Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

JOURNAL OF GEOPHYSICAL RESEARCH: OCEANS, VOL. 118, 1–13, doi:10.1002/jgrc.

20167, 2013

A comparison of methods to determine phytoplankton bloom initiation


Sarah R. Brody,1 M. Susan Lozier,1 and John P. Dunne2
Received 10 January 2013; revised 14 March 2013; accepted 17 March 2013.

[1] Phytoplankton bloom phenology has important consequences for marine ecosystems
and fisheries. Recent studies have used remotely sensed ocean color data to calculate
metrics associated with the phenological cycle, such as the phytoplankton bloom
initiation date, on regional and global scales. These metrics are often linked to physical or
biological forcings. Most studies choose one of several common methods for calculating
bloom initiation, leading to questions about whether bloom initiation dates calculated
with different methods yield comparable results. Here we compare three methods for
finding the date of phytoplankton bloom initiation in the North Atlantic: a biomass-based
threshold method, a rate of change method, and a cumulative biomass-based threshold
method. We use these methods to examine whether the onset of positive
ocean-atmosphere heat fluxes coincides with subpolar bloom initiation. In several
coherent locations, we find differences in the patterns of bloom initiation created by each
method and differences in the synchrony between bloom initiation and positive heat
fluxes, which likely indicate various physical processes at play in the study region. We
also assess the effect of missing data on the chosen methods.
Citation: Brody, S. R., M. S. Lozier, and J. P. Dunne (2013), A comparison of methods to determine phytoplankton bloom
initiation, J. Geophys. Res. Oceans, 118, doi:10.1002/jgrc.20167.

1. Introduction identify the date of bloom initiation in a manner that can


[2] Ocean primary productivity plays a fundamental role be efficiently and accurately applied to spatially extensive
in controlling the structure and health of marine ecosystems. ocean color data sets. Phenology studies currently use sev-
Phytoplankton bloom phenology is of particular interest eral methods to estimate the timing of a phytoplankton
because its coincidence with vulnerable phases of larval fish bloom. Ji et al. [2010] identify three broad categories of
[Platt et al., 2003] and zooplankton cycles [Koeller et al., methods. Threshold methods based on chlorophyll biomass
2009] may affect survival rates of those species [Quetin define bloom initiation as the time at which a chlorophyll a
et al., 1996; Koeller et al., 2009]. A recent study has time series [Siegel et al., 2002; Fleming and Kaitala, 2006;
suggested that climate change may affect the timing of Henson et al., 2006; Henson et al., 2009; Thomalla et al.,
bloom initiation [Kahru et al., 2011], leading to a pos- 2011; Racault et al., 2012; Cole et al., 2012] or a func-
sible decrease in the synchrony between phytoplankton tion or model fit to chlorophyll a data [Platt et al., 2009b;
blooms and upper trophic level life cycles [Edwards and Vargas et al., 2009; Wiltshire et al., 2008; Yamada and
Richardson, 2004; Mackas et al., 2007]. Such a possibil- Ishizaka., 2006; Zhai et al., 2011; Sasaoka et al., 2011;
ity emphasizes the importance of examining the drivers of Sapiano et al., 2012] crosses a set threshold. Threshold
phytoplankton bloom initiation and their vulnerability to methods based on cumulative chlorophyll biomass, which
interannual climate variability and change [Visser and Both, have mainly been used to investigate zooplankton phenol-
2005]. ogy, identify a bloom as the time at which a cumulative
[3] Remotely sensed ocean color data provide a valu- summation of chlorophyll biomass crosses a threshold per-
able tool for examining phytoplankton bloom initiation at centile of the total biomass [Greve et al., 2005; Mackas
the basin scale [Platt et al., 2009a]. However, before using et al., 2012]. Rate of change methods estimate bloom initi-
ocean color-derived chlorophyll a to investigate the con- ation from the point of most rapid increase on a chlorophyll
trols on phytoplankton bloom initiation, it is important to time series or function fit to that time series [Sharples et al.,
2006; Rolinski et al., 2007; White et al., 2009].
1
Division of Earth and Ocean Sciences, Duke University, Durham, [4] In the marine environment, biomass-based thresh-
North Carolina, USA. old methods have been most widely used in conjunction
2
Geophysical Fluid Dynamics Laboratory, Princeton, New Jersey, with ocean color data to identify bloom initiation and other
USA. phenological events in the marine environment at basin
Corresponding author: S. R. Brody, Division of Earth and Ocean and global scales. However, to date, we are aware of no
Sciences, Nicholas School of the Environment, Duke University, Box marine phenology studies critically comparing bloom initi-
90227, Durham, NC 27708, USA. (Sarah.Brody@duke.edu) ation dates derived from biomass-based threshold methods
©2013. American Geophysical Union. All Rights Reserved. with those derived from cumulative biomass-based thresh-
2169-9275/13/10.1002/jgrc.20167 old methods or rate of change methods. Here we compare
1
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

