The Arctic Ocean has experienced an unprecedented reduction in sea ice over the last 3 decades, increasing the potential for greater exchange of gases such as carbon dioxide (CO2) between the atmosphere and the upper ocean. The present study utilizes remotely sensed data on distributions of both sea ice and chlorophyll a, together with modeled temperature and salinity fields, to obtain high-resolution basin-scale estimates of the air-sea flux of CO2 (FCO2) in the Arctic Ocean for the years 1998–2003. Concentrations of dissolved inorganic carbon (DIC) were derived from multiple linear regression relationships with sea surface temperature, salinity, and chlorophyll a. The partial pressure of carbon dioxide (pCO2) in surface waters was computed from DIC and alkalinity, the latter of which varied with salinity. FCO2 was calculated from the air-sea difference in pCO2 and wind speed. Annual FCO2 was highest in the Atlantic-dominated Greenland and Barents sectors due to their lower sea ice cover, although area-normalized FCO2 in these sectors was low. Only the Siberian sector exhibited a significant increase in annual FCO2 during the time of our study, due to a corresponding increase in ice-free water. Overall, the Arctic Ocean was a net atmospheric sink for CO2, with annual FCO2 averaging 118 ± 7 Tg C yr−1 during 1998–2003.
 Recent changes in global climate have been attributed to an increase in atmospheric carbon dioxide (CO2), a major greenhouse gas that has increased from a preindustrial value of 280 ppm to 379 ppm in 2005 [Intergovernmental Panel on Climate Change, 2007]. Current atmospheric concentrations exceed the natural range over the last 650,000 years (180–300 ppm) and have resulted in a 20% increase in radiative forcing between 1995 and 2005. Because of the powerful effects of CO2 on global climate, it is of interest to assess the magnitude of CO2 sources and sinks.
 The ocean is a significant reservoir of CO2 [Siegenthaler and Sarmiento, 1993], including of postindustrial anthropogenic CO2 [Sabine et al., 2004]. There is substantial regional variability in the rate of air-sea CO2 exchange, with the waters between 40° and 60° latitude in both hemispheres being the most significant sinks [Takahashi et al., 2002, 2009]. To date, few studies have quantified the magnitude of air-sea CO2 exchange in the Arctic Ocean due to its relative inaccessibility and perennial sea ice cover. Nevertheless, the continued loss of sea ice over the last 3 decades [Comiso et al., 2008] has enhanced interest in quantifying the capacity of the Arctic Ocean to act as a sink for atmospheric CO2.
 Recent warming and loss of sea ice impact carbon (C) dynamics in the Arctic Ocean in a number of ways. First, the increased freshwater flux from both glacial melt [Dyurgerov and Carter, 2004] and the melting sea ice, which has increased in the Arctic from 8000 km3 in 1980 to 17,000 km3 in 1997 [Peterson et al., 2006], reduces surface salinity and increases the solubility of CO2. Melting sea ice also enhances vertical stratification, reducing the entrainment of both CO2 and nutrients from depth. Second, as dense water forms and sinks to form intermediate and deep water in the Nordic seas of the Arctic Ocean [Swift and Aagaard, 1981], it carries inorganic and organic C from the surface [Broecker and Peng, 1992; Skjelvan et al., 2005], allowing large quantities of anthropogenic CO2 to penetrate to intermediate and abyssal depths [Sabine et al., 2004]. The increased freshening of Arctic Ocean could reduce deep water formation [Aagaard and Carmack, 1989], potentially reducing the transport of CO2 into the deep ocean. Finally, by increasing light availability, reduced sea ice cover can enhance primary production and the uptake and export of CO2 [Bates et al., 2006].
 Given the unprecedented reduction in Arctic Ocean sea ice cover and its potential to act as a sink for atmospheric CO2, it is imperative that we quantify the pan-Arctic air-sea CO2 flux and estimate its interannual variability. Previous estimates of the Arctic CO2 sink were based either on extrapolation of regional measurements or by mass balance calculation [Bates et al., 2006; Anderson et al., 1998]. The objective of the present study was to assess CO2 dynamics throughout the Arctic Ocean and estimate the spatial and temporal variability in the net air-sea CO2 flux. Specifically, we set out to determine whether reasonable CO2 fluxes could be calculated using the high spatial and temporal resolution data provided by satellite remote sensing platforms or from numerical models. For this purpose, we used observed chlorophyll (Chl a) and sea ice concentrations, along with modeled sea surface temperature (SST), sea surface salinity (SSS), and wind speed, to calculate the air-sea CO2 flux over the Arctic Ocean during the period 1998–2003. Because of the logistical difficulties in sampling this hostile high-latitude environment in situ, remote sensing and modeling techniques provide the greatest potential for generating pan-Arctic estimates of many important biogeochemical fluxes.
 Briefly, our approach was to first calculate surface water pCO2 daily at 9 km resolution throughout the Arctic Ocean (all waters north of the Arctic Circle, 66°33′39″) and estimate the air-sea flux of CO2 (FCO2) from the pCO2 difference between the ocean and atmosphere (ΔpCO2), the CO2 solubility in seawater, and wind speed. We calculated pCO2 from dissolved inorganic carbon (DIC) concentration and alkalinity, which were derived empirically from modeled SSS and SST [Maslowski et al., 2000, 2004] and satellite-derived Chl a. SST could have been obtained from satellite data as well, but since satellite-based estimates of SSS are not yet available, we chose to use SSS and SST obtained from the same model runs for hydrographic consistency. In the future, should satellite-based SSS data become available at the appropriate spatial resolution, all input data needed to compute FCO2 should be available from satellites.
