Offshore wind data assessment near the Iberian Peninsula over the last 25 years

Numerous processes affecting coastal ocean dynamics and water properties occur at the air-sea interface as a result of wind blowing on the ocean surface. In Earth system research, it is crucial to appropriately characterize the ocean surface wind (OSW) field because of its significance in many academic and economic activities. This study aimed to evaluate the accuracy of the most recent OSW datasets based on numerical modeling and remote sensing products in estimating in situ observations along the Atlantic coast of the Iberian Peninsula. The results are three-fold: (1) when high temporal resolutions are not necessary, remote sensing products are an excellent choice because they provide reliable OSW estimates; (2) for analyses that require high temporal resolution, numerical weather models are the best choice because they can statistically reproduce the main trend; (3) fifth generation of European ReAnalysis (ERA5) showed that, despite having a lower spatial resolution than the dynamically downscaled weather research and forecasting simulation, it captures the spatial and temporal dynamics and variability of coastal winds and may be used as forcing of the atmosphere-ocean interface modeling without compromising its accuracy.


Introduction
Accurate numerical modeling of the Earth system relies on the precise representation of the oceanographic and atmospheric processes at the air-sea interface, such as heat, moisture content, and momentum exchanges. The ocean surface wind (OSW) directly impacts these processes and is also responsible for wave generation [1], coastal processes such as upwelling, ocean currents, sea surface temperature, and chlorophyll filaments advection [2]. Therefore, the correct representation of OSW is crucial for several academic, industrial, and economic activities that are based on the results of atmospheric and oceanic numerical models [3,4], offshore wind energy resource assessment and power production estimation [5,6].
The OSW data are available from several sources such as in-situ, remotely sensed, numerical weather predictions (NWP) models and atmospheric reanalysis, each with different strengths and weaknesses. The in-situ surface wind is measured at varying heights (usually 2 and 3 m) above sea level by oceanographic buoys in moorings, aboard ships, gliders, among others, being a direct measure of the wind field at a specific location. However, due to instrumental failures and the difficulty in replacing them, most time series have data gaps.
OSW can also be estimated from measurements performed by passive and active sensors on board of satellites that orbit the Earth, available at a near-global scale and in near real-time. However, passive sensor observations often suffer from significant data gaps due to various factors, such as cloud cover and the need for suitable sunlight conditions to obtain accurate readings. Additionally, equipment failure can also contribute to data gaps. Active sensors obtain measurements anytime, regardless of the time of day or season, and usually have a higher spatial resolution but smaller footprints compared to passive sensors. The OSW estimated by remote sensing is expected to have inaccuracies due to the influence of the local topography, the discontinuity between land and sea roughness, and also due to the thermal gradients resulting from land-sea

In-situ observations
Time series of observed wind data measured by 11 oceanographic buoys moored along the north, west and south of the Iberian Peninsula Atlantic coast (figure 1 and table 1), were collected. Six of these 11 buoys are operated by the Puertos del Estado Spanish Agency (PdE, Spain) and measure hourly wind speed and direction since the 1990s at 3 m above sea level. The remaining five buoys are operated by the Hydrographic Institute (IH, Portugal) agency of the Portuguese Navy and measure hourly wind data at 4 m above sea level since 2010. The data was collected from the Copernicus Marine Environment Monitoring Service In Situ Thematic Assembly Centre (CMEMS In Situ TAC, www.marineinsitu.eu/dashboard/), ensuring consistent and reliable access to a range of in situ data that can be used to service production and validation. The horizontal wind speed and direction is measured by a YOUNG Model 86106 Ultrasonic Anemometer sensor (more details at www.youngusa.com/product/ultrasonic-anemometer-5/). The wind data was converted to 10 m height considering the wind speed and wave height measurements and the neutral log profile corrected for the effects of low-level distortion of the wind profile by surface waves following [30] and given by equation: where U 0 is the surface wind velocity, u * if the friction velocity, κ = 0.4 is the von Kármán constant, z r is the reference height 10 m, z 0 is the height where the wind and the surface speed are equal U (z 0 ) = U 0 , H w is the wave height and Ω is a logarithm profile correction function due to wind distortion associated with surface waves (details in [30]).

