Evaluation of Satellite and Reanalysis Wind Products with In Situ Wave Glider Wind Observations in the Southern Ocean
1. Introduction
Mid- to high-latitude regions in the Southern Ocean are host to the strongest wind fields at the ocean surface. These strong winds (speeds > 20 m s−1; Yuan 2004) significantly impact upper-ocean properties and processes, such as mixed layer dynamics, Ekman processes, and air–sea exchange. Exchanges in heat, moisture, and momentum at the air–sea interface are facilitated by sea surface winds. In addition to driving physical processes at the sea surface, these winds also have implications for biological processes that extend below the surface. The flux of carbon dioxide (CO2) between the atmosphere and ocean is closely related to wind speed (Wanninkhof 1992; Wanninkhof and McGillis 1999; Wanninkhof et al. 2009; Nightingale et al. 2000; Ho et al. 2006), and as such sea surface winds are of interest to many scientists studying and modeling biogeochemical cycles within the ocean.
Surface ocean winds are measured using in situ techniques or remote sensing instruments. Satellite instruments interpret variability in lidar laser energy reflection as a result of sea surface roughness to infer wind vectors generally at a height of 10 m MSL, while in situ measurements are generally taken at varying heights above sea level dependent on the platform (e.g., ship, mooring, or glider). Because of the varying methods of collecting and synthesizing wind measurements, the uniformity of sea surface wind data is often questioned. Such inconsistencies require studies reliant on satellite-derived wind data (such as Yuan et al. 2009) to perform a validation prior to the start of their analysis in order to determine the ability of satellite-blended datasets to accurately represent wind fields at the air–sea interface and to determine whether postprocessing is required to reduce errors once compared to in situ measurements. There is a great need to reduce or eliminate this time costly step in wind data analysis. Further, the dearth of in situ wind stress observations in the Southern Ocean, particularly in winter (practically absent; Gille et al. 2016), is resulting in poorly constrained ocean and climate models, where wind forcing is reliant on unvalidated scatterometer or reanalysis wind products. This particular problem and air–sea flux uncertainties in general were highlighted at a recent international workshop that brought together scientists to discuss this issue and how it could be addressed (Gille et al. 2016). Through the Southern Ocean Observing System (SOOS), a new capability working group has been formed in an effort to enhance parameterizations of air–sea fluxes in the Southern Ocean. It is hoped this group will attempt to address, discuss, and engage with the larger community on the severe lack of flux observations and the associated research issues emanating from this problem (http://soos.aq/activities/capability-wgs/soflux).
Several studies have made attempts to compare and validate wind products in the Southern Ocean (see Table 1). Yuan (2004), Yuan et al. (2009), and Patoux et al. (2009) highlight the implications of misrepresented wind fields in the Southern Ocean in their evaluations of extreme wind events. Yuan (2004) compared Quick Scatterometer (QuikSCAT) wind data with reanalysis wind data from the National Centers for Environmental Prediction (NCEP), the National Center for Atmospheric Research (NCAR), and the European Centre for Medium-Range Weather Forecasts (ECMWF). This study showed that both weather station and QuikSCAT winds were stronger than reanalysis data, albeit only during high wind events. Yuan’s (2004) approach required the input of many different wind products to investigate the spatial and seasonal variability of high wind characteristics. Because of the remote area of the study, the reanalysis products that Yuan (2004) used were limited with regard to in situ data input. It was shown that monthly mean wind observations from satellite averages over the Southern Ocean were in agreement with model simulations except for the period from May to October (6 months centered over austral winter), when scatterometer winds were stronger than simulated winds, thereby suggesting a strong winter bias in data. Similar to Hilburn et al. (2003), Yuan (2004) attributed this discrepancy to the NCEP–NCAR reanalysis missing some storms entirely, along with a lack of observational data assimilated into the reanalyses. Additionally, Yuan et al. (2009) further investigated a study by Patoux et al. (2009) where scatterometer-derived pressure fields were modified using reanalysis wind data. QuikSCAT data were blended with ECMWF forecast data to produce modified pressure swaths. Yuan et al. (2009) claim that even small mesoscale cyclones contribute significantly to heat and momentum fluxes at the air–sea interface as a result of the high frequency of their occurrence. As such, a correct representation of mesoscale cyclones in numerical weather prediction (NWP) models of the Southern Ocean is therefore critical (Yuan et al. 2009). Although these studies have aided in improving estimates of wind stress and in understanding the time and space variability of Southern Ocean wind events, significant gaps in our understanding of wind stress accuracy and uncertainty remain. For instance, at any given location the diurnal variability of wind stress is often subject to aliasing as a result of the geographical variability of sampling by satellite, and the higher harmonics of the fundamental diurnal cycle may be distorted or eliminated (Risien and Chelton 2008). New acquisitions of in situ data from robotic platform deployments, such as presented in this study, are likely to play an ever-increasing role in understanding wind stress (and other air–sea fluxes) characteristics and validating gridded wind stress products, especially in remote and harsh locations of the world’s ocean, where historically field data have a strong bias to the summer season (Gille et al. 2016).
Characteristics of the different wind products used in this study.
The aim of this paper is to determine which wind data product (satellite blended, reanalysis, and modeled) best represents the in situ wind field magnitude and variability of a study location in the subantarctic Southern Ocean. This is done by comparing wind products with a valuable set of time series of in situ sea surface wind data collected by Wave Glider (WG) technology deployed in a series of experiments in the Southern Ocean. The paper is organized as follows: An explanation of the surface wind datasets, their sources, and the transformations performed to collocate them in space and time is provided in section 2. A detailed description of the statistical analysis performed in this study is provided in section 3. Remotely measured winds are compared with in situ winds at a quasi-fixed location in the Southern Ocean in section 4. The final section presents a discussion of the results and their broader implications for other scientific disciplines.
2. Data and methodology
a. In situ wind observations and experimental setup
The in situ data used in this study were collected via a Liquid Robotics WG autonomous surface vehicle. The primary components of the WG include a surface float, a subsurface unit, and a 7-m-long umbilical cable, which connects the two (Fig. 1). The fins installed on the subsurface unit convert orbital motion of surface waves into a horizontal force that drives the subsurface unit forward. The subsurface unit tows the surface float at speeds ranging between 0.5 and 2 kt (1 kt = 0.51 m s−1). Solar panels and batteries power sensors and communication systems that are installed in the surface float, enabling it to remain at sea for an extended period of time (multiple months). The experiment was located in the domain of austral summer maximum wind speed (December 2015 average > 10 m s−1), associated with the westerly winds of the Southern Ocean, as indicated by the red shading band in Fig. 1a. Two different Wave Glider deployments took place in the austral winter to summer of 2015. Gliders CSIR2 and CSIR1 were deployed in July and November 2015, respectively, and were set to navigate to 43°S, 8.5°E. Upon arrival at the study site, a pseudomooring octagon sampling pattern (Fig. 2) with a diameter of 16 km was maintained (further details in Monteiro et al. 2015). At the center of the octagon, a Seaglider profiled the upper 1 km of the ocean in pseudomooring mode.
(a) Study region in the subantarctic Southern Ocean, with the main location of where the in situ data were collected at 43°00′S, 8°30′0″E (black square) and overlaid on the monthly averaged ECMWF wind speeds (m s−1) for December 2015. Mean locations of the Southern Ocean fronts are overlaid (solid black curves) according to Swart et al. (2010). (b) Liquid Robotics SV2 Wave Glider and Airmar WeatherStation are indicated (photo credit: Fred Fourie).
