Journal of Geophysical Research: Oceans

Primary productivity in the Arctic Ocean: Impacts of complex optical properties and subsurface chlorophyll maxima on large-scale estimates

Authors


Abstract

[1] The Arctic Ocean is an optically complex environment and presents unique challenges for ocean color satellite remote sensing. Phytoplankton pigment packaging, high concentrations of chromophoric dissolved organic matter (CDOM), and the frequent presence of subsurface chlorophyll a (Chl a) maxima (SCM), complicate satellite measurement of surface Chl a. However, the impact of likely errors in surface Chl a on satellite-based estimates of depth-integrated daily net primary production (NPP) have yet to be quantified. Here we use a large in situ Chl a and primary production database (ARCSS-PP) to calculate the magnitude of the error that likely results from both omission of SCM and overestimated phytoplankton biomass when satellite-based Chl a is used as input to an NPP algorithm. Results show that errors in pan-Arctic NPP, due to omission of the SCM, increase from 0.2% in January to 16% in July and are largest in the Beaufort and Chukchi Sea. Over an annual cycle, the error is approximately 8%. Errors in regional NPP resulting from overestimates of surface Chl a by Sea-viewing Wide Field-of-view Sensor can be larger, particularly when surface Chl a and NPP are low, but are partially offset by underestimates in NPP due to omission of NPP at the SCM. As a result, the combined effect of underestimates in NPP due to omission of the SCM and overestimates in NPP due to high satellite Chl a yields a total error in annual pan-Arctic, depth-integrated NPP of <1%. The reasons for this surprisingly low error are discussed.

1. Introduction

[2] The Arctic Ocean has undergone unprecedented change in recent decades, the most noteworthy of which may be the dramatic decrease in sea ice extent and a shift from a multiyear to a largely annual sea ice cover [Stroeve et al., 2007; Wang and Overland, 2009]. These reductions in both sea ice area and persistence have driven marked changes in remotely sensed phytoplankton biomass [Pabi et al., 2008; Arrigo et al., 2008a; Arrigo and van Dijken, 2011] and have likely altered the balance between pelagic-based and sea ice–based primary production [Arrigo et al., 2008a]. Because the Arctic pelagic environment is relatively inaccessible and difficult to sample, in situ data on phytoplankton abundance and rates of production are often sparse and unevenly distributed in both space and time. While some Arctic regions have received considerable attention, particularly in recent years (e.g., the Chukchi, Beaufort, and Barents seas), others have either been sampled very little or the data from those areas are not widely available (e.g., the East Siberian, Laptev, and Kara seas). Consequently, large-scale studies of Arctic Ocean primary production have relied primarily on numerical modeling and satellite remote sensing approaches that can compensate for the large in situ data gaps [Pabi et al., 2008; Arrigo et al., 2008a; Zhang et al., 2010].

[3] Studies of phytoplankton abundance and net primary production (NPP) using ocean color remote sensing require development of accurate algorithms and high-quality input data, which are not always available [Saba et al., 2010, 2011]. In the Arctic, these challenges are compounded by large solar zenith angles, persistent cloud cover, and high riverine fluxes into its coastal margins. Low light levels in polar waters require that phytoplankton increase their cellular pigment concentration to be able to absorb sufficient light for net growth. As a result, Arctic and Antarctic phytoplankton are highly packaged and absorb less light per unit pigment than their temperate and tropical counterparts [Mitchell, 1992; Wang et al., 2005; Matsuoka et al., 2007]. More importantly, due to the large riverine fluxes into the Arctic Ocean, concentrations of chromophoric dissolved organic matter (CDOM) in nearshore waters can be very high [Guéguen et al., 2007; Granskog et al., 2007, 2009; Retamal et al., 2008; Osburn et al., 2009; Hessen et al., 2010]. These compounds absorb strongly in the blue wavelengths but much less so in the green, as does the phytoplankton chlorophyll a (Chl a) molecule. As a result, waters where CDOM is abundant appear from space to have a high Chl a content. Because pigment packaging and high CDOM concentrations render the optical properties of Arctic waters different from the rest of the global ocean [Matsuoka et al., 2007, 2011], the standard OC4v4 algorithm of Sea-viewing Wide Field-of-view Sensor (SeaWiFS) (and OC3M of the Moderate Resolution Imaging Spectroradiometer (MODIS)) overestimates surface Chl a concentrations, exhibiting a root mean square error (RMSE) of 0.21 mg Chl a m−3 when compared to in situ data [Matsuoka et al., 2007]. Errors are largest at low Chl a concentrations and diminish to near zero at Chl a concentrations above 0.5 mg m−3. Although the magnitude of this error is likely to be smaller for the more recent OC4v6 SeaWiFS algorithm, which shows improved agreement with in situ values in coastal waters (http://oceancolor.gsfc.nasa.gov/REPROCESSING/R2009/ocv6/), this is yet to be rigorously tested in the Arctic Ocean.

[4] Furthermore, because the upwelling radiance received by satellite ocean color sensors originates almost exclusively from the upper optical depth of the surface ocean, features deeper in the water column cannot be resolved. For example, parts of the Arctic Ocean are characterized by a well-developed subsurface chlorophyll maximum (SCM) that usually forms at the depth of the nutricline after surface nutrients have been exhausted. Because these features can be sites of substantial water column primary production in spring and summer [Martin et al., 2010], depth-integrated rates of NPP that are calculated from satellite-based measurements of surface Chl a [e.g., Arrigo and van Dijken, 2004; Pabi et al., 2008; Arrigo et al., 2008a] will be underestimated in regions of the Arctic Ocean that have a well-developed SCM.

[5] Primary production algorithms that use remotely sensed surface Chl a concentrations as a proxy for phytoplankton abundance generally assume that surface values are representative of Chl a concentrations over the entire upper mixed layer (ML). Below the ML, some algorithms assume that Chl a concentrations remain unchanged from surface values, while others assume that they diminish progressively with depth [e.g., Arrigo et al., 2008b; Pabi et al., 2008]. While the consequences of the assumption of uniformity within the ML are minor when euphotic depths are shallower than the ML depth (MLD), this is not always the case in the Arctic Ocean where SCM can develop at depths with considerable light still available to drive photosynthesis [Martin et al., 2010]. Therefore, vertical profiles of NPP generated from surface Chl a will differ from those calculated from in situ profiles of Chl a when an SCM is present.