Sentinel-1 level 2 OCeaN (OCN) products
Wind speed and direction can be derived from the Sentinel-1 (S1) A and B satellites observations (S1 B is unavailable since 23 December 2021) made by the Synthetic Aperture Radar (SAR) instrument on board. Sentinel-1 is in a near-polar, sun-synchronous orbit with a 12 day repeat cycle and 175 orbits per cycle for a single satellite. If combined with other SAR missions, the repeat cycle can be reduced to a single day. Geophysical data, such as ocean wind field (OWI), ocean swell spectra (OSW), or surface radial velocity (RVL), is derived from the level 1 ground range detected products and made available as Level 2 OCN products by the European Space Agency (ESA, https://scihub.copernicus.eu/). However, the availability of these data (OWI, OSW, and RVL) depends on the Sentinel acquisition mode. The OWI product can be derived from the StripMap (SM, 170 km × 80 km), interferometric wide swath (IW, 170 km × 250 km), or extra-wide swath (EW, 400 km × 400 km) modes. The SAR sensor can operate in all weather conditions, independent of cloud cover or solar illumination conditions (day and night). The OWI product, which assumes neutral atmospheric stratification, is a ground range gridded estimate of the surface wind speed and direction at 10 m above the surface with a spatial resolution of 1 km. OCN products are estimated from the Level 1 SAR images by inversion of its associated normalized radar cross section (NRCS) values using Bayesian statistical inference. The spatial resolution of the vector is directly related to the area on which the NRCS is computed [31]. The IW swath is the primary Sentinel acquisition mode for the Iberian coast. We gathered all OCN products accessible along the Iberian coast between the first S1 A (the first satellite to be launched) SAR images acquired from January of 2017 to December of 2021. Combining both satellites (A and B) and different orbits (ascending and descending), the number of products for each site varies from 857 to 1113, except for Nazaré (469). The S1 wind and direction were bilinearly interpolated (using the closest four pixels) to the buoy site location. Only pixels rated as good (0: good, 1: medium, and 2: poor [31]) based on the quality indicator of the inversion algorithm are retained to ensure accuracy.

ASCAT wind products
The ASCAT is a real-aperture radar instrument launched on board the satellites MetOp-A in 2006 (finished in 2021), MetOp-B in 2012, and MetOp-C in 2018 by the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT). The ASCAT instrument is used to obtain wind measurements, for use, primarily in weather forecasting and climate research. This instrument uses radar (C-band) to measure the electromagnetic backscatter from the wind-roughened ocean surface. As a result of the small-scale disruptions that winds over the sea create to its surface, which alter its radar backscattering characteristics, wind speed and direction can be estimated [32]. The MetOp satellites have a polar orbit, measure sea surface winds in two 500 km wide swaths at 25 km and provide global coverage in just five days. Details about the mission and the wind retrieval algorithm can be found in [33]. Due to land contamination in the coastal zone, which limits the quality of the product with the distance to the coast (approx. 70 km for the 25 km product), a high-resolution ASCAT at 12.5 km, which uses a different way of averaging the full resolution backscatter measurements, is available and unrestricted by EUMETSAT. This product allows good quality winds over the sea as close as 15 km to the shore [34]. In this work, we used the ASCAT coastal (A, B, and C) product, integrated into the Physical Oceanography Distributed Active Archive Center (PO.DAAC) level 2 datasets available at the high-level tool for interactive data extraction (HiTIDE, available at https://hitide. podaac.earthdatacloud.nasa.gov/). The ASCAT-A coastal data expands from August 2010 to November 2021 (10 003 orbits with data over the Iberian coast), the ASCAT-B from September 2012 to December 2021 (8049 orbits), and ASCAT-C from October 2019 to December 2021 (2005 orbits). Analogous to S1 data, the ASCAT wind speed and direction were estimated to the buoy locations and observation time using bilinear interpolation in space and linear in time. After checking the ASCAT wind quality flags [35, p 28], only pixels without warnings are used in the interpolation process.