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
Map of sampling region in the Southern Ocean, centered at 43°00′S, 8°30′0″E. WG sampling locations, encompassing two separate deployments, are indicated with respect to the spatial distribution of gridded wind products.
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
The WG is designed to augment existing marine technology by providing autonomous real-time data collection of parameters near the air–sea interface. The Council for Scientific and Industrial Research (CSIR) uses this technology in its integrated multiplatform approach for a series of Southern Ocean Seasonal Cycle Experiments (SOSCEx) that use research vessels, WGs, profiling gliders, bio-optic floats, satellite data, and numerical models to explore the climate sensitivity of carbon and ecosystem dynamics (Swart et al. 2012). The WGs deployed during the research expeditions were equipped with an Airmar WX-200 Ultrasonic Weatherstation (Airmar Technology Corporation 2016). This instrument is a compact weather station designed for moving platforms, with abilities to dynamically correct winds using an internal compass and correct up to a 30° pitch in rough seas. The sensor outputs apparent and true wind speed and direction (via ultrasonic transducers that are able to measure wind speed readings up to 40 m s−1), barometric pressure, air temperature, and GPS location. The meteorological sensor was mounted on a mast attached to the surface float of the WG at 0.70 m MSL. The sensor sampled at 1 Hz and then averaged over 10-min bins before transmitting data back to shore via Iridium satellite communications.
As stated by the manufacturer, the sensor has the following wind speed measurement accuracies depending on wind speed categories: 0–5 m s−1: 0.5 m s−1 RMS; 5–40 m s−1: 1 m s−1 RMS. In wet conditions, which include rain, frost, snow, and severe sea spray, errors may increase to 2.5 m s−1 as a result of moisture flow through the wind channel. This is expected to have a reduced impact on the data accuracy because of the bin averaging in the field and in relating the in situ wind speed with longer-term 6-hourly satellite or reanalysis wind products.
b. Satellite/reanalysis wind data
The SeaWinds (SW) globally gridded and blended sea surface vector winds were generated from the multiple satellite observations of the U.S. Department of Defense (DOD), NOAA, and NASA, along with the input of satellite wind retrievals from Remote Sensing Systems, Inc. More information about the satellites, instruments, and blending scheme can be found online (at https://www.ncdc.noaa.gov/data-access/marineocean-data/blended-global/blended-sea-winds). The methods used to generate such data include objective analysis and simple spatiotemporally weighted interpolation. This product provides global ocean coverage with a spatial resolution of 0.25° × 0.25° and a 6-hourly temporal resolution. The data are provided at 10 m MSL.
NOAA uses a global NWP model called the Climate Forecast System, which represents the interaction between the air and sea interfaces throughout the world’s oceans. The CFS, version 2 (CFSv2), operational near-real-time wind vector dataset provides time series at a period of record from 1 January 2011 to 1 April 2016 at a 6-hourly temporal resolution (Saha et al. 2014). CFS time series products are available at a spatial resolution of 0.205
Interim ECMWF Re-Analysis (ERA-Interim) Global Reanalysis Sea Surface Winds are provided at a spatial resolution of 0.125
Last, the NCEP–NCAR reanalysis project has created a sea surface wind dataset that covers the period from 1 January 1979 to the present. Because of increasing user processing errors and systematic biases in NCEP-1, the development of the NCEP–DOE AMIP-II reanalysis (R-2), or NCEP-2, was initiated in the late 1990s (Kanamitsu et al. 2002). NCEP-2 is of the same spatial (1.9047
c. Data comparison approaches
Satellite winds are derived from sea surface backscatter observed by microwave sensors in orbit to generate wind in an equivalent neutrally stable state. This estimation is based on the variation of a normalized radar cross section as a function of local wind conditions and observation geometry (Freilich and Vanhoff 2003). When a lidar laser beam hits a calm water surface at near-normal incidence, roughly 2% of the energy is reflected with little divergence; however, with increasing wind speed and subsequent surface roughening, the divergence angle of the reflected energy increases and the intensity of lidar backscatter decreases (Hu 2009). This process makes the sensors sensitive to ocean surface roughness as a result of the stratification of the atmosphere above sea level. These sensors are thus calibrated to an equivalent neutral wind at a reference height of 10 m above the sea surface (Liu and Tang 1996; Bourassa et al. 1999, 2003; Chelton et al. 2004). These equivalent neutral winds are the winds that would exist if the atmospheric boundary layer were neutrally stratified (Chelton et al. 2004; Carvalho et al. 2013). An observation operator (Tardif and Laroche 2012B) is typically applied to vertically interpolate wind information through the surface layer to the specified 10-m height using the logarithmic wind profile and Monin–Obukov similarity functions (as described by Geleyn 1988). Consequently, to compare real-time stability-dependent winds to equivalent neutral winds, a transformation must be performed to shift the WG winds to the same reference level as satellite winds. Any extrapolation method to shift measured wind speed to a different height is a function of atmospheric turbulence. Wind shear and the buoyant forcing of the atmosphere must be taken into account; ultimately, the vertical density stratification of the atmosphere must be accurately represented (Ruti et al. 2008).
An instantaneous sampling approach may be appropriate in cases where the in situ time series matches exactly with the satellite time series. On the other hand, according to Ruti et al. (2008), an averaging method may require consideration of the phase velocities of cyclones in the region of interest. By determining the typical phase velocity for Mediterranean cyclones, they could determine a general time frame of 30–60 min over which in situ data from the Mediterranean should be averaged when compared with scatterometer winds. Other studies, such as one by Bourassa et al. (2003), state that the optimal averaging period varies with respect to wind speed at the moment of observation. During high wind periods, a shorter averaging period may result in a smaller RMSE, whereas during low wind speed periods, longer averaging periods could also have the same effect. This variability in optimum averaging time with respect to wind speed is anticipated from Taylor’s (1938) hypothesis. This study took an averaged sampling approach to binning the time series. Data were transmitted back to the base station in 10-min averages. Each time series consisted of more than 10 000 wind measurements. For analysis the data were binned twice, first to create an hourly product and then to create a 6-hourly product. Instead of sampling over the duration of each hour, two data points—before and after each desired interval—were averaged.
To compare blended wind fields from satellite data with the WG observations, a collocation procedure was necessary to match the datasets spatially and temporally. For all wind products, a linear interpolation procedure was performed to produce wind pairs from both datasets. This method was possible as a result of the aforementioned temporal gridding of the WG data. Because the WGs were set to circle a fixed point, the majority of the collocations were performed on a coordinate points close to 43°00′S, 8°30′E (Fig. 2).
d. Statistical comparison
WG meteorological records are provided every 10 min, while the satellite products and numerical weather products provide four data points per day (corresponding to the hours 0000, 0600, 1200, and 1800 UTC). Initial statistical analysis is intended to assess the quality of the simultaneous records with respect to temporal and spatial accuracy. To determine whether the error associated with the satellite- and reanalysis-derived wind data is related to a particular wind speed, the wind data were analyzed as a whole, as well as in low, medium, and high wind speed categories. The thresholds for these wind speed categories were determined by the lower and upper quartiles of the WG dataset. This varied slightly for each time series; the winter–spring deployment of the CSIR2 WG in 2015 had a low speed threshold of 7.3 m s−1 and a high speed threshold of 12.4 m s−1, while during austral summer CSIR1 WG had a low speed threshold of 9.0 m s−1 and a high speed threshold of 16 m s−1. The purpose of this approach was to show the error dependence on wind speed and to determine which product best represented each wind category.