2.1. Estimation of DIC
 A multivariate linear relationship was formulated to estimate surface DIC concentrations (μmol kg−1) from observations of Chl a, SST, and SSS. This approach is similar to that used to determine pCO2 in the Arabian Sea and the Pacific Ocean [Sarma et al., 2006]. The relation is of the form
Because the Greenland and Barents seas are dominated by Atlantic water, the relationship between SST and SSS in these regions differs from that of the rest of the Arctic, which are dominated by Pacific waters flowing in through the Bering Strait. Therefore, one set of regression coefficients for equation (1) was used to define the relationship between DIC and SST, SSS, and Chl a for the Greenland and Barents seas while a different set was used for rest of the Arctic Ocean (Table 1). Data collected in the Greenland Sea by Wallace et al.  were used to derive these coefficients for Atlantic-dominated waters (e.g., Greenland and Barents seas). The agreement between in situ and algorithm-derived DIC is good (Figure 1a), with R2 = 0.83 and a root mean squared error (RMSE) of only 33.4 μmol kg−1 (<2% of the mean) after removing four stations located near river mouths where measured DIC was much lower than that estimated by equation (1) (n = 326). Consequently, all nearshore waters adjacent to major rivers were removed from further analyses.
Table 1. Coefficients for the Best Fit of Equation (1) Derived for Atlantic- and Pacific-Dominated Arctic Waters
A0 (μmol kg−1)
A1 (μmol kg−1 °C−1)
A2 (μmol kg−1 mg−1 m3)
A3 (μmol kg−1 psu−1)
 For regions dominated by Pacific water (e.g., Chukchi, Beaufort, Baffin, Kara, Laptev, and East Siberian seas), data from the recent Shelf-Basin Interaction (SBI) project were used [Bates et al., 2006; Hill and Cota, 2005]. Because SBI data were collected in both the spring and summer, separate relationships for these two seasons were derived (Figures 1b and 1c), the coefficients for which are shown in Table 1. Although when combined, all data fall on the same trend line (Figure 1c, R2 = 0.93, RMSE = 48.7 μmol kg−1), we found that when the spring and the summer data were treated independently, the relationship between measured DIC and DIC calculated from SST, SSS, and Chl a was slightly different for the two seasons (Table 1). Therefore, we elected to use separate regressions for the spring and summer data, despite the fact that it makes very little difference to the results. The RMSE for the calculated versus the measured Pacific DIC data in summer (61.6 μmol kg−1, n = 313) was higher than that for the spring (17.3 μmol kg−1, n = 214), although the R2 was also slightly higher (0.92 and 0.90, respectively).
2.2. Estimation of Total Alkalinity
 Total alkalinity (TA, μmol kg−1) in Arctic surface waters was calculated from SST and SSS, similar to the approach used by Lee et al.  to derive TA in the Atlantic, Pacific, Indian, and Southern Oceans. In the Atlantic-dominated sectors of the Arctic Ocean, TA was calculated as
using in situ data from Wallace et al. . Again, the agreement between in situ and algorithm-derived TA is very good (Figure 2a) when stations located near river mouths were removed from the analysis (n = 326, R2 = 0.86, RMSE = 26.9 μmol kg−1). In the Pacific-dominated sectors, we used the TA relationship derived by Kaltin and Anderson 
which relates TA to SSS only (R2 = 0.92, SE = 75 μmol kg−1).
2.3. Partial Pressure of CO2
 The partial pressure of CO2 (pCO2) was calculated from DIC, TA, SST, and SSS using the iterative formulations described in the Ocean Carbon Cycle Model Intercomparison Project protocols [Najjar and Orr, 1999], which use carbonic acid dissociation constants by Mehrbach et al.  that were refitted by Dickson and Millero . The pH was computed using a combination of Newton-Raphson and Bisection iterative methods to solve for hydrogen ion concentrations.
 We performed an error analysis to determine the extent to which uncertainty in estimates of DIC, TA, SST, and SSS impact the calculation of pCO2 (Table 2). To do so, pCO2 was calculated over a range of values (±RMSE) for each of the four input variables. In the case of SSS, for which we have no estimate of RMSE, we assumed a value of 2, a value much larger than is typically found when comparing modeled and observed SSS. Because DIC and TA vary in tandem, we varied them this way in our analysis as well, i.e., when DIC was increased, TA was increased proportionally to conform to the relationship between DIC and TA observed in Arctic waters (Figure 2c). This resulted in variations in TA in our analysis (41 μmol kg−1) that were larger than the RMSE for TA in data shown in Figure 2a (27 μmol kg−1).
Table 2. Error Analysis for Calculation of pCO2 Accounting for Uncertainty in Estimates of SSS, SST, DIC, and TAa
DIC (μmol kg−1)
TA (μmol kg−1)
The range in the error estimates for each of the four input variables used to calculate pCO2 (SSS, SST, DIC, and TA) reflects variability in the other three variables within its uncertainty range (±RMSE).
Input data range
33 ± 2
0 ± 0.5
2037 ± 61
2203 ± 69
Percent error in pCO2
Input data range
33 ± 2
0 ± 0.5
2136 ± 17
2305 ± 25
Percent error in pCO2
Input data range
33 ± 2
0 ± 0.5
2048 ± 33
2214 ± 41
Percent error in pCO2
 Results of our analysis showed that uncertainty in our estimates of DIC, TA, SST, and SSS had a surprisingly small impact on the calculation of pCO2. Uncertainty in SST resulted in the smallest computed errors in pCO2. Because SST values derived from the model are in such good agreement with observations (Figure 2b), RMSE was small (0.49°C, see section 2.5 below) and varying SST by ± RMSE resulted in only a 1.2% range in pCO2 in both Pacific- and Atlantic-dominated waters. Errors in pCO2 resulting from uncertainty in SSS were only slightly higher (1.85%), reflecting the relative insensitivity of pCO2 to salinity. Uncertainty in DIC and TA together generated the largest errors in pCO2, which ranged from a low of 1.5% in Pacific waters in spring to 5.7% in Pacific waters in summer. This difference is due to the larger RMSE in DIC data from the Pacific in summer. When uncertainties in all four variables are combined in a single analysis, errors in pCO2 range from 4.6% in Pacific waters in spring to 9.1% in Pacific waters in summer (the pCO2 error in Atlantic waters is 4.6%).