WRF model
The WRF mesoscale model is designed to serve atmospheric research needs and operational forecasting. The system is a next-generation mesoscale forecast and assimilation model. WRF is a non-hydrostatic model that is fully compressible [36]. It uses a terrain-following hydrostatic pressure coordinate for the vertical coordinate and an Arakawa C-grid staggering. The model uses a third-order Runge-Kutta scheme for time integration, with smaller time steps for the acoustic and gravity waves. Fifth-order and third-order advection schemes are used in the horizontal and vertical directions, respectively. The WRF model predicts a large set of atmospheric variables, including three velocity components, perturbation potential temperature, perturbation geopotential, and perturbation surface pressure of dry air. The model version 4.2 is implemented in a two-nest grid configuration with 60 vertical levels. With a horizontal resolution of 9 km, the largest domain covers the entire Iberian Peninsula, south France and north Africa, with a large portion of the Atlantic Ocean. The higher resolution domain, with a horizontal resolution of 3 km, covers all of the Portuguese Mainland, parts of Spain and a smaller region of the Atlantic Ocean covering the north, south and western coastal regions of the Iberian Peninsula. The model simulation covers over 44 years, starting in January 1979 (00 UTC) and ending in October 2022 (00 UTC). The initial and the 3-hourly boundary conditions for the simulation are obtained from the ERA5 Reanalysis, and the model physics package follows the one described in [37]. Both hourly meridional and zonal wind components (u, v) at 10 m above sea level are extracted from the model hourly output at the closest grid point to the buoy location.

ERA5 reanalysis
The ERA5 model is the fifth generation ECMWF atmospheric reanalysis of the global climate, covering the period from January 1950 to the present [14]. It reveals major improvements over its predecessor, the ERA-Interim reanalysis with improvements in the horizontal grid spacing (from 79 to 31 km), in the number of vertical levels to resolve the atmosphere from the surface to 80 km height (from 60 to 137), and in the temporal resolution (from 6 to 1 h). These changes allow a better representation of convective systems, gravity waves, tropical cyclones, and other meso-to synoptic-scale atmospheric structures [38]. These improvements are not only due to the advance of the physical parameterizations but also, and if not the most important, to the massive increase of observations that are assimilated, which went from on average 0.75 million per day in 1979 to about 24 million per day in 2018, boosted mainly by the increase of satellite radiances throughout the period, and more recently, by the GNSS-Radio Occultation, ozone products, ground-based radar observations, among other products [39]. In this study, 25

Methodology
The available wind data of each data source considered in this study over the 1996-2021 period is illustrated in figure 2 at each oceanographic buoy location. The in-situ data has been available since the early 1990s for some of the locations, others only after the 2010s. The buoys' data time series have gaps throughout their operation period. The buoy located at Sines was operational for a short period of only six months, therefore, it will not be considered for this study. The data based on numerical modeling is continuous in time and therefore is available for the entire study period (gray, black for WRF and green for ERA5). The satellite products are available from 2010 (ASCAT-A), 2017 (ASCAT-B) and 2019 (ASCAT-C). The accuracy of the offshore wind sources was assessed through several statistical analyses considering the available data over 25 years. The assessment of the wind velocity accuracy is conducted based on several statistical parameters: mean, bias, STD (σ), RMSE and Pearson correlation coefficient.
For the wind direction analysis and computation of statistical parameters, an adjustment to the wind direction is performed to overcome the difficulty of the discontinuity between 0 • and 360 • [40], as follows: • being θ and θB the wind direction of the data sources to be analyzed and of the buoy, respectively. When the difference is negative, it indicates a counterclockwise rotation of the wind vector measured by the buoy with respect to the value estimated by models or satellites, while a positive indicates a clockwise rotation. In the following section, the results are presented from the northeastern station to the southeastern, in a counterclockwise direction.

Results
The OSW velocities recorded by the ten buoys, and averaged over the available period, range from approximately 6 ms −1 -9 ms −1 , with the mean lowest wind speeds observed at Cadiz and Nazaré Coastal and the strongest ones at the northwest buoys of Villano and Bares (table 2). Also, at Villano station, located at a southwest-northeast coastline orientation, the highest wind velocity variance is observed, while a narrower range of velocity values characterizes Faro, Cadiz and Nazaré Coastal.  All data sources considered underestimate the mean velocity measured by the buoys, except for the ASCAT products at Peñas and Bilbao stations in northern Iberia, for Nazaré Coastal in the west and Cadiz in the south. The wind velocity estimations' highest bias is obtained at Villano and Bares, where the strongest winds are recorded, and at both Nazaré stations.