[6] Despite recognition that its unique optical properties and SCM make space-based measurements of NPP in the Arctic Ocean challenging [Weston et al., 2005; Matsuoka et al., 2007, 2011; Ben Mustapha and Larouche, 2008; Hill and Zimmerman, 2010], the impact of these features on estimates of NPP has never been rigorously quantified on a large scale. The goal of the present study was to quantify the magnitude of the error in depth-integrated NPP in the Arctic Ocean estimated using satellite-based algorithms that result from (1) omission of the SCM from vertical profiles of Chl a and (2) overestimates in surface Chl a concentration by the OC4v6 (for SeaWiFS) algorithm due to pigment packaging and CDOM. Furthermore, differences in the magnitude of this error were quantified for eight different geographic regions within the Arctic Ocean and over monthly, seasonal, and annual time scales.

2. Methods

2.1. Assessing the Contribution of the SCM to Depth-Integrated NPP

[7] While in situ data frequently exhibit a pronounced SCM, these features are necessarily absent when vertical profiles of Chl a are determined from satellite data. To assess the potential error associated with satellite-based estimates of depth-integrated NPP that omit the SCM, we quantified the contribution of the SCM to total water column production using in situ Chl a data from the Arctic System Science primary production (ARCSS-PP) database (http://accession.nodc.noaa.gov/0063065) [Matrai et al., 2011] as input to the NPP algorithm of Arrigo et al. [2008b] (as modified in work by Pabi et al. [2008] for Arctic waters). This was done by first calculating mean vertical profiles of in situ Chl a for each time bin (monthly, seasonal, and annual) and geographic sector from the ARCSS-PP database (see below), keeping the SCM intact (Figures 1a1d, solid lines). NPP was then calculated at each depth for which Chl a data were available, and these depth-dependent values were integrated vertically (to 100 m) to calculate total water column NPP. Separately, we applied the mean surface Chl a concentration from the in situ database for a given time bin to the upper water column using four different methods (see next paragraph), essentially removing any SCM that might be present (Figures 1a1d, dashed lines). We then recalculated total water column NPP for each time bin and in each geographic location using the adjusted vertical Chl a profiles as input to our algorithm. The difference in vertically integrated NPP between these two approaches reflects the amount of NPP associated with the SCM and provides an estimate of the degree to which NPP algorithms underestimate vertically integrated NPP as a consequence of omission of the SCM.

Figure 1.

Example vertical profiles of Chl a illustrating the four different methods used in this study to propagate surface Chl a over depth to: (a–d) remove the SCM (see section 2.1 for description of the methods), and (e–h) increase surface Chl a by 0.21 mg m−3 (see section 2.2. for description of methods), an amount equivalent to the RMSE of SeaWiFS for Arctic waters [Matsuoka et al., 2007]. The solid lines represent the original in situ Chl a profiles from the ARCSS-PP database [Matrai et al., 2011] and the dashed lines are the redistributed Chl a values used as input to the NPP algorithm.

[8] The four methods we used to redistribute surface Chl a over depth to remove the SCM included: (1) Extending in situ surface Chl a values down to the base of the SCM and continuing to use in situ values from the ARCSS-PP database at greater depths (Figure 1a); (2) extending surface in situ Chl a values to the entire upper 100 m of the water column, thereby eliminating all vertical structure (Figure 1b); (3) extending surface in situ Chl a values to the upper 20 m of the water column (to simulate a 20 m ML) and decreasing them exponentially at greater depths (as was done by Arrigo et al. [2008b]) (Figure 1c); and (4) extending surface in situ Chl a values to the upper 40 m of the water column (to simulate a 40 m ML) and decreasing them exponentially at greater depths (Figure 1d). These last three methods are representative of ways that NPP algorithms typically assign vertical profiles of Chl a using satellite-based surface measurements.

2.2. Assessing the Impact on NPP of Overestimates in Chl a by SeaWiFS

[9] We quantified the impact that overestimates of Chl a by SeaWiFS [Matsuoka et al., 2007], resulting primarily from higher pigment packaging and elevated CDOM concentrations in Arctic waters, would have on NPP calculated by our algorithm. To do so, we generated new vertical profiles of Chl a from in situ data by increasing the mean surface value in each time bin and geographic location by an amount equivalent to the calculated RMSE of SeaWiFS for Arctic waters (0.21 mg m−3). To produce these adjusted profiles, the surface Chl a value was held constant throughout the upper water column using four different methods (similar to those described in section 2.1). The difference in NPP estimated by our algorithm using the in situ Chl a profile (Figures 1e1h, solid lines) and the adjusted profiles (Figures 1e1h, dashed lines) as input provides an estimate of the error in depth-integrated NPP for each time bin and each geographic location resulting from overestimates of satellite-based surface Chl a by SeaWiFS.

[10] The four ways we redistributed surface Chl a over depth in this analysis included: (1) Extending in situ surface Chl a values +0.21 mg Chl a m−3 to the base of the presumed ML, the depth of which was approximated from each vertical profile, and continuing to use in situ values from the database at greater depths (Figure 1e); (2) extending surface in situ Chl a values +0.21 mg Chl a m−3 to the entire upper 100 m of the water column (Figure 1f); (3) extending surface in situ Chl a values +0.21 mg Chl a m−3 to the upper 20 m of the water column and decreasing them exponentially at greater depths (Figure 1g); and (4) extending surface in situ Chl a values +0.21 mg Chl a m−3 to the upper 40 m of the water column and decreasing them exponentially at greater depths (Figure 1h).

