Corresponding author: Z. W. Brown, Department of Environmental Earth System Science, Stanford University, Y2E2 Bldg., 473 Via Ortega, Rm. 140, Stanford, CA 94305, USA. (firstname.lastname@example.org)
 In the Eastern Bering Sea, changes in sea ice have been implicated in recent major upper-trophic level shifts. However, the underlying relationships between sea ice and primary producers have not been well tested. Here, we combine data from multiple satellite platforms, reanalysis model results and biophysical moorings to explore the dynamics of spring and summer primary production in relation to sea ice conditions. In the northern Bering Sea, sea ice consistently retreated in late spring, leading to ice-edge phytoplankton blooms in cold (0–1 °C) waters. However, in the southeastern Bering Sea, sea ice retreat was far more irregular. Although this did not significantly alter bloom timing, late retreat led to blooms at the ice-edge while early retreat led to blooms in open waters that were warmer (≤5.4 °C) and >70% more productive. Early sea ice retreat also led to higher productivity in summer, likely due to weaker thermal stratification. Overall, annual net primary production during warm years of early sea ice retreat was enhanced by 40–50% compared to years with late sea ice retreat in the southeastern Bering Sea. These findings suggest the potential for future sea ice loss to enhance overall carrying capacity of the southeastern Bering Sea ecosystem. Consistently warm blooms in the future may also channel more energy flow toward the pelagic, rather than benthic, environment. To date, however, neither sea ice extent nor the timing of its retreat have undergone long-term changes in the Eastern Bering Sea.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Sea ice is the dominant factor controlling the physical and chemical environment of the Eastern Bering Sea [Stabeno et al., 2010]. Each year, sea ice advances ~900 km across the Bering Sea's eastern continental shelf, driven southward by prevailing winds from latent heat polynyas that form in the lees of islands and coasts [Niebauer et al., 1999]. Sea ice melts at the southern edge of the pack, cooling the water column and facilitating further ice advance [Pease, 1980]. At its peak in March–April, sea ice covers an average of >500,000 km2 of the Bering Sea [Brown et al., 2011], mainly limited to the shallow continental shelf regions, before retreating completely each summer. Large interannual variability in sea ice extent, as much as 40% around the mean, is superimposed on this regular annual cycle [Niebauer and Day, 1989]. This variability is largely in response to atmospheric circulation, which is in turn a function of the position and intensity of the Aleutian Low pressure system [Niebauer et al., 1999; Stabeno et al., 2001]. The year-to-year extent and duration of sea ice influence water column temperature, salt balance, stratification, and light and nutrient availability across the entire Eastern Bering Sea shelf.
 Given its overwhelming importance to the Eastern Bering Sea physical environment, sea ice also dominates patterns of primary production. For example, brine rejection during sea ice formation drives winter convection over the northern shelf [Stabeno et al., 2010], replenishing surface nutrients and setting the stage for an intense spring phytoplankton bloom. Enhanced light levels from meltwater stratification stimulate the bloom, which tends to follow the retreating ice edge northward across the Bering Sea shelf, beginning in April and lasting until surface nitrate is depleted [Niebauer et al., 1990]. Post-bloom production is also dependent on sea ice, which largely determines water column stratification throughout the year and hence, the wind-driven injection of nitrate into surface waters [Sambrotto et al., 1986; Stabeno et al., 2010]. Over half of the Bering Sea's annual net primary production (NPP) falls within a brief 3 month pulse from May to July, a pronounced seasonal cycle that is mediated by sea ice [Brown et al., 2011].
 Sea ice is thought to be especially influential to the dynamics of the spring phytoplankton bloom. For example, the Oscillating Control Hypothesis (OCH) [Hunt et al., 2002, 2011] postulates that in warm years when sea ice retreats early (before mid-March), there is insufficient light to support net production until May or June, after winter storms have ceased and thermal stratification has developed; hence, phytoplankton bloom relatively late in warm (3–5 °C) ambient waters. In contrast, in cold years when sea ice retreats late (April or May), meltwater rapidly stabilizes the water column and phytoplankton bloom relatively early in cold (<0 °C) waters at the ice edge [Alexander and Niebauer, 1981]. In this way, sea ice is thought to control both the timing and temperature of the spring phytoplankton bloom.
 Bloom timing and temperature are both thought to be critically important to zooplankton grazers, which may have cascading impacts on top predators. For instance, the copepod Calanus marshallae may depend upon early, cold phytoplankton blooms, both to avoid exhausting their overwintering lipid reserves [Coyle et al., 2011] and for the successful recruitment of copepodites in spring [Baier and Napp, 2003]. The success of this and other large, lipid-rich crustacean zooplankton may be critical to the recruitment of the economically important gadid walleye pollock (Theragra chalcogramma) [Hunt et al., 2011; Coyle et al., 2011]. Furthermore, ocean temperature during the spring bloom is thought to influence the proportion of organic matter accruing in the pelagic environment versus sinking to the benthos [Mueter et al., 2006; Hunt et al., 2011]. Thus, the OCH predicts that by altering bloom timing and temperature, changes in sea ice may profoundly alter food web structure in the eastern Bering Sea. However, Hunt et al. [2002, 2011] developed this hypothesis using data from a single mooring site, so the spatial applicability of these relationships is unclear. Moreover, relationships between sea ice retreat, bloom timing, and bloom temperature have not been quantitatively established; hence, the impact of sea ice on spring bloom dynamics in the eastern Bering Sea awaits a rigorous test.
 Sea ice has also been implicated in altering annual primary production in the eastern Bering Sea. For instance, it has recently been suggested that intense summer stratification during warm, low-ice years can reduce post-bloom primary production, creating unfavorable conditions for large zooplankton and weakening pollock recruitment [Coyle et al., 2008]. Conversely, Brown et al.  showed that warm years of diminished sea ice actually had higher average chlorophyll concentrations and were more productive overall, although the causes of this enhanced productivity were unclear. Thus, although it has been suggested that sea ice can modify the timing, temperature, magnitude, and ultimate fate of primary production in the Eastern Bering Sea (with implications for upper-trophic level predators), the mechanisms behind these modifications and their ecosystem impacts remain uncertain.
 Here we use the wide spatial and temporal scope of satellite-derived data to examine the impact of sea ice on Bering Sea phytoplankton during spring and summer. We focus on the timing of sea ice retreat, because this metric encapsulates both the southward extent and the persistence of the ice pack, and because it is thought to have crucial ecosystem consequences [Hunt et al., 2002, 2011]. We construct the first long-term records of the timing of sea ice retreat to assess secular and decadal-scale ecosystem changes. We then analyze time series of sea ice cover, sea surface temperature (SST), and surface chlorophyll a (Chl a) concentrations, together with NCEP/NCAR (National Centers for Environmental Prediction/National Center for Atmospheric Research) reanalysis and biophysical mooring data, to explore patterns of spring bloom development in the Bering Sea during years of either early or late sea ice retreat. Finally, we assess summertime NPP in relation to water column stratification during years with contrasting sea ice conditions. We discuss the implications of our results for several current hypotheses of Bering Sea ecosystem change, which were generally developed without the benefit of long-term, broad-scale time series of sea ice and primary production.
2.1 Satellite Remote Sensing, Reanalysis, and Mooring Data
2.1.1 Sea Ice Concentration
 We obtained gridded daily sea ice concentrations at 25 km resolution over the entire passive microwave satellite data record (1979–2010) from the National Snow and Ice Data Center, derived from the Scanning Multichannel Microwave Radiometer and the Special Sensor Microwave/Imager using the NASA Team algorithm (developed by the Oceans and Ice Branch, Laboratory for Hydrospheric Processes at NASA GSFC (Goddard Space Flight Center)) [Cavalieri et al., 2008].
2.1.2 Sea Surface Temperature
 Daily SST data (1982–2010) are based on the Reynolds Optimum Interpolation SST Version 2 product [Reynolds et al., 2002] derived from the Advanced Very High Resolution Radiometer at 0.25° resolution, obtained from NOAA/OAR/ESRL Physical Sciences Division.
2.1.3 Chlorophyll a Concentration
 We constructed a 14 year time series (1997–2010) of surface Chl a concentrations (http://oceancolor.gsfc.nasa.gov/) that employs data from three ocean color sensors, including the Ocean Color and Temperature Scanner (OCTS) for 1997, the Sea-viewing Wide Field-of-View Sensor (SeaWiFS) for 1998–2007, and the Moderate resolution Imaging Spectroradiometer (MODIS) for 2008–2010. We produced daily composites from Level 2 Global Area Coverage (L2 GAC) data [reprocessing 2009.1] at 4.6 km resolution, derived using the standard empirical global algorithms (OC4v4 for SeaWiFS and OC3M for MODIS; OCTS also uses the OC4 chlorophyll algorithm, but tuned for OCTS spectral bands) [O'Reilly et al., 1998].
