1. Introduction
Global reanalysis products have provided researchers in the fields of climate change and ocean–atmosphere modeling a uniform dataset for their investigations. Over the surface of the ocean these reanalyses provide gridded fields of both standard atmospheric parameters (e.g., winds, pressure, temperature, and humidity) and fluxes [e.g., momentum (stress), latent heat, sensible heat, etc.]. The air–sea fluxes play a critical role in ocean variability, and quality estimates of the flux values are essential for numerical simulations of the ocean–atmosphere system. Herein we quantify uncertainties, over a relatively wide range of atmospheric conditions, in both air–sea fluxes and the meteorological fields used to create those fluxes from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (NCEPR; Kalnay et al. 1996) using a newly developed archive of high-quality, underway meteorological observations collected by research vessels (R/V) during the World Ocean Circulation Experiment (WOCE).
Surface flux observations over the open ocean originate from in situ or remotely sensed sources. In situ observations have been the backbone of the vast majority of air–sea flux products. The Comprehensive Ocean–Atmosphere Data Set (COADS) has been the most widely used set of surface marine observations. Value added products such as the UMW/COADS (da Silva et al. 1994) and similar collections will likely become the new standard as the importance of metadata are realized, and improved metadata are included in newly issued collections of surface marine data. The concentration of in situ data in shipping lanes (and little elsewhere) is a distinct disadvantage to employing these data in producing global products (Kent et al. 1999, manuscript submitted to J. Climate). Remotely sensed data hold great promise in providing global flux products (Schulz et al. 1997), but limited-temporal extent, limited representation of the lowest atmospheric layers, and the need for careful evaluation of the products restrict (for now) their applicability.
Global atmospheric analyses prior to the NCEPR and other reanalyses, were often plagued by biases (e.g., Siefridt et al. 1999) many of which were unknown due to the relatively poor availability of higher-quality in situ observations. Additionally, changes in operational analysis systems compounded this problem by introducing a nonstationary element to these biases, thus making it very difficult to utilize operational products for climate variability studies. Global reanalyses of the historical data using a frozen analysis or assimilation system largely eliminates the nonstationary biases due to changes in the model. Variations in the observing system leads to significant nonstationary biases (e.g., Fiorino 2000); however, the frozen assimilation system makes these products more attractive for a variety of investigations. To date there have been several reanalysis efforts. We focus on the NCEP effort because it covers the period when the WOCE data were most plentiful (i.e., 1990–95).
Quantitative evaluation of air–sea flux products is critical for identifying suitable products for flux applications (WGASF 2000). Previous attempts to qualify uncertainties in air–sea fluxes include cumulative error estimates (e.g., Gleckler and Weare 1997; Cayan 1992) and comparisons to in situ buoy observations (Weller et al. 1998).
More focused assessments of reanalyses air–sea fluxes have been completed by Bony et al. (1997) and Shinoda et al. (1999). Bony et al. (1997) found latent heat fluxes to be overestimated in regions of subsidence. Shinoda et al. (1999) compared air–sea fluxes derived from in situ, European Centre for Medium-Range Weather Forecasts, and satellite observations over the western Pacific warm pool (see Shinoda et al. 1998 for details) to the NCEPR fluxes on seasonal timescales. They found only small differences between their stress estimates and the NCEPR stress; however, the latent heat flux from the NCEPR was found to be larger than the values calculated from in situ data, especially for periods of high wind speeds.
Our technique differs from that of Shinoda et al. (1999) primarily because we use in situ data with worldwide coverage of the oceans. Our method has the advantage of allowing both global and regional evaluation of NCEPR uncertainties. In addition, we compare the fluxes and meteorological fields at synoptic (6-h) timescales by matching WOCE observations to gridded values from the NCEPR. The range of atmospheric conditions in these comparisons is relatively large. Matches are created in time and space, both horizontal and vertical, for surface fluxes and the meteorological variables used to create those fluxes. For the ship observations, adjustment to the NCEPR field heights and computation of surface fluxes is completed using a stability-dependent, bulk flux model. Once created, the meteorological and flux matches are statistically evaluated to determine uncertainties between the ship observations and the NCEPR. Uncertainties are determined for individual ships, their combined observations, by latitude, and for regions of the ocean that have dense coverage of R/V data.
The results presented reveal significant differences between the WOCE ship values and the NCEPR. NCEPR winds are underestimated across the globe, and sea level pressures are underestimated for both high and low values. Globally, latent and sensible heat fluxes are overestimated in the NCEPR. Further investigation brings the accuracy of the NCEPR flux algorithm into question.