2.3. The ARCSS-PP Database

[11] We used in situ data from the ARCSS-PP database [Matrai et al., 2011] to characterize regional and temporal variability in vertical distributions of Chl a and primary production for eight geographic sectors of the Arctic Ocean (as defined in work by Pabi et al. [2008]). The geographic sectors were demarcated by longitude and include the Chukchi (180° to 160°W), Beaufort (160°W to 100°W), Baffin (100°W to 45°W), Greenland (45°W to 15°E), Barents (15°E to 55°E), Kara (55°E to 105°E), Laptev (105°E to 150°E), and Siberian (150°E to 180°) sectors. The database consisted of 9484 Chl a stations and 1931 primary production stations sampled over 50 years (1954–2007). Both data types were binned into 10 m depth increments at monthly (e.g., January, February, March…), seasonal (January–March, April–June…), and annual (January–December) time scales to produce vertical profiles for each geographic sector.

[12] Not surprisingly, the months of April–September had the largest number of Chl a measurements in the ARCSS-PP database (Figure 2), with more values near the ocean surface than at depth [Matrai et al., 2011; V. J. Hill et al., Synthesis of integrated primary production in the Arctic Ocean: II. In situ and remotely sensed estimates, 1999–2007, submitted to Progress in Oceanography, 2011]. The number of Chl a measurements also varied geographically, with the Barents, Beaufort, Chukchi, and Greenland sectors containing the most data. The Laptev sector had too few observations (two profiles in July and none at other times) and was excluded from most of the analyses. Some profiles were located under the ice and within the ice edge, areas that cannot be sampled by SeaWiFS (Hill et al., submitted manuscript, 2011).

Figure 2.

Vertical profiles of mean in situ Chl a from the Arctic Ocean for each month of the year and for each geographic sector. Also shown is the number of observations for each month and geographic sector that were used in the calculation of the mean Chl a concentration at each 10 m depth interval.

[13] The ARCSS-PP database also contained vertical profiles of primary production (Figure 3), although these were less numerous and more unevenly distributed in space and time than the Chl a data (Hill et al., submitted manuscript, 2011). Nevertheless, these data were useful in assessing the prevalence of subsurface productivity maxima (SPM) across the Arctic Ocean. These data also allowed us to calculate the contributions by the SPM to total water column NPP, which were compared to similar estimates made for the SCM.

Figure 3.

Vertical profiles of mean in situ primary production from the Arctic Ocean for spring (Apr–Jun), summer (Jul–Sep), and autumn (Oct–Dec) seasons and for the entire year (Annual) for each geographic sector. Also shown is the number of observations for each month and geographic sector that were used in the calculation of the mean primary production at each 10 m depth interval.

2.4. Net Primary Production Algorithm

[14] Daily NPP was calculated from Chl a profiles using the algorithm of Arrigo et al. [2008b] as modified by Pabi et al. [2008]. The algorithm calculates the rate of NPP (mg C m−3 hr−1) as a function of spectral downwelling irradiance, temperature (°C), and Chl a concentration (mg m−3) using the phytoplankton photosynthetic parameters provided in work by Arrigo et al. [2008b]. For the purposes of this study, incident surface irradiance (200 μmol photons m−2 s−1), temperature (°C), and phytoplankton photosynthetic parameters (P*max = 1.8 mg C mg−1 Chl a hr−1, Ek = 60 μmol photons m−2 s−1) were held constant for all calculations, allowing us to isolate errors in NPP to those due only to differences in the Chl a profiles. It should be noted that the magnitude of the error was insensitive to the values chosen for incident surface irradiance, temperature, and phytoplankton photosynthetic parameters, as long as the same values were used in all calculations. Downwelling irradiance was attenuated with depth in the upper 100 m of the water column as described in work by Arrigo et al. [2008b] assuming that all bio-optical properties other than Chl a concentration remain constant for all profiles.

3. Results

[15] We have focused our analyses on NPP calculated from Chl a profiles that were adjusted using method 1, whereby either in situ surface Chl a values were extended to the base of the SCM and in situ values from the ARCSS-PP database were used at greater depths (section 2.1) or in situ surface Chl a values +0.21 mg Chl a m−3 were extended to the base of the presumed ML and in situ values from the database were used at greater depths (section 2.2). This allows us to focus our analysis strictly on errors associated with omission of the SCM and overestimates of surface Chl a by SeaWiFS. We discuss how these results change when the other three Chl a redistribution methods are used in section 4.2 in the context of applicability to satellite-based NPP algorithms.

3.1. Contribution of the SCM to Depth-Integrated NPP

[16] Multiyear, monthly averaged vertical profiles of Chl a for the different geographic regions indicate that the SCM may be present in the Arctic Ocean as early as March, being most prominent in the Greenland and Chukchi sectors (Figure 2e) where it accounted for approximately 20% of depth-integrated NPP (Table 1). It should be noted, however, that this pattern is based on less than 10 vertical profiles (Figure 2f). In April, when data density was much greater (Figure 2h), the SCM was much less obvious, with Greenland and Barents being the only sectors exhibiting slight increases in Chl a in the subsurface (Figure 2g). The SCM in these regions accounted for 2.5% and 6.9% of depth-integrated NPP, respectively (Table 1). Concentrations of Chl a throughout much of the Arctic Ocean had increased dramatically by May, and a strong SCM was observed in the Chukchi Sea (Figure 2i). Surprisingly, however, this SCM accounted for <5% of depth-integrated NPP for that month, similar to values in the Greenland and Baffin sectors where the SCM was much less prominent (Table 1). This is most likely because surface Chl a concentrations in the Chukchi sector were so high (>5 mg m−3) that little light would be transmitted to the SCM to fuel NPP.

Table 1. Percent Change in Depth-Integrated Daily Net Primary Production due to Removal of the Subsurface Chl a Maximum for Different Geographic Sectors and Time Periodsa
SectorJanFebMarAprMayJunJulAugSepOctNovDec
  • a

    Chl a was distributed vertically using method 1.

  • b

    ND indicates no in situ data were available.