2.1.4 NCEP/NCAR Reanalysis Data
 We obtained the mean daily u and v components of 10 m wind velocities, as well as 2 m surface air temperatures (SAT), at 2.5° resolution over the entire record from 1948–2010 from the NCEP/NCAR Reanalysis project (NOAA-ESRL, Boulder, Colorado; http://www.esrl.noaa.gov/psd/data/reanalysis/). NCEP/NCAR reanalysis 10 m wind data have been validated in the Bering Sea and were found to accurately reproduce offshore wind speed and direction to within 5% [Ladd and Bond, 2002].
2.1.5 Mooring Data
 The M2 and M4 moorings are located on the southeastern Bering Sea middle shelf (M2 at 56.9°N, 164.1°W; M4 at 57.9°N, 168.9°W) at a depth of ~70 m (Figure 1). These moorings provide nearly continuous year-round temperature and salinity measurements at discrete depths. We obtained M2 (1995–2009) and M4 (1996, 1999–2009) data from NOAA PMEL.
2.2 Net Primary Productivity
 We used satellite-derived sea ice concentrations, SST, and ocean color data as input to the NPP algorithm of Arrigo et al. [2008a], which was recently modified for use in the Bering Sea [Brown et al., 2011]. At each pixel of a projected satellite grid of the Bering Sea, we calculated daily area-normalized NPP (mg C m–2 d–1) as
where z is depth (m), t is time (h), Chl(z) is the Chl a concentration at depth z, G(z,t) is the net biomass-specific phytoplankton growth rate (h–1) calculated as a function of SST and available photosynthetically active radiation, and C/Chl is the phytoplankton carbon to Chl a ratio (90 g:g) [Pabi et al., 2008; Brown et al., 2011].
2.2.1 Sources of Error in Satellite-Derived Spring and Summer Chl a and NPP
 There are several potential difficulties in utilizing satellite-derived Chl a to analyze spring bloom dynamics and summer productivity in seasonally ice-covered seas. The first is that algae growing within and under sea ice, which have been observed in the southeastern Bering Sea region [Stabeno et al., 1998, 2012b], are not visible to satellites. Hence, by neglecting production within and under sea ice, our spring NPP values are likely to be underestimates, particularly in years when sea ice remains late into the spring (see section 4). Unfortunately, the prevalence and magnitude of Bering Sea under-ice and ice-algal production are unclear at this time and therefore deserving of further study. Furthermore, due to high riverine input, substantial chromophoric dissolved organic matter (CDOM) is introduced to the eastern Bering Sea shelf and may lead to overestimates of satellite-derived surface Chl a [e.g., Matsuoka et al., 2007]. We minimize the risk of CDOM contamination by focusing on the Bering Sea middle shelf domain (the portion of the eastern shelf with water depth 50–100 m, as delineated in Brown et al. ), which is far from the Yukon-Kuskokwim River plume (where pixels were frequently flagged as contaminated during summer Brown et al. ).
 In addition, Chl a retrievals can be impaired in marginal ice zones [Bélanger et al., 2007], potentially compromising our satellite assessment of ice-edge blooms. The “adjacency effect” introduces significant error to water-leaving radiance in the blue portion of the spectrum (443 nm), leading to underestimates of Chl a for moderate- to high-biomass pixels within ~5–15 km of the sea ice edge [Bélanger et al., 2007]. Nevertheless, because this is much shorter than the characteristic length scale of the ice-edge bloom perpendicular to the ice front (~20–100 km [Niebauer et al., 1990, 1995]), valid pixels should dominate the Chl a signal. Conversely, marginal ice zone Chl a may be overestimated due to the “subpixel sea ice contamination” effect [Bélanger et al., 2007]. We minimize the latter error by neglecting Chl a in any pixels with ≥10% sea ice concentration (which we considered to be ice-covered pixels) [Brown et al., 2011]. Satellite-derived Chl a has been used successfully in the large-scale detection and quantification of ice-edge blooms in the Bering Sea [Saitoh et al., 2002] and Arctic Ocean [Perrette et al., 2011].
 Because the Bering Sea eastern shelf mixes completely each winter, all spring blooms are expected to be surface features, detectable by satellites provided they are not underneath the sea ice. Moving into summer, however, the depletion of surface nutrients may cause a significant portion of the Chl a biomass and primary productivity to be located below the first optical depth, beyond detection by ocean color sensors. Such subsurface Chl a maxima (SCM) have been observed in the Bering Sea [e.g., Springer and McRoy, 1993; Stockwell et al., 2001; Stabeno et al., 2012a] and may lead to underestimates of summer NPP by our algorithm. However, Arrigo et al.  have recently demonstrated that in Arctic waters, the omission of any SCM present introduces surprisingly little error to regional estimates of NPP calculated from satellite algorithms, for several reasons.
 First, it is important to recognize that satellite-based algorithms do not ignore NPP at the depth of the SCM as is commonly assumed; rather, they calculate NPP at all depths using an assumed Chl a profile (in our case, vertical within the mixed layer and decaying exponentially below that). Effectively, satellite-based algorithms replace Chl a concentrations in the SCM with surface values; hence, they tend to underestimate (rather than ignore) NPP at the depth of the SCM. Second, Arrigo et al.  observed that although errors in NPP can be very large for individual stations, these errors get smaller as NPP is averaged over larger areas and longer time scales. Third, the SCM does not develop until summer, after surface nutrients have been exhausted. Polar waters have their highest rates of NPP in spring when the SCM has generally not yet developed (although SCM have been observed in late spring in the northern Bering Sea [Cooper et al., 2012]), meaning that even if the SCM is ignored completely, the error over the annual cycle is relatively small. Fourth, because light is attenuated exponentially with depth, NPP at the depth of the SCM is generally small compared to that at the surface if there is even a modest amount of phytoplankton capturing light at the surface. The largest errors arise when surface Chl a is very low (~0.1 mg m–3) and the SCM is located very deep (~40 m). This situation was rare in the pan-Arctic analysis of Arrigo et al. , being mainly limited to the Beaufort Sea. This situation is also likely to be rare in the Bering Sea, which more closely resembles the Chukchi Sea to which it is directly linked. Chl a profiles typical of the Chukchi Sea led to a ~6% underestimate of depth-integrated NPP during summer, and a ~8% underestimate over the annual cycle [Arrigo et al., 2011].
 Most importantly, Stabeno et al. [2012a] recently demonstrated that SCM are rare in the southeastern Bering Sea, the main focus of our NPP study. Likewise, Lomas et al.  observed a strong relationship between surface and depth-integrated Chl a in this region, suggesting that SCM are not well-developed in the south. Although we present NPP numbers from across the Bering Sea eastern shelf, we focus on the southeastern shelf for two reasons: (1) this is where we obtained in situ mooring data for comparison with satellite data, and (2) this is where differences between early and late sea ice retreat years are most pronounced. In contrast, summer SCM are prevalent in the northern Bering Sea due to a thicker pycnocline that provides a subsurface habitat with sufficient light for SCM development [Stabeno et al. 2012a; Cooper et al., 2012]. Overall, the impact of not capturing the SCM is likely to be minimal; it should only affect our reported NPP numbers in the summer after surface nutrients are exhausted, and then only in the northern Bering Sea. Our summer and annual NPP numbers from the northern Bering Sea shelf should be considered ≤10% underestimates.
 Our algorithm was validated previously in the Bering Sea by comparing frequency distributions of in situ versus satellite-derived Chl a over the growing season, then assessing the relationship between surface Chl a and daily depth-integrated NPP in situ versus as generated by our algorithm (see Brown et al.  for details). Total algorithm error based on root-mean-square error of the four input variables (Chl a, SST, photosynthetically active radiation, and mixed layer depth) was ±6.3%. Our algorithm-derived annual primary production numbers over the Eastern Bering Sea shelf were in good agreement with those of Rho and Whitledge  using a ship-based approach, matching to within ±13% for the Inner, Middle, and Outer Shelf domains.
2.3 Defining Regions of Interest
 Although our analyses were performed at all pixel locations in the Bering Sea, we selected four regions in which to conduct more in-depth statistical analyses. These are located in the middle shelf domain (water depth 50–100 m), with regions 1 and 2 overlying the northern shelf and regions 3 and 4 overlying the southeastern shelf (Figure 1 grey circles). Each encompasses an area of approximately 40,000 km2 (~1.7% of the total area of the Bering Sea). Averaging over this large area ensures a high frequency of days with valid Chl a measurements despite frequent cloud cover. These four regions were used to characterize the north-south gradient in annual sea ice cover and persistence. Furthermore, regions 3 and 4 are centered on the location of the M4 and M2 biophysical moorings, respectively, allowing us to develop relationships between in situ water column structure and surface satellite and reanalysis data.