The paper is organized as follows. A description of our method for calculating fluxes from the R/V observations and matching the R/V meteorological parameters and fluxes to the NCEPR values is in section 2. Section 3 contains a comparison of the R/V versus NCEPR surface meteorological parameters, and turbulent flux comparisons are found in section 4. Potential sources of the identified NCEPR overestimation of heat fluxes are in section 5. A summary and conclusions are presented in section 6.
2. Methodology
High-quality WOCE R/V observations are matched in space and time to NCEPR values for the period 1990–95. Matches are created for both surface fluxes and the meteorological variables needed to compute these fluxes. Fluxes are computed from ship observations using a bulk flux algorithm (Smith 1988). Due to the various timing of both the ship and NCEPR values, multiple temporal averages are used to match ship observations to the NCEPR times. The authors note that a large percentage of these high-quality ship observations were not utilized in creating the NCEPR and thereby represent pseudoindependent ground truth for assessing NCEPR fields.
a. Data
Ship meteorological data are chosen from a pool of 23 R/Vs collecting data for WOCE. Of these, only eight: A. von Humboldt, Aurora Australis, Franklin, Hakuho Maru, Heincke, Polarstern, Meteor, and Thomas G. Thompson meet our requirements for data quality and availability of necessary metadata (e.g., instrument heights). Data quality is assessed using a two-level quality control procedure (Smith et al. 1996) that includes both automated and visual inspection of the data. Wind data are especially problematic; however, previous work established guidelines and requirements for calculating and validating true winds (Smith et al. 1999). Ship data retained for comparisons include those flagged as correct, marking an interesting feature (e.g., frontal passage, extreme low pressure, etc.), or being a statistical outlier. Outliers are realistic values that fall outside four standard deviations of climatology (da Silva et al. 1994), a typical occurrence at high latitudes. Eliminating all other problematic data results in significantly better matches than when no quality control is utilized. For example, when comparing the eight select ships to the NCEPR, use of the quality control flags improves the root-mean-square (rms) error for the sea level pressure (SLP) from a value of 148.5 hPa (no flags eliminated) to 2.7 hPa.
The ship-measured parameters used for comparison to the NCEPR include air temperature, atmospheric pressure, sea temperature, moisture, zonal wind, meridional wind, and wind speed. The ship observations are first height adjusted (see section 2b) and then latent heat, sensible heat, and momentum fluxes are calculated taking into account both thermal and moisture stability conditions (Smith 1988). For 85% of the cruises, we utilize the high temporal resolution (≤15 min) data available from the automated observing systems on all eight ships. To increase the number of matches, high-quality bridge observations are used from the Polarstern and Heincke when their automated data are not available.
The 6-h NCEPR data are chosen for comparison to the selected WOCE vessel data. The NCEPR parameters evaluated include air temperature, surface temperature (a proxy over the ocean for sea surface temperature), specific humidity, zonal wind, meridional wind, SLP, sensible heat flux, latent heat flux, zonal momentum flux, meridional momentum flux and the stress magnitude. The NCEPR are on a Gaussian grid (approximately 2° × 2° resolution) with the exception of SLP which are on a 2.5° × 2.5° grid.
b. Spatial and temporal matching
Comparison of WOCE vessel and NCEPR data requires each ship value to be matched with an NCEPR value in both space and time. Horizontal spatial matching is achieved using bilinear interpolation to create an NCEPR value that corresponds to each ship location. The four NCEPR values closest to the ship's location are chosen for the interpolation. If a land value would be used as one of the four points in the interpolation, the potential NCEPR–ship match is eliminated. In addition, any potential matches where the ship is in sea ice are eliminated since the NCEPR surface temperature no longer represents a sea temperature. When in the sea ice, the ship measures a water temperature at or below the ice, while the NCEPR surface temperature represents a temperature just above the ice. The result is a NCEPR surface temperature significantly less than the sea surface temperature measured by the ship (Fig. 1). Our rejection of matches due to sea ice is based upon the NCEPR criteria for ice occurring when the surface temperature is less than −1.8°C (Kalnay et al. 1996).
Matching the NCEPR and ship atmospheric data in the vertical coordinate is accomplished by adjusting the ship data to the heights of the NCEPR gridded fields. WOCE R/V data are measured at instrument heights that range from 5 to 40 m. These heights are included in our database of R/V observations. Shipboard air temperature is adjusted to the NCEPR height of 2 m, and wind components are adjusted to 10 m. As an added complication, ships typically record different moisture values (e.g., dewpoint, relative humidity, wet bulb temperature), which are initially converted to specific humidity and then adjusted to 2 m. All height adjustments for temperature, specific humidity, and winds are accomplished by the bulk flux routine (Smith 1988). In addition, atmospheric pressure measured upon the vessels is adjusted to sea level prior to model input. No effort is made to adjust sea temperatures, which are recorded at depths ranging from 1 to 7 m.
