Comparison of Multiple Surface OceanWind Products with Buoy Data over Blue Amazon (Brazilian Continental Margin)

Remote sensing data for space-time characterization of wind fields in extensive oceanic areas have been shown to be increasingly useful. Orbital sensors, such as radar scatterometers, provide data on ocean surface wind speed and direction with spatial and temporal resolutions suitable for multiple applications and air-sea studies. Even considering the relevant role of orbital scatterometers to estimate ocean surface wind vectors on a regional and global scale, the products must be validated regionally. Six different ocean surface wind datasets, including advanced scatterometer (ASCAT-A and ASCAT-B products) estimates, numerical modelling simulations (BRAMS), reanalysis (ERA5), and a blended product (CCMP), were compared statistically with in situ measurements obtained by anemometers installed in fifteen moored buoys in the Brazilian margin (8 buoys in oceanic and 7 in shelf waters) to analyze which dataset best represents the wind field in this region. +e operational ASCAT wind products presented the lowest differences in wind speed and direction from the in situ data (0.77ms<RMSEspd< 1.59ms, 0.75<Rspd< 0.96, −0.68ms< biasspd< 0.38ms, and 12.7°<RMSEdir< 46.8°). CCMP and ERA5 products also performed well in the statistical comparison with the in situ data (0.81ms<RMSEspd< 1.87ms, 0.76<Rspd< 0.91, −1.21ms< biasspd< 0.19ms, and 13.7°<RMSEdir< 46.3°). +e BRAMS model was the one with the worst performance (RMSEspd> 1.04m·s, Rspd< 0.87). For regions with a higher wind variability, as in the southern Brazilian continental margin, wind direction estimation by the wind products is more susceptible to errors (RMSEdir> 42.4°). +e results here presented can be used for climatological studies and for the estimation of the potential wind power generation in the Brazilian margin, especially considering the lack of availability or representativeness of regional data for this type of application.


Introduction
Ocean surface wind is one of the main drivers of several oceanic, atmospheric, and climate processes, thus being an important indicator of climate change [1]. Ocean surface wind vector variability is intrinsically linked to that of oceanic processes, such as coastal upwelling [2,3], primary productivity [4,5], deep water formation [6,7], advective volume transport, air-sea momentum, and heat fluxes [8][9][10][11]. Having the ability to measure wind speed and direction with high accuracy becomes essential to these processes' studies, also considering the impacts on weather forecasting, navigation, marine engineering, and offshore wind energy [1,12,13]. e vast extension of the global ocean imposes technical and financial limitations on in situ ocean surface winds sampling, which is necessary and valuable in remote sensing calibration and validation process. Direct measurements are acquired by anemometers on moored buoys, research cruises, and vessels of opportunity, or by light detection and ranging (Lidar) sensors, providing precise and accurate data. However, these techniques are limited to point measurements that are not able to provide satisfactory spatial and temporal coverage to resolve variability at different scales [14].
Today, more than 10 global and regional atmospheric reanalysis and wind products are available for the scientific community [15]. Numerical prediction models are capable of simulating ocean surface winds on a global scale but depend on the existing knowledge about the physical processes and on the availability of data used to parametrize the initial conditions of the simulations [16,17]. In weather prediction, the ability to make a skilful forecast requires both realistic representation of the atmosphere by numerical models and accurate initial conditions [18].
Reanalysis products integrate simulated data from numerical prediction models with atmospheric, oceanic, and other measurements, but with relatively low spatial resolutions (250 to ∼31 km).
is can be insufficient to adequately represent small and mesoscale meteorological features and accurately characterize local wind regimes [19]. Offshore wind blended products usually consider a first guess of the wind vector as the background and assimilate multiple sources of measured wind data from buoys, ships, and satellites [20].
Orbital remote sensors provide wind data on a global scale, covering remote regions normally characterized by the availability of sparse in situ data, with relatively good accuracy and high spatial and temporal resolution [21]. Ocean surface wind vectors with high spatial and temporal resolution obtained by spaceborne scatterometers are utilized in various fields of meteorology, oceanography, and climate studies [22,23]. Surface wind and stress fields derived from scatterometer observations can be applied to drive ocean circulation models on various scales and can also be assimilated into regional and global numerical weather prediction models [21,24]. ese data have also been used to infer about offshore wind energy potential [20,25].