 We defined “winter” as the months of January–March, “spring” as April–June, “summer” as July–September, and “fall” as October–December. The time of sea ice retreat is defined as the last winter or spring day (i.e., prior to July 1) when sea ice concentration at a given pixel or within a given region of interest dropped below 10% [Pabi et al., 2008; Perrette et al., 2011]. We defined a storm as having a mean daily wind speed cubed of >2500 m3 s–3 [Hunt et al., 2002].
 At each pixel, we defined the date of the spring bloom as the spring date when surface Chl a was at its maximum. For each region of interest, the date of the spring bloom was defined as the spring date when weighted mean Chl a was at its maximum (see below). We defined a bloom as an ice-edge bloom if it peaked within 20 days of sea ice retreat [Perrette et al., 2011]. Bloom initiation and termination were defined as the first and last spring days, respectively, when weighted mean Chl a was 50% above winter levels. It should be noted that the interannual trends we observed regarding the timing of the spring bloom were not sensitive to the definition of the spring bloom, whether its date was defined as the bloom initiation, peak, or termination.
 We segregated the Bering Sea and the four regions of interest into either “early retreat” or “late retreat” years based on whether the date of sea ice retreat in each region was earlier or later, respectively, than the time series mean. However, years with very low winter/spring sea ice cover can still be defined as “late retreat” due to a late sea ice advection/melt event. Hence, although late retreat years may be generally thought of as “cold,” they did not always exhibit the greatest sea ice cover or the coldest air or sea temperatures.
2.5 Data Processing
 We extracted daily mean values of wind speed, sea ice concentration, SST, and Chl a from each region of interest. Because daily Chl a data are available only at cloud-free pixels, mean values for each region of interest were computed from a different suite of pixels each day. Therefore, to characterize temporal changes in Chl a, we computed a 9 day moving average, weighting each daily mean Chl a value by n/s2, where n is the number of valid Chl a pixels in a given day, and s is the standard deviation. In this way, mean Chl a estimates incorporating a high number of pixels with low standard deviation are given more weight. Although this approach smoothes out the higher frequency variations, it provides a more reliable estimate of the date when maximum Chl a concentrations (and hence the date of the spring bloom) are reached. Daily NPP at each region of interest was computed from raw Chl a data, except on infrequent days when too few pixels (<1% of the region of interest) were available. We filled in these missing days by linearly interpolating between valid daily NPP values from before and after.
 We generated annual maps (1979–2010) of the timing of sea ice retreat at each pixel of our Bering Sea grid. These maps were used to determine the standard deviation and long-term trend in the timing of sea ice retreat at each pixel location. In addition, we generated annual maps (1997–2010) of the timing of the spring phytoplankton bloom and the SST on the day of the spring bloom for each pixel location. We segregated these maps into either early or late ice retreat years, averaged them, and calculated the difference between early and late means. This approach yielded maps of the mean difference in ice retreat timing, spring bloom timing, and spring bloom SST between early retreat years and late retreat years.
 We computed daily means of M2 and M4 mooring temperature data at each available depth. We then determined the mean daily water column temperature (TWC) by linearly interpolating daily temperature measurements and taking the mean over the whole water column. For M2, we also determined the daily difference in temperature between the surface (upper 10 m) and the bottom (lower 10 m) of the water column (ΔTM2). For each year of mooring data, we used the mean of these measures over the summer to assess summer stratification for that year. The difference between surface and bottom temperatures is a good proxy for summer stratification, which is dominated by temperature rather than salinity at M2 [Hunt et al., 2002; Stabeno et al., 2007, 2010, 2012a]. Furthermore, in summer, the Bering Sea middle shelf domain consists of a wind-mixed upper layer and a tidally mixed lower layer [Coachman, 1986], such that surface and bottom temperatures should represent waters above and below the seasonal thermocline, respectively.
 We processed all satellite, reanalysis, and mooring data using Interactive Data Language and R [R Development Core Team, 2011]. We fitted linear models (LMs) and ran Welch's two-sample t-tests using R. The NPP algorithm was encoded using Fortran 77.
3.1 The Physical Environment
 Over the eastern shelf of the Bering Sea, sea ice generally retreated from the south to the north, on average retreating from region 4 on 27 March and from region 1 on 18 May (Table 1 and Figure 2). However, the timing of sea ice retreat was highly variable from year to year, especially in the south (Figures 2 and 3a). For example, in region 4, sea ice was completely absent in some years, and retreated as early as February and as late as May in others (Figure 2 and Table 1). In contrast, region 1 to the north had far more consistent sea ice conditions, almost always retreating during the month of May (Table 1 and Figure 2). Sea ice extended southward into regions 1 and 2 every year between 1979 and 2010, while there were 2 years when sea ice did not reach as far south as region 3 (2001 and 2005), and seven years when it did not reach region 4 (1979, 1981, 1987, 2001, and 2003–2005). This north-south gradient in sea ice variability is reflected by the standard deviation of the date of sea ice retreat, which was highest in a broad band across the southerly extent of the annual Bering Sea ice pack but lower over the northern shelf, decreasing progressively from 35.0 days in region 4 to 10.6 days in region 1 (Figure 3a and Table 1).
Table 1. Means ± SD of Various Properties of the Spring Phytoplankton Bloom Dynamics in Early Versus Late Sea Ice Retreat Years at Four Regions Across the Bering Sea Middle Shelf Domain Over the Period 1997–2010a, b
Sea Ice Retreat Timing
Spring Bloom Timing
SST During Spring Bloom (°C)
Years in which sea ice never reached >10% for the given region of interest are not included in the calculation of the mean or standard deviation of sea ice retreat timing.
The range for all years is shown in parentheses.
May 8 ± 7.9
May 15 ± 7.1
–0.22 ± 0.14
May 25 ± 5.5
May 28 ± 11.2
0.29 ± 0.65
May 18 ± 10.6
May 22 ± 11.4
0.07 ± 0.55
(27 April to 3 June)
(4 May to 9 June)
April 30 ± 10.8
May 16 ± 21.6
0.88 ± 1.26
May 26 ± 7.2
May 23 ± 9.1
0.06 ± 0.38
May 11 ± 16.2
May 19 ± 17.2
0.53 ± 1.04
(24 March to 3 June)
(18 April to 13 June)
March 19 ± 14.1
May 21 ± 28.7
2.72 ± 0.87
April 28 ± 16.3
May 13 ± 8.6
0.29 ± 0.79
April 15 ± 24.9
May 16 ± 19.4
1.34 ± 1.48
(ice-free – May 17)
(4 April to 17 June)
February 16 ± 12.8
May 17 ± 15.8
3.36 ± 1.39
April 21 ± 12.3
May 5 ± 12.5
0.28 ± 0.37
March 27 ± 35.0
May 11 ± 15.3
2.04 ± 1.89
(ice-free – May 6)
(14 April to 4 June)
 Despite this strong interannual variability, there were no long-term trends in the timing of sea ice retreat in any of our four regions of interest (Figure 2; LM: p = 0.64, 0.25, 0.31, and 0.99 for regions 1–4, respectively). In region 4, the 2000s were marked by a series of exceptionally early retreat (or ice-free) years in 2000–2006 followed by very late retreat years in 2007–2010. Apart from this pattern, however, 5 year running means for each of the four regions were fairly consistent, especially for regions 1–3 (Figure 2). This is reflected in a pixel-by-pixel examination of long-term change, which revealed no clear patterns for the Bering Sea middle shelf domain, most of which has experienced somewhat later sea ice retreat since 1979 (~7 days decade–1, Figure 3b).
 During the NCEP/NCAR reanalysis period (1948–2010), winds, like sea ice, were highly variable. For example, on the Bering Sea middle shelf, 8 years experienced no storms (as defined in section 'Definitions') during winter/spring, while three years (1989, 1991, 2001) experienced ≥7 storms (mean: 2.4 ± 1.8). The final winter/spring storm could be as early as January (22 years) or as late as mid-May (1997, 1998, 2002, 2005). Wind speed exhibited a regular annual cycle, being strongest in fall (8.62 ± 0.37 m s–1) and weakest in spring (5.99 ± 0.44 m s–1), with an annual mean of 7.17 ± 0.28 m s–1. We observed no significant difference in mean wind speed between years of early versus late sea ice retreat for any season (two-tailed t-tests: winter p = 0.19; spring p = 0.17; summer p = 0.58; fall p = 0.36). No significant long-term changes in wind speed were evident for spring (p = 0.98), summer (p = 0.61), or fall (p = 0.26), but mean winter wind speed did rise by 0.063 ± 0.032 m s–1 decade–1 (p = 0.049).