The statistical analysis for each OSW data source and buoy location presented in table 2 are illustrated through normalized Taylor diagrams [41] in figure 3. These normalized diagrams show the correlation, RMSE and σ for each buoy location. The RMSE and σ are normalized, considering as reference the STD of the buoy data. The more accurate the data source is, the closer the markers are to the reference.
The statistical analysis at Nazaré offshore buoy shows, in general, very small correlation values and high RMSE and bias for all data sources considered, which can be justified by the significant amount of gaps and potential artifacts introduced by the interpolation method due to the coast proximity. Among the remaining stations, both Silleiro and Villano present higher mean correlation values of 0.81 and 0.77, respectively, followed by Cadiz (0.75), Peñas (0.72), Bilbao (0.66), Faro (0.65) and Leixões (0.64). The lowest three correlations are Bares (0.59) and both Nazaré stations, offshore and coastal, with 0.43 and 0.42, respectively.
The Sentinel-1, ASCAT-C and ERA5 show the highest mean correlations, 0.77, 0.67 and 0.65, respectively, while the WRF simulation shows the lowest correlation, 0.59 (WRF 3 km) and 0.60 (WRF 9 km). The RMSE values are generally higher at Nazaré for all data sources of around 4 ms −1 and around 7 ms −1 at Bares for the model's outputs, probably due to the high proximity to the shore.
In order to compare the mean values and trends estimated and measured, two statistical tests were performed. The two-tailed t-test [42] was used to compare the OSW mean of each pair of datasets (buoys vs. other sources) to inspect if they might be derived from the same population. The null hypothesis states the equality of means, i.e. H0: {µ0 = µ1}, and the alternative hypothesis H1: {µ0 ̸ = µ1}. The other statistical test applied is based on a slightly modified version of the t-test [43] and inspects whether the trends (slopes) of each pair of datasets are equal. As for the first test, the null hypothesis states for equality of slopes, i.e. H0: {s0 = s1}, and the alternative hypothesis H1: {s0 ̸ = s1}. The significance level for both statistical tests was set at 5%. In table 2, the underlined bold values on the mean indicate acceptance of the null hypothesis (i.e. means are equal at a 5% significance level). Likewise, the underlined bold values on the slope indicate acceptance of the null hypothesis (i.e. the trends are equal at a 5% significance level). The first test's results clearly show that the satellite products and buoy observations are in good agreement, with an advantage for

(blue), ASCAT-A (yellow), ASCAT-B (magenta) and ASCAT-C (cyan). In red is the Weibull distribution for the in-situ buoy data (right y-axis).
Sentinel-1 and ASCAT-C, where seven out of ten buoy stations show the same statistical mean. The number is reduced to five stations for ASCAT-A and ASCAT-B. Most of the stations that disagree with satellite measurements are found along the west coast of the Iberian Peninsula (Silleiro, Leixões, Nazaré Offshore, and Nazaré Coastal). The second statistical test shows that the numerical models can reproduce the main trend observed by all the stations. The satellite product that better replicates the buoys trend signal is S1, followed by ASCAT-A, ASCAT-B, and ASCAT-C. This latest satellite source, ASCAT-C, also exhibits the most pronounced negative trends (except Bares), which contributed to the rejection of the null hypothesis (equal slopes). We attribute this fact to the short period of ASCAT-C, less than two years, which prevented the development of a trustworthy trend.
The velocity distribution frequency for each data source is analyzed through a Weibull distribution and illustrated in figure 4, where in red is the Weibull distribution for the in-situ and the remaining colors show the Weibull distribution bias computed as the difference between the probability density function (PDF) of each source and the in-situ one.