The mean and the standard deviation of all wind products per wind speed category are calculated with the intent to show the standard variance from the mean. Statistical parameters such as the root-mean-square error (RMSE) and the correlation coefficient R2 are calculated to assess the ability of the wind products to represent the observed winds. Variables Ui and Uj represent the satellite product wind speed and the observed wind speed by the WG, respectively. Term N is the total number of paired simulation–observed records. More specifically, these parameters will determine the accuracy of the varying wind product’s ability to represent the temporal variability of the wind. Bias is calculated to evaluate the tendency of the data and is intended to estimate the differences in the mean state of the wind field. The intent of observing this parameter is to determine whether wind products tend to over-/underestimate the WG-measured wind speed.
Last, Weibull probability density functions (PDFs) were used to evaluate the various wind products’ ability to describe the WG-measured wind regime. This method was similarly used by Liu et al. (2008) and Carvalho et al. (2013) in their assessment of QuikSCAT and cross-calibrated multiplatform (CCMP) surface winds’ ability to characterize buoy-measured wind regimes closest to reality.
3. Results
Wind speeds measured by the WG are compared against the corresponding gridded 10-m surface winds available from various products. Initially, the comparison was made between the entireties of all datasets. After hourly and 6-hourly binning methods were applied, each WG deployment rendered the following number of data points in the 6-hourly bins to compare to 6-hourly wind products: CSIR2—27 July 2015–2 November 2015: n = 394; CSIR1—7 December 2015–7 February 2016: n = 252 (Fig. 3).
Comparison of wind products (colored curves) with in situ WG data (black line) and upper and lower limits of each product (dotted lines in their respective colors). (a) WG CSIR2, July–November 2015. (b) WG CSIR1, December 2015–February 2016. Data gaps in (a) represent periods when the meteorology sampling was stopped in order to save battery power during poor solar charging periods in winter–spring.
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
To understand the degree of variance from the mean of each product, the mean and standard deviation of all wind products per wind speed category are provided in Table 2. The mean for the varying wind speed categories increases with respect to wind speed. Additionally, the magnitude of the increase in the mean is fairly uniform between the low and medium wind speeds (on average an increase of 2–3 m s−1, except for both CSIR series, which have an increase of ~5–6 m s−1). However, this magnitude varies significantly between the medium and high wind speed categories for all wind products. For instance, during both time series the differences in the medium and high wind means of the SW product increased on the order of 2–3 m s−1, while the differences for other wind products were on the order of 5–7 m s−1.
Mean and standard deviation of all wind products and WG data per wind speed category, as defined in section 2c. All boldface product values indicate the closest match to the WG data. The time series are defined as July 2015–November 2015 for CSIR2 and December 2015–February 2016 for CSIR1. Low wind speeds are defined as CSIR2 ≤ 7.3 m s−1 or CSIR1 ≤ 9.0 m s−1, medium wind speeds are defined as 7.3 m s−1 ≤ CSIR2 ≤ 12.4 m s−1 or 9.0 ≤ CSIR1 ≤ 16.0 m s−1, and high wind speeds are defined as CSIR2 ≥ 12.4 m s−1 or CSIR1 ≥ 16.0 m s−1.
During the CSIR2 WG time series in 2015 the SW, ECMWF, and CFSv2 products underestimated the mean state by approximately 0.3 m s−1, while the NCEP-2 product overestimated the mean state by approximately 1.1 m s−1. All wind products underestimated the mean state of the winds recorded by the CSIR1 WG by approximately 2.9 m s−1 in 2015–16, although the underestimation of the NCEP-2 product was on average just 1.3 m s−1 less than the in situ data. The variance of the mean of the WG data, as depicted by the standard deviation, is consistent at low and medium wind speeds, and increases slightly (approximately 20% increase) for high wind speeds. The only wind product that depicted this trend during both time series was NCEP-2—the product’s low wind speed category had a lower variance (averaged ±3.2 m s−1) with respect to the variance of the high wind speed category (averaged ±4.0 m s−1). The SW product exhibited a greater variance (±2.7 m s−1) with low wind speeds during the CSIR2 time series; however, the opposite was observed during the CSIR1 time series, as this wind speed category exhibited a reduced variance (±2.2 m s−1) with respect to the other wind speed categories. The ECMWF product consistently exhibited a trend of decreasing variance with an increase in wind speed (averaged low wind variance ±2.6 m s−1 compared to the averaged high wind variance of ±2.2 m s−1). The CFSv2 product exhibited the most consistent variance across time series and wind speed categories with an average variance of ±2.9 m s−1 for the CSIR2 time series and an average variance of ±2.3 m s−1 for the CSIR1 time series.
A further statistical comparison was conducted using root-mean-square error, mean bias, and the correlation coefficient (Table 3). Globally, there is a tendency for satellite product wind speed errors, as depicted by RMSE, to be greater in the presence of low and high winds (Carvalho et al. 2013). This trend was especially apparent for the SW product during the CSIR2 time series. However, also during this time series the ECMWF, CFSv2, and NCEP-2 products exhibited the highest RMSE for the low wind speed category. For the CSIR1 time series, the SW, ECWMF, and CFSv2 products exhibited the opposite trend where the high wind speed category had the highest RMSE. Only the NCEP-2 product exhibited the trend described by Carvalho et al. (2013) during this time series. The ECMWF product exhibited the lowest RMSE across all wind speed categories during the CSIR2 time series, whereas during the CSIR1 series the NCEP-2 product had the lowest RMSE for the all, medium, and high wind speed categories (outperformed by CFSv2 and ECMWF for the low wind speed category).
Comparison between in situ WG and satellite/reanalysis products, including the wind speed error per wind speed category, RMSE, bias, and correlation coefficients. Bold values highlight the wind product providing the closest match with the in situ data (i.e., the smallest value of error or bias and the highest value of the correlation coefficient). All correlation coefficients > |0.24| exhibit a confidence interval > 99%. The number of comparative observations is shown by N. The time series for WG CSIR1 is December 2015–February 2016 and July–November 2015 for WG CSIR2. Low wind speeds are defined as WG CSIR1 ≤ 9.0 m s−1 or WG CSIR2 ≤ 7.3 m s−1, medium wind speeds are defined as 9.0 m s−1 ≤ WG CSIR1 ≤ 16.0 m s−1 or WG CSIR2 ≤ 7.3 m s−1 ≤ 12.4 m s−1, and high wind speeds are defined as WG CSIR1 ≥ 16.0 m s−1 or WG CSIR2 ≥ 12.0 m s−1.
The time series correlation using data collected by WG CSIR2 from July 2015 to November 2015 (Fig. 4) represents an extensive austral winter to spring dataset, while the time series collected by WG CSIR1 from December 2015 to February 2016 (Fig. 5) represents an austral summer dataset of in situ observations. Related confidence levels of the correlations can be found in Table 3. For both time series, the SW product performed poorly across all wind speed categories and the majority of the correlations fell outside of confidence intervals. The SW product was observed to consistently have the highest error, especially so during high winds. Overall, the lowest error varies but is shown by the ECMWF wind product during the CSIR2 time series (all wind speed RMSE = 2.62 m s−1) and by the NCEP-2 product during the CSIR1 time series (all wind speed RMSE = 2.52 m s−1). For the all wind speed and medium wind speed categories, the ECMWF product had the highest correlation coefficients overall with respect to the WG in situ data.
