contrast to the shallower Yellow Sea (~380,000 km²) and Korea Strait (~980,000 km²). Such bathymetric variability, even across relatively short horizontal scales, can influence spatially heterogeneous surface wind conditions. Moreover, the East Sea is well known for its frequent mesoscale eddy activity [34,35], which generates complex current and wave fields, further modifying surface wind patterns. This complexity is evident in Figure 6 and Figure 7, where the East Sea exhibits more spatially intricate wind structures than the other regions. The high bathymetric importance captured by our DNN model likely reflects this indirect encoding of spatial context.

4.3. Seasonal Characteristics

Winter winds were stronger than in other seasons across the three regions: the Yellow Sea, Korea Strait, and East Sea. During winter, the Siberian High develops very strongly over the cold continental interior of northern China and Mongolia, with central pressures often exceeding 1050 hPa. At the same time, a low-pressure system (the Aleutian Low) typically forms over the western Pacific, east of Japan. This pressure distribution creates a strong pressure gradient from west to east, driving cold northwesterly winds from the Asian continent toward the Korean Peninsula and the surrounding seas. Moreover, Korea’s geography plays an important role. The Korean Peninsula acts as a pathway for cold continental air flowing toward the sea. The Yellow Sea has broad, flat coastal plains that allow cold air to move freely southward. The Korea Strait is a narrow passage between Korea and Japan, producing a channeling effect that accelerates the wind. In the East Sea, the Taebaek Mountains generate downslope (foehn) winds, further enhancing wind speeds along the coast. In autumn, although the sea surface temperature remains relatively warm, the air temperature gradually decreases, leading to the onset of cold-air advection over the Korea Strait. When the sea surface is warmer than the overlying air, the air is heated from below and begins to rise, making the atmosphere unstable. This instability enhances vertical air circulation, which in turn strengthens the surface winds over the sea [36,37,38].

4.4. Wind Energy Development

From an energy perspective, surface wind speeds of ≥5 m/s are generally considered suitable for the economic operation of offshore wind turbines. In Figure 5, the areas shaded in pink, red, and purple exceed this threshold and include the southern parts of the East Sea, offshore regions of the Korea Strait beyond the coastal islands, and the western offshore zones of the Yellow Sea. However, the buoy-based observations used for model training are typically recorded at 3 m above sea level [20], whereas wind turbines are installed at hub heights of 35 m or more. Using the Deacon power-law equation [39], a 5 m/s wind speed at 3 m corresponds to approximately 6.4 m/s at 35 m, 6.75 m/s at 60 m, 7.1 m/s at 100 m, and 7.2 m/s at 120 m. Conversely, a 5 m/s wind speed at 35 m translates to about 3.9 m/s at 3 m. Therefore, areas with surface wind speeds of 4 m/s—shown in yellow in Figure 4 and Figure 5—may reach 5 m/s at turbine hub heights, making them potentially viable for offshore wind energy development. Moreover, wind turbines must remain operational even during low-wind summer periods. Figure 7c shows July wind speeds, where light and dark green areas fall below 4 m/s, indicating regions that are generally less favorable for development. However, increasing the hub height to 120 m would allow a 5 m/s wind speed at that height to correspond to roughly 3.46 m/s at 3 m, thereby expanding the feasible development area—particularly in the Yellow Sea, where the dark green regions could meet this adjusted threshold.

The region between Jeju Island and Tsushima Island in the Korea Strait exhibited the strongest offshore winds, with an annual mean of approximately 7 m/s, a summer mean of 5 m/s (Figure 7c), and a winter peak of 9 m/s (Figure 7a). These conditions indicate that the area is highly favorable for offshore wind energy development. However, it lies along a frequent typhoon path, where wind speeds can exceed 33 m/s [40]. Therefore, wind turbine installations in this region will likely require countermeasures against typhoons, such as reinforced structural designs or preemptive evacuation systems.