patterns of bloom initiation dates observed from these three bloom during the year [D’Ortenzio et al., 2012; Martinez
methods, investigate the drivers of discrepancies, and sug- et al., 2006]. In this study, “primary” blooms denote the
gest questions in the field of phytoplankton phenology high-magnitude blooms caused by the lifting of the major
research for which each method is best suited. We also exam- limiting factor on phytoplankton growth, while “secondary”
ine how the method we use to fill gaps in the chlorophyll blooms denote smaller blooms created by transient effects.
data affects our conclusions. For instance, at some midlatitude locations, after a light-
[5] Additionally, we present an application to the study driven primary bloom in the spring exhausts the surface
of bloom initiation dates by examining a possible driver of nutrient supply, a deepening mixed layer in the fall can pro-
subpolar phytoplankton blooms, as has been done in many vide a new source of nutrients, leading to a secondary bloom
of the previously cited studies [Siegel et al., 2002; Henson [Cushing, 1959; Platt et al., 2009b; Sapiano et al., 2012].
et al., 2006, 2009; Thomalla et al., 2011; Sasaoka et al., [8] For all the methods in our study, we use level 3
2011; Kahru et al., 2011; Racault et al., 2012]. For light- SeaWiFS chlorophyll a concentrations for the years
limited (subpolar) regions, the mechanics of phytoplank- 1998–2007 (all full years of SeaWiFS data), obtained from
ton phenology are generally thought to be consistent with http://oceancolor.gsfc.nasa.gov/. We downloaded the 9 km,
the Sverdrup hypothesis [Sverdrup, 1953]. The Sverdrup 8 day resolution SeaWiFS data, then spatially averaged the
hypothesis states that during winter months, deep mixed original data to half-degree resolution. We chose 8 day data
layers and low surface irradiance result in phytoplankton to circumvent the large amount of missing data and noise
consistently mixed to depths with too little light to sus- inherent in daily chlorophyll data yet still maintain suffi-
tain growth. A spring bloom can initiate when the winter cient temporal resolution to make meaningful connections
mixed layer shoals above a critical depth, where integrated between the timing of positive ocean-atmosphere heat fluxes
photosynthesis is equal to integrated respiration. and bloom initiation. We dealt with missing data in the 8 day,
[6] Recent studies have challenged the idea that season- half-degree data set by first filling in the small and ubiq-
ally shoaling mixed layers create the necessary conditions uitous data gaps due to cloud cover by averaging the four
to promote phytoplankton growth, and have proposed other nearest-neighbor points. We performed this spatial averag-
mechanisms that may play a role in subpolar bloom ini- ing twice, so that the farthest distance from which a missing
tiation. For example, Taylor and Ferrari [2011] posit that pixel could draw information was 1°. We then filled remain-
the cessation of turbulent convection in the upper mixed ing data gaps, associated with high-latitude, winter low sun
layer, caused by a reversal in atmosphere-ocean heat fluxes angles, with the annual minimum chlorophyll value for each
at the end of winter, may stabilize the upper mixed layer pixel’s yearly time series.
and that this stabilization, rather than the seasonal shoal- [9] We examine the relationship between ocean-
ing of the entire mixed layer, can initiate a phytoplankton atmosphere heat flux changes and bloom initiation using
bloom. They provide support for this theory in the subpolar daily net surface heat fluxes for the years 1998–2007,
North Atlantic by noting the synchrony between increases in obtained from NCEP/NCAR reanalysis 2 (http://www.esrl.
chlorophyll biomass and the onset of positive atmosphere- noaa.gov/psd/). The heat flux data was plotted on a T62
ocean heat fluxes or net warming of the ocean surface. Gaussian grid, which we averaged to half-degree resolution,
However, Mahadevan et al. [2012] have challenged these then temporally averaged to 8 day resolution. We employed
conclusions by noting that in their detailed study of the North the same method of filling small data gaps to the regridded
Atlantic Bloom Experiment, chlorophyll began increasing heat fluxes as in the chlorophyll data but did not fill in large
while heat fluxes were still negative. Here we assess the rela- data gaps with the time series minimum.
tionship between heat fluxes and phytoplankton blooms by
comparing the time at which heat fluxes become positive 2.2. Method 1: Rate of Change Method (ROC Method)
to the date of bloom initiation determined using the three [10] We designed a rate of change method for find-
different methods. ing bloom start dates (BSDs). We based this method on
HANTS-FFT (Harmonic analysis of time series-fast Fourier
2. Data and Methods transform) method [Roerink et al., 2000], which performed
well in a terrestrial phenology study [White et al., 2009].
2.1. Study Site and Data Our method first performs a discrete FFT on the 10 year
[7] We conduct our study of bloom phenology methods in SeaWiFS data set to obtain a set of Fourier coefficients
the North Atlantic over the spatial domain 30°N–65°N and at each location. The first 20 Fourier coefficients, corre-
80°W–0°W. Chlorophyll in the North Atlantic has a complex sponding to the 20 largest sinusoidal periods, are used
and variable temporal and spatial structure. Convention- to reconstruct the data set (Figure 1). The first frequency
ally, productivity in this region is divided into subtropi- has a period spanning the 10 year time series; the twen-
cal (nutrient-limited) and subpolar (light-limited) regimes tieth frequency has a 6 month period. To find the yearly
[Follows and Dutkiewicz., 2001; Henson et al., 2006]. The blooms, we first identify all of the local maxima in the
main bloom in subtropical areas occurs in fall or winter, reconstruction. We select the 10 largest maxima as the pri-
when deep mixed layers entrain nutrients into the euphotic mary bloom peaks, then check to see if any of those 10
zone, whereas the main bloom in subpolar areas occurs maxima occur less than 8 months apart. If so, one of the
in spring and summer, by the mechanisms explained in 10 largest maxima is a large secondary bloom, rather than
section 1. The transition zone between these two regimes a primary bloom, and we substitute the next-largest local
can experience blooms due to the lifting of light and/or maxima for the secondary bloom peak. After determin-
nutrient limitation [Henson et al., 2009]. Conditions at tran- ing the 10 primary bloom peaks in the reconstruction, we
sition zone latitudes may also permit a second, smaller define the bloom start date for the yearly blooms as the dchl
dt

2
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

a.
0.3

chlorophyll (mg m−3)


0.25
0.2
0.15
0.1
0.05
0
Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07
chl 10 coef. reconstruction 20 coef. reconstruction 30 coef. reconstruction
10 coef. BSD 20 coef. BSD 30 coef. BSD 20 coef. max 21−25 coeff. BSDs

b.
0.3
chlorophyll (mg m−3)

0.25
0.2
0.15
0.1
0.05
0
Sept’03 Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug

Figure 1. Chlorophyll time series (36°N, 59°W) (a) for all 9 years and (b) for 2003–2004, showing the
chlorophyll data (blue line) and the reconstruction using 10 (red line), 20 (black line), and 30 (green line)
Fourier coefficients. Diamonds show BSDs calculated for each reconstruction; filled black circles show
the bloom peaks selected using the 20-coefficient reconstruction. At the bottom of Figure 1a, the BSDs
calculated using the 20-coefficient reconstruction (black diamonds) are shown on the x axis, and the BSDs
calculated using the 21–25 coefficient reconstruction (red, green, blue, magenta, and cyan crosses) are
shown above. Dashed lines denote year start dates.

maximum prior to the bloom peak (Figure 1a). Because the the reconstruction (i.e., BSDs calculated using 20 coeffi-
ROC method determines BSDs based on the yearly chloro- cients and BSDs calculated using 21–25 coefficients are very
phyll maxima, results using this method have the advantage similar), but as increasing numbers of coefficients are used,
of not being sensitive to the time series start date or the the reconstruction will tend to over-emphasize secondary
season in which the primary bloom occurs. blooms. For example, the reconstruction using 30 coeffi-
[11] Using the first 20 Fourier coefficients, we produce cients resolves the secondary spring bloom in this time series
reconstructions of chlorophyll time series that display both to such a degree that it identifies the BSD as occurring imme-
spring and fall blooms, accurately reproduce the timing diately prior to the secondary bloom (Figure 1). Thus, we
of the chlorophyll peak, and remove variability at sub- conduct the remainder of our investigation with an ROC
seasonal and lower scales (Figure 1). We show an exam- method that uses 20 Fourier coefficients to reconstruct the
ple of the problems associated with using too few or too chlorophyll time series.
many Fourier coefficients in the reconstruction by plot-
ting a chlorophyll time series from 36°N, 59°W, with the 2.3. Method 2: Threshold Method (TH)
reconstructions and associated BSDs using 10, 20, and 30 [12] The threshold bloom initiation method was intro-
coefficients (Figure 1). This time series represents a sub- duced for marine phenology studies in Siegel et al. [2002].
tropical regime, in which deepening mixed layers in fall This method finds the yearly or climatological median of a
(September–November) entrain nutrients from below the chlorophyll time series, then identifies the bloom start date
mixed layer, creating a primary bloom that extends through as the first point at which chlorophyll levels rise a certain
the winter [Follows and Dutkiewicz, 2001]. For most years, percentage above the median. The BSDs identified by the
the time series also shows a secondary chlorophyll peak threshold method have been found to be relatively insensi-
occurring near the end of the primary bloom, possibly due to tive to the percentage used [Siegel et al., 2002]. Here we
increased spring irradiance prior to mixed layer depth shoal- used a threshold of 5% above the median to be consis-
ing and consequent nutrient depletion. The primary bloom tent with previous threshold-based phenology studies [Siegel
still begins in fall for these years, but the peak of the pri- et al., 2002; Henson et al., 2006, 2009; Thomalla et al.,
mary bloom occurs in spring. Reconstructions of the time 2011; Racault et al., 2012; Sapiano et al., 2012; Cole et al.,
series that use fewer than 20 coefficients do not resolve the 2012]. The threshold is calculated after filling in the missing
secondary peak; therefore, the bloom peak falls in winter data.
rather than in spring. While Figure 1 shows the results of the [13] The threshold method is often modified to mini-
10-coefficient reconstruction, the same problems occur for mize the misidentification of secondary blooms as primary
any reconstruction using fewer than 20 coefficients. Using blooms by shifting the chlorophyll time series to begin
more than 20 coefficients has a less significant impact on close to an estimated bloom season, for example, by starting
3
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