Comparison of wind products with in situ (CSIR2 WG) data collected July 2015–November 2015 (austral winter–spring). Wind speed categories are determined using the lower and upper quartiles of the in situ data. Linear fit for each wind speed category is shown.
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
Comparison of wind products with in situ (CSIR1 WG) data collected December 2015–February 2016 (austral summer). Wind speed categories are determined using the lower and upper quartiles of the in situ data. Linear fit for each wind speed category is shown.
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
ECMWF exhibited the following correlation coefficients for the all wind speed category: 0.76 and 0.93, and the following coefficients for the medium wind speed category: 0.40 and 0.81 (Table 3). Low wind speeds are best represented overall by the CFSv2 product, with correlation coefficients of 0.21 and 0.58, respectively. Statistically, the NCEP-2 product outperformed all other products in the high wind speed category with coefficients of 0.74 and 0.82, even though it significantly overestimated the magnitude of low and high winds during the CSIR2 time series (Fig. 4). This overestimation positively skews the overall correlation of the NCEP-2 with the in situ data, which would suggest that it is not best suited to represent the temporal variability of this particular time series even though the statistics in Table 3 would suggest otherwise. Therefore, the CFSv2 and ECMWF products similarly best represent the high wind speed category and in general best represent the temporal variability of the wind field during this sampling period.
The bias exhibited by all wind products with respect to the data collected by the WG ranged from as small as −0.18 m s−1 (CFSv2 vs CSIR2 2015 all wind speed category) to as large as −7.44 m s−1 (SW vs CSIR1 2015–16 high wind speed category). Similar to the RMSE, the trends in bias are different during each time series. During the CSIR2 time series (July–November 2015), the bias for the ECMWF, CFSv2, and NCEP-2 products remained fairly consistent across wind speed categories but exhibited a positive bias for low winds and a negative bias for medium and high winds (exception: NCEP-2 high winds). The SW product during this time series exhibited a positive bias of +2.20 m s−1 for the low speed category, −0.36 m s−1 for the medium wind speed category, and −3.36 m s−1 for the high wind speed category. This trend in bias was exhibited only by the SW and ECWMF products during the CSIR2 time series (July–November 2015) and causes the SW product to appear better in performance than the NCEP-2 product in this respect (SW all wind speed bias = −0.47 m s−1, NCEP-2 all wind speed bias = +1.13 m s−1). During the CSIR1 time series (December 2015–February 2016), all products exhibited a systematic increase in bias with an increase in wind speed; however, the magnitude of increase was largest for the SW and smallest for the NCEP-2 product. Additionally, during this time series, all products exhibited a negative bias, indicative of an underrepresentation of the measured wind field (exception: SW low wind speed category).
Figure 6 visualizes the wind speed residuals from the wind products. There is an observed offset between the linear fit for each time series across all wind products; the CSIR2 time series exhibits residuals on average 2 m s−1 higher than the CSIR1 time series. The SW product is observed to consistently overestimate low wind speeds and significantly underestimate high wind speeds. The ECMWF and CFSv2 products exhibit very similar trends in residuals and only slightly overestimate low wind speeds and slightly underestimate high wind speeds. The NCEP-2 product performs differently per time series by having the smallest residuals at the highest wind speeds during the CSIR2 time series and the largest residuals at highest wind speeds during the CSIR1 time series. Although each product performs slightly different from one another, they all follow the same trend: overestimating low wind speeds and underestimating high wind speeds.
Wind speed residuals (wind product minus in situ data) for both WG time series. Linear fit for each time series and product is displayed in their respective colors.
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
To assess which of the wind products offers a characterization of the WG wind regime closest to reality, a Weibull PDF is used (Fig. 7). For the CSIR2 time series (July 2015–November 2015), the ECMWF product wind speed frequency distribution is almost identical to the WG frequency distribution, with the CFSv2 product being a close second. The SW product exhibits a similar curve but significantly overestimates the frequency of the medium wind speeds while underestimating frequencies of high wind speeds. The NCEP-2 product underestimates the frequencies of low and medium wind speeds and overestimates the frequencies of high wind speeds during austral spring. However, during the CSIR1 time series (December 2015–February 2016), the NCEP-2 product best represents the frequency distribution of the in situ wind speeds. During austral summer the SW, ECMWF, and CFSv2 products overestimate the frequency of low and medium wind speeds, and underestimate the frequency of high wind speeds. While the in situ wind speed frequency distribution varies between the two time series, the products consistently estimate similar frequency distributions. This could be a result of sensor calibration offsets between the Airmar sensors and/or a seasonal change in the wind distributions between the two time series. When compared to the other products, the SW product systematically represented the wind fields having higher frequencies of lower wind speeds. The ECMWF and CFSv2 products are characterized by a more even frequency distribution across all wind speed categories, while the NCEP-2 product tended to introduce reduced variability in the wind speed frequency distribution. It was observed that the higher the maximum wind speed of the wind product, the greater the variability and the more even the distribution of wind speeds.
Wind speed frequency distribution (Weibull PDF) for all products (SW, ECMWF, CFSv2, and NCEP-2) compared to in situ wind collection by WG (a) CSIR2 and (b) CSIR1.
Citation: Journal of Atmospheric and Oceanic Technology 34, 12; 10.1175/JTECH-D-17-0079.1
4. Discussion
a. Comparing field data to global wind products
At global scale, numerous gridded scatterometer and/or reanalysis wind products, such as QuikSCAT, OSCAT, and CCMP, tend to overestimate the true wind speed exponentially with respect to wind intensity (Carvalho et al. 2013), and they have the highest margins of error during low and high wind events (Sudha and Prasada Rao 2013; Alvarez et al. 2014; Jayaram et al. 2014). However, because various wind products have not been rigorously intercalibrated, any comparison between satellite- and model-derived wind data with in situ data will give only relative information on accuracy (Vogelzang et al. 2012).