Temporal matches to the NCEPR meteorological variables, which represent instantaneous values, are determined differently for bridge and automated data. For bridge data from the Polarstern and Heincke, the ship value closest to the NCEPR 6-h time is chosen to represent an instantaneous value for that 6-h period. Automated ship data (time resolution <30 min) are averaged within a ±15 min window centered on the corresponding NCEPR 6-h time. A sensitivity test using 1-min observations from the Meteor and an averaging window ranging from 4 to 60 min indicates no dependence of the representative 6-h ship value on the averaging period.
NCEPR flux values are not instantaneous, but instead represent a 6-h average beginning at each 6-h NCEPR time. Therefore, R/V fluxes are calculated at every observation time and then are averaged over the 6 h beginning with the NCEPR observation time. For a ship flux value to be determined over a 6-h period, 70% of the ship's data must be nonmissing values and 80% of the 6-h time span must be covered by these nonmissing data. These criteria are chosen to maximize the number of matches for which fluxes can be calculated.
3. Surface meteorology comparison
Meteorological data from 4773 ship observation times are matched to NCEPR values (Fig. 2). These matches are distributed over the global oceans with concentrations in the Arabian Sea and the Atlantic, western Pacific, and Southern Oceans. Note that not every variable was available for all of these ship times (see Table 1). Statistics (e.g., rms, bias, etc.) are calculated based upon NCEPR minus ship differences for each pair of matched surface meteorological values. These statistics are calculated for all ships combined to assess overall differences and for each ship individually to determine whether the overall differences are heavily influenced by a specific vessel. NCEPR minus ship differences are also calculated in 20° latitude bands and by ocean basins (regions outlined in Fig. 2) to determine if differences have latitudinal or regional bias. Uncertainties in the differences are calculated to determine their significance.
a. Pressure and wind speed
Rms and mean (bias) differences between NCEPR and the WOCE ship data reveal some surprisingly large mismatches (Table 1). The largest differences in the combined matches from all eight vessels occur with the sea level pressure and the wind speed. The combined pressure rms is 2.7 hPa and ranges from as high as 3.7 hPa for the Meteor down to 1.2 hPa for the Polarstern. The bias in the pressure, though negative when averaged over the eight vessels, ranges from 1.0 to −1.3 depending upon the vessel. For the wind speed, the bias for all ships is negative and rms differences range from 2.0 to 3.9 m s−1. These results indicate that the NCEPR generally underestimates both the near-surface winds and the sea level pressure.
Verification that the NCEPR is underestimating the winds is accomplished by averaging the NCEPR minus WOCE ship bias over 2.5 m s−1 bins of ship wind speed (Fig. 3). A trend toward larger biases at higher wind speeds is revealed, and this trend may be related to biases in the SLP. Plotting the pressure bias averaged over 10-hPa bins of ship pressure reveals a clear split between high and low pressures (Fig. 4). When the ship pressure is low (<990 hPa) the NCEPR values are on average 1 hPa higher than the ship pressure. The opposite is true for ship pressures greater than 990 hPa, where the NCEPR pressure averages 0.5–1.0 hPa lower than the ship value. From this, we infer NCEPR is underestimating the amplitude and/or position of highs and lows over the oceans, resulting in weaker pressure gradients and underestimated winds.
Analysis of the pressure and wind speed matches by latitude further clarifies the relationship between NCEPR and ship values. Rms differences for pressure vary from less than two hectopascals in the Tropics and midlatitudes to over four hectopascals in the polar region (Fig. 5). In addition, differences between mean NCEPR and ship pressures are largest in the tropical latitudes (Fig. 6a). From 30°N to 30°S the average NCEPR pressure is one hectopascal lower than the ship pressure. Standard errors less than 0.2 hPa for the NCEPR and ship pressure in these tropical latitudes indicate that the pressure differences are significant. The NCEPR wind speeds are less than those from the ships in all the latitude bands (Fig. 6b). Rms differences are just under 4 m s−1 in the polar latitudes and average 2 m s−1 in the Tropics, where wind speeds are generally light (Fig. 5). The standard errors for the NCEPR and ship data show the wind speed differences to be significant at all latitudes, with the exception of the north polar and southern subtropical ranges (Fig. 6b).