Scatterometers measure the radar cross section of the ocean surface, and numerical inversion of the geophysical model function (GMF) yields the scatterometer wind measurement. e present constellation of scatterometers maps the surface wind field globally, with a typical spatial resolution of 25-12.5 km [26], and has been successfully used in weather forecasting applications [27], long-term climate studies [28], and air-sea interactions [29]. e main limitations of scatterometers are contamination by rain (depending on the frequency of the transmitted signal), lack of data near the coast (typically within 15 km), and poor temporal sampling [21].
Although the number of space-borne scatterometers is increasing in recent years, regional validation of wind products from different scatterometers [30,31] and producers [32] remains a challenge [33]. Moored ocean buoys provide the absolute calibration reference for satellite wind retrievals [22]. While the development of GMF and wind retrieval algorithms rely on many inputs (numerical models, wind retrievals for other satellites, statistical constraints, etc.), finalized satellite wind retrievals always need to be verified by comparisons with buoys.
Several studies evaluated the accuracy of wind products over the ocean comparing them with in situ measurements globally and regionally [19,24,[34][35][36][37][38][39][40][41][42][43][44][45][46]. In general, these authors observed good agreement (mean square error, bias, and correlations) between wind products and in situ measurements. However, significant systematic differences do exist in the wind products derived from different scatterometers [20,30,31] and from different producers [32], which may result in detrimental impacts on different applications, if unknown. erefore, characterizing errors and inconsistencies among different wind products is important, not only for guiding wind product users, but also to motivate research by the producers to solve inconsistencies problems.
e Brazilian oceanic area represented by the "Territorial Sea" and the "Exclusive Economic Zone" corresponds to approximately 3.6 million km 2 . If the "Extended, External, Jurisdictional" or "Legal Continental Shelf" claimed to the Commission on the Limits of the Continental Shelf of the United Nations in three "Revised Partial Proposals" for the outer limit of the Brazilian Continental Shelf (2.1 million km 2 ) is added to this area, the total oceanic area called the "Blue Amazon" may be approximately 5.7 million km 2 . is immense oceanic area contains an enormous amount of economically, socially, and strategically relevant resources and is extremely important for climate stability and environmental quality of the adjacent coast [47].
e South Atlantic Subtropical Anticyclone (SASA) is the main feature of the atmospheric circulation over the South Atlantic Ocean, with a seasonal variability extending over south-eastern Brazil during the austral winter and retracting eastward during the summer [48]. SASA contributes with northeast winds to the Brazilian continental margin that are more constant during austral spring and summer [49].
Over the Tropical Atlantic, the Intertropical Convergence Zone (ITCZ) associated with the confluence zone of the northeast and southeast trade winds migrates seasonally from its most northerly position, around ∼10°N in August-September, to its southerly position, around ∼2°S in March-April in the Atlantic Ocean [50,51]. Associated with the mesoscale oscillations, the frontal systems move with cyclones and post-frontal anticyclones, changing the fields of atmospheric pressure, wind, and other atmospheric variables along their trajectory [52]. Between 30°S and 50°S, the zones with the highest occurrence of cold fronts are concentrated on the South-Southwestern Atlantic, mainly in the coastal areas of Argentina and Uruguay [53]. e south and south-eastern coasts of Brazil and the region between the extreme south of Brazil and the coast of Argentina are favourable regions for the formation of cyclones [54]. ese systems cause strong winds, precipitation and decrease in temperature [55]. In this region, the atmospheric circulation is characterized, on a synoptic scale, mainly by the movement of transient systems such as cyclonic vortices, frontal systems, and the South Atlantic Convergence Zone (SACZ) [56].
For better understanding the wind regimes on the Brazilian continental margin, and to aid in decision-making for the installation of offshore wind turbines, there is a need to conduct a study that contemplates this vast oceanic area. e dataset used must be adequate to characterize the spatial and temporal variability of the local wind fields, and also to estimate the potential offshore wind energy production.
is study aims to evaluate and compare different alternative sources of offshore wind data, including the Advanced SCATterometer (ASCAT) on board MetOP-A (ASCAT-A) and MetOP-B (ASCAT-B) European satellites, the Cross-Calibrated Multi-Platform wind product, the European Centre for Medium-Range Weather Forecasts ERA5 Reanalysis, and numerical simulations from the Brazilian Regional Atmospheric Modelling System (BRAMS/CPTEC) with in situ ocean surface wind measurements, aiming to analyze the capacity of these wind data sources in describing the wind field in the Blue Amazon region.