Based on our results, the performance of each product varied with respect to time of year and wind strength. The SW product consistently underestimated the wind fields and exhibited poor correlation across all wind speeds. As such, it is not considered the preferred wind product for use in process understanding of upper-ocean dynamics or realistic numerical analysis in the Southern Ocean. This may in part be attributed to the fact that while SW data are available at a relatively high spatial resolution (0.25° × 0.25°), it is a gridded satellite product. This product uses a simple objective analysis method (spatially and temporally weighted interpolation) to generate a gridded and blended product (Zhang 2006). This method of blending satellite observations minimizes but does not completely eliminate data gaps. As such, the high resolution of spatial and temporal sampling conducted by WG technology may not be represented accurately by the SW dataset. There is a dearth of in situ observations in this remote ocean, which leads to difficulties in validation of any given wind product. This is further amplified by our poor understanding of the surface boundary layer, where accurate estimates of both momentum and heat fluxes are poor and likely contribute to inaccurate algorithms and bulk formulas that are used to derive satellite wind estimates (Gille et al. 2016). Furthermore, such a gap in observational input may lead to inadequate assimilation and parameterization of the boundary layer and surface frictional processes. This can ultimately impact how we derive and model (reanalyze) wind estimates to produce gridded wind products. The SW data also poorly represent wind estimates at different wind speed thresholds, particularly at high wind states. This indicates that there may be issues with how satellite scatterometer data are interpreted during extreme events, which could be a result of high sea state (swell), surface wave breaking, and surface turbulence—all affecting the backscatter observed by the satellite. The strong performance of all reanalysis wind products in this study is indicative of improved algorithms and model parameterizations for different wind speeds related to different sea states. The improved performance of both the ECMWF and CFSv2 products may be attributed to the high spatial resolution of the datasets (0.125° × 0.125° and 0.205° × 0.204°, respectively) that take into account finer-scale variability in the wind field and meteorological processes. In this study, four of the ECMWF sampling locations fall either directly on or within a tenth of a degree (~10 km) away from the circumference of the octagon sampling pattern (Fig. 2) of the WGs for both time series. For the NCEP-2 product, the nearest point of reference to the majority of the in situ measurements is roughly 100 km (see Fig. 2). Such low-resolution data could lead to larger uncertainties between NCEP-2 and the in situ measurements. Overall, the ECMWF product performs best (with “all” wind speed correlations of 0.76 and 0.93) at consistently representing the temporal wind field variability as observed by the WG.
During the CSIR2 time series (July–November 2015), measures of error and bias for ECMWF and CFSv2 were lower than for NCEP-2 for all wind speed categories. However, during the CSIR1 time series (December 2015–February 2016), the NCEP-2 product outperforms all other products with regard to error and bias. While the CFSv2 product is better than ECMWF at representing the mean wind state of low and medium wind speeds, ECMWF is better at representing the temporal variability of the wind field magnitude. Upon closer inspection of the performance of the NCEP-2 product, it best represented the mean wind state by exhibiting similar trends in standard deviation across wind speed categories. For example, across time series the NCEP-2 product had similar deviations from the mean of the low and medium wind speed categories, and a greater deviation for the high wind speed category. This trend is also depicted by the in situ data (Table 2); however, all other products exhibit opposing trends in deviation from the mean. Because of these factors, the NCEP-2 product is best at representing the mean wind state and deviation from the mean state, particularly during periods of high wind speeds. However, this product should be used with caution, given that NCEP-2 derives its overall high correlation through its overestimation of low and high wind speeds (Fig. 6).
The ability of a product to best represent the wind speed frequency distribution is illustrated in Fig. 7. Findings by Pickett et al. (2003) and Tang et al. (2004) suggest that satellites measure higher frequencies of strong winds and lower frequencies of weak winds. However, during this study the frequency distribution of the in situ wind speeds varied between times series, while the frequency distribution of the satellite product wind speeds remains consistent. As such, a different product for each time series best represented the frequency distribution of observed wind speeds (Fig. 7). The SW product consistently showed higher frequencies of low wind speeds, which did not reflect the frequency distribution of the in situ winds for either time series. ECMWF and CFSv2 similarly showed a more uniform distribution of wind speed frequencies that almost perfectly matched the frequency distribution measured by the CSIR2 WG during the July–November 2015 time series. There was an observed 2.52 m s−1 increase in the mean wind speed of the CSIR1 time series (December 2015–February 2016) with respect to the CSIR2 time series (Table 2). NCEP-2 consistently showed a broader distribution of speed frequencies and thus this product best represented the frequency distribution of wind speeds during the CSIR1 time series.
Satellite-derived wind data as well as reanalysis and model products can be very different from in situ anemometer data in that they tend to represent synoptic-scale wind perturbations. A study by Atlas et al. (1999) showed that satellite products can detect mesoscale features; however, this representation is limited, since satellite and NWP products are of lower spatial and temporal resolution (Carvalho et al. 2013). Anemometer wind data are closer to a temporal average of instantaneous moments, while satellite winds represent a spatial average of instantaneous moments (Pensieri et al. 2010). Studies have shown that in cases where a limited amount of data existed, such as from a single satellite, any interpolations or extrapolations were inaccurate in representing the true strength of mesoscale events (Isaksen and Stoffelen 2000). Bearing this in mind, over most SW grids the number of observational input data points is over 40 (Zhang 2006). As such a multisatellite blending scheme can be advantageous in representing finescale wind perturbations in the Southern Ocean. ECMWF, on the other hand, is a reanalysis product that uses a sequential data assimilation scheme, advancing forward in time using 12-hourly analysis cycles (Dee et al. 2011). A temporally short-scale assimilation scheme could give ECMWF an advantage over other reanalysis products, such as NCEP-2, in representing the temporal variability on shorter temporal scales, as is a dominant scale of this study. CFSv2 is a fully coupled ocean–land–atmosphere dynamical model that uses a Global Ocean Data Assimilation System (GODAS) operational at NCEP (Saha et al. 2014). This product is unique in that it makes retrospective forecasts to calibrate operational subsequent real-time subseasonal and seasonal predictions (Saha et al. 2014). This calibration scheme could be why CFSv2 is a more rounded product and performs well in representing both the temporal variability and the mean wind state. The NCEP-2 product is also a model simulation (forecast/hindcast reanalysis); however, the data assimilation scheme is slightly different from ECMWF, which is forced to conform to observational input. Using 4D data assimilation methodology, observational input from ships, satellite, radiosonde, aircraft, and meteorological station observations are incorporated into NCEP-2 (Kalnay et al. 1996). An assimilation methodology with high input gives NCEP-2 the ability to represent regional climate dynamics (Kanamitsu et al. 2002), even though it is of the coarsest spatial resolution compared to all other wind products.
b. Glider meteorological sampling assumptions and potential shortcomings
Compared to conventional in situ meteorological observing platforms, such as ships and moorings, the WG has a low profile with respect to sea level with the Airmar sensor located just 0.7 m above the sea surface. As stated in the methodology section, all wind observations in this study are extrapolated to 10 m above the sea surface for comparison to the wind products. However, this proximity of sampling so close to the sea surface is likely impacted by surface friction and wave sheltering of the true wind field, especially in the case of extreme wind events or storms where large swell conditions of the Southern Ocean may impact the WG to gridded product comparison. This scenario may explain the greater observed bias between the WG and product winds as a function of increasing wind speed and the impacts felt by the associated increase in sea state. This would suggest that wind observations from near-surface platforms measure the winds between waves rather than the wind field “above” the ocean surface. As such WGs are likely prone to underestimating wind speeds, particularly in the presence of large waves, and so are more suited to characterizing wind variability versus wind speed magnitude. In future deployments, ship-based meteorological observations could be used to compare with the WG observations to assess the magnitude and behavior of these potential inaccuracies in varying sea states and wind regimes. In addition, the Airmar sensor is prone to larger errors at higher wind speeds (see the methodology section), which could contribute to the enhanced bias with higher wind speeds.