0.8
standard deviations above the yearly median and replacing
chl time series
these points with the quality-controlled time series max-
0.7 threshold
imum. The time at which this curve rises above a set
0.6
maximum point percentage of the total biomass is identified as the BSD. We
points below designed the CS method (illustrated in Figure 3) to be as
threshold
0.5 1 insensitive as possible to the time period in which the pri-
BSD mary bloom occurs by identifying the minimum point of
0.4 each yearly time series (cross in Figure 3) then shifting the
time series to put the minimum at the beginning (dashed
0.3 line). The rationale for this shift is the expectation that the
lowest chlorophyll level generally occurs prior to the pri-
2
0.2 mary bloom and will create a cumulative biomass curve
(dotted line) that does not significantly increase until the
0.1 primary bloom begins.
[16] BSDs identified by the CS method are sensitive to the
0
Sept Oct Nov Dec Jan Feb threshold used to define them, in contrast to the TH method.
We determined that a threshold of 15% of the total biomass
Figure 2. Half year of chlorophyll data (42°N, 33°W, (cross) best predicted the bloom initiation date for our study
1997) showing the threshold method, using a threshold of area using two techniques. First, we visually inspected mul-
5% above the yearly median chlorophyll. The straight line tiple chlorophyll time series over the entire study area with
denotes the threshold, the open circle shows the chloro- BSDs determined using six thresholds (5–30%). From these
phyll maximum or start point of the method, closed circles time series, we found that thresholds of 10–15% best pre-
show the points below the threshold used to determine the dicted the bloom initiation date in subtropical regions, while
BSD, and the diamond denotes the BSD. Arrows show threshold of 15–20% best predicted the bloom initiation date
the direction in which the TH algorithm searches for the in subpolar regions. We then plotted, for the six thresholds,
below-threshold points (arrow 1) and BSD (arrow 2). the number of occurrences in which chlorophyll levels at
each threshold’s BSD exceeded chlorophyll levels prior to
the time series in late summer for regions with fall pri- the BSD but were not larger than the yearly chlorophyll
mary blooms [Siegel et al., 2002; Henson et al., 2006, 2009; median plus one standard deviation. We found that the 15%
Thomalla et al., 2011]. Our method, illustrated in Figure 2, threshold had the largest number of points with increasing
calculates the BSD based on the time series maximum, rather chlorophyll levels at the BSD, which confirmed the results
than on an imposed time series start date. This calculation of the visual inspection and led to the choice of the 15%
is generally consistent with recent phenology studies [e.g., threshold.
Racault et al., 2012; Cole et al., 2012]. Our method finds the Dec Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov
maximum point of the time series (open circle in Figure 2), chl time series
1.2
then works backward from the maximum (arrow 1) to find minimum point
where chlorophyll levels go below the threshold for two con- shifted time series
secutive weeks (closed circles), in order to ensure that the 1
cumulative sum
chlorophyll (mg m−3)

identified decrease in chlorophyll is robust over the season threshold crossing


rather than a transient effect of data noise. We subsequently 0.8
BSD
identify the BSD as one point closer to the maximum than
the below-threshold points (arrow 2, diamond). 0.6
[14] While some previous studies implementing the
threshold method with remotely sensed chlorophyll data 0.4
have first fit a function or model to the data for smoothing
purposes before identifying the BSD [Platt et al., 2009b;
Vargas et al., 2009; Wiltshire et al., 2008; Yamada and 0.2

Ishizaka., 2006; Zhai et al., 2011; Sasaoka et al., 2011;


Sapiano et al., 2012], we find that using chlorophyll data 0
Sept Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug
that has already been temporally and spatially averaged and
requiring chlorophyll levels to stay below the threshold for
two consecutive weeks provides sufficient filtering of noisy Figure 3. Chlorophyll time series (21° W, 55° N, 2001)
data. To confirm this, we calculated TH BSDs using the showing the CS method. The thick line is the chlorophyll
Fourier reconstruction of the chlorophyll data described in time series, the cross is the minimum point of the time series,
section 2.2 and found very little difference between the TH the dashed line shows the shifted time series, the dotted line
BSDs calculated with reconstructed chlorophyll data and shows the cumulative sum of the chlorophyll time series
original chlorophyll data. (with the maximum value of the cumulative sum normalized
to the maximum value of the time series), the plus shows the
2.4. Method 3: Cumulative Sum Method (CS Method) threshold (15% of the total biomass), and the diamond shows
[15] Our third method is based on a cumulative biomass the BSD. The bottom axis represents the original chlorophyll
curve of the chlorophyll time series, created after screen- time series; the top axis represents the shifted chlorophyll
ing out points in each yearly time series greater than three time series and cumulative sum curve.
4
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

ROC TH CS
80°W 40°W 0° 80°W 40°W 0° 80°W 40°W 0°

a. 60°N b. 60°N c. 60°N


50° N 50° N 50° N

40° N 40° N 40° N

Sept Oct Nov Dec Jan Feb Mar Apr May Jun

TH minus ROC TH minus CS CS minus ROC


80°W 40°W 0° 80°W 40°W 0° 80°W 40°W 0°
6 5 60°N 6 5 60°N 6 5 60°N
d. 4 e. 4 f. 4
50° N 50° N 50° N
3 40° N
3 40° N
3 40° N
2 2 2
1 1 1

−20 −15 −10 −5 0 5 10 15 20

Figure 4. Bloom start dates calculated using (a) ROC, (b) TH, and (c) CS methods, and difference plots
of (d) TH minus ROC BSDs, (e) TH minus CS BSDs, and (f) CS minus ROC BSDs. All plots show the
average of the 10 year time series, with BSDs calculated using a time series start date of 1 September.
The scale of the bottom axis is in 8 day periods. Diamonds denote the locations of the time series in
Figures 6,9,11, and12.