Buoy Wind Data.
Time series of in situ wind data measured by anemometers installed in 15 oceanographic buoys moored on the Brazilian continental shelf and in deeper oceanic waters were obtained from the Brazilian National Buoy Program (PNBOIA) and the Prediction and Research Moored Array in the Tropical Atlantic (PIRATA) project, respectively. e geographical locations and summarized information regarding these buoys are presented in Figure 1 and Table 1, respectively.
Eight PIRATA buoys used in this study are moored at depths between 1,500 and 6,000 m and are equipped with a R. M. Young mechanical anemometer, model 05103 (http:// www.youngusa.com/products/11/5.html), or with a Gill WindSonic ultrasonic anemometer (http://gillinstruments. com/data/datasheets/windsonic-1405-027-iss7.pdf), installed 4 m above sea level. Wind direction is measured with a Sparton SP3004D compass (https://sparton.com/) or an EG&G fluxgate compass model 63764 or KVH LP101-5 (https://www. pmel.noaa.gov/gtmba/sensor-specifications). Data with a 10minute temporal resolution, containing measurements since 1997, are available from NOAA's Pacific Marine Environmental Laboratory (https://www.pmel.noaa.gov/tao/drupal/ disdel/). Seven PNBOIA buoys used in this study are moored close to the continental shelf break (∼200 m isobath). e PNBOIA data, maintained by the Brazilian Navy Hydrography Centre, are available on the Internet (https://www. marinha.mil.br/chm/dados-do-goos-brasil/pnboia-mapa). Time series of wind data are provided with hourly time resolution. e temporal coverage of the data ranges from 04/2009 to 11/2018. e buoys contain 2 anemometers, one mechanical installed at 4.7 m above sea level (used as the main reference) and another ultrasonic sensor installed at 3.7 m. More information about the sensors of each buoy can be found at the same Internet address for data access.
To adjust the wind measured by the buoys to 10 m, a steady-state neutrally stable logarithmic vertical wind profile relation was considered. In this case, the wind speed (U) at height z is given by where k is the von Kármán constant (k � 0.41) and z 0 is the surface roughness that depends on the friction velocity (u * ), which in turn is calculated from an iterative process using the wind speed measured by the buoys. For more details, please refer to References [57,58]. ree ASCAT products were acquired for this study: (i) ASCAT-A at processing level 2 (L2), with 12.5 km of spatial resolution, from the period 08/2010 to 10/2018 (https://projects.knmi.nl/ scatterometer/ascat_osi_co_prod/ascat_app.cgi); (ii) ASCAT-B, at L2 processing level, spatial resolution of 12.5 km, from the period 10/2012 to 10/2018 (https://projects.knmi.nl/ scatterometer/ascat_b_osi_co_prod/ascat_app.cgi); and (iii) ASCAT Coastal Wind Data Record (ASCAT-CDR), L2 processing level, 12.5 km spatial resolution, from the period 01/2007 to 03/2014 (https://projects.knmi.nl/scatterometer/ ascat_cdr/). e ASCAT-A and ASCAT-B products were processed using the CMOD-5n geophysical model function [59] until September 2018. From then on, the GMF CMOD-7 function [60] started to be used (Verhoef, personal communication).

Scatterometer Wind
e ASCAT-CDR product is the result of reprocessing ASCAT-A data using the latest GMF CMOD-7. e three products, regardless of the GMF used, are optimized for coastal regions [23,61,62], containing wind data at 10 m above sea level in along-track grids (along the satellite path) with a spatial resolution of 12.5 km, coming from 14 orbital cycles per day. Table 2 summarizes the main characteristics of all alternative offshore wind products used in this study.