At overall stations, the most frequent wind velocity is in the range of 3-8 ms −1 , around 48% of the complete OSW regime at Bilbao and Peñas, around 45% at Silleiro, Leixões and Nazaré offshore, and around 53% at Nazaré coastal, Faro and Cadiz buoys. The exception occurs at Bares and Villano, where approximately 42% of the wind velocity records are within the range 8-16.5 ms −1 , showing a right shift of the frequency peak in the Weibull distribution compared to the remaining stations' frequency distributions (illustrated in red in figure 4). Weaker winds, with velocities under 3 ms −1 , are measured during less than 11% of the time at Bares, Villano, Leixões and Nazaré offshore, between 12% and 14% at Faro and Silleiro, around 18% at Nazaré Coastal and Cadiz, and around 21% at Bilbao and Peñas. The stations at the northwestern Iberia, Bares and Villano, are more prone to be influenced by strong wind velocities, recording velocities higher than 16.5 ms −1 for approximately 7% and 9% of the time, respectively, resulting in a more elongated tail in the frequency distribution. At Silleiro buoy, these strong winds are detected at 5.6% of the entire time series. At the same time, in the remaining northern stations, Bilbao and Peñas, and the western stations of Leixões and Nazaré offshore, it occurs approximately 3% of the time. The Nazaré coastal buoy and the buoys located at south, Faro and Cadiz recorded these velocities around 1% of the time series. At most stations, all data sources analyzed have similar biases, underestimating the weaker and strongest winds and overestimating the most frequent wind velocities in the range of 3-8 ms −1 . This is visible by the colored lines in figure 4, showing negative bias in the lowest and highest velocities and positive bias in the most frequent velocities. A common feature is the variability among data sources bias signals observed at most coastal stations, Nazaré coastal and Peñas, showing that the short distance to the land leads to significant uncertainties. At these stations, as well as at Bilbao, Faro and Cadiz, the ASCAT data strongly underestimates the frequency of wind velocities smaller than 5 ms −1 , overestimating higher velocities when compared to the remaining data sources.
The frequency bias of the wind data provided by the analyzed sources results in narrower Weibull distributions, estimating more occurrences of moderate winds and fewer occurrences of the strongest wind velocities.
The wind direction is analyzed based on wind roses diagrams which illustrate the direction from where the wind blows at each station and for each data source ( figure 5). Overall, the coastal winds are aligned to the orientation of the coastline, approximately zonal at stations located in the north and the south, meridional at those located in the west and with a nearly 45 • rotation at Villano. At Bilbao, the westerlies are slightly more frequent than easterlies, 12% and 8%, respectively, also significant the winds blowing from the south (∼6%). At Peñas, a predominance of westerlies and easterlies winds is observed, as well as winds blowing from SSW. In northwestern Iberia, winds are predominant from SW at Bares and NE-SSW at Villano. Along the west coast, northerlies are dominant, representing 16% of the frequency at Silleiro and nearly 24% at both Leixões and Nazaré stations. At these four west coast locations, the remaining wind directions occur on no more than 8% of the records. At Faro, the second most offshore buoy, the wind direction is predominantly from NW (12%), also occurring from SE during 8% of the time series and approximately 6% from WSW and NE. The wind at Cadiz blows equitably from almost all quarters with a frequency of around 8%, except from the  The wind direction bias, computed as the difference between the wind directions estimated and observed, is analyzed considering four velocity bins based on the Beaufort wind force scale, which considers the sea state (table 3).
The signal and magnitude of the wind direction bias vary along the wind velocity spectrum ( figure 6). On average of all data sources, the highest mean differences are obtained for the weakest and the strongest wind bins, 7 • and 9 • for 0-3 ms −1 and >16.5 ms −1 , respectively. Among all data sources, the highest mean differences in the wind direction are ∼15 • for wind velocities 0-3 ms −1 and 10 • for 3-8 ms −1 given by Sentinel-1, and ∼8 • for wind velocities 8-16.5 ms −1 and 19 • for >16.5 ms −1 given by ASCAT-C. Closer to the observed directions with differences of around 4 • are ASCAT-B for all velocity bins and ERA5 for the velocity range 3-16.5 ms −1 .
Spatially, the highest divergences occur at both Nazaré and Faro stations, mainly for the weaker and stronger wind velocities, and at Leixões for moderate winds. Moreover, at these stations, the differences surpass 20 • for Sentinel-1 and ASCAT-A/C. Except for both Nazaré stations, most data sources show a clockwise rotation of the weak winds estimated with respect to the value measured by the buoy (positive differences).