 Our conclusion from this analysis is that the calculation of pCO2 from DIC, TA, SST, and SSS is prone to relatively small errors, due primarily to the large data sets used to generate the relationships given in equations (1)–(3) that resulted in a small RMSE. It must be noted, however, that while the data sets we used to derive our statistical descriptions of DIC and TA were relatively large, their spatial distribution is somewhat restricted, coming mainly from the Beaufort, Chukchi, and Greenland seas. Errors may increase when data from other locations are included in the analysis.
2.4. Air-Sea CO2 Flux
 For ice-free pixels, the net air-sea CO2 flux (FCO2) was determined from the relationship
where k is the gas transfer (or piston) velocity, σ is the CO2 solubility in seawater, which is computed as a function of temperature and salinity [Weiss and Price, 1980], and ΔpCO2 is the air-sea difference in pCO2. Because FCO2 scales linearly with ΔpCO2, estimates of FCO2 have the same error associated with them as do estimates of seawater pCO2 (4–9%). Atmospheric pCO2 was set at 380 μatm for this study. The sign of ΔpCO2 determines the direction of CO2 exchange; negative values indicate a flux of CO2 from the atmosphere into the ocean.
where W10 is the wind speed (m s−1) computed from daily mean wind fields at 10 m above sea level obtained from the NCEP/NCAR Reanalysis project [Kalnay et al., 1996] and Sc is the Schmidt number, which varies as a function of temperature
In the original formulation of Wanninkhof , the value for γ was set at 0.39. However, a recent analysis by Sweeney et al.  using a more complete bomb radiocarbon data set determined a new value for γ of 0.27, the value we have adopted here. Sweeney et al.  showed that this value was valid for both short-term and long-term averaged winds and that the resulting calculation of the gas transfer velocity had an error of ±30%, accounting for errors in NCEP wind speeds, bomb radiocarbon, and their inversion technique. Taking into account errors in estimates of TA, DIC, and gas transfer velocity, we estimate that the errors on our estimate of FCO2 are on the order of ±35%.
2.5. Arctic-Wide Application
 To calculate FCO2 over the entire Arctic Ocean (all waters north of the Arctic Circle, 66°33′39″), data for SSS, SST, and Chl a are required at high temporal and spatial resolution over the entire basin. Satellite data best fulfills this criterion and are available for both Chl a and SST. However, no satellite-based SSS data are yet available. Therefore, the next best option was to use SSS and SST fields derived from a coupled ice-ocean circulation model of the Arctic Ocean [Maslowski et al., 2000, 2004]. This model has a horizontal resolution of 9 km with 45 vertical layers and has been well validated for the Arctic Ocean using observations of SST, buoy-derived sea ice motion (International Arctic buoy data program) and satellite-derived sea ice concentration [Maslowski et al., 2000, 2004; Steiner et al., 2004]. Both SSS and SST were obtained for model runs that simulated Arctic Ocean dynamics for the years 1998–2003. Modeled SST fields during this time are in good agreement with corresponding satellite-based SST estimates made at the same time and location (n = 253145, R2 = 0.98, RMSE = 0.49°C, Figure 2b).
 To calculate the CO2 flux for the Arctic Ocean, only open water areas (ice free water) were considered. Open water area was determined from Special Sensor Microwave Imager (SSM/I) 37 and 85 GHz bands using the Polynya Signature Simulation Method (PSSM) algorithm [Markus and Burns, 1995], which allows determination of sea ice presence/absence at 6.25 km resolution. A given pixel is defined as being ice covered if the sea ice concentration is greater than approximately 10%.
 All input data had a temporal resolution of 1 day, with the exception of SeaWiFS Chl a, which were 8 day means for improved spatial coverage (to compensate for heavy cloud cover). Output fields (DIC, alkalinity, pCO2, and FCO2) were computed on a daily basis and projected onto a common polar stereographic grid. All data were processed using Interactive Data Language (IDL) and all analysis and visualization was done using Matlab.
2.6. Defining Regions of Interest
 For the purpose of characterizing regional differences in Chl a, DIC, pCO2, and FCO2, we divided the Arctic Ocean into eight geographic sectors demarcated by longitude (Figure 3a) as defined by Pabi et al. . These include the Chukchi (180°–160°W), Beaufort (160°W–100°W), Baffin (100°W–45°W), Greenland (45°W–15°E), Barents (15°E–55°E), Kara (55°E–105°E), Laptev (105°E–150°E), and East Siberian (150°E–180°) sectors.
3. Algorithm Input Fields
3.1. Sea Surface Salinity
 The large-scale spatial distribution of SSS in the Arctic Ocean is controlled by the balance between freshwater influx from river discharge and sea ice melt and the inflow of higher-salinity waters from the Atlantic Ocean (∼34.8) through the Fram and Nares straits as well as by Pacific waters (∼32.5) entering through the Bering Strait (Figure 3b). The surface waters above the deep central Arctic basin exhibit relatively low SSS, ranging from 28.5 to 32.5 and reflecting relatively high rates of seasonal sea ice melt and the dominating influence of low-salinity Pacific waters. SSS over the continental shelf is also highly influenced by the large number of rivers that drain into the Arctic Ocean; the effect of these freshwater inputs can be seen as low-salinity regions (21–26) along the coast of the Beaufort, Laptev, Kara, and East Siberian seas (Figure 3b). In the Greenland and Barents sectors, SSS is much higher (33.5–34.8) than elsewhere in the Arctic Ocean due to the advection of large quantities of high-salinity Atlantic water into these sectors. Eventually this water gets diluted as it moves deeper into the Arctic Ocean and mixes with the low-salinity waters of the central basin and continental shelf.