Regional analyses (see boxes in Fig. 1) indicate a similar pattern of pressure and wind differences (Table 2). Regardless of the region, the NCEPR wind speeds are less than those from the ship. Rms differences are largest for the more poleward regions of the Southern Ocean and the North Atlantic. Interestingly, these same two regions have pressure biases near zero while the regions with matches predominantly in the Tropics (South Atlantic, west Pacific, and Arabian Sea) show negative pressure biases similar to Fig. 6a.
Combining the latitudinal and regional variations of the wind speed and pressure with the shift in pressure biases from positive to negative when the ship pressure approaches 990 hPa (Fig. 4), suggests NCEPR is underestimating the strength of the subtropical highs. In addition, the NCEPR may not be capturing the depth or position of deep low pressure centers (Fig. 4). The underestimated NCEPR pressures may result in decreased pressure gradients which may, in part, explain the underestimation of NCEPR wind speeds.
b. Temperature and humidity
The NCEPR and ship temperatures (air and sea) and specific humidity are generally in better agreement than the winds and pressure. The sea temperature matches, after the removal of suspected sea ice values, show the smallest overall rms and bias (Table 1). The rms differences are 1.1°C and 1.3 g kg−1 for the air temperature and specific humidity, with the temperature having a negative bias on most vessels. The bias in specific humidity varies by ship from 0.6 to −0.7 g kg−1, but the overall humidity bias is positive (NCEPR too moist). The variations in the NCEPR versus ship comparisons by latitude reveals rms differences for specific humidity (air temperature) are largest (smallest) in the 30°N–30°S band, Fig. 5. This pattern is correlated to natural variability amplitudes. The biases show small differences with latitude (Figs. 6c,d); however, the pattern is significant. The temperature bias indicates NCEPR is too cold everywhere except poleward of 50°. Additionally, NCEPR humidity is generally too high in all bands except near the equator. The maximum bias is in the 10°–30°S region, a region of significant subsidence. The moist bias of the NCEPR analyses confirms findings from others (Bony et al. 1997), and narrowly defined areas may have even higher biases (e.g., Renfrew et al. 2001, hereafter RMGB).
4. Surface flux comparison
We will compare the in situ–based fluxes to the NCEPR model fluxes. Additionally, we calculate an alternative set of fluxes (labeled NCEPR/S) based on 6-h NCEPR meteorological variables (winds, temperatures, and humidity) and the flux parameterization utilized for the 30-min mean R/V observations (Smith 1988). This will help examine the role of the NCEPR flux algorithm in explaining the differences between the NCEPR and in situ–based fluxes. This calculation also provides an evaluation of the suitability of the NCEPR meteorological variables for estimating fluxes through independent flux parameterizations.
a. Wind stress
Differences between NCEPR and WOCE wind stress are consistent for most of the R/Vs (Table 3). The rms differences are smallest (less than 0.1 N m−2) between 50°N and 50°S, but increase rapidly with latitude poleward (Fig. 7). The biases are comparable across all the tested R/Vs. The NCEPR stresses underestimate at all latitudes, most significantly between 50°S and 30°N. Mean differences are near −0.01 N m−2, and poleward of 50°N they jumped fivefold (Fig. 8a). The distribution of wind stress errors across the various R/Vs as well as the spatial variations mesh well with those anticipated from examining solely the wind comparison (Fig. 6b). Additionally, the east–west wind stress components were found to have slightly higher biases and rms differences, but higher correlations to the ship values. This underestimation of stress by NCEPR extends previous conclusions based on comparisons to buoy observations in the Coupled Ocean–Atmosphere Response Experiment region (Shinoda et al. 1999). They found that NCEPR stress underestimates buoy stress in low-wind conditions, but overestimated buoy stress in higher-wind conditions. Our results indicate the underestimation occurs throughout a wide range of wind conditions.
Overall, the NCEPR/S stresses are weaker than the NCEPR stresses by 0.01 N m−2. In the Tropics the NCEPR/S stresses are of the same order as the NCEPR stresses suggesting the NCEPR flux algorithm for stress in this region cannot explain the discrepancy between the NCEPR and R/V fluxes. Evidence suggests that the weak tropical NCEPR winds result in weak NCEPR tropical stresses. Poleward of 30°, the NCEPR and NCEPR/S stresses are very different confirming that the NCEPR stress algorithm could be, in part, responsible for the differences.
b. Sensible heat flux
The amplitude of the sensible heat flux differences was surprisingly large. The rms differences were in the range of 10–30 W m−2. As expected from the NCEPR wind evaluation, the rms differences were smallest in the Tropics and increased poleward of 30° to maximums near 25 W m−2 (Fig. 7). The biases were fairly uniform with latitude (nearly 5 W m−2) except in the 50°–70°S (30°–50°N) region where the bias changed sign (increased) to −5 W m−2 (∼10 W m−2), Fig. 8b. White and da Silva (1999) also found larger NCEPR sensible heat flux values in comparison to climatological in situ values, but in contrast found relatively more agreement in the 30°–50°N region than in other latitudes. Further investigation of our results revealed the sensible heat flux bias increases with stability strength. Additionally, the biases are largest in high-wind conditions, but there is also a relative large maximum (5 W m−2) in the range of 7–9 m s−1.