ChukchiNDbND−19.70.0−4.5−20.6−14.10.00.00.0−20.8ND
Beaufort0.0−2.4−5.7ND0.00.0−35.3−10.1−20.40.0−13.50.0
Baffin−4.4−6.20.00.0−6.00.0−0.6−5.1−15.8−8.10.00.0
GreenlandNDND−20.4−2.5−6.70.0−5.8−2.5−16.6−6.50.00.0
BarentsND0.00.0−6.90.0−2.5−2.1−3.7−1.4−4.1NDND
KaraNDNDND−1.90.00.00.0−10.40.0−1.9NDND
LaptevNDNDNDNDNDNDNDNDNDNDNDND
SiberianNDNDNDNDNDND0.00.0−6.9NDNDND
All−0.2−0.8−2.0−3.0−9.1−10.5−15.6−1.2−2.5−2.6−15.40.0
 
 Jan–MarApr–JunJul–SepOct–DecAnnual
Chukchi−19.7−12.0−6.10.0−7.6
Beaufort−10.70.0−20.4−3.4−11.7
Baffin−6.6−0.2−4.1−7.5−4.1
Greenland−20.4−2.0−6.80.0−4.5
Barents0.0−2.1−2.5−4.1−1.2
KaraND0.00.0−1.90.0
LaptevNDND0.0ND0.0
SiberianNDND−0.5ND−0.5
All0.0−8.8−6.8−7.4−7.6

[17] By summer, the contribution of the SCM to depth-integrated NPP had increased, being highest in June and July, where it accounted for 10.5% and 15.6%, respectively, over the entire Arctic (Table 1). The contributions were greatest in the Chukchi and Beaufort sectors, with each exhibiting a prominent SCM and relatively low surface Chl a concentrations (Figures 2k and 2m), which provided for increased light transmission and, thus, resulted in higher rates of NPP at the SCM. The Greenland and Barents sector also harbored an SCM, although it accounted for a relatively small fraction of NPP (<6%) in these sectors (Table 1). As summer progressed, Chl a began to decline in all Arctic sectors except for the Baffin (Figure 2o), where a prominent SCM that accounted for 5.1%–15.8% of depth-integrated NPP had developed. The SCM in the Beaufort and Greenland sectors also contributed significantly to NPP, accounting for 20.4% and 16.6%, respectively, in September (Figure 2q).

[18] By October, Chl a concentrations had fallen even further (Figure 2s) and the amount of NPP associated with the SCM began to decline (Table 1). This proportion increased again in November, but only in the Beaufort and Chukchi sectors, which were not yet fully ice covered [Pabi et al., 2008] and, in the case of the Chukchi, still harbored relatively high Chl a concentrations (Figure 2u), albeit over a much smaller area [Pabi et al., 2008]. By December, Chl a concentrations had dropped even further, although data were available for only a few geographic sectors and none of those exhibited an obvious SCM (Figure 2w).

[19] In all regions, the contribution of the SCM to total depth-integrated NPP in the Arctic Ocean increased each month between winter and summer, rising from 0.2% in January to 15.6% in July (Table 1). Values dropped dramatically thereafter and remained <3% through December, with the exception of the month of November when the percent contribution by the SCM increased temporarily in the Beaufort and Chukchi sectors. In fact, the Beaufort and Chukchi sectors appear to harbor the most well developed and quantitatively important SCM in the Arctic Ocean, as can be more clearly seen when the data are binned seasonally and annually. These are the only two sectors where the SCM accounted for >10% of depth-integrated NPP in two of the four seasons (Table 1). More importantly, the annually averaged data show that the SCM in the Beaufort and Chukchi sectors accounts for 2–3 times the amount of depth-integrated NPP in any other Arctic sector.

[20] For the entire Arctic Ocean, the SCM accounted for only 7.6% of depth-integrated NPP averaged on an annual basis (Table 1). This value ranged seasonally from 0% in January–March to 8.8% in April–June. The surprisingly low contribution of the SCM to annual NPP is due in part to the fact that of the 62 sector/months for which Chl a data were available, 28 (45%) exhibited no SCM (Table 1). Thus, contrary to the notion that the SCM is a ubiquitous feature throughout the Arctic Ocean, it appears that its importance to depth-integrated NPP varies as a function of location and time, being most pronounced in summer months. Furthermore, this analysis suggests that the magnitude of the error in satellite-based estimates of NPP, due solely to their inability to resolve the SCM, is <12%, and in many sectors, is much less than that.

[21] To confirm these results, we also performed a similar analysis using vertical profiles of primary production (rather than Chl a) from the ARCSS-PP database. Although there were far fewer data available, particularly in January–March (no data) and October–December (Figure 3f), the critical months of April–September contained tens to hundreds of observations in many of the sectors (Figures 3b and 3d), a sufficient number to perform this analysis. For any time bin exhibiting a subsurface productivity maximum (SPM), we replaced elevated production values in the SPM with the lower surface productivity values, just as we did with the SCM, and calculated the change in depth-integrated primary production resulting from the removal of the SPM. Results of this analysis indicate that omission of the SPM had an even smaller impact on estimates of annual NPP than the removal of the SCM. Seasonally averaged data indicate that SPM were only present in July–September (Figure 3c) (see also Hill et al., submitted manuscript, 2011), and only in the Beaufort, Barents, and Siberian sectors. Surprisingly, when data were averaged over the entire year, no SPM was apparent (Figure 3g). On the basis of the seasonally averaged data shown in Figure 3, primary production between April and December was approximately 202 g C m−2 when the SPM was left intact. When the SPM was removed from all vertical profiles and primary production was recalculated, the value for April to December dropped to 196 g C m−2, a decline of only 2.9%, further demonstrating that while satellite-based NPP will be underestimated in waters with an SPM (or an SCM), the magnitude of the underestimate is relatively small.

3.2. Impact on NPP of Overestimates in Chl a by SeaWiFS

[22] The impact on NPP of increasing in situ surface Chl a concentrations by 0.21 mg m−3 throughout the Arctic (to simulate errors in satellite-derived Chl a) varied seasonally and by region (Table 2). When Chl a concentrations are low, increasing ML Chl a by 0.21 mg m−3 represents a large change in the vertical Chl a profile (Figure 4c) and NPP can be dramatically overestimated, with the error in depth-integrated NPP exceeding 100% in this example (Figure 4d). In contrast, when Chl a concentrations are high, the percent error in depth-integrated NPP is greatly reduced. This is because at high Chl a concentrations, increasing surface Chl a by 0.21 mg m−3 results in only a small change in the vertical profiles of both Chl a (Figure 4a) and NPP (Figure 4b).