 Over an annual cycle, SSS is highest during the months of April–May and then falls to its minimum value during the August–September period, coinciding with the time of peak open water area in the Arctic Ocean. Seasonal variability is highest in the Pacific-influenced waters due to a more variable river flow compared to that in the Atlantic-influenced sectors. During the period 1998–2003, the annual mean SSS of the Arctic Ocean decreased at the rate of 0.01 yr−1 (R2 = 0.69, p = 0.04).
3.2. Sea Surface Temperature
 The SST of the Greenland and the Barents sectors of the Arctic Ocean is dominated by the influx of warm Atlantic Ocean water, as evidenced by its elevated SST (Figure 3c). Annual mean SST in these waters ranges from 3.5 to 4.0°C and is at least 3°C warmer than the rest of the Arctic Ocean. SST over the deep central Arctic basin remains near the freezing point of seawater throughout the year due to its perennial sea ice cover. Waters are somewhat warmer over the continental shelves (−1°C), which are exposed to solar radiation when the annual sea ice melts in the spring and summer. Warm waters of 0.5–1.5°C can be found in the southern Chukchi Sea and in southern Baffin Bay.
 The seasonal pattern of SST is opposite that of SSS, with the maximum temperature (and lowest SSS) found during the months of August–September. During the period 1998–2003 there was a significant (R2 = 0.7, p = 0.03) increase in SST in the Arctic Ocean, which rose at the rate of 0.04°C yr−1.
3.3. Chlorophyll a
 A detailed description of Chl a dynamics in the Arctic Ocean is given by Pabi et al. . In short, the spatial distribution of surface Chl a is marked by high values on the continental shelf, which range from 0.7 to > 3.0 mg m−3 (Figure 3d), reflecting higher nutrient concentrations and rates of primary production. In the deep central Arctic Ocean, Chl a concentration generally ranges from 0.05 to 0.3 mg m−3.
 The annual cycle of Chl a is characterized by an initial spring bloom in April–May, and in some years a subsequent summer bloom during July–August. Between these two blooms, mean surface Chl a concentrations in the Arctic Ocean remain relatively high, generally exceeding 1.5 mg Chl a m−3. There is no significant secular trend in Chl a in any of the sectors or in the pan-Arctic basin during the period 1998–2003.
 The highest mean annual wind speeds are associated with the Greenland and Barents sectors, which range from 7 to 9 m s−1 and exhibit spatial patterns similar to that of SST (Figure 3e). Elsewhere over the Arctic Ocean, mean annual wind speed varies from 5 to 6 m s−1, except for a relatively large region around northwestern Greenland, where wind speeds average approximately 4 m s−1.
 The annual cycle in wind speed is similar Arctic-wide, with wind speeds peaking (∼5 m s−1) during October–March and decreasing in intensity from April to September (3.7 m s−1). Seasonal variability in wind speed is highest in the Barents and Greenland sectors, with annual amplitudes of 5.5 and 4.7 m s−1, respectively. Variability in the other sectors ranged from 2 to 4 m s−1, with the lowest variability observed in the Laptev sector.
3.5. Open Water Area
 The sectors with greatest amount of open water area during 1998–2003 were the Atlantic-influenced sectors of the Barents and the Greenland seas (Figure 3f) where open water persisted for >320 d yr−1, except for the northeastern Greenland coast and northern Barents sector. In the Chukchi and Baffin Bay sectors, open water lasts for approximately 160–200 d yr−1. Open water area is greatest during the months August–September, lagging the SST peak by about 20 days. A more detailed description of the changes in open water area in the Arctic Ocean is given by Pabi et al. . There was no secular trend in open water area in the Arctic Ocean during 1998–2003.
 Although spatial maps of DIC, pCO2 and FCO2 were all produced at 1 day temporal resolution, for the sake of simplicity and to save space, we have elected to describe spatial patterns in these quantities using annual means, just as we did above for algorithm input parameters.
 The spatial pattern of mean annual surface DIC concentrations calculated using equation (1) exhibits a great deal of variability within the Arctic Ocean (Figure 4a). Mean annual surface DIC was highest in the Atlantic-influenced waters of the Greenland and Barents sectors (salty, warm waters shown in Figures 3b and 3c), ranging from approximately 2200 to 2240 μmol kg−1. Areas of relatively high DIC concentration (>2100 μmol kg−1) also were found in the Baffin and Beaufort sectors, and in the southern Chukchi. On average, however, DIC values in these sectors were substantially lower than in regions dominated by Atlantic water. Surface DIC concentration also was lower in areas influenced by freshwater sources from rivers, consistent with observations made in the Beaufort Gyre by Hansell et al. . The lowest surface DIC concentrations were found in the northern Chukchi, East Siberian, Laptev, and coastal regions of the Kara sectors (Figure 4a). These DIC distributions are in good agreement with in situ measurements made previously in these waters [Bates, 2006; Bates et al., 2006; Miller et al., 1999; Kaltin et al., 2002; Omar et al., 2003].