 Over the Bering Sea middle shelf domain, colder winter/spring SATs were tightly correlated with later sea ice retreat (r = –0.70, p < <0.001) and more extensive winter/spring sea ice cover (r = –0.91, p < <0.001, Figure 4a). There was a striking SAT minimum between 1971 and 1976 (Figure 4b), with the latter year exhibiting the coldest winter/spring SATs on record (–7.2 °C). An unprecedented 3 year warming event followed, with 1979 experiencing the warmest winter/spring SATs on record (1.1 °C). However, winter/spring SATs have undergone no secular trend since 1948 (r = 0.017, p = 0.89) and mean SATs prior to 1971 are no different from mean SATs after 1977 (Figure 4b). We note that although winter/spring SATs have not changed since 1948, summer SATs over the Bering Sea middle shelf have warmed by 0.15 °C decade–1 during this period (p < 0.001).
3.2 Spring Bloom Dynamics
3.2.1 Spring Bloom Timing
 When averaged over the 14 year Chl a record, the phytoplankton bloom peaked in the month of May in all four regions of interest, with the bloom coming slightly earlier in the south than in the north (11 May in region 4 and 22 May in region 1; Table 1). However, the timing of the spring bloom varied interannually by more than one month in all four regions (Table 1), with region 1 exhibiting the narrowest range (4 May to 9 June) and region 3 the broadest (4 April to 17 June). The northern Bering Sea (regions 1 and 2) experienced almost exclusively ice-edge blooms (with the exception of 2000 and 2002 in region 2; Figures 5a and 5b). For this reason, earlier sea ice retreat was associated with significantly earlier phytoplankton blooms (LM region 1: p = 0.01; Figure 5a).
 In the southeastern Bering Sea (regions 3 and 4), the timing of sea ice retreat determined whether blooms were at the ice edge or in open water (Figures 5c and 5d). March 15 appeared to be an important threshold (as was also shown by Hunt et al. ): in region 4, all 8 years when sea ice retreated before 15 March experienced open-water blooms, while 5 of the 6 years when ice retreated after this date experienced ice-edge blooms. Because of the way ice-edge blooms are defined (here and elsewhere), the timing of ice-edge blooms are invariably controlled by the timing of sea ice retreat (LM region 3: p = 0.01; region 4: p = 0.05). In contrast, we observed that the timing of open-water blooms, which do not depend on the dynamics of a retreating ice edge, were not controlled by the timing of sea ice retreat.
 The net result is that when considering all years, the timing of sea ice retreat did not control the timing of the spring bloom in the southeastern Bering Sea (Figures 5c and 5d; LM region 3: p = 0.40; region 4: p = 0.52). Even when using one-tailed t-tests (allowing us to incorporate ice-free years that were excluded from the above linear models because sea ice retreat timing could not be quantified) we observed no significant difference in mean timing between open-water blooms of early retreat (or ice-free) years and ice-edge blooms of late retreat years (region 3: p = 0.26; region 4: p = 0.07). Phytoplankton bloomed in early-mid May on average regardless of the timing of sea ice retreat (Table 1).
 These relationships can be clearly seen in a pixel-by-pixel comparison of early retreat versus late retreat years (Figure 6). Over a broad swath of the southeastern Bering Sea shelf, sea ice retreated ~50–60 days earlier during early retreat years; however, this did not clearly alter bloom timing – most pixels in this swath had a slightly later bloom (~10–20 days), while some pixels had a slightly earlier bloom (Figure 6b). In contrast, over the northern Bering Sea shelf, sea ice retreated ~10–20 days earlier leading to a ~10–20 day earlier phytoplankton bloom (Figures 6a and 6b).
 Thus, on the northern Bering Sea shelf, the spring bloom inevitably formed at the ice edge and its timing was controlled by the timing of sea ice retreat. However, on the southeastern shelf, phytoplankton could bloom at the ice edge or in open water, and ice retreat timing was not a good predictor of spring bloom timing. Time series analysis of SST, water column temperature, fractional sea ice cover, wind speed, and Chl a demonstrate that on the southeastern Bering Sea shelf (regions 3 and 4), bloom timing is not controlled by the timing of sea ice retreat alone, but from a complex interplay between winds and sea ice. This interplay took on one of two major or two minor patterns:
Sea ice retreats late, phytoplankton bloom in salinity stratified ice-edge waters, and winds are unimportant (Figures 7a and 7b). This pattern played out in region 4 whenever sea ice retreated later than mid-March (Figure 5d). Once ice began to retreat, it appeared that the bloom could not be delayed or arrested even by a strong storm event (as observed in 1999, 2007, and 2010). However, it should be noted that a strong storm event after bloom termination, as observed in May 1997, could mix nutrients to the surface and cause a secondary increase in Chl a (Figures 7a and 7b; see also Stabeno et al. ).
Sea ice retreats early and phytoplankton bloom in open water after the end of winter storms at the onset of thermal stratification. This pattern characterized most of the early retreat years in region 4 (1998, 2000, 2001, 2003, 2004, and 2006, Figures 7c and 7d). Although open-water blooms could develop late, as in 1998 and especially 2006 (Figures 7c and 7d), on average these open-water blooms were not significantly delayed relative to ice-edge blooms. Open-water blooms after the relaxation of winds and the onset of thermal stratification could be as early as 14 May (2000). Because winter/spring winds were highly variable, spring bloom timing was unpredictable during years of early sea ice retreat, and was not correlated with the timing of sea ice retreat.
Sea ice retreats early and phytoplankton bloom early during a temporary relaxation of winds, without any apparent thermal stratification. We observed this pattern in region 4 in 2002 and 2005 (Figures 8a and 8b) and in region 3 in 2000 and 2005. Field studies have shown that blooms can develop prior to stratification in the southeastern Bering Sea middle shelf [Eslinger and Iverson, 2001; Stabeno et al., 2010].
Strong winds advect sea ice into the region, which melts in situ and stimulates a very early bloom during a temporary relaxation of winds. We observed this pattern only in region 3 in 2003 (Figures 8c and 8d). In this case, the indirect effect of a strong storm event was to stratify, rather than destratify, the water column by advecting fresh water into the region. This led to the earliest spring bloom observed in any year in any region (4 April). A similar pattern was observed by Stockwell et al.  in 1997 in the vicinity of M2.
3.2.2 Spring Bloom Temperature
 Over the 14 year Chl a time series, the mean SST during the peak phytoplankton bloom decreased moving northward on the Bering Sea eastern shelf, from 2.04 ± 1.89 °C in region 4 to 0.07 ± 0.55 °C in region 1 (Table 1). Mean bloom SSTs were coldest on the northern and eastern portions of the shelf (0–1 °C) and increased moving toward the deep Bering Sea basin, averaging 4–6 °C (Figure 9a). However, the range in bloom SST was largest (6–8 °C) over the same broad swath of the southeastern Bering Sea where the timing of sea ice retreat was highly variable (Figures 3a and 9b). Whereas the range in bloom SST in region 1 was only 1.7 °C, it was 5.4 °C in region 4, with a similarly large range in region 3 (Table 1 and Figure 9b).
 In the northern Bering Sea shelf (regions 1 and 2), bloom SST was always low and unrelated to the timing of ice retreat (LM region 1: p = 0.46, region 2: p = 0.39; Figures 6, 10a, 10b, and Table 1). However, in the southeastern Bering Sea, earlier sea ice retreat led to significantly warmer spring phytoplankton blooms (LM, region 3: p < 0.01; region 4: p = 0.03; Figures 10c and 10d). This was especially true when ice-free years were included (one-tailed t-test, regions 3 and 4: p < <0.01). Generally, warm blooms were in open water while cold blooms were at the ice edge: in region 4, SST in all but 2 open-water blooms was ≥3.2 °C, while SST for all ice-edge blooms was ≤0.6 °C. On average, open-water blooms during early retreat (or ice-free) years were 3.1 °C warmer in region 4 and 2.5 °C warmer in region 3 than in ice-edge blooms of late retreat years (Table 1). This can be clearly seen in our pixel-by-pixel comparison of early retreat versus late retreat years: over a broad swath of the southeastern Bering Sea, ~50–60 day earlier sea ice retreat led to ~3 °C warmer spring phytoplankton blooms on average (Figures 6a and 6c).
3.2.3 Spring Bloom NPP
 In the southeastern Bering Sea, early retreat (or ice-free) years had more productive spring blooms. In region 4, phytoplankton fixed >70% more carbon during the bloom in early retreat years (52.1 ± 19.7 g C m–2) than in late retreat years (30.1 ± 5.0 g C m–2; one-tailed t-test, p < 0.01), with similar results in region 3. The enhanced NPP of early retreat blooms was not due to blooms being longer in duration (one-tailed t-test: region 4, p = 0.11; region 3, p = 0.07), but was instead due to higher rates of production during the bloom (region 4 early: 988.0 mg C m–2 d–1; late: 738.2 mg C m–2 d–1). In both early retreat (or ice-free) years and late retreat years, the spring phytoplankton bloom accounted for an average of ~30–35% of annual NPP, within the range of 10–65% estimated by Niebauer et al. .