BRAMS Numerical Forecasting Model.
e Brazilian Regional Atmospheric Modelling System (BRAMS) is a numerical simulation system designed for atmospheric forecasting and research from regional to hemispheric scales. Based on the Regional Atmospheric Modelling System (RAMS), originally developed at Colorado State University, BRAMS was until recently developed and maintained by the Centre for Weather

Blended CCMP Data.
e Cross-Calibrated Multi-Platform (CCMP) is a level 3 (L3) wind product which uses satellite and in situ data, and numerical modelling simulations. e first CCMP version, described in the study of Atlas et al. [63], was discontinued in 2012. e second CCMP version used in this work is described in the study of Wentz et al. [64]. is dataset uses the variational analysis method (VAM) to fill gaps in the gridded product [65][66][67]. VAM generates a surface wind grid, minimizing an objective function, which measures the mismatch of the in situ and satellite wind data overlaid on reanalysis bottom base, considered as an initial estimate [63]. Wind data at 10 m from the CCMP product are available in regular grids for dates after 1987 with a spatial resolution of 25 km and a temporal resolution of six hours. Data were obtained for the period from 01/1998 to 12/2018 (http://www.remss.com/ measurements/ccmp/).

ERA5 Reanalysis.
e ERA5 Reanalysis product is developed by the Copernicus Climate Change Service, implemented by the European Centre for Medium-Range Weather Forecasts (ECMWF) (https://www.ecmwf.int/en/ forecasts/datasets/archive-datasets/reanalysis-datasets/ era5). Reanalysis data are produced by combining shortterm numerical predictions with observational data. ERA5 is produced using the assimilation of atmosphere and surface data in four dimensions through the ECMWF Integrated Forecast System with 137 vertical levels in sigma coordinates (which use atmospheric pressure on the surface as a reference) in grids with 31 km of resolution for atmospheric levels. Hourly ERA5 data are available for several atmospheric variables, including wind zonal and meridional components, and air temperature at 10 m above sea level. Monthly updates of ERA5 hourly data on single levels from 1979 to present are published within 3 months of real time. Data were obtained for the period from 01/2000 to 12/2017.

Statistical Analysis.
To compare the different sources of offshore wind speed data with the in situ buoy measurements, the root mean square error (RMSE spd ), bias spd , Pierson correlation coefficient (R spd ), and standard deviation (σ spd ) were calculated considering the simultaneous and valid wind speed data records among all datasets, according to the following equations: where x refers to the wind speed values at instant i, r refers to the in situ buoy data used as a reference, and e refers to the data estimated by the alternative sources of wind data. Regarding wind direction data, the circular statistics was applied. e mean direction (θ) is estimated as follows: Considering each angle (θ) in an instant i as a vector, C � N i�1 cos(θ i ) and S � N i�1 sin(θ i ) are the respective xand y-axis representations of the resultant vector length (r). e circular standard deviation (σ dir ) is estimated by the following equation: where r � (r/N) � ( /N) would be the mean resultant length. e circular correlation (R dir ) is defined as Please, refer to the study of Jammalamadaka and Sengupta [68] for more information about circular statistics. e bias and RMSE formulas for wind direction data are the same as for wind speed data. However, the minor angle difference (θ d ) was derived as follows: erefore, RMSE and bias for direction can be read as follows: e comparisons considered the buoys geographic coordinates and the respective grid point or pixel of the other datasets. Each buoy is identified by its name (PNBOIA, shelf waters) or geographic coordinates (PIRATA, oceanic waters), as adopted by the respective monitoring programs. An interval of 30 minutes was considered around the measurement time of the in situ PNBOIA data, and of 5 minutes in the case of PIRATA buoys, according to the respective temporal resolution of the in situ data recording (hourly or 10 minutes, respectively). Two approaches were taken for data comparison, prioritizing different criteria: (i) the first approach compared data that were obtained simultaneously by all the five datasets-ASCAT, CCMP, ERA5, BRAMS, buoys-(ii) a second approach compared all available paired data separately for each buoy. When comparing For shelf water buoys PNBOIA, the higher mean wind speeds were observed in Fortaleza (8.91 m·s −1 ), Rio Grande (8.34 m·s −1 ), Itajaí (7.68 m·s −1 ), and Cabo Frio 2 (7.52 m·s −1 ) ( Figure 2, Table 3). Fortaleza, Recife, and Cabo Frio 2 buoys showed a lower variability in wind direction, with dominance from the SE-E and NE quadrants, respectively. However, in Cabo Frio 2, the standard deviation of the mean wind direction was 75.0° (Table 3), while for Fortaleza and Recife, the standard deviations were lower (23.1°and 33.3°, respectively).