 The temporal trend in surface DIC concentration is similar for all Arctic Ocean sectors during the April–September open water season (Figure 5). The highest values for surface DIC are found in the spring, particularly in the deep basins, due to entrainment of DIC-rich deep waters into the surface layer during deep winter mixing. As the season progresses, increased light availability stimulates phytoplankton photosynthesis and DIC is drawn down markedly. At the same time, higher surface temperatures increase surface water pCO2 and accelerate the loss of CO2 from the ocean to the atmosphere, diminishing surface DIC. In sectors with high biological production that were not heavily impacted by river runoff, such as the southern Chukchi, DIC fell from a spring maximum of 2100 μmol kg−1 to 1850 μmol kg−1 in August (Figure 5). In river-dominated sectors such as the Beaufort, Laptev, Kara, and East Siberian, the seasonal amplitude in DIC is further exaggerated by dilution of surface waters by low-DIC water from rivers, which reach peak discharges in spring and early summer [Pavelsky and Smith, 2004]. In these sectors the seasonal amplitude of DIC ranged from 500 to 700 μmol kg−1 (Figure 5).
 By late summer and early autumn, surface DIC concentrations begin to increase again (Figure 5) as surface waters cool, wind speeds increase, and mixed layers begin to deepen, entraining higher DIC water into the surface. This is most easily seen in the Greenland sector, due to its relatively long open water season (Figure 3f). As sea ice begins to form in the other Arctic Ocean sectors, salinization of surface waters aids in vertical mixing and restores surface DIC concentrations to the early spring values. However, because of the large freshwater fluxes in many regions of the Arctic (due primarily to river flow), surface waters remain stratified for much of the year [Harms et al., 2000; Dmitrenko et al., 2001], resulting in little entrainment of DIC-rich deep waters into the surface layer and relatively low DIC concentrations in surface waters that persist throughout the winter (e.g., Laptev, East Siberian).
 The temporal trends in DIC reported here (Figure 5) are consistent with previous measurements of DIC in Arctic waters. Our estimates of surface DIC concentration for the Chukchi sector vary from 2073 to 2166 μmol kg−1 in spring and from 997 to 1930 μmol kg−1 in summer, similar to measurements made by Bates  for the same seasons (2100–2200 μmol kg−1 in spring and 526–1900 μmol kg−1 in summer). In the Barents sector, we estimated a mean DIC concentration in July of 1998–2003 of 2105 μmol kg−1, with a value of 2130 μmol kg−1 in July of 1999. This is in excellent agreement with the July 1999 measurements of DIC in the Barents Sea of 2136 μmol kg−1 by Kaltin et al. . In the Beaufort sector, measurements of DIC made during September of 2008 were within 3% of our estimate of 2044 ± 42 μmol kg−1 averaged over the same time period (H. Thomas, personal communication, 2010). Finally, in the Greenland sector during the period 1993–1995, in situ DIC concentrations dropped from a mean of 2143 μmol kg−1 in spring to 2064 μmol kg−1 in summer, but remained as high as 2100 μmol kg−1 in several regions [Miller et al., 1999]. In our study, the average DIC concentration in the Greenland sector during the spring and summer of 1998–2003 was similar at 2186 ± 23 μmol kg−1 and 2119 ± 23 μmol kg−1, respectively. The slightly higher values in our study for the Greenland sector may be due to an increase in the influx of DIC from the atmosphere since 1995 [Sabine et al. 2004]. A similar anthropogenic increase in DIC was reported for the Barents sector by Omar et al.  between the years 1967 and 2000–2001.
4.2. Partial Pressure of CO2
 The spatial pattern of surface water partial pressure of CO2 (pCO2) (Figure 4b) broadly mimics the spatial pattern of surface DIC (Figure 4a), with higher pCO2 in the regions dominated by Atlantic waters and relatively lower values in the other geographic sectors. This spatial pattern is consistent with observations by Kaltin and Anderson  which showed that the flux of atmospheric CO2 into the Atlantic-influenced Barents Sea was about half that in the Chukchi Sea. The relatively high pCO2 in surface waters closer to the deep central basin is due to a combination of deep mixing entraining high DIC and reduced surface alkalinity from dilution by sea ice melt.
 In the Chukchi sector, surface pCO2 was higher over the central Chukchi shelves than in coastal waters of the western Beaufort Sea (Figure 4b), similar to the spatial pattern reported by Bates . Also in agreement with Bates et al. , we estimate low pCO2 (Figure 4b) along the coast in the vicinity of freshwater discharge from the Mackenzie River. In the Greenland sector, the annual mean pCO2 was 313 ± 4.1 μatm during 1998–2003 (Table 3). Measurements archived by Takahashi et al.  suggest that the mean surface pCO2 measured in this region was 282 ± 31 μatm in 1995. The relatively higher values of pCO2 in our study are most likely due to equilibration of the ocean water in this region with increasing atmospheric CO2 values between 1995 and 2003. This explanation is supported by the observed 28 ± 21 μatm increase in surface water pCO2 from 1994 to 2001 in the Greenland Sea reported by Nakaoka et al. .
Table 3. Annual Mean Partial Pressure of CO2 (pCO2) in Surface Waters (μatm)a
Uncertainty in the calculation of pCO2 from SSS, SST, TA, and DIC is 5-9%.
 The temporal trend of pCO2 (Figure 6) also is similar to that of DIC (Figure 5), with high pCO2 in early spring that decreases throughout the spring and summer. Also like DIC, pCO2 begins to increase in early autumn in several sectors of the Arctic Ocean, including the Greenland, East Siberian, and Chukchi. The correspondence between the temporal trends in DIC and pCO2 demonstrates that pCO2 in these Arctic waters during spring and summer is controlled more by decreases in DIC concentration than by decreases in CO2 solubility (which would increase pCO2) that result from increasing temperatures. The relatively small impact of SST on pCO2 is probably due to low seasonal variability of SST in these polar waters compared to the large seasonal changes in DIC.