 Spring blooms were also more productive during early retreat years in region 2 of the northern Bering Sea; this relationship was not significant in region 1 due to high interannual variability (region 1: early 23.5 ± 15.5 g C m–2, late 11.5 ± 9.3 g C m–2, p = 0.13; region 2: early 38.7 ± 13.1 g C m–2, late 14.7 ± 7.4 g C m–2, p < 0.01).
3.3 Summer NPP and Stratification
 In the southeastern Bering Sea, early retreat years also had significantly enhanced NPP during summer. In region 4, during summer of early retreat years phytoplankton fixed 65.0 ± 10.1 g C m–2, while during summer of late retreat years phytoplankton fixed only 44.5 ± 21.1 g C m–2 (one-tailed t-test: p = 0.048). We obtained similar results in region 3 (early 63.0 ± 9.0; late 41.5 ± 9.6 g C m–2; p < 0.001). We assessed summer stratification in the southeastern Bering Sea as a possible mechanism for these changes in summer NPP.
 Over the 15 year M2 mooring record, late retreat years were more thermally stratified than early retreat years. M2 summer SSTs were similar (~9–10 °C) during both early and late retreat years (one-tailed t-test: p = 0.12; Table 2 and Figure 12). However, M2 summer bottom temperatures were >2 °C colder during late retreat years due to trapping of cold waters under the seasonal thermocline (p < <0.01; Table 2). This resulted in a steeper vertical temperature gradient during summer of late retreat years. In late retreat years, the surface-bottom temperature difference in summer (ΔTM2) was 8.5 °C, while in early retreat years, it was only 6.9 °C (p = 0.004; Table 2). Thus, although late retreat years had colder mean water column temperatures in summer (p < <0.01; Table 2), they were more thermally stratified.
Table 2. Mean (±s.d.) Summer Temperatures Measured at the M2 Mooring in Years of Late Versus Early Sea Ice Retreat, 1995–2009a
Late retreat years were 1995, 1997, 1999, 2007–2009; early retreat (or no ice) years were 1996, 1998, 2000–2006.
The difference between surface and bottom temperatures provides a measure of thermal stratification.
Early Retreat (warm) Years
10.11 ± 1.22
3.16 ± 0.73
6.96 ± 0.87
5.69 ± 0.78
Late Retreat (cold) Years
9.44 ± 0.89
0.94 ± 0.51
8.50 ± 0.91
3.44 ± 0.54
 To expand estimates of thermal stratification beyond the limited spatial and temporal coverage of observational moorings, we assessed whether satellite-based SST can be used to infer summer thermal stratification. This requires knowing both surface and bottom temperatures during summer. Obviously, surface temperatures can be measured directly by satellite, and satellite-based summer SSTs in region 4 were highly correlated (R = 0.97) and in excellent agreement with SSTs measured at M2, having a slope of 1.01 ± 0.07 (p < <0.01, Figure 11a). Summer bottom temperature, however, must be inferred indirectly from satellite observations. Fortunately, because the water column over the Bering Sea shelf is well mixed in early spring [Hunt et al., 2002], SST can be used to estimate bottom temperature at this time. When the shelf waters stratify in summer, these early spring bottom waters are trapped under the seasonal thermocline, where they are isolated from surface heat exchange and do not warm or cool appreciably thereafter [Stabeno et al., 2007; Zhang et al., 2012; Stabeno et al., 2012b]. Thus, satellite-derived early spring SST (1 April to 15 May) can be used as a proxy for summer bottom temperature.
 Early spring satellite-derived SSTs at region 4 were highly correlated (R = 0.94) with summer bottom temperatures measured at M2 (p < <0.01, Figure 11b), with a slope of 1.05 ± 0.11 and an intercept of ~1 °C (which likely reflects a small amount of mixing of warm surface waters into the bottom layer after mid-May). Hence, the summer – early spring difference in satellite SSTs (ΔTSAT) measured at region 4 was an good predictor of the surface – bottom temperature difference measured at M2 in summer (ΔTM2; p < < 0.01; Figure 11c). We obtained similar results from an identical analysis of region 3 overlying mooring M4 (i.e., ΔTSAT at region 3 was correlated with ΔTM4; p = 0.044). In this way, satellite-based SSTs provide an excellent proxy for thermal stratification on the southeastern Bering Sea shelf.
 ΔTSAT was greater during late retreat years (~7.5–9.0 °C) than early retreat years (~6.0–7.5 °C) not just in the proximity of the M2 and M4 moorings, but over the entire Eastern Bering Sea shelf (data not shown). This suggests that enhanced thermal stratification during late retreat years is widespread. However, it is unclear how far from the M2 and M4 moorings ΔTSAT can provide an accurate assessment of thermal stratification. For example, the inner shelf and deep basin have different summer mixing regimes than the middle shelf domain where the moorings are located [Coachman, 1986] and salinity increasingly controls stratification to the north [Stabeno et al., 2010]. Nevertheless, it is apparent that thermal stratification in summer is stronger in years characterized by late sea ice retreat across much of the southeastern Bering Sea.
 Expanding ΔTSAT over the satellite SST record (1982–2010), we observed that years with more extensive winter sea ice cover exhibited significantly enhanced thermal stratification at both region 3 (p = 0.013) and region 4 (p = 0.047; Figure 11d). A notable exception to this trend was 2004, which had the warmest summer SSTs on record and intense thermal stratification despite minimal sea ice cover (Figures 11a, 11c, and 11d). Furthermore, thermal stratification has been increasing since 1982 in the southeastern Bering Sea (region 3 p = 0.011; region 4 p = 0.006, Figure 11e). This is partly due to a secular increase in summer SST [Brown et al., 2011] and partly due to the cold period at the end of our satellite record, with the most recent years trapping very cold waters under the summer thermocline. Finally, a comparison of summer NPP with ΔTSAT over the satellite record showed that stronger thermal stratification (typical of cold years with late sea ice retreat) was significantly related to reduced summer NPP in both region 3 (p = 0.037) and region 4 (p = 0.05, Figure 11f).
3.4 Annual NPP
 Mean annual NPP over the eastern Bering Sea shelf decreased from south to north, with the mean in region 4 being nearly 40% greater than region 1 (region 4: 139.2 ± 35.2; region 3: 132.6 ± 31.5; region 2: 105.8 ± 28.5; region 1: 100.7 ± 31.2 g C m–2 a–1). These annual NPP results for the Bering Sea middle shelf domain are in good agreement with previous estimates by Springer et al.  and Rho and Whitledge  (see also Brown et al. ).
 Early retreat (or ice-free) years exhibited 40–50% greater annual NPP than late retreat years in the southeastern Bering Sea (region 3: early 159.2 ± 12.9, late 107.0 ± 23.1, p < <0.01; region 4: early 152.8 ± 20.8, late 106.8 ± 35.6 g C m–2 a–1, p = 0.04). Annual NPP during early retreat years was also enhanced in region 2 (early 126.5 ± 11.5, late 81.6 ± 22.0 g C m–2 a–1, p < 0.01), but not in region 1 to the north (early 103.1 ± 33.1, late 98.6 ± 21.6, p = 0.78), where there was far less interannual variability in the timing of sea ice retreat (Table 1). In regions 2–4, the enhanced annual NPP of early retreat years was a result of greater NPP in both spring and summer (fall NPP did not differ between early and late retreat years in any region of interest).
4.1 Long-Term Change
 We observed no long-term trend in the timing of annual sea ice retreat in the Bering Sea. This was true both over the satellite record (1979–2010) and using reanalysis SATs stretching back to 1948, which we have shown are an excellent proxy for both the timing of sea ice retreat and for sea ice cover. The long-term SAT record strongly suggests that while sea ice cover has fluctuated over the Bering Sea middle shelf domain, its extent and timing of retreat have undergone no secular trend in the past six decades (similar to Bond and Adams ; Figures 4a and 4b).