Furthermore, the WG measures absolute wind conditions at a particular location, while scatterometers measure wind relative to a moving ocean surface (Sudha and Prasada Rao 2013). Dickinson et al. (2001) observed that during alignment (in both the same and opposite directions) of both wind and surface ocean currents, scatterometer data tend to be under- or overestimated. This could partially explain the bias observed by products using scatterometer data. Shown in Table 2, the greatest standard deviation of wind products is often associated with low wind speeds (except for NCEP-2), potentially indicative of the scatterometer’s inability to appropriately detect weak wind speeds as a result of the limited detection limits of the sensor in measuring surface scatter caused by wind–surface interaction. Additionally, the relation between wind and backscatter is provided by an empirical geophysical model function. However, because the quantity of backscatter cannot be obtained during in situ sampling, model functions have varying methods of interpretation and subsequent assumptions, which may not translate as intended for remote locations such as the Southern Ocean. An example of such an assumption is of the neutral stratification of the marine atmospheric boundary layer that is, on average, weakly unstable and the global average neutral equivalent wind is ~0.2 m s−1 stronger than real wind (Hersbach 2010). Ebuchi et al. (2002) note the generation of small-scale waves via wind stress also varies with respect to atmospheric stability, which in turn affects radar backscatter used to derive wind speed. In addition, the difference between surface air temperature and SST is largely responsible for variability in stability. In the condition of a stable boundary layer, neutral equivalent winds can be lower than real winds by as much as 0.5 m s−1 (Hersbach 2010). A similar argument can be made for the WG, where in this study the vertical transformation of WG data did not include information on the atmospheric stability because we have no observations of the relative humidity from the WG. It is possible that data transformation could be in the order of 0.7 m s−1 greater in magnitude (Singh et al. 2013). In future deployments and glider experiments, the measurement of the atmospheric stability via relative humidity—not yet available on sensors located in such proximity to the sea surface—should be included in order to reduce uncertainty in the WG-derived winds.
c. Broader implications
Ocean circulation modeling, climate change estimation, and numerical weather predictions are heavily dependent on the accuracy of input of meteorological information. Upper-ocean dynamics are closely linked to atmospheric variability, particularly in the Southern Ocean, where wind magnitudes and heat flux variations are great. As mentioned previously, the midlatitude regions of the Southern Ocean are host to the strongest wind fields at the ocean surface. If these winds are not accurately represented, models will incorrectly simulate a range of processes and parameters, such as air–sea exchanges of heat and moisture, lateral advection, stratification and mixed layer dynamics (frontal formation and instabilities), and Ekman pumping and transport. Given the importance of this, the present study and others show that in many cases the wind fields can be poorly represented.
Upper-ocean models that include satellite-derived wind speed input do not accurately represent variance in mixed layer depth (MLD) and SST. Ocean circulation models have shown the sensitivity of the MLD to the gustiness in the wind (Lee et al. 2008). Turbulent mixing and subsequent buoyancy forcing caused by sea surface winds strongly influence the MLD. As the MLD shallows, high wind events are observed to have a greater impact on the deepening of the mixed layer. Carranza and Gille (2015) observed that during high wind events, water from below the mixed layer is entrained in the upper ocean as the mixed layer deepens. Swart et al. (2015) observe with high-resolution glider datasets that this variability of the MLD is driven by synoptic storms (4–9 days) where turbulent mixing deepens the surface mixed layer, while quiescent wind episodes allow the upper ocean to restratify and subsequently shoal the MLD (du Plessis et al. 2017). This wind-driven mixing also influences SST, where cold water from below the MLD is entrained in the upper ocean. This in turn influences the mixed layer heat budget (Bonekamp et al. 1999) and produces cooler SSTs. Carranza and Gille (2015) found a statistically significant negative correlation between wind speed and SST anomalies, and that wind speed alone can explain as much as 80% of the variance observed in SSTs. The accurate representation of these upper-ocean dynamics has been proven to have important implications for numerous chemical, biogeochemical, and biological processes (e.g., Thomalla et al. 2011; Fauchereau et al. 2011; Swart et al. 2015; Carranza and Gille 2015), as well as biogeochemical models (e.g., Nicholson et al. 2016), that are not elaborated on in this study.
5. Conclusions
Predictions of heat, moisture, and momentum exchanges between the ocean and the atmosphere remain uncertain in data-sparse regions, such as the Southern Ocean (Gille et al. 2016). We use high-frequency observations of wind stress obtained from WG deployments in the subantarctic Southern Ocean to compare with four available satellite scatterometer and reanalysis global wind products. It was found that the NCEP-2 product best represented the mean wind state in certain conditions, namely, with respect to the high wind speed category correlations for CSIR2 and CSIR1 (0.74 and 0.82, respectively). Overall, ECMWF most consistently represented the temporal wind field variability as observed by the glider, by representing the highest “all” wind speed correlations with coefficients of 0.76 and 0.93 for the independent time series. CFSv2 was a close performer to ECMWF in its representation of the temporal wind field variability with correlations differing on average <0.05 from the ECMWF correlations with the WG. However, on average CFSv2 had a slightly higher bias and RMSE (on average <0.2 m s−1) compared to ECMWF. The results clearly showed that the SW product performed poorly at representing the mean or wind stress variability (majority of the correlations <0.3) compared to those observed by the WG.
The overall comparison between WG winds and gridded products in this study provide confidence that autonomous surface vehicles, such as the WG, can be used to understand and validate satellite-derived wind estimates. Further use of such vehicles should be supported to provide improved time- and space-scale datasets and data volumes, particularly in remote and harsh regions, such as the high-latitude oceans. There is a need to correctly parameterize transient and synoptic-scale wind fields in order to improve depictions of decadal and century-scale changes in ocean–atmosphere relationships.
Acknowledgments
The Wave Glider wind data were provided by the Southern Ocean Carbon and Climate Observatory, Council for Scientific and Industrial Research (CSIR). The authors of this work would like to thank SANAP and the captain and crew of the S. A. Agulhas I and S. A. Agulhas II for their contribution in the deployment and retrieval of the Wave Gliders. We acknowledge the work of SAMERC and Sea Technology Services for housing, managing, and piloting the gliders. The South African ocean glider program is supported by the Department of Science and Technology (DST) through the Sustainability and Innovation in Southern Ocean Observational Infrastructure grant. The SeaWinds blended altimetry data, CFS climate forecast, and NCEP-2 reanalysis wind product were downloaded from NOAA’s National Centers for Environmental Information. The ECMWF reanalysis wind product was downloaded from the European Centre for Medium-Range Weather Forecasts’ open-access database. The authors also thank two anonymous reviewers of K. Schmidt’s M.S. dissertation, which has been used to present this study, as well as two anonymous reviewers of this paper. S. Swart thanks the following funding agencies: the South African National Research Foundation (NRF), Grant SNA14071475720; and a Wallenberg Academy Fellowship (WAF 2015.0186).
REFERENCES
Airmar Technology Corporation, 2016: WeatherStation 200WX fact sheet. 2 pp., http://www.airmar.com/uploads/brochures/WX-OFFSHORE.pdf.
Alvarez, I., M. Gomez-Gesteira, M. deCastro, and D. Carvalho, 2014: Comparison of different wind products and buoy wind data with seasonality and interannual climate variability in the southern Bay of Biscay (2000–2009). Deep-Sea Res. II, 106, 38–48, doi:10.1016/j.dsr2.2013.09.028.
Atlas, R., S. C. Bloom, R. N. Hoffman, E. Brin, J. Ardizzone, J. Terry, D. Bungato, and J. C. Jusem, 1999: Geophysical validation of NSCAT winds using atmospheric data and analyses. J. Geophys. Res., 104, 11 405–11 424, doi:10.1029/98JC02374.