Figure 4.

Vertical profiles of Chl a from in situ data (solid lines) and with surface Chl a increased (dashed line) by an amount equal to the RMSE of the SeaWiFS OC4v4 algorithm in Arctic waters (0.21 mg Chl a m−3; Matsuoka et al. [2007]). Shown here are examples of waters with (a) high Chl a concentrations and (c) low Chl a concentrations and the corresponding vertical profiles of NPP (Figures 4b and 4d, respectively).

Table 2. Percent Change in Depth-Integrated Net Primary Production Caused by Increasing in Situ Chl a by the RMSE of Satellite-Derived Chl a for Different Geographic Sectors and Time Periodsa
SectorJanFebMarAprMayJunJulAugSepOctNovDec
  • a

    Chl a was distributed vertically using method 1.

  • b

    ND indicates no in situ data were available.

ChukchiNDbND−12.111.6−2.4−14.8−8.66.914.016.90.6ND
Beaufort384.3635.9125.8ND25.012.6−24.73.64.143.9127.0276.3
Baffin4.717.1131.656.44.68.93.0−2.4−6.44.836.4158.8
GreenlandNDND−8.85.92.212.410.413.1−11.621.994.5118.1
BarentsND1014.686.321.87.111.014.715.120.0102.3NDND
KaraNDNDND46.913.56.312.60.917.010.8NDND
LaptevNDNDNDNDNDNDNDNDNDNDNDND
SiberianNDNDNDNDNDND10.214.641.5NDNDND
All10.013.133.010.6−4.0−3.7−9.04.010.513.745.0105.1
 
 Jan–MarApr–JunJul–SepOct–DecAnnual
Chukchi−11.0−7.5−0.516.8−1.9
Beaufort252.814.1−7.061.00.4
Baffin7.810.4−1.19.2−0.5
Greenland−8.88.76.856.94.9
Barents87.88.915.4119.210.0
KaraND12.214.79.613.3
LaptevNDNDNDNDND
SiberianNDND8.6ND7.1
All30.9−2.5−0.511.8−0.9

[23] Note that Table 2 contains mostly positive but also some negative values, indicating that in most cases, NPP was overestimated when surface Chl a was increased by the RMSE, while in others, it was underestimated. When there is no SCM present, overestimates of surface Chl a by SeaWiFS (and MODIS) would always be expected to result in higher depth-integrated NPP (Figure 4). However, when an SCM is present, overestimates of surface Chl a by SeaWiFS and MODIS may either fully or partially offset the fact that satellites cannot account for enhanced production at the SCM, in which case depth-integrated NPP may either be overestimated or underestimated, depending on how much production is associated with the SCM. Consequently, large negative values in Table 2 (situations where increased surface Chl a were associated with lower NPP) are always associated with high values in Table 1 (situations where the SCM accounts for a large fraction of depth-integrated NPP).

[24] Not surprisingly, estimated errors in NPP were largest for those months and locations where surface Chl a concentrations were low, and an absolute increase of 0.21 mg Chl a m−3 represented a large percentage change in Chl a concentration. For example, increasing surface Chl a by the RMSE in the Barents sector in February resulted in a 10-fold overestimate of NPP (Table 2). Although this was the most extreme case, overestimates in depth-integrated NPP exceeded 100% when surface Chl a was increased by 0.21 mg m−3 in most sectors during the low productivity months of January–March and December (Table 2). It must be remembered, however, that the largest errors were invariably associated with sectors and months having low Chl a concentrations (and hence low rates of daily NPP), thereby minimizing the impact of these errors on annual NPP.

[25] Fortunately, during the most productive months of May–August [Pabi et al., 2008; Arrigo and van Dijken, 2011], errors in NPP for the entire Arctic Ocean, resulting from likely overestimates of surface Chl a by satellites, ranged from only 3.7% to 9.0% (Table 2). In specific sectors, errors during these months were as high as 25% (Beaufort sector) or as low as 2% (Greenland sector), although most were in the range of 5%–15%. Patterns were similar in the seasonally averaged data, with errors in April–September (0.5%–2.5%) being much smaller than in January–March and in October–December (12%–31%). Errors also tended to be smaller in the seasonal data than in the monthly data because errors in the latter largely offset each other. For example, over the entire Arctic Ocean, increasing surface Chl a by 0.21 mg m−3 and propagating it throughout the upper ML resulted in a 9% underestimate of NPP in July (due to loss of NPP associated with the SCM) but a 10% overestimate in September (due to increased Chl a at a time when there was no SCM). Consequently, the error over the July–September interval was relatively small (−0.5%). This phenomenon also explains the low error in NPP during April–June (Table 2).

[26] Over an annual cycle for the entire Arctic Ocean, propagating in situ surface Chl a + 0.21 mg m−3 over the depth of the ML and using it as input to the NPP algorithm, rather than using in situ Chl a profiles, results in an error of only about 1%, although within specific geographic sectors, errors are as large as 13.3% (Kara sector). Errors are smaller in the western Arctic (Chukchi, Beaufort, Baffin, and Greenland sectors) than they are in the east, due to the greater prevalence of SCM in the west (recall that in regions where the SCM is important, overestimates of surface Chl a when SeaWiFS Chl a is propagated over the ML are partially offset by loss of the SCM). It must be noted, however, that the eastern Arctic is not well represented in the ARCSS-PP database and, as more data become available, estimated errors associated with omission of the SCM and use of remotely sensed surface Chl a as algorithm input may change.

4. Discussion

[27] The unprecedented spatial and temporal coverage of the global ocean surface afforded by satellite remote sensing has revolutionized the way we view and study ocean processes. This technology is particularly important in regions, like the Arctic Ocean, that are changing rapidly but are logistically difficult to sample in situ. Unfortunately, satellite remote sensing in polar waters is also challenging for a variety of reasons, including persistent cloud cover, often heavy sea ice, high solar zenith angles, and complex bio-optical characteristics. Therefore, it is essential that satellite products from the Arctic Ocean are rigorously validated, and if necessary, suitable regional algorithms be developed [e.g., Cota et al., 2004]. While efforts are currently underway to increase the amount of in situ data from the Arctic Ocean that can be used for algorithm development and validation of satellite imagery, coincident matchups between in situ measurements and ice-free and cloud-free satellite overpasses in this region are still relatively rare.