 Many studies of ecosystem change in the Bering Sea have focused on sea ice loss connected with the well-documented North Pacific “regime shift” of 1976–1977 [Wyllie-Echeverria and Wooster, 1998; Napp et al., 2000; Stabeno et al., 2001; Hunt and Stabeno, 2002; Hunt et al., 2002; Schumacher et al., 2003; Overland and Stabeno, 2004; Stabeno et al., 2007]. Each of these studies utilizes a sea ice time series generated by the National Ice Center beginning in 1972, which shows a step-like decline in sea ice concentration in the southeastern Bering Sea bridging this “regime shift.” Unfortunately, this record captures only a few years prior to the dramatic warming event of 1976–1977, and it is generally assumed that the cold years of the early 1970s represent pre-“regime shift” conditions over the Bering Sea middle shelf domain. However, although the long-term SAT record corroborates the unprecedented warming event of the late 1970s, viewing this event in the context of the long-term SAT record makes clear that this 1976–1977 “regime shift” was not a permanent (or even decadal scale) transition from cold to warm conditions. Instead, the period 1971–1976 appears as a deep local SAT minimum within a long-term record characterized by a few multiyear cold and warm temperature deviations and no secular change (Figure 4b), similar to the long-term St. Paul Island temperature records of Overland et al. . This underscores the hazard of utilizing time series that may have inadequate temporal coverage. Although the National Ice Center record has been used widely in the Bering Sea literature, it is prone to misinterpretation because its earliest few years fall within a brief, exceptionally cold period.
 It is important to understand secular trends in sea ice given the overwhelming importance of sea ice to the rich Bering Sea eastern shelf environment. In the Chirikov Basin north of St. Lawrence Island, sea ice extent and duration are declining [Grebmeier et al., 2006; Brown et al., 2011], similar to the rapid sea ice decline observed in the Arctic Ocean [Arrigo et al., 2008b; Comiso et al., 2008]. However, this is not the case on the Eastern Bering Sea shelf, where the annual cycle of sea ice extent has remained steady over time, albeit with great interannual variability. This may be cause for reevaluation of long-term Bering Sea ecosystem changes that have been ascribed to the secular loss of sea ice, such as ecosystem reorganization favoring pelagic over benthic communities [Overland and Stabeno, 2004]. In addition, northward shifts of sub-Arctic species on the Bering Sea shelf have been attributed to a contraction of the subsurface “cold pool” (<2 °C) that overlies much of the Bering Sea eastern shelf each year [Mueter and Litzow, 2008]. However, the year-to-year extent of this cold pool is thought to reflect the southward penetration of the ice pack [Wyllie-Echeverria and Wooster, 1998], which has not contracted over time. Hence, other factors, such as population density, may be driving biogeographical shifts over the Bering Sea eastern shelf [Mueter, 2011].
 It is important to note that although there is no secular trend in the environmental forcing from sea ice [Overland et al., 2012], major events such as the warming of the late 1970s could elicit longer-term nonlinear biological responses. For example, the extremely large pollock year-class of 1978 may have been related to this warming event [e.g., Wyllie-Echeverria and Wooster, 1998; Bailey, 2000], and the impact of its increased population on the marine ecosystem likely persisted long after the warming period was over [e.g., Springer, 1992].
4.2 Sea Ice and Net Primary Productivity
 It remains uncertain how Bering Sea NPP will trend under future sea ice loss. For example, Grebmeier et al.  postulated that sea ice loss has led to reduced NPP, thus reducing organic matter flux to the benthos, a possible mechanism for observed benthic biomass loss over time in the northern Bering Sea. Schumacher and Alexander  also suggest that future sea ice loss will reduce Bering Sea production due to diminished eddy-driven nutrient flux onto the shelf. More recently, Lomas et al.  similarly suggested that warm years may be less productive based on a review of historical studies and field data from recent cold years. Conversely, Loeng et al.  predict that Bering Sea primary production will rise due to a prolonged ice-free growing season, similar perhaps to increasing NPP with sea ice loss observed [Arrigo and Van Dijken, 2011] and modeled [Lavoie et al., 2010; Zhang et al., 2010] in the adjacent Arctic Ocean. Because Bering Sea primary producers currently sustain world-class stocks of pelagic and benthic predators as well as major fisheries, the issue of how they will respond to future sea ice loss demands particular attention. Our observation that early retreat years were more productive during both spring and summer supports the view that NPP may rise under future conditions of sea ice loss and earlier retreat.
4.2.1 Spring Bloom Productivity
 We observed that in the southeastern Bering Sea, the open-water blooms of early retreat years were >70% more productive than the ice-edge blooms of late retreat years. There are a few possible reasons for this dramatic difference in spring bloom NPP. First, phytoplankton growth rate is sensitive to temperature [Eppley, 1972], which is included in our algorithm [see Brown et al., 2011]. Therefore, because open-water blooms in the Bering Sea were ~3 °C warmer than ice edge blooms, they are approximately 26% more productive, which accounts for about one third of the difference in NPP between the two systems. Second, in the southeastern Bering Sea, warm years tend to have warm bottom waters in winter [Stabeno et al., 2012b], possibly enhancing nutrient remineralization rates from the sediments and providing additional nutrients to the upper water column (we would not expect wind mixing to be a factor in enhancing spring nutrient loads, as wind speeds were similar between early and late retreat years; Figure 12; see also Stabeno et al. [2010, 2012b]). Finally, during late retreat years, surface nutrients may be more heavily depleted by under-ice or ice-algal production prior to sea ice retreat, making the eventual ice-edge bloom less productive. Although under-ice and ice-algal production have never been quantified in this region, they have been observed previously in the southeastern Bering Sea [Stabeno et al., 1998; Hunt et al., 2011; Stabeno et al., 2012b]. Hence, it is possible that early and late retreat years actually have similar amounts of total spring NPP, with the under-ice portion of NPP during late retreat years being invisible to satellites. Understanding the relative magnitude of NPP within and under sea ice, at the ice edge, and in open water will be a crucial step in understanding bloom dynamics in the southeastern Bering Sea.
 Surprisingly, spring blooms were also more than twice as productive during early retreat years in the northern Bering Sea, even though sea ice retreat here was much more consistent and blooms were always cold, ice-edge blooms. Again, this could be partially a result of more time available for under-ice or ice-algal production during late retreat years, which may deplete surface nutrients prior to the ice-edge bloom. However, physical processes may also reduce the pre-bloom nutrient load available during late retreat years. Cooper et al.  observed that strong northeasterly winds can drive oligotrophic Alaska coastal water westward across the shelf prior to late sea ice retreat, reducing nutrient concentrations and spring Chl a. They also point to the possible importance of late winter brine formation in setting spring nutrient quotas in the northern Bering Sea.
4.2.2 Summer Stratification and NPP
 Our finding that early retreat years in the southeastern Bering Sea were less stratified and more productive during summer than late retreat years contrasts with the conventional wisdom that warmer conditions reflect intensified stratification and hence, reduced NPP. For example, Coyle et al.  posit that warm, low ice years have intensified summer stratification and depressed post-bloom NPP, which they hypothesized to account for shifts in the zooplankton community in favor of smaller species. Coyle et al.  developed this hypothesis by comparing 1999 (a cold year) to 2004 (a warm year). In 2004, surface waters of the Bering Sea were extremely warm, stability was threefold higher, and the large zooplankton community was much reduced compared to 1999. Remarkably, in terms of thermal stratification, these 2 years examined by Coyle et al.  are the two major outliers over the M2 record (1995–2009). 2004 is the only early retreat year with intense summer thermal stratification (Figures 11c and 11d) and 1999 is the only late retreat year with low thermal stratification and enhanced summer NPP (Figures 11c and 11f; similar to the results of Ladd and Stabeno ). Hence, the conclusions drawn by Coyle et al.  from these two years should not be considered the general pattern for warm versus cold years. In the southeastern Bering Sea, warm conditions can no longer be assumed to reflect stronger stratification and lower NPP (see also Ladd and Stabeno ).
 The reduced summer stratification and enhanced NPP of warm years in the southeastern Bering Sea is likely a direct consequence of sea ice conditions the previous winter. Sea ice cools the water column as it advances across the Bering Sea shelf [e.g., Stabeno et al., 2010]; thus, warm years of diminished sea ice have warmer water column temperatures during winter [Wyllie-Echeverria and Wooster, 1998]. In addition, diminished sea ice imposes less of a physical cap to the water column and introduces less low-density surface meltwater, thereby permitting more effective wind mixing during spring of warm years, which redistributes surface heat gains throughout the water column and warms shelf bottom waters (e.g., Figures 7c and 7d). For these reasons, warm years in the southeastern Bering Sea lack an extensive subsurface cold pool, making them less temperature stratified. This likely accentuates the wind-driven injection of nitrate into surface waters, enhancing summer NPP in the southeastern Bering Sea [Sambrotto et al., 1986; Stabeno et al., 2010] (Figure 11f). Reduced summer stratification over much of the southeastern Bering Sea shelf, in addition to a longer ice-free growing season, probably helps explain greater annual pan-Bering Sea NPP in warm, low-ice years [Brown et al., 2011].