3. Results 3.2. Comparison of Heat Flux Changes and BSDs


3.1. Comparison of Bloom Phenology Patterns [18] We expand upon the idea introduced in Taylor and
Ferrari [2011] by subtracting the 8 day period on which
[17] We begin our comparison by noting the similarities heat fluxes become positive (the “zero-crossing” or ZC)
and differences in the patterns of bloom initiation dates from the BSDs calculated using each method (Figure 5).
calculated using the three methods (Figures 4a–4c). All The difference plots are shown with a discretized colorbar
methods show the same major pattern in BSDs: a sharp tran- and a highlighted zero-contour to facilitate the comparison.
sition between subtropical latitudes, where blooms occur in In the subtropics, blooms occur well before the heat flux
the fall and winter, and subpolar latitudes, where blooms ZC, as expected, since phytoplankton growth in these areas
occur in the spring and summer, indicating that this feature is not light-limited. The largest difference between BSDs
of North Atlantic bloom phenology is robust to the method and ZCs in the subtropics occurs at the Gulf Stream, where
used to identify blooms. However, in several areas, bloom consistently warm sea surface temperatures delay the zero-
phenology patterns differ between the methods, with dif- crossing. Over most of the subpolar region, ROC and CS
ferences that can exceed 2 months. We have highlighted BSDs (Figures 5a and 5c) are approximately synchronous
these areas by plotting the difference between BSDs calcu- with ZCs, occurring within four 8 day periods of each other.
lated using each method (Figures 4d–4f). First, ROC BSDs Both of these methods show patchiness in whether the BSD
show a band of early blooms centered at approximately occurs before or after the zero-crossing. Almost all subpo-
36°N–39°N, which TH and CS BSDs do not show. Sec- lar TH BSDs (Figure 5b) occur after the zero-crossing, with
ond, CS BSDs shift the transition between subtropical and some areas occurring over four 8 day periods, or 32 days,
subpolar regimes northward, especially in the central and after the zero-crossing. We use the same time series as in
eastern North Atlantic. Third, in the eastern North Atlantic, Figures 4d–4f (locations plotted in Figures 5a–5c as purple
between approximately 55°N and 60°N, the TH method diamonds) to examine the relationship between the heat flux
calculates late blooms relative to the rest of the subpolar ZC and increases in chlorophyll.
study area, while the ROC and CS methods show blooms
in this area occurring at the same time as, or slightly ear-
lier than, the rest of the subpolar North Atlantic. Finally, 3.3. Time Series 1 and 2
all three methods show an area of earlier blooms west of [19] We use time series 1 and 2 (Figure 4), both located
Greenland, but this feature is exaggerated in ROC blooms. in the nutrient-limited subtropics, to determine why the
In sections 3.3–3.6, we will explore the reason for these ROC method shows early blooms in a band centered at
differences using time series located at the diamonds in 36°N–39°N (time series 2), compared with the rest of the
Figures 4d–4f. subtropical region (time series 1). Figure 6 shows the time
5
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

ROC minus ZC difference between the two BSDs (TH minus ROC) is –2.7
80° W 40° W 0° periods. In contrast, at 36°N (time series 2), TH BSDs often
a. 60° N 12 occur when the bloom is well underway, with an average
°
50 N
difference between the two BSDs of 8.5 periods.
[20] Time series 1 and 2 also show that at 36°N, the yearly
40° N 8 median threshold is elevated relative to the time series, lead-
ing to later TH blooms than at 32°N. The elevated median
4
appears to be caused by the longer duration of the blooms
TH minus ZC at 36°N, as compared with 32°N. Indeed, an overlay of the
80° W 40° W 0° Fourier reconstructions of the two time series (Figure 7)
0 shows that at 36°N the bloom both begins earlier and lasts
b. 60° N
longer than at 32°N. The earlier blooms are likely not iden-
50° N tified by the TH method because the longer bloom duration
−4 elevates the median threshold relative to the chlorophyll time
40° N
series.
−8
[21] We generalize this observation for the entire band of
early ROC BSDs at 36°N–39°N (Figure 4a) by defining a
CS minus ZC metric called the relative median position:
80° W 40° W 0° −12
c. 60° N
chlmax – chlmin
50° N . (1)
−16 chlmax – chlmedian
40° N

−20
As seen in the schematic (Figure 8a), long-duration blooms
(solid curve) tend to have higher relative median positions
Figure 5. (a) ROC, (b) TH, and (c) CS BSDs minus the
than short-duration blooms (dashed curve). Plotting relative
8 day period during which heat fluxes become positive
median positions for the study area (Figure 8b) shows that
(zero-crossing or ZC). The zero-contour is shown in black.
the band of ROC-calculated early blooms at 36°N–39°N
Monthly heat flux data was downloaded from http://www.
corresponds with high relative median positions. Thus, our
esrl.noaa.gov/psd/data/gridded/data.godas.html and linearly
conclusion from time series 1 and 2 that high relative median
interpolated to match the 8 day SeaWiFS data.
positions create earlier blooms likely holds for the entire
band of early blooms.
series plotted for the nine full years of data. At both time [22] CS BSDs generally show slightly earlier blooms than
series, ROC BSDs occur as chlorophyll levels are beginning either ROC or TH BSDs for the subtropical regions in the
to increase. At 32°N (time series 1), TH BSDs occur at the domain but, like TH BSDs, do not show a band of espe-
same time as or slightly earlier than ROC BSDs. The average cially early blooms at 36°N–39°N for similar reasons: the

Time Series 1
chlorophyll (mg m−3)

0.3
0.25
0.2
0.15
0.1
0.05
Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Time Series 2
chlorophyll (mg m−3)

0.3
0.25
0.2
0.15
0.1
0.05
Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Figure 6. Time series with BSDs calculated using all three methods at 32°N and 36°N. The values of
the CS threshold have been normalized to the values of the chlorophyll time series. Dashed lines indicate
September of each year.
6
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

° °
0.25 32 N 36 N

chlorophyll (mg m−3)


0.2

0.15

0.1

0.05

Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Figure 7. Fourier reconstructions of both time series in Figure 6.

longer bloom duration, which increases the total cumula- well with the band of early CS BSDs, supporting the
tive biomass, raises the cumulative sum threshold compared explanation that where CS BSDs occur much earlier than TH
with other subtropical latitudes. The band of early blooms at and ROC BSDs, they identify the start of secondary blooms.
36°N–39°N coincides approximately with the Gulf Stream,
which creates increased eddy kinetic energy and mixing in 3.5. Time Series 4 and 5
this area [Richardson, 1983; Fratantoni, 2001]. Therefore, [25] Time series 4 and 5 (Figure 11) are both located in a
increased mixing, and the consequent increase in entrained relatively large area within the light-limited, subpolar North
nutrients, likely create the conditions necessary to support a Atlantic, from approximately 55°N to 60°N, where TH
longer phytoplankton bloom period. BSDs occur later than either CS or ROC BSDs (Figure 4).
Examining these time series shows that blooms in this area
3.4. Time Series 3
[23] Time series 3 (Figure 9) represents an area in which
the CS method shows fall and winter BSDs, consistent with a.
subtropical areas, while TH and ROC BSDs show spring
BSDs, consistent with subpolar regions. Figures 4e and 4f maximum
show the degree to which the location of the subtropical-
subpolar transition differs between CS BSDs versus TH
and ROC BSDs. The shape of the chlorophyll data and its
Fourier reconstruction in time series 3 indicate that in most maximum
years this area experiences a primary bloom in the spring and
a secondary bloom in the fall, likely due to the mechanisms
described in section 2.1 and consistent with the findings of median
Martinez et al. [2011] and D’Ortenzio et al. [2012]. The tim-
ing of the heat flux ZC just prior to or coincident with the median
primary bloom provides further evidence for a light-limited
spring bloom and a nutrient-limited fall bloom in this region.
Further, the minimum chlorophyll levels occur between the 80° W 40° W 0°
blooms during the highly stratified, nutrient-depleted sum-
b.
mer months. These two observations explain the early CS 60° N
BSDs: for the time series used in the CS method, which is
shifted to start at the minimum chlorophyll point, the sec-
ondary bloom occurs before the primary bloom, and the 50° N
chlorophyll increase during the secondary bloom is large
enough to raise the cumulative chlorophyll sum above the
prescribed threshold. The other two methods, which search 40° N

for the BSD in relation to the yearly maximum of the


chlorophyll time series, identify the BSD very close to the
heat flux ZC.
[24] The conclusions drawn from time series 3 can be
extended to the entire area of early CS BSDs by plotting 1 1.2 1.4 1.6 1.8 2
the magnitude of the 20th Fourier coefficient, which corre-
sponds to the bi-yearly bloom period, divided by the sum of Figure 8. (a) Schematic diagram illustrating the relative
the magnitudes of all the Fourier coefficients, over the study median position metric (vertical solid line divided by vertical
domain (Figure 10). Areas on the map where the bi-yearly dashed line). The solid curve is an example of a low relative
bloom period is large compared with the other periods can be median position; the dashed curve is an example of a high
used to estimate areas where there are prominent spring and relative median position. (b) Relative median positions,
fall blooms. The band of spring and fall blooms corresponds averaged over the 9 year time series.
7
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