Ocean Surface Wind Statistical Comparisons.
roughout the statistical comparisons presented below, a negative systematic bias was observed for the BRAMS wind speed data, which was corrected by applying the scaled distribution mapping method [71]. For oceanic waters (OW) buoys comparisons, the correction of the systematic bias resulted in lower RMSE spd values (a decrease of 1.8 m·s −1 in the best case), while for shelf waters (SW) buoys, there was no relevant change with slightly higher RMSE spd values for two buoys ( Figure S1 in Supplemental File). erefore, BRAMS wind speed data with systematic bias correction were used only for comparisons with OW buoys. e simultaneous comparisons of the ocean surface wind products with in situ measured buoys data did not differ much with respect to wind speed (Figures 3-5 and Table S1 in Supplemental File). Considering the comparison of all buoys together, the RMSE spd values ranged between 0.98 m·s −1 and 1.82 m·s −1 . Considering the OW and SW buoys separately, the RMSE spd values were relatively lower (0.91 m·s −1 and 1.02 m·s −1 in the best cases, respectively). e highest values of RMSE spd were obtained for the product BRAMS, even after removing the systematic bias. For all buoys, the bias spd varied between ±0.30 m·s −1 . For OW buoys, the bias spd varied between ±0.19 m·s −1 . For SW, the bias spd varied between −0.18 m·s −1 and −0.25 m·s −1 for ASCAT and CCMP products and between −0.70 m·s −1 and −0.48 m·s −1 for BRAMS and ERA5. e highest correlations (R spd > 0.87) were obtained for ASCAT products considering all and SW buoys, or CCMP considering OW buoys.
In general, relatively higher differences were observed in relation to wind direction, especially for shelf water buoys. Best results were obtained for OW with RMSE dir values <23.2°, bias dir between −0.09°and 0.19°, and correlations >0.83, except for the BRAMS product (RMSE dir � 31.2°, bias dir � −1.40°, R dir � 0.72, in the best cases). Considering all buoys, the values of RMSE dir and bias dir varied between 46.8°a nd 53.4°, and between −1.06°and 4.88°, respectively. Also, the best correlations (R dir > 0.87) were obtained for the CCMP product. On the contrary, high RMSE dir values >73.5°, bias dir between −14.34°and −0.09°and lower correlations (<0.35) were obtained for SW. e comparisons of all available paired data for each oceanic water buoy did not differ much with respect to wind speed ( Figure 6 and Table S2

Error Dependence on Measured Wind Speed.
Scatterometers limitations in representing low and high wind speeds are well documented in the literature [72,73]. To analyze if wind speeds from the alternative wind data sources show any variation or dependence with the in situ buoy data, four different bins were considered for statistical comparison: 3-6 m·s −1 ; 6-9 m·s −1 ; 9-12 m·s −1 and >12 m·s −1 (Figures 8 and 9 and Table S4 in Supplemental File). In general, the lowest RMSE spd (0.76 m·s −1 ) and bias spd (−0.05 m·s −1 ), and highest R spd values (0.70) were observed       in the 6-9 m·s −1 interval, while higher differences were observed for winds >12 m·s −1 (1.11 m·s −1 < RMSE spd > 4.59 m·s −1 , −3.55 m·s −1 < bias spd > −0.59 m·s −1 ). For all speed intervals and all datasets considered in the analysis, the lower RMSE spd values and higher R spd values were obtained for ASCAT-A and ASCAT-B products. Regarding wind direction, higher differences were observed in the 3-6 m·s −1 interval for OW buoys, considering all wind products. For SW buoys, higher differences were also observed for the lower wind speed interval, considering ASCAT-A, ASCAT-B, and ERA5 products. Best comparisons were obtained for CCMP product with oceanic water buoys (RMSE dir < 23.3°, −0.01°< bias dir < 0.61°, R dir > 0.91), and for ASCAT-B with shelf water buoys (RMSE dir < 56.6°, −0.47°< bias dir < 7.66°, R dir > 0.59), in all speed intervals.