The rms differences between the NCEPR/S and R/V sensible heat fluxes are smaller than those for NCEPR sensible heat fluxes (Table 4). What is more interesting (but anticipated) is the relatively large change in the biases. The sensible heat flux bias is nearly 6 W m−2 for NCEPR and less than 1 W m−2 for the NCEPR/S. The decrease in the NCEPR/S bias is more consistent with the weaker NCEPR winds and marginally cooler NCEP air temperatures. The NCEPR/S sensible heat flux compares more favorably to ship fluxes for all stability conditions. This is the case as well for nearly all wind speeds except in calm conditions where the NCEPR/S biases are larger. The correctness of the parameterizations under calm conditions is of course subject to much uncertainty.
c. Latent heat flux
For latent heat flux, the rms deviations for all R/Vs exceeded ∼25 W m−2, and averaged nearly 50 W m−2 (Table 3). The biases again indicated NCEPR overestimates latent heat flux by 10–30 W m−2. This bias is largest in the Northern Hemisphere, especially in the 10°–50°N band (Fig. 8c). Under the most stable conditions, latent heat flux is underestimated (−25 W m−2) by NCEPR. The bias in latent heat increases nearly linearly with (dry) stability [i.e., SST minus air temperature (AT)]. Additionally, this bias is increasingly pronounced with increasing wind speed, but there is some suggestion that at the highest winds (greater than 14 m s−1) the bias becomes smaller (Fig. 9). This must be further explored, as the quality of the higher wind speed data may be less than the other observations. At speeds below ∼3 m s−1 the NCEPR values underestimate the latent heat flux (Fig. 9). The overestimation of the latent heat flux (and to a lesser extent sensible heat fluxes) has been noted previously in comparisons of various reanalyses (e.g., Bony et al. 1997). The reasons were not apparent, but seemed related to subsidence and stability.
There is a dramatic change in latent heat flux biases from 17 W m−2 for NCEPR to −15 W m−2 for NCEPR/S. The NCEPR/S latent heat fluxes are too weak compared to the R/V fluxes. This negative bias for the NCEPR/S latent heat flux is consistent with the previously noted weaker NCEPR winds and higher humidity. At higher latitudes, the difference between the latent heat flux biases is less (compared to the Tropics). The changes in bias are not as easy to categorize with regards to environmental conditions. In wind conditions between 5 and 10 m s−1 (covering a significant portion of wind observations), the magnitude of the underestimation of latent heat fluxes by NCEPR/S is larger than the amplitude of the NCEPR estimates (Fig. 9). For speeds in excess of 10 m s−1, the NCEPR/S flux values still underestimate observed values, but the magnitude of the bias is smaller than NCEPR by 15 W m−2. Thus simply employing the NCEPR/S values may be attractive for higher wind speeds, but for a significant portion of the wind observations will introduce latent heat fluxes with larger (5 W m−2) biases than the original NCEPR products.
In conclusion, the NCEPR fluxes have significant biases that map into modest variations with latitude, wind speed, and stability. The use of NCEPR meteorological variables and an alternative flux parameterization resulted in improved comparisons to the observationally based sensible heat fluxes. For latent heat flux, this alternative approach results in underestimates with smaller bias magnitudes in high wind speed cases; however, for conditions near mean wind speeds the amplitude of the NCEPR/S biases are slightly larger than for NCEPR fluxes.
5. Discussion
The authors note in light of evidence from wind, temperature, and humidity comparisons that the NCEPR turbulent heat fluxes should be smaller than the R/V fluxes; however, the turbulent heat fluxes are overestimated by NCEPR. A comparison of 30-min (as opposed to the 6-h period) integrated R/V fluxes to the NCEPR flux values addresses the dependence of the flux comparison on the availability of ship observations in the 6-h averaging period. This comparison uses times coincident with the meteorological variables from the automated R/V data. The 30-min fluxes are very close to those for the 6-h periods. The rms differences and biases differ from the 6-h results by small amounts (less than 10%) for all R/Vs. In the 30°N–30°S band, the 30-min biases are nearly the same as the 6-h values, and the rms differences are smaller. Poleward of 30°, the 30-min biases for the latent heat flux become smaller, but increase for stress with increasing southern latitude. These results suggest the evaluation of NCEPR fluxes is only somewhat (∼10%) dependent on the availability of comparison data over the entire 6-hourly period.