Discussion
The OSW along the Iberian Atlantic coast is heterogeneous and highly modulated by the coastal geometry, as observed over several coastal regions [9,10,44,45]. As in previous works, which analyzed the surface winds based on PdE buoys data for the period 2000-2009 [10] and 2008 [9], the dominance of westerlies at Bilbao, of easterlies winds at Peñas, of west-southwest sector winds at Bares, of northeasterly winds at Villano, of northerlies at Silleiro and of north-west sector winds at Cadiz, is obtained. For the IH buoys along the west coast, the northerlies are clearly dominant, while in the south, the wind direction observed at Faro is predominantly from the southeast. Therefore, it is observed that the coastal winds are approximately zonal at stations located in the north and the south, meridional at those located in the west and with a nearly 45 • rotation at Villano.
The lower mean wind speeds are observed at Cadiz and Nazaré Coastal (around 18% below 3 ms −1 ), and the stronger ones at the northwest buoys of Villano and Bares (42% is within the range of 8-16.5 ms −1 and 9 and 7% of the time higher than 16.5 ms −1 ). The higher occurrence of strong winds at Villano and Bares, velocities above 9-10 ms −1 , was also reported when analyzing the surface wind in the southern Bay of Biscay over the 2000-2009 period [9,10]. At Bilbao and Peñas, the moderate OSW velocities in the range of 3-8 ms −1 are the most frequent of all ranges, occurring in nearly 50% of the complete OSW data available since 1996. Despite this, the frequency observed in this study is slightly smaller than the 69% and 62% respectively, obtained for 2000-2009 by [10]. The velocity dispersion seems modulated by the station's location, with a wider range of velocities at Villano, while a narrower range of velocities characterizes Faro, Cadiz and Nazaré Coastal.
Over the analyzed period, the data sources based on remote sensing, numerical model and reanalysis provide similar wind regimes showing generally good agreement between the wind estimations and in-situ measurements. This agreement is statistically significant for the satellites' mean velocity and slope estimated by numerical models. Models are often better than satellites at reproducing the tendency of atmospheric data since they are designed to simulate the physical processes that govern atmospheric behavior. As a result, they can capture the underlying mechanisms that drive changes in atmospheric conditions over time and space, which can result in more accurate predictions of the slope of the data [36]. In contrast, satellites provide a snapshot of atmospheric conditions at a specific point in time and space. While they help detect changes in atmospheric conditions over broad areas, they may not be as effective at capturing the underlying mechanisms that drive those changes. As a result, satellite measurements may not be as reliable for estimating the tendency of atmospheric data over longer periods or across different locations.
Overall, satellite surface winds provide relatively higher accuracies, either regionally [6,9,10,45] or globally [7]. Sentinel-1 shows the best normalized combined statistics among all sources presenting, on average, the lowest RMSE and high correlation with respect to observations. All three ASCATs produce better results than ERA5, as observed in other regions of the globe [e.g. 45]. This better performance of ASCAT products when compared to ERA5, is likely due to its higher spatial resolution [45].
It is observed that the mean OSW is underestimated for most stations by Sentinel-1, ERA5 and WRF and overestimated by ASCAT. The mean velocity underestimation by reanalysis as ERA5 was already stated by [45][46][47] over the Brazilian Continental Margin and coastal regions of Mexico, reporting that the reanalysis tends to underestimate the wind speeds at a majority of the sites. In the WRF configuration used in this study, the mean wind velocity is underestimated, however [9], showed a tendency of a WRF configuration to overestimate the wind speed. Therefore, it is believed that the accuracy of the wind estimations by the numerical model is dependent on the configuration setup. An overestimation of the mean wind velocities based on ASCATs products was also observed in 2013 [6] at most of the stations analyzed in this study. In fact, the ASCAT tends to overestimate the wind speeds in the Northern Hemisphere [6,34].
Moreover, all data sources generally underestimate the weakest and strongest in-situ winds and overestimate the moderate ones, resulting in narrower Weibull distributions with higher frequency peaks and shorter right tails compared to those obtained by in-situ data. In fact, the velocity bias dependency on the wind speed intensity was also observed by [6,10,45], with less accurate results for wind velocities below 4 ms −1 and above 12 ms −1 than for intermediate velocities.