 In the Pacific-dominated waters of the Chukchi Sea, we calculated surface pCO2 values ranging from 346 ± 132 μatm in May–June and dropping to 222 ± 152 μatm in July–August and 209 ± 125 μatm in September (Figure 6). Similarly, Bates et al.  reported high pCO2 in the Barrow Canyon region and western Beaufort Sea during May–June (∼300–350 μatm) that fell to 180–220 μatm in July–August. Also sampling the Chukchi along the Central Channel west of Cape Lisburne to Barrow Canyon, both Pipko et al.  and Murata and Takizawa  reported somewhat higher pCO2 levels in late September, ranging from 280 to 320 μatm and from 290 to 350 μatm. However, Murata and Takizawa  reported lower pCO2 values in this region in September of 1999 (240–280 μatm) and 2000 (180–220 μatm), similar to values calculated here. In the Atlantic-dominated waters of the Greenland sector, our results show that pCO2 is highest in April (335 ± 1 μatm) and decreases thereafter until reaching its seasonal minimum (297 ± 11 μatm) in July–August (Figure 6). This seasonal difference of 38 μatm agrees well with the range of 42 μatm calculated from observed seasonal pCO2 changes in the Greenland Sea between April and July [Takahashi et al., 2002].
4.3. Air-Sea Flux of CO2
 Air-sea flux of CO2 (FCO2) is primarily controlled by ΔpCO2 as well as winds that drive turbulence in the marine atmospheric boundary layer. Consequently, the spatial pattern of mean daily FCO2 (Figure 4c) was broadly similar to that of mean annual pCO2 (Figure 4b). The Pacific-influenced waters exhibited a higher FCO2 from the atmosphere into the ocean (more negative FCO2) than the other sectors. It is notable that the mean FCO2 was negative in most places, indicating that the Arctic Ocean is on average a net sink of atmospheric CO2.
 The temporal pattern of FCO2 illustrates that the Pacific-dominated sectors experience outgassing in spring (Figure 7). This is most likely due to a combination of high pCO2 conditions under the retreating sea ice (Figure 6) coupled with high wind speeds and low rates of primary production (limited by light). In the Pacific-dominated sectors, the influx of Pacific waters supersaturated in CO2 [Kaltin and Anderson, 2005] also enhances outgassing of CO2 in these regions. During summer and early autumn, the increase in both freshwater content and primary production in ice-free waters leads to depletion of CO2 in the surface layers [Kaltin and Anderson, 2005]; consequently, the flux of CO2 from the atmosphere into the surface ocean increases.
 In contrast, no outgassing of CO2 was observed during spring in the Atlantic-dominated Barents and Greenland sectors (Figure 7) since surface waters in these sectors remained undersaturated with CO2 throughout the year, i.e., surface water pCO2 was below atmospheric levels (Figure 6). Unlike the other sectors, the air-sea flux of CO2 into surface waters in the Barents and Greenland sectors decreased slightly during the months of June and July (Figure 7). This is due to a steep decrease in wind speed in these sectors combined with low seasonal variability of pCO2 (280–340 μatm) relative to other sectors. However, the flux of CO2 into the ocean intensified in September as the wind speed increased at a time when pCO2 in these waters was still quite low. This temporal pattern matches that derived by Nakaoka et al.  for the Greenland and the Barents Sea, who also emphasized the effect of wind speed on the air-sea CO2 flux. The pan-Arctic FCO2 trend is heavily weighed by that of the relatively large Barents and Greenland sectors and consequently looks similar to that of these sectors.
 In the Chukchi sector, the temporal trend in FCO2 produced by our algorithm is consistent with that reported by Bates  for the same region. Bates  measured values for FCO2 in May–June that ranged from < −0.1 to −1.0 mmol CO2 m−2 d−1; in our study we estimated FCO2 in the Chukchi to be −0.8 ± 0.7 mmol CO2 m−2 d−1. During July–August, the FCO2 reported by Bates  ranged from −30 to −90 mmol CO2 m−2 d−1 in the Central Valley, Hanna Valley, and Barrow Canyon and < −2.0 mmol CO2 m−2 d−1 over the Chukchi Sea slope. Here we calculated a spatial mean FCO2 of −9.1 ± 7 mmol CO2 m−2 d−1 for the Chukchi sector during July–August, with a maximum FCO2 of −21 mmol CO2 m−2 d−1 in mid-August, consistent with estimates of Bates . Our mean September FCO2 of −10 ± 7 mmol CO2 m−2 d−1 also agrees with the values observed by Murata and Takizawa , who reported FCO2 ranging from −17 mmol CO2 m−2 d−1 over the shelves to −11 mmol CO2 m−2 d−1 on the slopes. In the Greenland sector, the FCO2 calculated by Takahashi et al.  ranged from −16 to −48 mol m−2 month−1 during March–August, in agreement with our estimate of monthly FCO2 (−20 mol m−2 month−1) for the Greenland sector during the same period.
4.4. Annual FCO2
 The annual FCO2 was negative in almost all regions of the Arctic Ocean, indicating that the Arctic Ocean is a net sink of atmospheric CO2 (Figure 4d). Annual FCO2 was most negative in the Greenland and the Barents sectors because of their relatively low pCO2 throughout the year (Figure 6) and their longer open water season (>320 d yr−1) (Figure 3f). In general, the annual FCO2 from the atmosphere into the ocean was lower (i.e., less negative) in waters remote from the continental shelves (Figure 4d). This was due to higher DIC concentrations in these waters (Figure 4a) resulting from deeper mixing as well as to increased pCO2 (Figure 4b) as a result of lower alkalinity caused by melting of sea ice.