Time Series 3

chlorophyll (mg m−3)


0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Figure 9. Time series with BSDs calculated using all three methods shown for 31°W, 43°N. The values
of the CS threshold have been normalized to the values of the chlorophyll time series. Dashed lines
indicate September of each year.

begin with several small increases in chlorophyll prior to the study, the heat flux ZC also occurs at the beginning of the
chlorophyll maximum. The difference in average TH ver- phytoplankton bloom, indicating that stabilization due to
sus CS and ROC BSDs is driven by large differences in net ocean surface warming may also play a role in driv-
BSDs for a few years out of each time series (shaded years ing blooms here. However, BSDs calculated with the ROC
in Figure 11) rather than systemic differences in BSDs. For method are earlier than blooms calculated using TH and CS
the years that do contain late-occurring TH BSDs, the dif- methods. Figure 12 shows that the chlorophyll time series
ference is due to an especially large decrease (two or more in this area is composed of long periods of very low chloro-
weeks below the threshold) in chlorophyll levels between phyll punctuated by brief, high-magnitude blooms. The low
the small chlorophyll spikes and the chlorophyll maximum, chlorophyll levels throughout the fall, winter, and spring
which the TH algorithm identifies as the seasonal minimum months are likely due to a combination of low irradiance and
prior to the bloom. This effect is likely accentuated by high the large amount of missing data in this region, which is then
relative median positions in this region of the North Atlantic, filled with the yearly time series minimum (see section 3.7).
as compared to the western North Atlantic at the same lati- Comparing the Fourier reconstruction of the data to the
tudes (Figure 8). By contrast, the Fourier reconstruction used chlorophyll time series shows that the 20 coefficients chosen
to identify ROC BSDs smooths the multiple chlorophyll to create the reconstruction cannot replicate the extremely
increases into one continuous bloom, and, in most years, the short duration of blooms at this latitude. While the timing
low-magnitude chlorophyll increases are still large enough of the chlorophyll maxima in the reconstruction is consistent
to raise the cumulative chlorophyll sum used to calculate CS with the timing of the maxima in the time series, the recon-
BSDs above the 15% threshold. structed blooms are much longer and thus start earlier than
[26] There are several possible explanations for the the blooms in the time series, leading to ROC BSDs that
observed low-magnitude increases in chlorophyll prior to occur during the late winter.
the main bloom. This pattern could be driven by biology,
where each chlorophyll spike represents a successive phy- 3.7. Effect of Filling Missing Data
toplankton species that gets quickly grazed down until a
larger phytoplankton species establishes a long-duration, [28] While remotely sensed ocean color data provides a
high-magnitude bloom [Sommer, 1985; Taylor et al., 1993]. valuable resource for examining phytoplankton bloom phe-
Alternatively, this pattern could be driven by physical pro- nology, gaps in the data, especially at high latitudes, can
cesses. The mechanism for bloom initiation described in affect the quality of phenology estimates. Cole et al. [2012]
Taylor and Ferrari [2011] and in section 1 could allow for
more transient chlorophyll increases due to mixed layer sta-
bilization before the onset of seasonal stratification. This
explanation finds support in our results: for all years in which
TH blooms are delayed relative to ROC blooms, the heat flux
ZC occurs at roughly the same time as the small chlorophyll
increases identified by the ROC method as the BSD.

3.6. Time Series 6


[27] Time series 6 (Figure 12) is located west of
Greenland, also a light-limited area for phytoplankton
growth. In this area, all three methods show earlier BSDs
than at the same latitude in other locations (Figure 4). This
pattern is also seen in the phenology study by Henson et al.
[2009] and is attributable to early stratification induced by Figure 10. Magnitude of the 20th Fourier coefficient (cor-
westward advection of relatively light freshwater melt from responding to a 6 month or bi-yearly bloom period) divided
Greenland [Frajka-Williams et al., 2009]. In the present by the sum of the magnitudes of all coefficients.
8
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

Time Series 4

chlorophyll (mg m−3)


1.5

0.5

0
Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Time Series 5
1.2
chlorophyll (mg m−3)

0.8
0.6

0.4

0.2

Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Figure 11. Time series with BSDs calculated using all three methods shown for 21°W, 55°N and 38°W,
59°N. The values of the CS threshold have been normalized to the values of the chlorophyll time series.
Shading indicates years in which TH BSDs occur later than ROC and CS BSDs. Dashed lines indicate
September of each year.

provided a comprehensive study of the effect of missing section 2.1. However, the BSDs calculated using a data set
data on the TH method of estimating BSDs by comparing with only spatial filling were virtually identical to those cal-
BSDs and bloom peak dates derived from a SeaWiFS data culated using a data set where no filling was applied, leading
set containing missing data with those derived from a SeaW- to the conclusion that the filling of large data gaps with the
iFS data-assimilating biogeochemical model. Similar studies time series minimum has the largest effect on BSDs. When
conducted for other methods of determining BSDs would be we created the same difference plots as in Figures 4d–4f
a useful addition to phenology research. However, in this using unfilled data (Figure 13), very similar general pat-
study, our aims in examining missing data are twofold: one, terns appear, although the magnitudes of those patterns
to ensure that our method of filling the missing data does not differ (compare Figure 4d with 13a, Figure 4f with 13b,
strongly affect our comparisons between the methods; and and Figure 4e with 4c and 4d). In fact, when comparing
two, to briefly comment upon the effect of filling data gaps BSDs using the unfilled data, the correspondence between
with the time series minimum on the BSDs we observe for areas with early CS BSDs (Figures 13b and 13d) and areas
the TH and CS methods. with prominent secondary blooms strengthens. We therefore
[29] To address our first aim, we calculated BSDs using conclude that our findings regarding the reason for the dif-
the TH and CS methods applied to a data set with no data ferences in BSDs in sections 3.3 through 3.6 are independent
filling. We could not conduct this calculation for the ROC of the amount of missing data in each time series and the
method, because performing a Fourier reconstruction on the method of filling missing data.
chlorophyll data requires a gap-free data set. We also calcu- [30] We address our second aim by examining the differ-
lated BSDs using the TH and CS methods for a data set that ences between TH and CS BSDs using filled and unfilled
had been filled using only the spatial averaging described in data. Figure 14a shows the average percentage of missing