The comparison of the 30-min fluxes does not sufficiently explain the unexpected (positive) bias in the NCEPR turbulent heat fluxes. Another possible explanation includes the role of nonlinear effects when averaging the fluxes over 30 min–6 h, though this is likely to be very small based on previous studies (Hanawa and Toba 1987). Additionally, there is disparity between the spatial areas sampled by the R/V fluxes (a sequence of point measurements) over a relatively small portion of a model grid box. Prior works show this error due to subsampling is limited to ∼10% (Esbensen and McPhaden 1996) in the Tropics.
The large changes in sensible and latent heat bias for NCEPR/S also suggests the NCEPR flux algorithm is at least partly responsible for the positive NCEPR sensible and latent heat flux biases. Recent work by others (e.g., Zeng et al. 1998) explored the algorithms in use in the NCEPR system and, using tropical observations, determined the NCEP algorithm does generate much higher sensible and latent heat flux values, particularly for strong (>6 m s−1) wind conditions. This overestimation was noted to be due to overly large temperature and humidity roughness lengths in the NCEPR model. RMGB compared reanalysis analyses and fluxes to in situ from a winter cruise in the Labrador Sea. They likewise found the NCEPR turbulent heat fluxes to be much higher than those based on observations (especially for stable conditions and in areas of moderate to high winds) and furthermore concluded the overestimation was largely due to the NCEPR flux parameterization. They note a new operational NCEP flux algorithm improves these fluxes. Finally, comparisons of NCEPR fluxes to those based on a different bulk algorithm applied to recent buoy observations in the North Atlantic also suggest the NCEPR sensible and latent heat fluxes are too large (Josey 2001).
6. Summary and conclusions
Comparisons between pseudoindependent, high-quality, underway meteorological observations collected by R/Vs during WOCE and the gridded surface fields of the NCEPR reveal some surprisingly large differences. Foremost from the meteorological variables is the underestimation of near-surface wind speeds in all regions of the globe. In addition, the surface pressures are significantly weaker in the tropical latitudes indicating an underestimation of the strength of the subtropical highs in the NCEPR. The NCEPR also underestimates the depth and/or position of deep low pressures (<990 hPa). The underestimation of SLP is likely related to the general underestimation of NCEPR wind speeds.
The other WOCE component variables (air and sea temperature, specific humidity) also deviate from the NCEPR. Of these, the most significant differences are the NCEPR temperatures (slightly lower in the Tropics) and the specific humidity (NCEPR is higher except in the 10°–30°N lat band).
Comparisons of 6-h NCEPR wind stresses and sensible and latent heat fluxes to continuous flux estimates integrated over the same time period from eight R/Vs produce quantitative evaluations of the NCEPR fluxes. The NCEPR stresses underestimate the observed stress (mean bias is −0.01 N m−2). The stress underestimation is consistent across most latitudes and wind speeds. It does appear to be larger in the northern high latitudes. The underestimation in the tropical regions is due to weak winds. In the extratropics, the underestimation of the stresses is related to both weaker NCEPR winds and other factors. The NCEPR sensible and latent heat fluxes are both surprisingly overestimated. The mean overestimate for sensible (latent) heat flux is ∼6 (20) W m−2 and is of the same sign at all latitudes. This bias increases with increasing unstable conditions. Both biases also increase with wind speed; however, results suggest the latent heat bias decreases slightly at the highest wind speeds. The latent heat bias varies significantly with latitude reaching a maximum (∼28 W m−2) in the region from 10° to 30°N. The weakness of NCEPR winds must be countered by parameterization effects to result in these large positive heat biases. As suggested by previous works, unusually large temperature and humidity roughness lengths in the NCEPR parameterizations for moderate to high wind speed conditions may be the cause. Using the NCEPR meteorological variables and an independent flux parameterization, a revised NCEPR flux improves the comparison to the observational sensible heat fluxes. However, the biases of the revised NCEPR latent heat flux change sign (NCEPR underestimates the in situ data), and also become more consistent with underestimated winds and higher humidity found in the NCEPR fields. Furthermore, while the revised latent heat flux values reduce the magnitude of the bias at higher wind speeds, they increase the bias at (more frequently occurring) moderate wind speeds and thus may not be suitable for many applications.