[28] In the absence of such data, it is necessary to at least quantify the errors in satellite retrievals of ocean color and in the products that are calculated from them [Carr et al., 2006]. In this light, the analysis presented here is not meant to be a validation exercise of a particular algorithm per se. We did not compare algorithm-based estimates of NPP to similar measurements made in situ at the same place and time (e.g., Hill et al., submitted manuscript, 2011). Rather, we were interested in quantifying the likely errors in estimates of NPP given our knowledge of the shortcomings in retrievals of surface Chl a by ocean color sensors such as SeaWiFS and MODIS. The fact that global Chl a algorithms overestimate low Chl a concentrations in polar waters is well known [Carder et al., 2004; Matsuoka et al., 2007; Ben Mustapha and Larouche, 2008], as is the fact that ocean color sensors are incapable of quantifying Chl a associated with the SCM [Weston et al., 2005; Martin et al., 2010; Hill et al., submitted manuscript, 2011]. What was not known was how these shortcomings would impact satellite-based estimates of NPP. We have shown here that although the percent error in satellite-based estimates of NPP can occasionally be large on a regional and monthly basis, particularly in autumn and winter months characterized by low rates of NPP, calculated errors are surprisingly small during the spring and summer. What is particularly noteworthy is that this is true even though we opted not to use the more accurate regional Chl a algorithm for the Arctic Ocean [e.g., Cota et al., 2004] that had a RMSE of only 0.13 mg Chl a m−3, significantly smaller than the 0.21 mg Chl a m−3 used here. Had we used the Arctic-specific OC4L Chl a algorithm [Cota et al., 2004], errors in NPP would have been even smaller than those reported here. However, we chose to base our analysis on the standard SeaWiFS algorithm (OC4v4; O'Reilly et al. [1998, 2000]) because data generated using this algorithm are more readily available and have been much more widely used.

4.1. Why Are the Errors in NPP So Low?

[29] There are a number of factors that contribute to the relatively low errors in depth-integrated NPP calculated during our analysis as a result of both overestimates in surface Chl a by SeaWiFS and the omission of the SCM. One of the most important is that depth-integrated NPP is not linearly related to surface Chl a but, rather, is proportional to the square root of surface Chl a [Eppley et al., 1985]. Light is attenuated exponentially with depth and as surface Chl a increases, the growing phytoplankton population attenuates an ever increasing proportion of the light. As a result, as Chl a increases, NPP increases ever more slowly due to phytoplankton self-shading. As a consequence of this nonlinear relationship between depth-integrated NPP and surface Chl a, the error in NPP will be much smaller than the error in Chl a from which it is calculated. For example, given that depth-integrated NPP is proportional to the square root of surface Chl a, a 50% error in surface Chl a translates into only a 22% error in depth-integrated NPP.

[30] Another reason for the generally small error in NPP is that the Arctic is often characterized by relatively high Chl a concentrations (Figure 2), particularly in the spring and summer when the bulk of the annual NPP takes place. As a result, errors in satellite retrievals of surface Chl a on the order of 0.21 mg m−3 are often small compared to absolute Chl a concentrations in situ. Furthermore, Matsuoka et al. [2007] showed that the agreement between SeaWiFS and in situ Chl a is quite good at concentrations above 0.5 mg m−3, which represent the highest frequency of surface Chl a data available in the ARCSS-PP data set [Matrai et al., 2011]. This means that the errors in Chl a are even smaller than the 0.21 mg m−3 assumed here at times when the Arctic is most productive, thereby minimizing the error in satellite-based estimates of annual NPP.

[31] With respect to errors in NPP caused by the inability of ocean color sensors to detect phytoplankton biomass at the SCM, a simple sensitivity analysis shows that the SCM accounts for the largest fraction of depth-integrated NPP when the SCM is located deeper in the water column and surface Chl a is low (Figure 5). Under these conditions, satellite-based estimates of NPP would be expected to differ most strongly from in situ values. For example, in the case where a relatively well-developed SCM is positioned 20 m below surface waters that contain 1.5 mg Chl a m−3, lowering Chl a concentrations in the SCM to those characteristic of surface values (Figure 5a) reduces depth-integrated NPP by only 5% (Figure 5b). Performing the same analysis, but with the SCM positioned at 30 m (Figure 5c), results in a 20% reduction in depth-integrated NPP (Figure 5d). In a much more extreme case, where Chl a in the upper 20 m is reduced to 0.1 mg m−3 and the SCM is positioned at 40 m (Figure 5e), removing the SCM by using surface Chl a to define the vertical Chl a profile reduces NPP by 72% (Figure 5f). Data from the ARCSS-PP database suggest that this latter case is relatively rare (Figure 2), which explains the relatively small errors in NPP due to omission of the SCM by the NPP algorithm presented in Table 1. Interestingly, vertical profiles similar to those in Figure 5e have been observed in the Beaufort Sea [Martin et al., 2010], the geographic sector in our analysis that showed the largest errors in NPP due to omission of the SCM. Regional variability should thus be more closely examined.

Figure 5.

Vertical profiles of Chl a containing subsurface maxima at three different depths (Figures 5a, 5c, and 5e) and corresponding vertical profiles of NPP (Figures 5b, 5d, and 5f). Examples shown are for SCM at depths of (a) 20 m, (c) 30 m, and (e) 40 m. The solid lines show the original Chl a profiles with the SCM intact and the dashed lines show the vertical profiles after the Chl a concentrations at the depth of the SCM were replaced with surface Chl a values. Also shown in Figures 5b, 5d, and 5f are the percent differences in depth-integrated NPP between the vertical profiles with the SCM intact and with the SCM removed.