Discussion
In the present study, data from six different ocean surface wind products were compared with in situ measurements from 15 moored buoys in oceanic and shelf waters of the Brazilian margin. In general, there is a good correspondence between the wind speeds from the simultaneous records of wind products and buoy observations (Figures 3-5 and Table S1). Without considering the BRAMS model, comparisons of simultaneous records are even better. It can be noted that ASCAT and CCMP products had slightly better results than ERA5. e spatial resolution of the databases seems to influence their performance, with the finest resolution products (satellitederived databases) showing better results than the coarser ERA5 database. Despite the removal of the systematic bias in BRAMS wind speed data, the highest RMSE spd continued to be obtained for this dataset (RMSE spd > 1.56 m·s −1 in simultaneous comparisons). Indeed, if analysis is restricted to OW buoys, best performance is observed for CCMP product ( Figure 6 and Table S2), showing the lowest errors in terms of RMSE dir and the best values of R spd and R dir for 5 or more buoys among the 8 OW buoys compared. Although the data from the PIRATA buoys are assimilated in the generation of the CCMP dataset, these results agree with previous studies that compared CCMP and PIRATA wind data [41][42][43]. Suzuki et al. [42] obtained an RMSE spd value of 1.04 m·s −1 , and Yan et al. [43] obtained RMSE spd , bias spd , and R spd values of 1.29 m·s −1 , 0.24 m·s −1 , and 0.90, respectively. For wind direction, these authors [43] obtained RMSE dir � 22.4°, bias dir � 0.89°, and R dir � 0.97. In the present study, ASCAT products showed the best results after CCMP, with a performance close to other recent comparisons [44]. e lowest errors in terms of RMSE spd for 5 of the 8 OW buoys were obtained for ASCAT-A. Verhoef et al. [61] validated the ASCAT-A product optimized for coastal regions using in situ buoy data from the tropical oceans and along North American and European coasts. For buoys located more than 50 km from the coast, these authors obtained RMSE spd of 2.2 m·s −1 and bias spd of −0.29 m·s −1 . In the present study, RMSE spd < 1.25 m·s −1 and bias spd between −0.33 m·s −1 and 0.25 m·s −1 were computed for ASCAT-A in oceanic waters. In the ASCAT-B validation report [62], bias values of 0.02 m s −1 (CMOD5.n) and 0.04 m s −1 (CMOD7) were reported, whereas in the present study, bias values varied between −0.24 m·s −1 and 0.26 m·s −1 . In the ASCAT-CDR validation report [23], bias   In terms of wind direction, relatively higher differences were observed in the present study, even without considering BRAMS comparisons. Most oceanic water buoys are located in the Tropical Atlantic region where winds are more stable [50,51]. However, highest statistical differences were observed for shelf water buoys ( Figure 2 and Table S1). Wind direction comparisons were considered more satisfactory only for SW buoys located in the northern part of the Brazilian continental margin (Figure 7 and Table S3). For Santos, Itajaí, and Rio Grande buoys located in the southern part of the study region, wind direction comparisons resulted in relatively higher differences. According to Lin et al. [74], when comparing the wind vectors classified as variable or stable, the estimated error variances of the variable winds are much higher than those of the stable winds.
e southern Brazilian continental margin is well known in the literature for frequent incursions of frontal systems and relevant cyclogenetic activity, which increases the variability of the wind field in this region [52][53][54][55][56]. is variability pattern for these 3 buoys can be depicted from the wind roses shown in Figure 2. As an example, Figure 10 shows the differences between the wind roses of Santos buoy and ASCAT-A paired series, which are also observed for the other datasets considered in this study (Figures S2-S16 in Supplemental File). It is worth remembering that PNBOIA data (SW) are hourly averages, while PIRATA data (OW) are 10 minutes averages, which can also contribute to the relatively greater differences observed for southern shelf water buoy comparisons [75]. For satellite-derived wind products, land masking limitations can be an issue [45]. In addition, more coastal winds are very dependent on the surrounding topography, land-ocean gradients, and discontinuities.