4.5. Application to Other Regions

This study aimed to develop a wind energy resource map for the seas surrounding the Korean Peninsula. Accordingly, supplementary meteorological and oceanographic datasets, together with SAR satellite data, were collected for the study area. To generate wind resource maps for other marine regions, it would be necessary to collect region-specific data. The Sentinel-1 SAR and OSTIA SST datasets are globally available and can therefore be applied to various oceanic regions. The regional atmospheric model (LDAPS) used in this study can be substituted with global reanalysis models, such as ECMWF [30], for applications in other areas. While high-resolution bathymetric data are preferable, the GEBCO dataset—which provides a spatial resolution of approximately 500 m [41]—can be used when detailed local bathymetry is unavailable. Importantly, in situ offshore wind observations can be obtained either from national meteorological agencies or from the NOAA National Data Buoy Center (NDBC) [42].

4.6. Limitations and Future Work

Only the wind speeds measured at the time of the Sentinel-1 satellite overpass were aggregated to compute the monthly mean, which was then considered a representative value. However, because the satellite overpass times are fixed (09:31 UTC and 21:32 UTC for this region), these wind measurements are not randomly sampled but temporally biased. Moreover, wind speeds can exhibit a much wider range over the full diurnal cycle due to solar-driven atmospheric expansion and contraction. Therefore, orbital SAR-derived wind speeds have inherent limitations in fully satisfying the statistical requirements for representative sampling over the diurnal period.

The training dataset used for model development contained wind speeds up to approximately 16 m/s. In contrast, the maximum wind speed observed in the study region during the same period reached 60 m/s due to typhoon events. These extreme winds were excluded from the training dataset for two reasons: Sentinel-1 imagery was unavailable during those periods, and the limited number of buoy observations under such extreme conditions exhibited questionable quality. As a result, model predictions exceeding the 0–16 m/s training range should be interpreted with caution. Nonetheless, given that the wind-resource maps published by the KMA show maximum wind speeds of about 14 m/s [43], defining the operational wind-speed domain as 0–16 m/s is considered reasonable for the waters surrounding the Korean Peninsula. Nevertheless, for offshore wind industry purposes, wind resource assessments should encompass the extreme wind speeds induced by typhoons. Further studies are therefore required to incorporate these extreme conditions into wind resource mapping.

Complex coastal geography can significantly influence wind speed. Vogelzang et al. [44] demonstrated that incorporating coastal distance into scatterometer-based wind models improves both the accuracy and magnitude of wind estimates. Therefore, follow-up studies are needed to develop AI models that incorporate geographic information—such as coastal topography and elevation—as explanatory variables.

Although our model demonstrated stable performance across the three regions surrounding the Korean Peninsula, future work should include several advancements to ensure greater reliability. These should encompass analyses of model uncertainty, better methods for handling extreme wind speeds, and considerations for complex coastal geography. Additionally, the inclusion of more spatially and temporally high-resolution SAR imagery is required. This will enable more localized, high-resolution, and realistic assessments of wind resources in coastal areas.

5. Conclusions

In this study, DNN models were developed to estimate offshore wind energy resources around the Korean Peninsula, using Sentinel-1 SAR imagery, numerical weather prediction (NWP) data, and bathymetry. The DNN framework comprised six sub-models, each designed to predict eastward and northward wind components across three regions—the Yellow Sea, Korea Strait, and East Sea—to account for spatial heterogeneity. The proposed models outperformed existing approaches, achieving the MAE ranging from 1.31 to 1.69 m/s and the CC between 0.827 and 0.913. These DNN models were then applied to generate offshore wind energy maps with a 150 m spatial resolution, effectively capturing both seasonal and regional variability

The resulting high-resolution maps enable preliminary assessments of the abundance and variability of regional wind resources without the need for in situ measurements. During the early stages of site selection, they can serve as valuable tools for identifying areas with high-quality and abundant wind energy potential across broad marine regions. As offshore wind power continues to emerge as one of the most promising sources of renewable energy, the developed maps are expected to serve as essential resources supporting the strategic growth of the offshore wind industry.