Time Series 6
10
chlorophyll (mg m−3)

0
Sept’98 ’99 ’00 ’01 ’02 ’03 ’04 ’05 ’06 ’07

Figure 12. Time series with BSDs calculated using all three methods shown for 53°W, 62°N. The values
of the CS threshold have been normalized to the values of the chlorophyll time series. Dashed lines
indicate September of each year.
9
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

unfilled TH minus ROC unfilled CS minus ROC


80° W 40° W 0° 80° W 40° W 0°
rather than the late summer, and thus the bloom will occur
later (farther from 1 September). This explains why the
a.a. 60° N b.
b. 60° N
unfilled CS BSDs have an even higher propensity than the
50° N 50° N filled BSDs to identify secondary, fall blooms as primary
40° N 40° N blooms. North of 50°N, the large amount of missing data
that has been given a value in the filled data set, rather than
unfilled TH minus CS TH minus unfilled CS
treated as a missing, zero value in the unfilled data set,
80° W 40° W 0° 80° W 40° W 0° causes the filled cumulative sum curve to start increasing
c.c. 60° N d. 60° N
during the period of missing data, and thus rise above the
d.
50° N 50° N
threshold earlier than in the unfilled case.
40° N 40° N

4. Discussion
[32] Our examination of the differences between the
−20 −10 0 10 20 ROC, TH, and CS methods for calculating BSDs has
revealed areas with interesting chlorophyll seasonal cycles
Figure 13. Difference between BSDs calculated with the and raised several questions about the biological and phys-
three methods, with either (a and c) the TH BSDs or (b and ical underpinnings of those cycles. We find that the TH
d) the CS BSDs calculated using unfilled chlorophyll data. method, as we are implementing it, is likely to identify BSDs
The scale of the colorbar axis is 8 day periods. as occurring during or immediately after the largest increases
in chlorophyll concentrations, especially when the relative
median position of the chlorophyll time series is large (loca-
data per pixel for the study area, which, as expected,
increases at higher latitudes. Comparing filled and unfilled
missing data 1
TH-calculated BSDs (Figure 14b) shows that south of
approximately 45°N, filled TH BSDs are slightly later than 80° W 40° W 0°

unfilled BSDs, especially off the coast of Cape Cod, where


60° N
the percentage of missing data per pixel is somewhat higher a.
than the surrounding areas. North of 50°N, filled TH BSDs 50° N 0.5
are generally earlier than unfilled TH BSDs. Time series
from these areas (not shown) reveal that south of 45°N, data 40° N
filling tends to occur in winter, when chlorophyll levels are
high. While the gaps at these latitudes are small enough that
data filling does not noticeably affect the median thresh- 0
old, the gaps can occur as chlorophyll levels are increasing
prior to a bloom. If the gaps are two weeks or longer, filling TH minus TH unfilled
the gaps with the minimum chlorophyll value can delay the 80° W 40° W 0°
filled BSDs until after the gap, even if the chlorophyll time 20
series has increased above the threshold prior to the gap. 60° N
North of 50°N, where the amount of missing data per pixel b.
is higher, filled TH BSDs occur earlier than unfilled BSDs 50° N
because filling large data gaps with the time series minimum
10
lowers the median threshold for filled data as compared to 40° N
unfilled data. The tendency of filled data to produce earlier
blooms than unfilled data at subpolar latitudes is consistent
with Cole et al. [2012], while the tendency of filled data to
produce later blooms than unfilled data at subtropical lati- 0
tudes differs from Cole et al. [2012]. In that study, the filled CS minus CS unfilled
data set, which came from a data-assimilating biogeochemi- 80° W 40° W 0°
cal model, did not replace gaps in the chlorophyll data with
60° N
the yearly minimum chlorophyll value. c. −10
[31] Filled versus unfilled CS-calculated BSDs
50° N
(Figure 14c) show the same general patterns as TH BSDs.
In this case, filled CS BSDs occur later than unfilled BSDs 40° N
south of 50°N because filling the data set with the time
−20
series minimum can change the position of the first occur-
rence of the chlorophyll minimum after 1 September, which
has been set as day 1 of the year for all methods. Since the Figure 14. (a) Percentage of 8 day periods with missing
chlorophyll time series minimum generally occurs in the data per year, averaged for the 10 year time series. (b and
late summer or early fall in these regions and data gaps c) Comparison of TH BSDs minus TH BSDs calculated
occur in the winter, using filled data will cause the time using the unfilled data set (b) and CS BSDs minus CS BSDs
series used to calculate the CS BSD to start in the winter calculated using the unfilled data set (c).
10
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