 Annual FCO2 over the entire Barents sector averaged −26 ± 10 g C m−2 yr−1, somewhat less than previous estimates of −46 ± 27 g C m−2 yr−1 made by Nakaoka et al.  and −44 ± 10 g C m−2 yr−1 by Fransson et al. , although these studies covered a much smaller spatial extent. In the Greenland sector, we estimated annual FCO2 to be −27 ± 10 g C m−2 yr−1, which compares well with the estimate by Takahashi et al. , who calculated an annual FCO2 of −25 to −50 g C m−2 yr−1 (after correction using 10 m winds from http://www.ldeo.columbia.edu/res/pi/CO2). After these sectors, the highest FCO2 was in the Chukchi sector (−19 g C m−2 yr−1) while the lowest annual FCO2 per unit area was calculated for the Kara (−12 g C m−2 yr−1) and Siberian sectors (−13 g C m−2 yr−1).
 The spatially integrated annual FCO2 in the combined Atlantic-dominated Barents and Greenland sectors was −62 ± 5 Tg C (1012 g C) yr−1 (Table 4), somewhat below the value of 90 Tg C yr−1 reported by Skjelvan et al. [1999a, 1999b]. Of this, the Greenland sector accounted for −38.0 ± 2.8 Tg C yr−1 and the Barents sector another −24 ± 2.8 Tg C yr−1 (Table 4). These two sectors accounted for the largest fraction of spatially integrated annual FCO2 in the Arctic Ocean. Annual FCO2 in the Barents sector peaked in 2000 (28.2 Tg C yr−1), the year exhibiting the maximum open water area (1,207,785 km2, Table 5) as well as the highest mean annual wind speed (5.4 m s−1, Table 6) of the 6 year time series. In the Greenland sector, annual FCO2 was greatest in 1999 (−41.6 Tg C yr−1), which, while not the year of highest open water area (open water area was 1,665,777 km2 in 1999, slightly less than the maximum of 1,677,023 km2 attained in 2003), was the year having both the highest wind speed (5.8 m s−1) and lowest surface water pCO2 (308 μatm, Table 3).
Table 4. Spatially Integrated Annual FCO2 (Tg C yr−1)a
Negative sign denotes a flux from the atmosphere into the ocean. Uncertainty in the calculation of FCO2 from pCO2 using NCEP winds is approximately 35%.
Table 5. Annual Mean Open Water Area (106 km2)
Table 6. Annual Mean Wind Speed (m s−1)
 In the Chukchi sector, we computed a mean annual FCO2 of −9.0 ± 1.1 Tg C yr−1 (Table 4), about one fourth of the total flux of −38 ± 7 Tg C yr−1 reported by Bates . However, the Chukchi Sea boundaries used by Bates  encompassed an area threefold larger (595,000 km2) than those used in the present study (192,887 ± 39,639 km2). Correcting for this difference in areas brings the two estimates much more in line. The highest flux in the Chukchi Sea was in 2003 (−10.0 Tg C yr−1, Table 4), which coincided with the year of maximum open water area during 1998–2003 (Table 5).
 After the Greenland and Barents sectors, the highest spatially integrated annual FCO2 was in the Kara sector (−12.4 ± 3.5 Tg C yr−1), followed by the Baffin (−9.7 ± 1.8 Tg C yr−1) and Laptev (−9.6 ± 2.5 Tg C yr−1) sectors (Table 4). Although the Baffin sector had more open water area than the Kara (Baffin: 415,973 km2, Kara: 268,962 km2, Table 5), the higher FCO2 in the Kara was due to lower pCO2 values (Baffin: 310 ± 9 μatm, Kara: 294 ± 21 μatm, Table 3) and higher wind speed (Baffin: 3.8 m s−1, Kara: 4.2 m s−1, Table 6). The spatially integrated annual FCO2 in the Beaufort sector (−9.1 ± 2.3 Tg C yr−1) was similar to that of the Laptev (Table 4). Although both of these sectors had open water areas comparable to the Kara (Kara: 268,962 km2, Laptev: 222,143 km2, Beaufort: 247,136 km2, Table 5), and comparable pCO2 (Kara: 294 ± 21 μatm, Laptev: 210 ± 18 μatm, Beaufort: 254 ± 16 μatm, Table 3), these sectors also had lower annual wind speed than Kara (Kara: 4.2 m s−1, Laptev: 3.4 m s−1, Beaufort: 3.8 m s−1, Table 6). The lowest spatially integrated annual FCO2 was in the Siberian sector (−5.6 ± 2.0 Tg C yr−1), which had the lowest open water area among all Arctic sectors (159,593 km2). However, the last 2 years of the time series (2002–2003) showed a sharp increase in spatially integrated annual FCO2 in this sector (Figure 8) due to a similar sharp increase in open water area (Table 5).
 The spatially integrated annual FCO2 for the pan-Arctic Ocean (north of 66°N) was calculated here to be −118 ± 7 Tg C yr−1. This is much lower than the estimate of −310 Tg C yr−1 for waters north of 50°N by Takahashi et al. . However, the Arctic Ocean estimate by Takahashi et al.  included the highly productive Bering Sea and North Atlantic, which were not part of our study. On the other hand, our value is higher than that of Bates , who estimated pan-Arctic FCO2 to be −66 Tg C yr−1. However, given that annual FCO2 for just the Barents and the Greenland Seas has been estimated at −90 Tg C yr−1 [Skjelvan et al., 1999a, 1999b], the pan-Arctic estimate by Bates  may be on the low side.
 Our basin wide estimate of FCO2 for the Arctic Ocean of 118 ± 7 Tg C yr−1 is 20–25% of recent estimates of net primary production over the same region [Pabi et al., 2008; Arrigo et al., 2008] and represents 6–8% of the global net air-sea CO2 flux [Sweeney et al., 2007; Takahashi et al., 2009]. Given that the Arctic Ocean (here north of ∼66°N) occupies only 2.6% of the total area of world's ocean and <1% of the volume [Jakobsson et al., 2004], this flux is significant. With the Arctic Ocean expected to become ice free in the summer sometime this century [Holland et al., 2006], its role as a sink for atmospheric CO2 could increase, particularly since recent declines in sea ice have been accompanied by an increase in primary production [Pabi et al., 2008; Arrigo et al., 2008].