More rigorous comparisons are needed to elucidate the respective role of each element (parameterization, stability, observations) of this comparison and thereby lead to a clearer picture of where deficiencies may lie. Such a study should carefully consider the relative contribution of the spatial and temporal sampling. Our archive of underway surface meteorological data continues to grow as more data (e.g., non-WOCE) as well as metadata are obtained. Already we have begun processing these data to increase our comparative data holdings fourfold. With such a rich and largely independent dataset available for comparison purposes, we hope to pursue this more rigorous comparison of the R/V observational suite to several reanalysis products and expand the scope of our comparisons to possibly explore variables not discussed here (e.g., precipitation).
Acknowledgments
Support provided by the National Science Foundation for the WOCE Surface Meteorology/Flux Center, Grants OCE-9314515, OCE-9529384, and OCE-9818727. COAPS receives its base funding from the Physical Oceanography Section of the Office of Naval Research. Glenn White and an anonymous reviewer contributed helpful advice.
REFERENCES
Bony, S., Y. Sud, K. M. Lau, J. Susskind, and S. Saha, 1997: Comparison and satellite assessment of NASA/DAO and NCEP–NCAR reanalyses over tropical ocean: Atmospheric hydrology and radiation. J. Climate, 10 , 1441–1462.
Cayan, D., 1992: Variability of latent and sensible heat fluxes estimated using bulk formulae. Atmos.–Ocean, 30 , 1–42.
da Silva, A. M., C. C. Young, and S. Levitus, 1994: Algorithms and Procedures. Vol. 1, Atlas of Surface Marine Data 1994, NOAA Atlas NESDIS 6.
Esbensen, S. K., and M. J. McPhaden, 1996: Enhancement of tropical ocean evaporation and sensible heat flux by atmospheric mesoscale systems. J. Climate, 9 , 2307–2325.
Fiorino, M., 2000: The impact of the satellite observing system on low-frequency temperature variability in the ECMWF and NCEP reanalyses. Proc. Second WCRP Int. Conf. on Reanalyses, WCRP-109 (WMO/TD 985), Wokefield Park, United Kingdom, WCRP, 65–68.
Gleckler, P., and B. Weare, 1997: Uncertainties in global ocean surface heat flux climatologies derived from ship observations. J. Climate, 10 , 2764–2781.
Hanawa, K., and Y. Toba, 1987: Critical examination of estimation methods of long-term mean air–sea heat and momentum transfers. Ocean–Air Interact, 1 , 79–93.
Josey, S., 2001: A comparison of ECMWF, NCEP–NCAR, and SOC surface heat fluxes with moored buoy measurements in the subduction region of the northeast Atlantic. J. Climate, 14 , 1780–1789.
Kalnay, E., and Coauthors,. . 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc, 77 , 437–471.
Renfrew, I. A., G. W. K. Moore, P. S. Guest, and K. Bumke, 2001: A comparison of surface layer and surface turbulent-flux observations over the Labrador Sea with ECMWF analyses and NCEP reanalyses. J. Phys. Oceanogr., in press.
Schulz, J., J. Meywerk, S. Ewald, and P. Schlüssel, 1997: Evaluation of satellite-derived latent heat fluxes. J. Climate, 10 , 2782–2795.
Shinoda, T., H. H. Hendon, and J. Glick, 1998: Intraseasonal variability of surface fluxes and sea surface temperature in the tropical western Pacific and Indian Oceans. J. Climate, 11 , 1685–1702.
Shinoda, T., H. H. Hendon, and J. Glick, 1999: Intraseasonal surface fluxes in the tropical western Pacific and Indian Oceans from NCEP reanalyses. Mon. Wea. Rev, 127 , 678–693.
Siefridt, L., B. Barnier, K. Béranger, and H. Roquet, 1999: Evaluation of operational ECMWF analyses surface heat fluxes: Impact of parameterization changes during 1986–95. J. Mar. Syst, 19 , 113–135.
Smith, S. D., 1988: Coefficients for sea surface wind stress, heat flux, and wind profiles as a function of wind speed and temperature. J. Geophys. Res, 93 , 15467–15472.
Smith, S. R., C. Harvey, and D. M. Legler, 1996: Handbook of quality control procedures and methods for surface meteorology data. WOCE Rep. 141/96, COAPS Rep. 96-1, WOCE Data Assembly Center, Center for Ocean Atmospheric Prediction Studies, The Florida State University, 56 pp. [Available from COAPS, The Florida State University, Tallahassee, FL 32306-2840.].