tions 1 and 2, Figure 6) or the start of the bloom consists the relevant feature of the chlorophyll times series. The CS
of successive, low-magnitude chlorophyll increases (loca- method is also very sensitive to the date used for the start of
tions 4 and 5, Figure 11). Based on these findings, the TH the time series and thus cannot be implemented at the basin
method might be especially appropriate for investigating scale using a fixed start date. Our method of shifting each
the match or mismatch between phytoplankton and upper time series to start at its minimum chlorophyll level, based
trophic levels [e.g., Edwards and Richardson, 2004; Mackas on the assumption that the yearly chlorophyll minimum
et al., 2007; Koeller et al., 2009], because the match- precedes the primary bloom, accounts for that sensitivity.
mismatch hypothesis is based on the timing of the high However, location 3 (Figure 9) shows an area where the
phytoplankton biomass period [Cushing, 1959]. The way in yearly chlorophyll minimum precedes the secondary bloom,
which we chose to implement the TH method contributed causing the CS method to misidentify secondary fall blooms
somewhat to these conclusions, especially at locations 4 and as primary blooms and thus falsely extend the subtropical
5 and in areas where data gaps filled with the time series region of the North Atlantic. Shifting the time series in this
minimum occurred near the beginning of the bloom, by manner, combined with our method of data filling, also cre-
potentially increasing the likelihood that the identified BSD ated differences in filled and unfilled CS BSDs. However,
would occur during a period of high chlorophyll growth. when examining phytoplankton blooms using basin-scale,
It would be possible to modify the method we used by remotely sensed chlorophyll data, the need to efficiently
increasing the number of 8 day periods below the thresh- process large data sets necessitates the use of simplifying
old required to find a bloom, although this heightens the assumptions such as the one we made. Future use of the CS
risk of identifying isolated chlorophyll spikes as the main method in remotely sensed phytoplankton phenology stud-
bloom. Starting the search for points above the threshold ies should be accompanied by a careful treatment of the time
from the beginning of the time series is a more commonly series start date, and it may be found that the CS method
used TH method and might also provide somewhat differ- works better as a way of identifying the first bloom in a
ent results with chlorophyll time series similar to those at time series regardless of whether that bloom is a primary or
locations 4 and 5. Implementing the threshold method in this secondary bloom.
manner increases its sensitivity to the time series start date [35] The discussion of positive heat fluxes as a possible
and, depending on the start date chosen, could also be used driver of subpolar phytoplankton blooms in section 3.2 illus-
to identify the timing of secondary blooms. Finally, the dis- trates the importance of carefully considering which method
crepancies in our comparison of TH BSDs from filled and to use for an investigation into phytoplankton bloom phe-
unfilled data with the findings of Cole et al. [2012] indi- nology, since comparing the heat flux zero-crossing to the
cate that temporal interpolation of missing data may be a ROC-, TH-, and CS-calculated BSDs (Figure 5) produced
more appropriate method of filling data gaps in subtropical different results. Comparing subpolar ROC or CS BSDs to
areas, where data gaps often occur as chlorophyll levels are the zero-crossing (Figures 5b and 5c) leads to the conclu-
increasing prior to the bloom. sion that blooms begin very near the time that heat fluxes
[33] The ROC method identifies blooms as starting when become positive, consistent with the theory and results of
chlorophyll is increasing rapidly from the pre-bloom min- Taylor and Ferrari [2011]. Additionally, because in some
imum, but biomass levels may still remain low. The ROC areas BSDs can occur before heat fluxes become positive,
method, therefore, could be useful in examining the sea- ROC and CS BSDs support the findings of [Mahadevan
sonal physical or biological mechanisms that create condi- et al., 2012]. In contrast, comparing subpolar TH BSDs to
tions in which a bloom can occur [e.g., Behrenfeld, 2010]. the zero-crossing (Figure 5d) leads to the conclusion that,
The Fourier reconstruction method we used to smooth the because BSDs occur well after the zero-crossing, positive
chlorophyll data to implement the ROC method negatively heat fluxes alone are not sufficient to initiate a bloom. These
impacted the ROC BSDs at location 6 (Figure 12) and may two conclusions can be reconciled by the patterns seen in
have had similar impacts at other high-latitude locations. time series 4 and 5 (Figure 11): during years with sev-
Alternative methods of reconstructing the chlorophyll data, eral low-magnitude chlorophyll increases prior to the main
such as the Gamma Generalized Linear Model method in chlorophyll increase, the TH method identifies BSDs as
Vargas et al. [2009] and Sapiano et al. [2012] or the method occurring after the low-magnitude increases, while the ROC
of filtering coefficients using a Lanczos filter [Duchon, and CS methods identify BSDs as occurring at the beginning
1979] could be tested in the future to determine whether of the low-magnitude chlorophyll increases. It is possible
they produce different ROC BSD,s and which smoothing that the low-magnitude chlorophyll increases identified by
technique best reconstructs the chlorophyll time series. the ROC and TH methods could be caused by a decrease in
[34] While the CS method, like the TH method, depends turbulent convection driven by heat fluxes becoming posi-
on a biomass-based threshold, the sensitivity of the CS BSDs tive, while the high-magnitude chlorophyll increase identi-
to the particular threshold creates flexibility in the feature fied by the TH method could be caused by the seasonally
of the chlorophyll time series that the CS method identifies shoaling mixed layer, consistent with the Sverdrup hypothe-
as a bloom. The 15% of the cumulative biomass thresh- sis. However, while the results of our investigation support
old we used produced BSDs more closely aligned with the the theory advanced by Taylor and Ferrari [2011], these
first increases in chlorophyll biomass, while a 30% threshold results do not provide definitive evidence for the onset of
would produce BSDs associated with the largest increases positive ocean-atmosphere heat fluxes as opposed to sea-
in biomass. Thus, the CS method could be used for either sonal mixed layer shoaling as a driver of subpolar bloom
of the purposes we identified for the TH and ROC methods. initiation. First, the onset of positive ocean-atmosphere heat
However, this feature of the CS method also necessitates fluxes may coincide with seasonal mixed layer shoaling. If
a careful choice of the threshold to ensure that it identifies this is the case, it would not be clear whether one factor was
11
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