 There are a number of pathways that control the exchange of CO2 between the Arctic Ocean and its surroundings. CO2 in the surface ocean exchanges with the atmosphere via direct air-sea flux as discussed in this study; it also enters the surface waters as a result of the entrainment of CO2 from the deep ocean. CO2 is removed from the surface ocean by photosynthesis and subsequent conversion to organic C. Part of this biogenic C sinks below the surface ocean and is exported to depth. A substantial amount of C also leaves the surface ocean via deep water formation, as dense waters sink and carry dissolved organic and inorganic C with them. Surprisingly, the horizontal transport of this deep ocean C between Arctic basins often is larger than the flux of C by all the other mechanisms. For, example, Skjelvan et al.  summarized C transport via these mechanisms for the Nordic Ocean (Greenland and Barents Sea). They estimated that 140 Tg C yr−1 entered the surface water, 90 Tg C yr−1 of which was via air-sea flux and 50 Tg C yr−1 from entrainment of deep waters high in DIC. In total, 170 Tg C yr−1 was calculated to exit these surface waters toward the deep ocean, including fluxes of 120 Tg C yr−1 from deep water formation and 50 Tg C yr−1 via export production of particulate organic C (the difference between C entering and exiting surface waters is due to the large uncertainty in the export production estimate). The net lateral transport of C from the deep Nordic seas was calculated to be 100 Tg C yr−1 [Skjelvan et al., 2005].
 The future capacity of the polar oceans to take up atmospheric CO2 is currently a matter of debate. Some studies suggest that this capacity is decreasing [Skjelvan et al., 2005; Le Quere et al., 2007] as surface waters equilibrate with increasingly high atmospheric CO2 concentrations. For the Arctic Ocean, these conclusions are based on the long-term assessment of annual surface fugacity of CO2 [Lefèvre et al., 2004] as well as from time series measurements of DIC in the Norwegian sea [Skjelvan et al., 2008]. On the other hand, Bates et al. , report an increase in CO2 uptake capability based on the reduction in sea ice extent observed over the last 3 decades. In our 5 year study, no temporal trend in DIC or pCO2 was found, but given that the increase in open water area in the East Siberian sector was associated with increasing annual FCO2, we can speculate that an Arctic-wide increase in open water area may lead to an increase in FCO2 into the ocean. This is consistent with recent studies showing that biological removal of CO2 from Arctic Ocean surface waters increased by 26% between 2003 and 2007 associated with the dramatic loss of sea ice [Arrigo et al., 2008].
 However, increased CO2 uptake by the Arctic Ocean will not continue if rapidly melting sea ice and large influxes of freshwater from rivers reduce the salinity enough to adversely affect the buffering capacity of the Arctic Ocean. Increased freshwater flux also can reduce deepwater formation, weakening the meridional circulation and reducing the deep transport of C between the Arctic and other ocean basins [Aagaard and Carmack, 1989]. Also, increased inflow of terrigenous DOC from these rivers can impede primary production by reducing light availability, which will reduce CO2 uptake and increase DIC. Remineralization and photooxidation of terrigenous DOC will further increase the DIC pool within the Arctic Ocean basin, possibly reducing its capacity to act as a sink for atmospheric CO2. However, most of the terrigenous DOC flowing into the Arctic Ocean is refractory and is remineralized to CO2 on very long timescales [Amon and Benner, 2003].
 Finally, if the wind speeds were to increase over the Arctic Ocean, as is predicted for the Antarctic [Russell et al., 2006], then changes in wind-driven mixing also could have pronounced effects on the air-sea flux of CO2. Large mixing events caused by intense storms have been reported to reach not only the halocline, but the deeper thermocline as well [Yang et al., 2004]. Thus, increases in wind speed or an increase in the frequency of storms, as has been witnessed in Alaska [Arctic Climate Impact Assessment, 2005], should result in entrainment of CO2 from deep waters while at the same time, increasing the flux of nutrients into surface waters and enhancing phytoplankton production [Pickart et al., 2009]. The balance between these two processes will determine the extent to which changes in the wind field will impact the ability of the Arctic Ocean to act as a sink for atmospheric CO2.
 While no temporal change in annual FCO2 was observed in the Arctic over the relatively short time period of our 5 year study, it is important to continue to monitor changes in FCO2 in light of the many dramatic changes happening in the Arctic environment. Moreover, as significant trends in annual primary production related to changes in open water area have been observed during the years 1998–2007 [Pabi et al., 2008; Arrigo et al., 2008], it is likely that a similar temporal trend in FCO2 will emerge. This study sets the path for future work in this direction by not only quantifying the interannual flux of CO2 into the Arctic Ocean, but also providing a framework for future assessments in this biogeochemically important, but largely inaccessible, region using remotely sensed data from satellites. Moreover, the application of this work will become even more relevant after the launch of two future salinity-measuring satellites, Aquarius from the National Aeronautics and Space Administration (NASA) that is to be launched in 2011 and the Soil Moisture and Ocean Salinity (SMOS) sensor from European Space Agency (ESA) that was launched on 2 November 2009. Although the data from these platforms will be relatively coarse (50–100 km resolution), they will provide a better understanding of sea surface salinity distributions and temporal variability that is critical to estimates of surface water pCO2 and FCO2.
 We thank all the members of the Arrigo research group for their helpful discussions and suggestions to improve the manuscript. We also thank the Shelf Basin Interaction (SBI) project for providing some of the data used to develop the algorithms presented in this paper. This research was supported by NASA grant NNG05GC92G to K. Arrigo.