 Our results corroborate those of Ladd and Stabeno , who found that September mean Chl a near M2 was inversely correlated with the August stratification index (suggesting that enhanced production results from weaker stratification). However, whereas we found that early retreat years were less thermally stratified than late retreat years during summer at M2, Ladd and Stabeno  found no difference in stratification index between warm and cold years. The inconsistency of these results is unlikely to be explained by the fact that we quantified stratification using the difference between surface and bottom temperatures, which ignores the effect of salinity. Temperature dominates summer stratification in the vicinity of M2 [e.g., Stabeno et al., 2012a], and the minor contribution of salinity to stratification is enhanced during cold years [Ladd and Stabeno, 2012], which would reinforce our results. More likely, the inconsistency is due to the fact that we assessed stratification over the whole summer (July-September) while Ladd and Stabeno  focused on the month of August. More importantly, we divided years into early versus late based on the timing of sea ice retreat, whereas Ladd and Stabeno  utilized average water column temperatures to separate years into warm, cold, or average. Hence, our early (late) retreat years do not exactly coincide with their warm (cold) years. Nevertheless, our results agree with those of Ladd and Stabeno  in that warm, early sea ice retreat conditions are not associated with enhanced stratification, and weaker summer stratification is correlated with enhanced summer Chl a and NPP in the southeastern Bering Sea.
4.2.3 Annual NPP
 The enhanced NPP we observed on the Bering Sea eastern shelf during early retreat years supports results of Mueter et al.  and Brown et al.  and agrees with observations of higher Chl a concentrations over the eastern shelf during warm years [Lomas, 2012]. This implies that under conditions of future sea ice loss and earlier retreat – barring long-term reductions in nutrient concentrations over the shelf or unanticipated feedbacks associated with climate change – NPP may be expected to rise. However, the fate of this additional organic carbon within the ecosystem is difficult to predict and may not enhance upper-trophic level carrying capacity or be beneficial to fisheries. Shifts in dominant species, phenology, and environmental factors such as temperature may alter the flow of energy through the lower trophic levels of the ecosystem [e.g., Lomas et al., 2012]. For example, anomalous atmospheric conditions in the summer of 1997 led to intense stratification and an unprecedented bloom of coccolithophores in the southeastern Bering Sea [Napp and Hunt, 2001]. This suggests that the secular increase in summer stratification that we observed (Figure 11e) could partially shift the Bering Sea ecosystem away from traditional diatom-dominated food webs. Indeed, coccolithophorid blooms are thought to be becoming more prevalent over the Bering Sea eastern shelf [Harada et al., 2012]. Furthermore, recent warm years were characterized by low populations of large zooplankton that may be crucial to recruitment of predators such as pollock [Hunt et al., 2011; Coyle et al., 2011; Mueter et al., 2011].
4.3 Spring Bloom Dynamics
4.3.1 Spring Bloom Timing
 In the Bering Sea, as in high-latitude seas in general, the phenology of many organisms is tied to the timing of the spring phytoplankton bloom. For example, certain overwintering zooplankton species time their reproductive cycles such that the rapid growth of juvenile stages takes place during this major pulse of NPP [e.g., Smith and Vidal, 1986]. The migrations and/or breeding of many upper-trophic level predators also coincide with the spring bloom [e.g., NRC report, 1996]. Therefore, substantially altering bloom timing can potentially lead to predator/prey mismatches [Cushing, 1990] and disrupt life history strategies that have evolved over thousands of years.
 In the southeastern Bering Sea, bloom timing is thought to be a consequence of the timing of sea ice retreat, with early retreat leading to a delayed bloom and vice-versa [Hunt et al., 2002, 2011]. In turn, changes in bloom timing have been proposed to lead to mismatches with zooplankton grazers, with cascading impacts on upper-trophic level predators. For example, under the assumption that the bloom is delayed during warm years, it has recently been hypothesized that populations of large, overwintering zooplankton species (such as Calanus marshallae) are reduced in warm years because the delayed bloom deprives them of early food resources (possibly impairing lipid accumulation and overwintering survival of age-0 pollock) [Baier and Napp, 2003; Coyle et al., 2011].
 However, we found no evidence of an inverse relationship between the timing of sea ice retreat and the timing of the bloom in the southeastern Bering Sea (Figures 5c and 5d). Although late sea ice retreat (after mid-March) consistently led to ice-edge blooms, these blooms were not always early, with the timing varying by >1 month (i.e., very late retreat could lead to late ice-edge blooms). Similarly, although early sea ice retreat consistently led to open-water blooms, these were not always late, with the timing also varying by >1 month. A hiatus in wind mixing appeared to be the most important trigger of open-water blooms during early retreat years (e.g., Figures 7c and 7d). Although previous work has pointed to the importance of winds in controlling bloom timing [Sambrotto et al., 1986; Stabeno et al., 2001; Stockwell et al., 2001; Hunt and Stabeno, 2002], it has generally been assumed that the timing of sea ice retreat is the dominant factor determining spring bloom timing [Hunt et al., 2002]. Moreover, the marked interannual variability in both sea ice and winds has precluded the identification of any significant relationship between the timing of sea ice retreat and the timing of the bloom. Overall, phytoplankton bloomed in early-mid May in both early and late retreat years (Table 1 and Figure 12). Similarly, Niebauer et al.  reported no difference in timing between ice-edge and open-water blooms between 1975–1988 despite large variations in sea ice extent. Therefore, it can no longer be assumed that the timing of sea ice retreat controls the timing of the spring bloom in the southeastern Bering Sea.
 One possible reason that our results do not agree with those of Hunt et al. [2002, 2011] with respect to bloom timing is that satellite ocean color sensors do not capture production within and under sea ice. The studies of Hunt et al. [2002, 2011] utilized fluorometers mounted on mooring M2 as a proxy for the presence of phytoplankton blooms, potentially allowing them to observe early production under sea ice during late retreat years. Indeed, Stabeno et al. [2012b] observed under-ice fluorescence at M2 during recent cold years (2007–2009), which they attribute to ice algae sloughing off and sinking through the water column. Ice-algae are also thought to provide early forage to euphausiids during cold years [Lessard et al., 2010]. Under-ice (or ice-algal) production may therefore be very important to the timing of spring NPP, and therefore to grazers, in the southeastern Bering Sea. Specifically, it is possible that during late retreat years, grazers have access to a continuum of under-ice (or ice-algal) food throughout the spring leading up to the ice-edge bloom. In contrast, during early retreat years, grazers may face a crucial “hiatus in food availability” between sea ice retreat and the eventual open-water bloom (George Hunt, personal communication). In 2006 near M2, for instance, that hiatus would have lasted nearly four months, from sea ice retreat on 7 February to the open-water bloom on 4 June. Thus, although we show no significant difference between the timing of open-water versus ice-edge blooms in the southeastern Bering Sea (Figures 5c and 5d), under-ice production may begin significantly earlier than either, providing early forage prior to the ice-edge bloom during late retreat years. Because this early forage may be crucial to certain large, overwintering grazers and therefore to age-0 pollock [Baier and Napp, 2003; Coyle et al., 2011], its timing and contribution to annual NPP clearly warrants further study.
 In the northern Bering Sea, we observed the opposite of the pattern generally assumed in the south, with earlier sea ice retreat leading to earlier blooms and vice-versa (Figure 5a). This is due to the fact that sea ice consistently retreated during a narrow window in late spring and blooms were always at the ice edge, such that bloom timing closely followed retreat timing (Figure 5a). Earlier retreat also leads to earlier ice-edge blooms in the Arctic Ocean [Kahru et al., 2011], so this pattern probably dominates wherever sea ice generally retreats in late spring, at a time when light is sufficient for meltwater stratification to promote NPP.
4.3.2 Spring Bloom Temperature
 In addition to the timing of the spring bloom, the water temperature during the bloom is thought to be critical to energy flow and pelagic-benthic coupling in the Bering Sea. Because many grazers are temperature sensitive [Huntley and Lopez, 1992], it has been suggested that when blooms develop in warm waters, grazers are more productive and transfer more carbon to pelagic predators, whereas when the bloom is cold, grazers are less productive and more carbon sinks to the benthos [Walsh and McRoy, 1986; Hunt et al., 2002]. Recent work supports this paradigm that more carbon accrues to the pelagic environment with warm spring conditions, possibly enhancing early-season survival of larval pollock [Hunt et al., 2011]. However, warm spring water temperatures may also accelerate grazer metabolic rates, thereby depleting lipid stores and reducing populations of large, overwintering zooplankton species that pollock may depend upon to survive their first winter [Coyle et al., 2011]. Variability in water temperature during phytoplankton blooms therefore has implications for recruitment of pollock and for the flux of fixed C to the benthos.