Previous studies reported overestimations of wind speed by coastal ASCAT-A and ASCAT-B products in the Iberian, Irish, and Japanese continental shelf regions [20,45,46]. Takeyama et al. [46] proposed that this overestimation is inversely proportional to the distance from the coast due to increased backscatter by land contamination. However, Verhoef et al. [61] observed that the ASCAT-A product optimized for the coastal region underestimated buoy wind speed measurements in the northern hemisphere. In the present study, not only ASCAT products but all wind products underestimated buoy wind speed for shelf waters. Payan [76] observed that the ASCAT-A product overestimated buoy wind speeds in the northern hemisphere and  Figure 9: Barplots representing the statistical metrics computed for oceanic and shelf water buoys in four wind speed bins (m·s −1 ). RMSE spd and RMSE dir : mean square error for wind speed and direction; bias spd and bias dir : bias for wind speed and direction; R spd and R dir : Pierson's correlation coefficient for wind speed and direction. underestimated speeds measured in the southern hemisphere. In general, the comparisons of wind speed obtained in the present study for shelf waters agree with previous studies. e best performances are distributed between ASCAT products (Figure 7, Table S3). CCMP and ERA5 also show good performance for wind speed, with exception for Recife buoy, which is the closest to the coast (35 km).
Regarding the comparisons by wind speed intervals (Figures 8 and 9 and Table S4), ASCAT-A and ASCAT-B showed a better statistical performance in terms of RMSE and R, considering all datasets. In general, the lowest RMSE spd and higher R spd values were observed in the 6-9 m·s −1 interval. e poor wind speed accuracy for interval >12 m·s −1 for OW buoys can be related to the relatively low amount of data recorded for this wind speed interval, as seen in the wind roses shown in Figure 2. Almost all wind products overestimated speeds between 3 m·s −1 and 6 m·s −1 and underestimated speeds >6 m·s −1 , so that as wind speed increases, the negative bias increases proportionally. According to Carvalho et al. [19], weaker winds produce low amounts of backscatter, which scatterometers have difficulties in measuring. Stronger winds can only produce proportional backscatter in the ocean surface until a certain threshold, after which no more backscatter is produced even if the wind speed keeps increasing. Regarding wind direction, best comparisons with OW buoys were obtained for CCMP product. High values of RMSE dir and low correlations for SW buoys can be related to subcell wind variability in southern Brazilian shelf, as mentioned above [74]. In general, the statistical performance for wind direction increases towards the higher speed intervals. According to Chakraborty et al. [37], the scatterometer errors associated with wind direction are highly random and are dependent on wind speed. In general, the lower the wind speeds, the higher the wind direction errors.

Conclusions
is study evaluates and compares satellite-derived ocean surface wind data (ASCAT-A and ASCAT-B scatterometers), reanalysis (ERA5), BRAMS-modelled surface winds, and a satellite-reanalysis-buoy-blended product (Cross-Calibrated Multi-Platform ocean wind vectors), to analyze the performance of these datasets in representing the in situ measurements obtained by moored buoys along the Brazilian margin. e results presented here can be of great value for climate and oceanographic studies, among others, especially where in situ measured wind data are not available or are insufficient, therefore requiring alternative sources of wind data to be considered.
Ocean surface wind estimations obtained with ASCAT scatterometer optimized for coastal regions and with a spatial resolution of 12.5 km are the ones that best represent the in situ wind speed, with the lower RMSE spd (<1.59 m·s −1 ), lower bias spd (±0.68 m·s −1 ), and higher correlation (>0.75) in relation to the other products considered in the study. However, CCMP product also showed a good statistical comparison with in situ buoy data (RMSE < 1.87 m·s −1 , bias ± 0.96 m·s −1 , R > 0.82), as well as ERA5 (RMSE spd < 1.84 m·s −1 , bias spd ± 1.21 m·s −1 , R spd > 0.76). Among all the compared ocean surface wind products, BRAMS is the least able to represent the wind field (RMSE spd > 1.04 m·s −1 , bias spd ± 2.89, R spd < 0.87). In regions with more variable winds, as in the southern Brazilian continental margin, the estimation of wind direction is more susceptible to errors (RMSE dir > 42.4°).
Considering the relatively higher temporal resolution of ERA5 and CCMP products (1 and 6 hours, respectively), and the extension of the available time series (>30 years), it is suggested to use these datasets to build a regional climatology and to analyze the space-time variability of wind fields in the study region. Nevertheless, there are seven orbital scatterometers operating currently [77][78][79][80][81]. is number of operational scatterometers has been increasing in recent years and in near future. Once the scatterometers can give winds within 6 hours over the regional ocean, it may be good to use combined winds from multiple scatterometers. e derived information for such satellite constellation and alternative sources could be useful for several applications, as on the estimation of the potential wind power generation in the Brazilian margin region.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.