Smith, S. R., M. A. Bourassa, and R. J. Sharp, 1999: Establishing more truth in true winds. J. Atmos. Oceanic Technol, 16 , 939–952.
Weller, R. A., M. F. Baumgartner, S. A. Josey, A. S. Fischer, and J. Kindle, 1998: Atmospheric forcing in the Arabian Sea during 1994–1995: Observations and comparisons with climatology and models. Deep-Sea Res. II, 45 , 1961–1999.
WGASF, 2000: Intercomparison and validation of ocean–atmosphere energy flux fields. Final report of the Joint WCRP/SCOR Working Group on Air–Sea Fluxes, WCRP-112, WMO/TD-No. 1035, Geneva, Switzerland, 308 pp.
White, G. H., and A. da Silva, 1999: A comparison of fluxes from the reanalyses with independent estimates. Proc. Second WCRP Int. Conf. on Reanalyses, WCRP-109 (WMO/TD 985), Wokefield Park, United Kingdom, WCRP, 123–128.
Zeng, X., M. Zhao, and R. E. Dickinson, 1998: Intercomparison of bulk aerodynamic algorithms for the computation of sea surface fluxes using TOGA COARE and TAO data. J. Climate, 11 , 2628–2644.
NCEPR sea temperature plotted vs the sea temperature measured on the Aurora Australis. Note the dramatic improvement in linear statistics when all matches with NCEPR sea temperature <−1.8°C are removed
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Map showing the location of the 4773 NCEPR − WOCE ship matches for the period 1990–95. Matches are color coded by ship according to the legend. Boxes outline matches used for regional studies of the (a) North Atlantic, (b) South Atlantic, (c) Southern, and (d) west Pacific Oceans and the (e) Arabian Sea
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Wind speed bias (NCEPR − WOCE ship) averaged in 2.5 m s−1 bins of wind speed reported by the research vessel. Biases are plotted at the center of the wind speed bins. A standard error for the biases in each bin is plotted with an error bar on the mean bias
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
SLP bias (NCEPR − WOCE ship) averaged in 10-hPa bins of pressure reported by the research vessel. Biases are plotted at the center of pressure bins. A standard error for the biases in each bin is plotted with an error bar on the mean bias
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Rms differences between NCEPR and WOCE ship values of SLP, air temperature, wind speed, and specific humidity. Rms calculated for all matches in 20° lat bins
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Mean (a) SLP, (b) wind speed, (c) air temperature, and (d) specific humidity for paired WOCE ship (solid line) and NCEPR (dotted line) observations averaged in 20° lat bins. Standard error for the ship and NCEPR values in each latitude bin are noted with bold and thin error bars, respectively. Tick marks on the x axis denote the center of the latitude bin represented by labels on either side of the tick
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Rms differences between NCEPR and WOCE ship values of latent heat, sensible heat, and momentum (stress) flux. Rms calculated for all matches in 20° lat bins
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Mean (a) momentum, (b) sensible heat, and (c) latent heat flux for paired 6-h WOCE ship (solid line) and NCEPR (dotted line) observations averaged in 20° lat bins. Standard error for the ship and NCEPR values in each latitude bin are noted with bold and thin error bars, respectively. Tick marks on the x axis denote the center of the latitude bin represented by labels on either side of the tick
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
Latent heat flux bias for NCEPR − WOCE ship (triangles) and NCEPR/S − WOCE ship (circles). Biases are averaged in 1 m s−1 bins of NCEPR wind speed. Biases are plotted at the center of the wind speed bins. A standard error for the biases in each bin is plotted with an error bar on the mean bias
Citation: Journal of Climate 14, 20; 10.1175/1520-0442(2001)014<4062:QUINRU>2.0.CO;2
NCEPR minus WOCE ship rms difference and mean bias for eight high-quality ship datasets. An rms and bias are also presented for all ships combined along with the number of matches for each variable
NCEPR minus WOCE ship rms differences and mean bias for five geographic regions. Regions were chosen for their good coverage of WOCE data (see Fig. 2)
Rms difference and mean bias between matched NCEPR and integrated 6-hourly turbulent fluxes for all R/Vs. The differences are all NCEPR − R/V. For the Heincke and Polarstern, high temporal resolution (hires) automated observations are analyzed separately from manual bridge observations
Rms difference and mean bias between NCEPR/S fluxes computed with 6-h NCEPR meteorological variables and the Smith (1988) flux parameterization and integrated 30-min fluxes for all R/Vs using the same flux parameterization. The differences are NCEPR/S − R/V. The time period is the same as in Table 3. For the Heincke and Polarstern, only the high temporal resolution (hires) automated observations are analyzed