[32] Care must be taken when comparing our estimates of the impact of removing the SCM on depth-integrated NPP to estimates of the contribution of the SCM to depth-integrated NPP reported elsewhere. For example, the SCM has been reported to account for 55% of depth-integrated NPP in the Beaufort Sea [Retamal et al., 2008], 37% of annual new production in the central North Sea [Weston et al., 2005], and 30%–50% of depth-integrated NPP in spring and early summer and 20% in autumn in the Southern Ocean [Parslow et al., 2001]. In nonpolar waters, the impact of the SCM is somewhat smaller [Saba et al., 2010], accounting for 17% of NPP in the Solomon Sea [Satoh et al., 1992] and 28.5% of NPP in the western North Pacific [Yamaguchi and Ichimura, 1980]. Although these numbers appear large compared to those calculated here, it must be remembered that their values are based on total NPP within depths occupied by the SCM and the fraction of depth-integrated NPP accounted for by those depths (Figure 6a). In contrast, we calculate NPP over the entire water column with the SCM intact and again with the Chl a concentration of the SCM replaced with surface Chl a values. The difference between these values, while representative of the amount of production associated with removal of the SCM (and an estimate of the satellite-based estimate of NPP), is always less than the total amount of NPP within depths occupied by the SCM (Figure 6b).

Figure 6.

Schematic illustration showing calculations of the contribution of the SCM to depth-integrated NPP. Studies that report the fraction of depth-integrated NPP contributed by the SCM base their estimates on total NPP within the SCM, depicted in (a) as the shaded area. On the other hand, (b) shows that our estimates of the contribution of the SCM to depth-integrated NPP include only the amount of production above what would have been produced if Chl a were constant over the mixed layer (dashed line). This is why the contributions of the SCM to depth-integrated NPP reported here are lower than other reported values.

[33] Finally, as mentioned earlier, errors in depth-integrated NPP associated with satellite-based overestimates in surface Chl a are of the opposite sign compared to errors in NPP associated with not accounting for the SCM. As a result, these two errors can offset each other. This was seen in our analysis where we increased surface Chl a by an amount equivalent to the RMSE of SeaWiFS (to simulate satellite retrieval errors) and propagated this value to the base of the ML (which also removes any SCM that might have been present). Because both offsetting errors were included in this analysis, the total error was relatively small. Furthermore, the error became smaller as the analysis was based on more heavily averaged data, both temporally and spatially. When the analysis was done for the entire Arctic over a complete annual cycle, errors in NPP became negligible, being reduced to <1% (Table 2).

4.2. What Does This Mean for Satellite-Based Estimates of NPP?

[34] Errors in satellite-based NPP resulting from both omission of the SCM and overestimates of surface Chl a were calculated using adjusted in situ vertical profiles of Chl a from the ARCSS-PP database. As a result, the level of adjustment required, for example to remove the SCM, varied for every profile. In some profiles, no adjustment was needed because no SCM was present. In other profiles, the SCM was relatively thick and the upper ML was markedly altered. However, in all cases, unaltered depths not associated with the SCM retained the original Chl a values from the ARCSS-PP database. This approach ensured that the calculated errors were attributable only to either the omission of the SCM or overestimates of surface Chl a.

[35] However, in practice, satellite-based NPP algorithms must specify vertical profiles of Chl a from remotely sensed surface values, potentially compounding errors resulting from the inability of satellites to resolve the SCM and their tendency to overestimate surface Chl a. To assess the additional error resulting from algorithm-based specification of the vertical Chl a profile, we compared results from our initial analysis (method 1, sections 2.1 and 2.2) to those from three additional analyses where surface in situ Chl a was redistributed over depth as is typically done in satellite-based NPP algorithms. In these analyses, surface in situ Chl a values were applied to either the entire upper 100 m of the water column (method 2, sections 2.1 and 2.2), the upper 20 m of the water column but decreased exponentially at greater depths (method 3, sections 2.1 and 2.2), or the upper 40 m of the water column but decreased exponentially at greater depths (method 4, sections 2.1 and 2.2).

[36] Surprisingly, the errors resulting from the use of methods 2–4 for Chl a redistribution were similar to those from our initial analysis that specified vertical Chl a profiles using in situ data from the ARCSS-PP database (method 1). Errors due to either omission of the SCM (Table 3) or overestimated surface Chl a (Table 4) were largest during months when data were sparse and NPP was low (January–March and October–December). During spring and summer, errors were much smaller and not markedly different from the errors generated using method 1, with the exception of July–September when the SCM was removed and the MLD was set at 20 m (Table 4). In this case, errors were approximately 2-fold higher than for method 1, but still well under 20% in all geographic sectors but the Beaufort.

Table 3. Percent Change in Depth-Integrated Daily Net Primary Production Due to Removal of the Subsurface Chl a Maximum for Different Geographic Sectors and Different Time Periodsa
 Jan–MarApr–JunJul–SepOct–DecAnnual
  • a

    Chl a was distributed vertically using methods 2–4.

  • b

    ND indicates no in situ data were available.

Chlorophyll a Constant in the Upper 100 m (Method 2)
Chukchi−18.9−12.0−6.13.0−7.6
Beaufort−76.98.4−2018.3−10.8
Baffin3.12.2−3.8−2.9−3.7
Greenland−16.80−5.113.7−3.0
Barents18.80.9−0.17.51.9
KaraNDb8.69.12.48.9
LaptevNDNDNDNDND
SiberianNDND0.3ND0.3
All10.3−9.0−6.6−5.9−7.4
 
Chlorophyll a Constant in Upper 20 m and Declines Exponentially With Depth (Method 3)
Chukchi−25.9−15.5−10.7−9.1−12.3
Beaufort−85.42.1−30.7−12.8−20.8
Baffin−10.8−5.2−5.4−16.4−6.1
Greenland−26.6−8.4−16.6−10.9−13.0
Barents−10.9−8.2−14.6−23.4−9.9
KaraND6.31.9−5.94.7
LaptevNDNDNDNDND
SiberianNDND−7.9ND−7.9
All−7.3−14.7−12.0−19.7−13.2
 
Chlorophyll a Constant in Upper 40 m and Declines Exponentially With Depth (Method 4)
Chukchi−19.9−12.3−6.50.7−8.0
Beaufort−81.07.8−22.36.6−12.6
Baffin0.11.3−3.9−5.9−3.8
Greenland−18.7−1.2−7.45.9−4.7
Barents8.1−0.5−3.4−4.9−0.3
KaraND8.58.31.28.6
LaptevNDNDNDNDND
SiberianNDND−0.9ND−0.9
All6.0−9.6−7.1−9.1−8.1
Table 4. Percent Change in Depth-Integrated Net Primary Production Caused by Increasing in Situ Chl a by the RMSE of Satellite-Derived Chl a for Different Geographic Sectors and Different Time Periodsa
 January–MarchApril–JuneJuly–SeptemberOctober–DecemberAnnual
  • a

    Chl a was distributed vertically using methods 2–4.