Regarding the spatial variability of the wind bias, the proximity to land seems to lead to significant uncertainties in velocity estimations (see figure 4). In most coastal locations, ASCAT data strongly underestimates the frequency of wind velocities smaller than 5 ms −1 , overestimating higher velocities compared to the remaining data sources. In fact, when validating ASCAT-A in coastal and non-coastal buoys spread over the oceans a higher RMSE and bias in coastal buoys were obtained [34]. The proximity to land also impacts the numerical model and reanalysis uncertainty due to the difficulty of numerically simulating the strong land-sea gradients and discontinuities at these regions [6,10].
Among all data sources, Sentinel-1 and ASCAT-C show the highest mean differences in the wind direction, while ASCAT-B for all velocity bins and ERA5 for moderate winds are closer to the observed directions. Except for Nazaré stations, most data sources show a clockwise rotation relative to the values measured by the buoys of the weakest winds estimated (positive differences). The counterclockwise rotation bias at Nazaré was also obtained by [6] when analyzing the single year of 2011. Overall, excluding Silleiro and Cadiz, the models-based data show smaller deviations from what is observed. It is observed that for some locations, the highest errors are obtained for lower velocities, which is in accordance with the overall idea that the weaker the winds are, the higher errors are obtained due to the greater variance of wind direction and higher STDs [10,45]. However, the three PdE stations on the northern coast present high differences concerning observations for higher velocity bins. This might be related to errors associated with bad weather conditions that affect the measurements or the small number of occurrences leading to more errors [10].
The ERA5 global spatial coverage and temporal resolution of 1 h make this reanalysis a suitable database for research and applications that require large domains and high temporal resolutions. Moreover, the increased spatial resolution of ERA5, when compared to other reanalysis, improve, on average, the correlation coefficient of around 0.1 [46], making ERA5 reanalysis a suitable alternative to observational datasets.

Conclusions
This study evaluates the accuracy of several data sources in estimating the OSW regimes along the Iberian Peninsula Atlantic nearshore over the last 25 years of available data. OSW estimations based on the latest releases of remote sensing (Sentinel-1 and ASCAT coastal) and numerical modeling (WRF and ERA5) were compared to in-situ observations collected by ten coastal buoys operated by Portuguese and Spanish Institutions.
All data sources reveal similar wind regimes with generally good agreement between observations and estimations. The stations closest to the coast (Bilbao, Peñas, Nazaré Coastal) and the most sheltered buoy, Cadiz, are characterized by weaker OSW, while the northwest buoys of Villano and Bares record the strongest wind velocities. Overall, the estimated velocities at most stations have similar bias, underestimating the weaker and strongest winds and overestimating the moderate ones. The OSW directions nearly parallel to the shoreline orientation reflect the buoy location's influence on the wind regime. The highest differences between the observed and estimated wind directions are obtained at Faro and Nazaré coastal buoys.
It was found that remote sensing products produce accurate estimations of the OSW and should be considered a good option when high temporal resolutions are not imperative. However, it has to be taken into account that the statistical tests showed that these data sources present higher confidence estimations for the mean than for the trend. In turn, the numerical weather models are able to reproduce the main trend statistically and are the best option for studies or analyses that rely on high temporal resolution. Also, ERA5 revealed that despite its lower spatial resolution when compared to dynamical downscaled WRF simulation, it captures the spatial and temporal dynamics and variability of coastal winds and can be used without compromising the surface wind field statistical distribution to constrain the atmosphere-ocean interface modeling.
Additionally, OSW estimates from data acquired by Sentinel-1A&B around the Iberian coast were analyzed for the first time. Sentinel-1 is the best product statistically, displaying, on average, the lowest RMSE and high correlation coefficients regarding buoys. In fact, it outperformed the other products by a significant margin, with an average improvement of about 23%, however, in terms of bias, it does not show significant improvements over the other products. The high spatial resolution (1 km) can help to explain this outcome. However, it has a small spatial footprint and irregular temporal resolution, making it ineffective as an atmospheric driver for ocean modeling proposes. Nevertheless, the positive outcome suggests that this product can be used to calibrate other products or study regional coastal wind patterns and could even be of great value in offshore wind energy studies.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary information files).