Bentamy, A., D. Croize-Fillon, and C. Perigaud, 2008: Characterization of ASCAT measurements based on buoy and QuikSCAT wind vector observations. Ocean Sci., 4, 265–274, doi:10.5194/os-4-265-2008.
Bonekamp, H., A. Sterl, and G. J. Komen, 1999: Interannual variability in the Southern Ocean from an ocean model forced by European Centre for Medium-Range Weather Forecasts reanalysis fluxes. J. Geophys. Res., 104, 13 317–13 331, doi:10.1029/1999JC900052.
Bourassa, M. A., D. G. Vincent, and W. L. Wood, 1999: A flux parameterization including the effects of capillary waves and sea state. J. Atmos. Sci., 56, 1123–1139, doi:10.1175/1520-0469(1999)056<1123:AFPITE>2.0.CO;2.
Bourassa, M. A., D. M. Legler, J. J. O’Brien, and S. R. Smith, 2003: SeaWinds validation with research vessels. J. Geophys. Res., 108, 3019, doi:10.1029/2001JC001028.
Carranza, M. M., and S. T. Gille, 2015: Southern Ocean wind-driven entrainment enhances satellite chlorophyll-a through the summer. J. Geophys. Res. Oceans, 120, 304–323, doi:10.1002/2014JC010203.
Carvalho, D., A. Rocha, M. Gómez-Gesteira, I. Alvarez, and C. Silva Santos, 2013: Comparison between CCMP, QuikSCAT and buoy winds along the Iberian Peninsula coast. Remote Sens. Environ., 137, 173–183, doi:10.1016/j.rse.2013.06.005.
Chelton, D. B., M. G. Schlax, M. H. Freilich, and R. F. Milliff, 2004: Satellite measurements reveal persistent small-scale features in ocean winds. Science, 303, 978–983, doi:10.1126/science.1091901.
Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553–597, doi:10.1002/qj.828.
Dickinson, S., K. A. Kelly, M. J. Caruso, and M. J. McPhaden, 2001: Comparisons between the TAO buoy and NASA scatterometer wind vectors. J. Atmos. Oceanic Technol., 18, 799–806, doi:10.1175/1520-0426(2001)018<0799:CBTTBA>2.0.CO;2.
du Plessis, M., S. Swart, I. J. Ansorge, and A. Mahadevan, 2017: Submesoscale processes promote seasonal restratification in the Subantarctic Ocean. J. Geophys. Res. Oceans, 122, 2960–2975, doi:10.1002/2016JC012494.
Ebuchi, N., H. C. Graber, and M. J. Caruso, 2002: Evaluation of wind vectors observed by QuikSCAT/Seawinds using ocean buoy data. J. Atmos. Oceanic Technol., 19, 2049–2062, doi:10.1175/1520-0426(2002)019<2049:EOWVOB>2.0.CO;2.
Fauchereau, N., A. Tagliabue, L. Bopp, and P. M. S. Monteiro, 2011: The response of phytoplankton biomass to transient mixing events in the Southern Ocean. Geophys. Res. Lett., 38, L17601, doi:10.1029/2011GL048498.
Freilich, M. H., and B. A. Vanhoff, 2003: The relationship between winds, surface roughness, and radar backscatter at incidence angles from TRMM Precipitation Radar measurements. J. Atmos. Oceanic Technol., 20, 549–562, doi:10.1175/1520-0426(2003)20<549:TRBWSR>2.0.CO;2.
Geleyn, J. F., 1988: Interpolation of wind, temperature and humidity values from model levels to the height of measurement. Tellus, 40A, 347–351, doi:10.1111/j.1600-0870.1988.tb00352.x.
Gille, S., S. Josey, and S. Swart, 2016: New approaches for air-sea fluxes in the Southern Ocean. Eos, Trans. Amer. Geophys. Union, 97, doi:10.1029/2016EO052243.
Herrera, J. L., S. Piedracoba, R. A. Varela, and G. Roson, 2005: Spatial analysis of the wind field on the western coast of Galicia (NW Spain) from in situ measurements. Cont. Shelf Res., 25, 1728–1748, doi:10.1016/j.csr.2005.06.001.
Hersbach, H., 2010: Comparison of C-band scatterometer CMOD5.N equivalent neutral winds with ECMWF. J. Atmos. Oceanic Technol., 27, 721–736, doi:10.1175/2009JTECHO698.1.
Hilburn, K. A., M. A. Bourassa, and J. J. O’Brien, 2003: Development of scatterometer-derived surface pressures for the Southern Ocean. J. Geophys. Res., 108, 3244, doi:10.1029/2003JC001772.
Ho, D. T., C. S. Law, M. J. Smith, P. Schlosser, M. Harvey, and P. Hill, 2006: Measurements of air-sea gas exchange at high wind speeds in the Southern Ocean: Implications for global parameterizations. Geophys. Res. Lett., 33, L16611, doi:10.1029/2006GL026817.
Hu, Y., 2009: Ocean, land and meteorology studies using space-based lidar measurements. Recent Advances in Remote Sensing: Proceedings of the Fifth WSEAS International Conference on Remote Sensing (REMOTE’09), R. Revetria, V. Mladenov, and N. Mastorakis, Eds., WEAS Press, 47–50.
Isaksen, L., and A. Stoffelen, 2000: ERS scatterometer wind data impact on ECMWF’s tropical cyclone forecasts. IEEE Trans. Geosci. Remote Sens., 38, 1885–1892, doi:10.1109/36.851771.
Jayaram, C., T. Udaya-Bhaskar, D. Swain, E. Pattabhi, R. Rao, S. Bansal, D. Dutta, and K. Hanumantha Rao, 2014: Daily composite wind fields from Oceansat-2 scatterometer. Remote Sens. Lett., 5, 258–267, doi:10.1080/2150704X.2014.898191.
Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437–471, doi:10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.
Kanamitsu, M., W. Ebisuzaki, J. Woollen, S. K. Yang, J. J. Hnilo, M. Fiorino, and L. Potter, 2002: NCEP–DOE AMIP-II Reanalysis (R-2). Bull. Amer. Meteor. Soc., 83, 1631–1643, doi:10.1175/BAMS-83-11-1631.
Lee, T., O. Wang, W. Tang, and W. T. Liu, 2008: Wind stress measurements from the QuikSCAT-SeaWinds scatterometer tandem mission and the impact on an ocean model. J. Geophys. Res., 113, C12019, doi:10.1029/2008JC004855.
Liu, W. T., and W. Tang, 1996: Equivalent neutral wind. Jet Propulsion Laboratory Publ. 96-17, 22 pp.
Liu, W. T., W. Tang, and X. Xie, 2008: Wind power distribution over the ocean. Geophys. Res. Lett., 35, L13808, doi:10.1029/2008GL034172.
Mears, C. A., D. K. Smith, and F. J. Wentz, 2001: Comparison of Special Sensor Microwave Imager and buoy-measured wind speeds from 1987 to 1997. J. Geophys. Res., 106, 11 719–11 729, doi:10.1029/1999JC000097.
Monteiro, P. M. S., L. Gregor, M. Lévy, S. Maenner, C. L. Sabine, and S. Swart, 2015: Intraseasonal variability linked to sampling alias in air-sea CO2 fluxes in the Southern Ocean. Geophys. Res. Lett., 42, 8507–8514, doi:10.1002/2015GL066009.