  • b

    ND indicates no in situ data were available.

Chlorophyll a Constant in Upper 100 m (Method 2)
Chukchi−10.7−7.5−0.418.1−1.9
Beaufort274.015.8−6.0121.81.6
Baffin21.510.7−0.915.20.0
Greenland−4.59.79.266.19.1
Barents104.011.420.114516.5
KaraNDb12.417.511.914.5
LaptevNDNDNDNDND
SiberianNDND9.7ND9.7
All36.4−2.4−0.313.4−0.7
 
Chlorophyll a Constant in Upper 20 m and Declines Exponentially With Depth (Method 3)
Chukchi−17.4−10.8−4.86.4−6.3
Beaufort156.29.9−16.477.6−8.1
Baffin7.93.8−2.42.0−2.3
Greenland−14.01.8−1.837.9−0.5
Barents65.02.85.793.05.2
KaraND10.210.74.110.4
LaptevNDNDNDNDND
SiberianNDND2.0ND2.0
All18.6−7.7−5.3−0.3−6.1
 
Chlorophyll a Constant in Upper 40 m and Declines Exponentially With Depth (Method 4)
Chukchi−11.6−7.7−0.816.1−2.3
Beaufort209.915.3−7.9108.6−0.0
Baffin19.09.9−0.912.7−0.1
Greenland−6.08.77.358.87.6
Barents92.810.217.3128.614.7
KaraND12.316.811.014.2
LaptevNDNDNDNDND
SiberianNDND8.8ND8.8
All32.7−2.9−0.810.7−1.2

[37] Our analysis shows that algorithms that specify a shallow MLD (e.g., 20 m) will generally underestimate NPP to a greater degree when an SCM is present than algorithms that propagate surface Chl a more deeply in the water column (Table 3). Surprisingly, errors were virtually the same when surface Chl a was distributed uniformly down to 100 m (method 2) and when surface Chl a was distributed over a 40 m ML (method 4). In contrast, algorithms that specify a shallower ML will do better under conditions where already low surface Chl a concentrations have been underestimated by ocean color satellites (Table 4). This can be seen by the smaller errors in January–March and October–December in all geographic sectors when the ML was set to 20 m. On an annual basis, however, the errors were consistently largest when the ML was 20 m, although they were still only 10% or less in all geographic sectors.

[38] Results of this analysis suggest that errors in depth-integrated daily NPP due to omission of the SCM and overestimation of surface Chl a are not seriously compounded in algorithms that specify a uniform depth over which remotely sensed surface Chl a is applied. Although in our analysis, algorithms that specify a deeper ML (or uniform Chl a throughout the upper 100 m) performed better, even the algorithm that employed shallow MLs exhibited relatively small errors in depth-integrated daily NPP (Table 3). These results show that satellite-based NPP algorithms have the potential to perform quite well in the Arctic Ocean when applied to sufficiently large spatial and temporal scales, despite its complex bio-optical characteristics and widespread SCM.

[39] Unfortunately, because the Arctic Ocean is a bio-optically complex, high-latitude environment characterized by pervasive cloudiness and sea ice cover, relatively few studies have attempted to utilize ocean color satellite data to assess its spatial and temporal variability. The few regional studies that have made the attempt demonstrated poor agreement between satellite-based and in situ estimates of Chl a and NPP [Matsuoka et al., 2007, 2011; Ben Mustapha and Larouche, 2008; Hill and Zimmerman, 2010; Martin et al., 2010]. These studies largely focused on the Beaufort and Chukchi seas, where we show that discrepancies between satellite retrievals of Chl a and in situ values are the greatest and errors in NPP resulting from the inability of satellites to resolve the SCM is the largest (Table 1).

[40] However, even in these western Arctic regions, over an annual cycle, errors in regional estimates of NPP were reduced to <12% due to the fact that the SCM is a transient feature and does not always account for a large fraction of depth-integrated NPP. Furthermore, errors in satellite-based NPP in other Arctic sectors are much smaller, particularly in the highly productive months of May–August. This is similar to observations by Hill et al. (submitted manuscript, 2011), who showed that over the entire Arctic Ocean, monthly averaged, depth-integrated, daily NPP could be accurately calculated from satellite-based observations when Chl a was vertically homogeneous within the ML, or when most of the NPP was restricted to the upper optical depth of the water column. Perhaps the most important lesson learned from our analysis is that, while in any given month in any given geographic sector, errors in NPP can be very large (Table 2), errors diminish substantially when NPP is integrated over broader space and longer time scales. This suggests that annual estimates of NPP that are based on satellite retrievals of surface Chl a are sufficiently reliable at regional scales to be useful for monitoring long-term changes in the Arctic marine environment.

[41] Our results further show that errors in NPP due to overestimates in satellite retrievals of Chl a, resulting primarily from absorption by high concentrations of CDOM in Arctic waters, are larger than errors associated with omission of the SCM. This is clear from Table 2 which shows that NPP is generally overestimated whenever surface Chl a is increased by the RMSE of SeaWiFS (0.21 mg m−3), even when an SCM is present. Fortunately, because CDOM concentrations decline rapidly with distance from riverine sources [Granskog et al., 2007; Cisek et al., 2010; Lund-Hansen et al., 2010], satellite pixels most seriously impacted by CDOM can be removed from further analysis with little impact on large-scale estimates of NPP [Pabi et al., 2008].

Acknowledgments

[42] This research was supported by NASA grants NNG05GC92G and NNX10AF42G and by NSF grant ARC-0731524 to K. Arrigo.

Ancillary