primarily responsible for phytoplankton bloom initiation. research and future directions, J. Plankton Res., 32(10), 1355–1368, doi:
Second, net warming of the ocean surface is not the only 10.1093/plankt/fbq062.
Kahru, M., V. Brotas, M. Manzano-Sarabiaz, and B. Mitchell (2011), Are
factor that could stratify the upper mixed layer of the ocean. phytoplankton blooms occurring earlier in the Arctic? Global Change
For example, the mechanism for bloom initiation described Biol., 17, doi:10.1111/j.1365-2486.2010.02312.X.
in Mahadevan et al. [2012] relies on eddy-induced stratifi- Koeller, P., et al. (2009), Basin-scale coherence in phenology of shrimps
and phytoplankton in the North Atlantic ocean, Science, 324 (5928),
cation of the mixed layer. Factors such as wind-generated 791–793, doi:10.1126/science.1170987.
mixing and freshwater fluxes from seasonal sea ice melt, Mackas, D., S. Batten, and M. Trudel (2007), Effects on zooplankton of
as we briefly discussed in section 3.6, may also play a role a warmer ocean: Recent evidence from the Northeast Pacific, Prog.
in upper mixed layer stratification. While a more detailed Oceanogr., 75, 223–252.
Mackas, D., et al. (2012), Changing zooplankton in a changing ocean:
investigation of the potential mechanisms that create and Comparing time series of zooplankton phenology, Prog. Oceanogr., 97,
break down chlorophyll increases prior to the main bloom 31–62, doi:10.1016/j.pocean.2011.11.005.
would need to address these questions, the heat flux study Mahadevan, A., E. D’Asaro, C. Lee, and M. Perry (2012), Eddy-driven
stratification initiates North Atlantic spring phytoplankton blooms,
we present here is an example of why comparisons of mul- Science, 337, 54–58, doi:10.1126/science.1218740.
tiple BSD methods can be useful and why the selection of Martinez, E., D. Antoine, F. D’Ortenzio, and D. de Boyer Montégut (2011),
a definition for the onset of a phytoplankton bloom appro- Phytoplankton spring and fall blooms in the North Atlantic in the 1980s
priate to the question being asked is important to consider and 2000s, J. Geophys. Res., 116, C11029, doi:10.1029/2010JC006836.
Platt, T., C. Fuentes-Yaco, and K. Frank (2003), Spring algal bloom and
when selecting a method for determining the date of bloom larval fish survival, Nature, 423, 398–399, doi:10.1038/423398b.
initiation. Platt, T., S. Sathyendranath, G. White, C. Fuentes-Yaco, L. Zhai, E. Devred,
and C. Tang (2009a), Diagnostic properties of phytoplankton time series
from remote sensing, Estuaries Coasts, 33, doi:10.1007/s12237-009-
[36] Acknowledgments. SeaWiFS chlorophyll a data were obtained 9161-0.
using the NASA ocean color database. Heat flux data were obtained from Platt, T., G. White III, L. Zhai, S. Sathyendranath, and S. Roy
the ESRL/PSD database. This work was funded by a NASA Earth and (2009b), The phenology of phytoplankton blooms: Ecosystem indi-
Space Science Fellowship and by a NOAA MPOWIR fellowship. We would cators from remote sensing, Ecol. Model., 220, 3057–3069, doi:
like to thank R. Asch, who provided helpful suggestions and comments dur- 10.1016/j.ecolmodel.2008.11.022.
ing the development of this work. We would also like to acknowledge three Quetin, L., R. Ross, T. Frazer, and K. Haberman (1996), Factors affecting
anonymous reviewers, whose helpful suggestions and comments greatly distribution and abundance of zooplankton with an emphasis on Antarctic
improved this paper. krill, Euphausia superba, Antar. Res. S., 70, 357–371.
Racault, M., C. Le Quéré, E. Buitenhuis, S. Sathyendranath, and T. Platt
(2012), Phytoplankton phenology in the global ocean, Ecol. Indic., 14,
152–163.
References Richardson, P. (1983), Eddy kinetic energy in the North Atlantic from
surface drifters, J. Geophys. Res., 88(C7), 4355–4367.
Behrenfeld, M. (2010), Abandoning Sverdrup’s Critical Depth Hypothesis Roerink, G., M. Menenti, and W. Verhoef (2000), Reconstructing cloud-free
on phytoplankton blooms, Ecology, 91, 977–989, doi:10.1890/09-1207.1. NDVI composites using Fourier analysis of time series, Int. J. Remote
Cole, H., S. Henson, A. Martin, and A. Yool (2012), Mind the gap: The Sens., 21, 1911–1917.
impact of missing data on the calculation of phytoplankton phenology Rolinski, S., H. Horn, T. Petzoldt, and L. Paul (2007), Identifying cardinal
metrics, J. Geophys. Res., 117, C08030, doi:10.1029/2012JC008249. dates in phytoplankton time series to enable the analysis of long-term
Cushing, D. (1959), The seasonal variation in oceanic production as a trends, Oecologia, 153, 997–1008, doi:10.1007/s00442-007-0783.
problem in phytoplankton dynamics, Journ. Conseil., 24, 455–464. Sapiano, M., C. Brown, U. Schollaert S., and M. Vargas (2012), Establish-
D’Ortenzio, F., D. Antoine, E. Martinez, and M. R. d’Alcalá (2012), Phe- ing a global climatology of marine phytoplankton phenological charac-
nological changes of oceanic phytoplankton in the 1980s and 2000s teristics, J. Geophys. Res., 117, C08026, doi:10.1029/2012JC007958.
as revealed by remotely sensed ocean-color observations, Global Bio- Sasaoka, K., S. Chiba, and T. Saino (2011), Climatic forcing and phyto-
geochem. Cycles, 26, GB4003, doi:10.1029/2011GB004269. plankton phenology over the subpolar North Pacific from 1998 to 2006,
Duchon, C. (1979), Lanczos filtering in one and two dimensions, J. Appl. as observed from ocean color data, Geophys. Res. Lett., 38, L15609, doi:
Meteorol., 18, 1016–1023. 10.1029/2011GL048299.
Edwards, M., and J. Richardson (2004), Impact of climate change on marine Sharples, J., O. Ross, B. Scott, S. Greenstreet, and H. Fraser (2006), Inter-
pelagic phenology and trophic mismatch, Nature, 430, 881–884, doi: annual variability in the timing of stratification and the spring bloom
10.1038/nature02808. in the North-western North Sea, Cont. Shelf Res., 26, 733–751, doi:
Fleming, V., and S. Kaitala (2006), Phytoplankton spring bloom inten- 10.1016/j.csr.2006.01.011.
sity index for the Baltic Sea estimated for the years 1992 to 2004, Siegel, D., S. Doney, and J. Yoder (2002), The North Atlantic spring phyto-
Hydrobiologia, 554, doi:10.1007/s10750-005-1006-7. plankton bloom and Sverdrup’s Critical Depth Hypothesis, Science, 296
Follows, M., and S. Dutkiewicz (2001), Meteorological modulation of the (5568), 730–733, doi:10.1126/science.1069174.
North Atlantic spring bloom, Deep Sea Res. Part II, 49, 321–344, doi: Sommer, U. (1985), Seasonal succession of phytoplankton in Lake
10.1016/S0967-0645(01)00105-9. Constance, BioScience, 35(6), 351–357.
Frajka-Williams, E., P. Rhines, and C. Eriksen (2009), Physical controls Sverdrup, H. (1953), On conditions for the vernal blooming of phytoplank-
and mesoscale variability in the Labrador Sea spring phytoplankton ton, Journ. Conseil., 18, 287–295.
bloom observed by Seaglider, Deep Sea Res. Part I, 56, 2144–2161, doi: Taylor, A., D. Harbour, R. Harris, P. Burkill, and E. Edwards (1993), Sea-
10.1016/j.dsr.2009.07.008. sonal succession in the pelagic ecosystem of the North Atlantic and the
Fratantoni, D. (2001), North Atlantic surface circulation during the 1990’s utilization of nitrogen, J. Plankton Res., 15(8), 875–891.
observed with satellite-tracked drifters, J. Geophys. Res., 106 (C10), Taylor, J., and R. Ferrari (2011), Shutdown of turbulent convection as a
22,067–22,093. new criterion for the onset of spring phytoplankton blooms, Limnol.
Greve, W., S. Prinage, H. Zidowitz, J. Nast, and F. Reiners (2005), On Oceanogr., 56, 2293–2307, doi:10.4319/lo.2011.56.6.229.
the phenology of North Sea ichthyoplankton, ICES J. Mar. Sci., 62, Thomalla, S., N. Fauchereau, S. Swart, and P. Monteiro (2011), Regional
1216–1223, doi:10.1016/j.icesjms.2005.03.011. scale characteristics of the seasonal cycle of chlorophyll in the Southern
Henson, S., I. Robinson, J. Allen, and J. Waniek (2006), Effect of meteoro- Ocean, Biogeosciences, 8, 2849–2866, doi:10.5194/bg-8-2849-2011.
logical conditions on interannual variability in timing and magnitude of Vargas, M., C. Brown, and M. Sapiano (2009), Phenology of marine phy-
the spring bloom in the Irminger Basin, North Atlantic, Deep Sea Res. toplankton from satellite ocean color measurements, Geophys. Res. Lett.,
Part I, 53, 1601–1615, doi:10.1016/j.dsr.2006.07.009. 36, L01608, doi:10.1029/2008GL036006.
Henson, S., J. Dunne, and J. Sarmiento (2009), Decadal variability in Visser, M., and C. Both (2005), Shifts in phenology due to global climate
North Atlantic phytoplankton blooms, J. Geophys. Res., 114, c04013, change: The need for a yardstick, P. Roy. Soc. B, 272, 2561–2569, doi:
doi:10.1029/2008JC005139. 10.1098/rspb.2005.3356.
Ji, R., M. Edwards, D. Mackas, J. Runge, and A. Thomas (2010), Marine White, M., et al. (2009), Intercomparison, interpretation, and assessment of
plankton phenology and life history in a changing climate: Current spring phenology in North America estimated from remote sensing for

12
BRODY ET AL.: METHODS TO DETERMINE BLOOM INITIATION

1982–2006, Global Change Biol., 15, 2335–2359, doi:10.1111/j.1365- Yamada, K., and J. Ishizaka (2006), Estimation of interdecadal change of
2486.2009.01910.X. spring bloom timing, in the case of the Japan Sea, Geophys. Res. Lett.,
Wiltshire, K., A. Malzahn, K. Wirtz, W. Greve, S. Janisch, P. Mangelsdorf, 33, L02608, doi:10.1029/2005GL024792.
B. F. J. Manly, and M. Boersma (2008), Resilience of North Sea phy- Zhai, L., T. Platt, C. Tang, S. Sathyendranath, and R. Walls (2011), Phy-
toplankton spring bloom dynamics: An analysis of long-term data at toplankton phenology on the Scotian Shelf, ICES J. Mar. Sci., 68,
Helgoland Roads, Limnol. Oceanogr., 53, 1294–1302. 781–791, doi:10.1093/icesjms/fsq175.

13

You might also like