 On the northern Bering Sea shelf, the bloom was consistently cold and ice-associated (Figures 5, 9, and 10), likely leading to a consistently high flux of fixed C to the benthos. This is supported by rapid sedimentation of organic matter shortly after sea ice retreat [Cooper et al., 2002], as well as observations of a primarily benthic food web in the northern Bering Sea [Walsh and McRoy, 1986; Grebmeier et al., 2006]. However, in the southeastern Bering Sea, we observed that bloom temperature was a strong function of the timing of sea ice retreat. Earlier sea ice retreat by 40–60 days led to spring blooms in open waters that were an average of ~3 °C warmer than the ice-edge blooms of late retreat years (Table 1 and Figures 6 and 12). According to the relationship of Huntley and Lopez , the range in bloom temperatures in region 4 (–0.2 to 5.2 °C) is sufficient to increase copepod production by ~90%. Of course, Bering Sea grazers may have different temperature responses than this global relationship, particularly because recent work suggests that microzooplankton, rather than copepods, may be the dominant grazers in this region [e.g., Strom et al., 2007; Sherr et al., 2009; Moran et al., 2012]. Nonetheless, our results indicate that interannual variations in southeastern Bering Sea bloom temperature may be large enough to strongly affect grazers (consistent with Hunt et al. [2002, 2011]). A ~3 °C rise in bloom temperature is also likely to reduce the amount of carbon reaching the benthos, as observed by Townsend et al.  in the shallow temperate North Atlantic. This is consistent with reduced survival and recruitment of benthic-feeding flatfish species during early retreat years in the southeastern Bering Sea [Mueter et al., 2006; Hunt et al., 2011].
4.4 Implications for the Oscillating Control Hypothesis
 The OCH predicts that early sea ice retreat leads to a delayed, warm open-water spring phytoplankton bloom, while late sea ice retreat leads to an early, cold ice-edge bloom [Hunt et al., 2002, 2011]. In turn, the dynamics of the spring bloom affect the ability of grazers to crop the bloom and transfer fixed carbon to upper-trophic levels, altering energy flow and the ultimate fate of organic carbon in the Bering Sea ecosystem. Our understanding of the response of grazers to bloom timing and temperature is still evolving, but recent work suggests that large, overwintering species such as C. marshallae require an early, cold phytoplankton bloom [Baier and Napp, 2003; Coyle et al., 2011; Hunt et al., 2011]. On the other hand, more fixed carbon accumulates in the pelagic environment during warm years [Mueter et al., 2006; Hunt et al., 2011], implying that certain grazers (possibly microzooplankton [Lomas et al., 2012] but see Strom and Frederickson ) fare well during warm conditions, as more bloom carbon appears to be grazed before sinking to the benthos. Teasing out the effects of warm versus cold conditions on different species of grazers and upon upper-trophic level predators such as pollock will require further investigation, but it is clear that the timing and temperature of the spring phytoplankton bloom are important to the ecosystem dynamics of the Bering Sea eastern shelf.
 Here we have used satellite-derived data to provide the first rigorous test of the relationship between sea ice retreat timing and spring bloom dynamics in the Bering Sea. We observed that in the southeastern Bering Sea, earlier sea ice retreat indeed led to significantly warmer spring phytoplankton blooms (by ~3 °C), as predicted by the OCH. However, inconsistent with the OCH, the timing of sea ice retreat did not significantly impact the timing of the spring bloom (Figures 5c and 5d). This aspect of the OCH should be reevaluated and modified, especially in light of the possibility that under-ice or ice-algal production may be very important to the timing of NPP in the southeastern Bering Sea.
4.4.1 Spatial Applicability of the OCH
 Because the OCH is a broad and potentially very powerful hypothesis, linking the physico-chemical environment through several trophic levels to the top predators of the ecosystem, it has been invoked in many diverse marine environments, as distant as the Arctic Ocean [Moline et al., 2008], the Black Sea [Katara et al., 2011], and the Oyashio Current [Shida et al., 2007]. This makes it important to clearly define where the fundamental mechanisms of this hypothesis apply.
 The OCH is predicated on large year-to-year fluctuations in sea ice retreat, which then alter bloom timing and temperature [Hunt et al., 2002, 2011]. In turn, large interannual variability in bloom timing and/or temperature significantly changes zooplankton grazing from year to year, the key mechanism for altered organic matter transfer to upper-trophic levels and partitioning between the pelagic and benthic environments. Although bloom timing does not correlate with sea ice retreat (Figures 5 and 6b), bloom temperature fluctuates widely across a broad swath of the southeastern Bering Sea (Figures 6c and 9b) where the timing of sea ice retreat was highly variable (Figure 3a) and where blooms develop either at the ice edge or in open water (Figures 5c and 5d). Further to the south, beyond the shelfbreak and the reach of the annual sea ice pack, ice-edge blooms cannot develop and blooms are invariably warm (~5–6 °C; Figure 9). Further to the north, the timing of sea ice retreat was restricted to a narrow window in late spring and the bloom was always cold (~0-1 °C) and located at the ice-edge (even in 2001–2005, among the warmest years on record since 1948) (Figures 4, 5, 10). Thus, in the Bering Sea, the physical mechanisms of the OCH appear spatially constrained to the region of the fluctuating southern extent of sea ice (although different mechanisms may cause broader marine regions to alternate between bottom-up and top-down control, the postulated effect of the OCH). The fundamental mechanisms of the OCH likely do not function poleward of 60°N at this time, nor are they likely to be applicable to the Arctic Ocean, where ice-edge blooms are also proposed to be ubiquitous [Perrette et al., 2011].
 Using a combination of remote sensing, NCEP/NCAR reanalysis, and biophysical mooring data, we compiled long-term records to assess the impacts of sea ice on Bering Sea spring bloom dynamics and NPP. The impacts were especially pronounced in the southeastern Bering Sea due to large interannual fluctuations in sea ice extent and persistence (Figure 13). In this region, the timing of sea ice retreat did not alter the timing of the spring bloom but did alter bloom temperature by up to 5.4 °C. Early sea ice retreat (before mid-March in region 4) led to warm (>3 °C) open-water blooms that were triggered by a relaxation of winds. Late sea ice retreat led to cold (≤0.6 °C) ice-edge blooms whose timing was controlled by the timing of sea ice retreat. Furthermore, open-water blooms after early sea ice retreat were far more productive (>70%) than ice-edge blooms after late retreat. The lower NPP of ice-edge blooms may be partially due to reduced nutrients resulting from prior under-ice or ice-algal production not captured by satellites – quantifying these components of spring production is an important avenue of future research. Early retreat years were also more productive during summer, likely due to reduced thermal stratification. Overall, years of early sea ice retreat sustained 40-50% greater NPP in the southeastern Bering Sea. Stabeno et al. [2012a] predict that sea ice is likely to decline in this region (while continuing to be highly variable for several decades). This may lead to a more productive southeastern Bering Sea shelf with generally warmer open-water blooms, but is unlikely to precipitate any consistent shifts in bloom timing. The fate of this additional NPP will depend upon a complex mix of physical factors such as water temperature and stratification, and biological factors such as phytoplankton community shifts and zooplankton phenology.
 In the northern Bering Sea, sea ice retreat was far more consistent and blooms were invariably cold (~0-1 °C) and located along the retreating ice edge (Figure 13). For this reason, earlier sea ice retreat led to earlier ice-edge blooms. Earlier retreat also led to more productive ice-edge blooms, which again may be partly attributable to the shorter time available for under-ice or ice-algal production to consume nutrients that are required by the subsequent ice-edge bloom. In this region, Stabeno et al. [2012a] predict that substantial sea ice will remain, but that it may retreat earlier. Such a change will likely lead to earlier ice-edge blooms in cold waters, and potentially a lengthened growing season leading to enhanced annual NPP. However, Cooper et al.  point to a number of factors such as winds and late winter brine rejection that may make it difficult to predict how primary producers will respond to sea ice changes in the north.
 Sea ice extent and the timing of sea ice retreat on the Bering Sea eastern shelf have undergone no secular changes to date, making them poor candidates for explaining documented long-term ecosystem shifts [e.g., Hunt et al., 2002; Overland and Stabeno, 2004; Grebmeier et al., 2006]. In particular, the proposed impacts of the North Pacific “regime shift” of 1976–1977 should be reevaluated in light of the long-term SAT record, an excellent proxy for sea ice extent and timing of retreat, which exhibits no secular trend since 1948. Nevertheless, our results suggest that year-to-year sea ice variations can impact both the fate of organic carbon (pelagic versus benthic) and the overall carrying capacity of the ecosystem. Therefore, future climate change and sea ice reductions clearly have vast potential for ecosystem restructuring in the Eastern Bering Sea.
 This research was made possible by NSF grant ARC-0731524 and NASA grants NNX10AF42G to K. Arrigo and NNX11AL52H to Z. Brown. We thank the NSIDC for providing access to satellite-derived sea ice concentration data and the NASA GSFC for providing ocean color data. NOAA OI/SST V2 data and NCEP/NCAR reanalysis data were provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their website at http://www.esrl.noaa.gov/psd/. We thank P. Stabeno and D. Kachel (NOAA PMEL) for generously providing M2 and M4 mooring data.