Nicholson, S. A., M. Lévy, J. Llort, S. Swart, and P. M. S. Monteiro, 2016: Investigation into the impact of storms on sustaining summer primary productivity in the Sub‐Antarctic Ocean. Geophys. Res. Lett., 43, 9192–9199, doi:10.1002/2016GL069973.
Nightingale, P. D., G. Malin, C. S. Law, A. J. Watson, P. S. Liss, M. I. Liddicoat, J. Boutin, and R. C. Upstill-Goddard, 2000: In situ evaluation of air-sea gas exchange parameterizations using novel conservative and volatile tracers. Global Biogeochem. Cycles, 14, 373–387, doi:10.1029/1999GB900091.
Patoux, J., X. Yuan, and C. Li, 2009: Satellite-based midlatitude cyclone statistics over the Southern Ocean: 1. Scatterometer-derived pressure fields and storm tracking. J. Geophys. Res., 114, D04105, doi:10.1029/2008JD010873.
Peixoto, J. P., and A. H. Oort, 1992: Physics of Climate. American Institute of Physics, 520 pp.
Pensieri, S., R. Bozzano, and M. E. Schiano, 2010: Comparison between QuikSCAT and buoy wind data in the Ligurian Sea. J. Mar. Syst., 81, 286–296, doi:10.1016/j.jmarsys.2010.01.004.
Pickett, M. H., W. Tang, L. K. Rosenfeld, and C. H. Wash, 2003: QuikSCAT satellite comparisons with nearshore buoy wind data off the U.S. West Coast. J. Atmos. Oceanic Technol., 20, 1869–1879, doi:10.1175/1520-0426(2003)020<1869:QSCWNB>2.0.CO;2.
Risien, C. M., and D. B. Chelton, 2008: A global climatology of surface wind and wind stress fields from eight years of QuikSCAT scatterometer data. J. Phys. Oceanogr., 38, 2379–2413, doi:10.1175/2008JPO3881.1.
Ruti, P. M., S. Marullo, F. D’Ortenzio, and M. Tremant, 2008: Comparison of analyzed and measured wind speeds in the perspective of oceanic simulations over the Mediterranean basin: Analyses, QuikSCAT and buoy data. J. Mar. Syst., 70, 33–48, doi:10.1016/j.jmarsys.2007.02.026.
Saha, S., and Coauthors, 2014: The NCEP Climate Forecast System Version 2. J. Climate, 27, 2185–2208, doi:10.1175/JCLI-D-12-00823.1.
Satheesan, K., A. Sarkar, A. Parekh, M. R. R. Kumar, and Y. Kuroda, 2007: Comparison of wind data from QuikSCAT and buoys in the Indian Ocean. Int. J. Remote Sens., 28, 2375–2382, doi:10.1080/01431160701236803.
Singh, P., A. Parekh, and R. Attada, 2013: Comparison of a single logarithmic and equivalent neutral wind approaches for converting buoy-measured wind speed to the standard height: Special emphasis to North Indian Ocean. Theor. Appl. Climatol., 111, 455–463, doi:10.1007/s00704-012-0674-2.
Sudha, A. K., and C. V. Prasada Rao, 2013: Comparison of Oceansat-2 scatterometer winds with buoy observations over the Indian Ocean and the Pacific Ocean. Remote Sens. Lett., 4, 171–179, doi:10.1080/2150704X.2012.713140.
Swart, S., S. Speich, I. J. Ansorge, and J. R. E. Lutjeharms, 2010: An altimetry-based gravest empirical mode south of Africa: 1. Development and validation. J. Geophys. Res., 115, C03002, doi:10.1029/2009JC005299.
Swart, S., and Coauthors, 2012: Southern Ocean Seasonal Cycle Experiment 2012: Seasonal scale climate and carbon links. S. Afr. J. Sci., 108, 1089, doi:10.4102/sajs.v108i3/4.1089.
Swart, S., S. J. Thomalla, and P. M. S. Monteiro, 2015: The seasonal cycle of mixed layer dynamics and phytoplankton biomass in the Sub-Antarctic Zone: A high-resolution glider experiment. J. Mar. Syst., 147, 103–115, doi:10.1016/j.jmarsys.2014.06.002.
Tang, W., W. T. Liu, and B. W. Stiles, 2004: Evaluation of high-resolution ocean surface vector winds measured by QuikSCAT scatterometer in coastal regions. IEEE Trans. Geosci. Remote Sens., 42, 1762–1769, doi:10.1109/TGRS.2004.831685.
Tardif, R., and S. Laroche, 2012B: An equivalent neutral wind observation operator for variational assimilation of scatterometer ocean surface wind data. Quart. J. Roy. Meteor. Soc., 138, 2086–2104, doi:10.1002/qj.1953.
Taylor, G. I., 1938: The spectrum of turbulence. Proc. Roy. Soc. London, 164A, 476–490, doi:10.1098/rspa.1938.0032.
Thomalla, S. J., N. Faucherau, S. Swart, and P. M. S. Monteiro, 2011: Regional scale characteristics of the seasonal cycle of chlorophyll in the Southern Ocean. Biogeosciences, 8, 2849–2866, doi:10.5194/bg-8-2849-2011.
Vogelzang, J., and A. Stoffelen, 2012: Scatterometer wind vector products for application in meteorology and oceanography. J. Sea Res., 74, 16–25, doi:10.1016/j.seares.2012.05.002.
Wanninkhof, R., 1992: Relationship between wind speed and gas exchange over the ocean. J. Geophys. Res., 97, 7373–7382, doi:10.1029/92JC00188.
Wanninkhof, R., and W. R. McGillis, 1999: A cubic relationship between air-sea CO2 exchange and wind speed. Geophys. Res. Lett., 26, 1889–1892, doi:10.1029/1999GL900363.
Wanninkhof, R., W. E. Asher, D. T. Ho, C. Sweeney, and W. R. McGillis, 2009: Advances in quantifying air-sea gas exchange and environmental forcing. Annu. Rev. Mar. Sci., 1, 213–244, doi:10.1146/annurev.marine.010908.163742.
Wu, Q., and G. Chen, 2015: Validation and intercomparison of HY-2A/ MetOp-A/ Oceansat-2 scatterometer wind products. Chin. J. Oceanol. Limnol., 33, 1181–1190, doi:10.1007/s00343-015-4160-4.
Yang, X., G. Liu, Z. Li, and Y. Yu, 2014: Preliminary validation of ocean surface vector winds estimated from China’s HY-2A scatterometer. Int. J. Remote Sens., 35, 4532–4543, doi:10.1080/01431161.2014.916049.
Yuan, X., 2004: High-wind-speed evaluation in the Southern Ocean. J. Geophys. Res., 109, D13101, doi:10.1029/2003JD004179.
Yuan, X., J. Patoux, and C. Li, 2009: Satellite-based midlatitude cyclone statistics over the Southern Ocean: 2. Tracks and surface fluxes. J. Geophys. Res., 114, D04106, doi:10.1029/2008JD010874.
Zhang, H.-M., 2006: Blended and gridded high resolution global sea surface winds from multiple satellites. 14th Conf. on Satellite Meteorology and Oceanography, Atlanta, GA, Amer. Meteor. Soc., P2.23, https://ams.confex.com/ams/Annual2006/webprogram/Paper100004.html.