MARINE ECOLOGY PROGRESS SERIESMar Ecol Prog Ser

Vol. 544: 1–14, 2016doi: 10.3354/meps11601

Published February 18

© The authors 2016. Open Access under Creative Commons byAttribution Licence. Use, distribution and reproduction are un -restricted. Authors and original publication must be credited.

Publisher: Inter-Research · www.int-res.com

*Corresponding author: jrgraff@science.oregonstate.edu

FEATURE ARTICLE

Physical and optical properties of phytoplankton-rich layers in a coastal fjord: a step toward prediction

and strategic sampling of plankton patchiness

Jason R. Graff1,2,*, Susanne Menden-Deuer1

1Graduate School of Oceanography, University of Rhode Island, South Ferry Road, Narragansett, RI 02882, USA2Present address: Botany and Plant Pathology, Cordley Hall 2082, Oregon State University, Corvallis, OR 97333, USA

OPENPEN ACCESSCCESS

ABSTRACT: Dense aggregations of phytoplankton inlayers or patches alter the optical and physical propertiesof the water column and result in significant hetero-geneity in trophic and demographic rates of local plank-ton populations. Determining the factors driving patchformation, persistence, intensity, and dissipation is keyto understanding the ramifications of plankton patchi-ness in marine systems. Regression and multi-parametricstatistical analyses were used to identify the physicaland optical properties associated with 71 phytoplank-ton-rich layers (PRLs) identified from 158 CTD profilescollected between 2008 and 2010 in East Sound, Wash-ington, USA. Generalized additive models (GAMs) wereused to explore water column properties associated withand characterizing PRLs. Patch presence was associatedwith increasing water column stability represented bythe Brunt-Väisälä frequency (N 2), Thorpe scale (Lt), andturbulent energy dissipation rate (ε). A predictive re-gression identified patch presence with 100% accuracywhen log10(N 2) = −1 and 70% of the cases when log10(ε)= −3. A GAM of passively measured variables, which didnot include fluorescence, was able to model patch inten-sity with considerable agreement (R2 = 0.58), and the fitwas improved by including fluorescence (R2 = 0.69). Flu-orescence alone was an insufficient predictor of PRLs,due in part to the influence of non-photochemicalquenching (NPQ) in surface waters and the wide rangeof fluorescence intensities observed. The results showthat a multi-parametric approach was necessary to char-acterize phytoplankton patches and that physical struc-ture, resulting in steep gradients in bio-optical proper-ties, hold greater predictive power than bio-opticalproperties alone. Integration of these analytical ap-proaches will aid theoretical studies of phytoplanktonpatchiness but also improve sampling strategies in thefield that utilize autonomous, in situ instrumentation.

KEY WORDS: Plankton · Patchiness · Water columnstability · Fluorescence · Beam attenuation · General-ized additive model · Nonphotochemical quenching

INTRODUCTION

The heterogeneous nature of plankton distribu-tions have long fascinated and puzzled scientists,including Darwin on the ‘Beagle’ (Martin 2005).Plankton patches are a visual documentation thatassumptions of homogeneity in the marine environ-ment are invalid. Discrete patches with elevatedplankton concentrations have important ramifica-tions for the physical, optical and acoustical proper-ties of the water column (reviewed in Durham &Stocker 2012). Plankton patches are characterizedby taxonomically diverse communities of auto- and

Measuring, modeling, and predicting phytoplankton pat -ches in the coastal ocean.

Graphic: Jason Graff

Mar Ecol Prog Ser 544: 1–14, 2016

heterotrophic, uni- and multicellular organisms, and,at times, detritus (Alldredge et al. 2002, Rines et al.2002, Menden-Deuer 2008, Holliday et al. 2010,Rines et al. 2010). Consequently, the mechanismsleading to patch creation are equally diverse. Pro-posed mechanisms have included physical concen-tration of cells in areas of neutral buoyancy (Franks &Anderson 1992, Alldredge et al. 2002), shear (Franks1995, Birch et al. 2008), the balance of phytoplanktongrowth and zooplankton predation (Menden-Deuer2008), and swimming behaviors interacting with fluidflow (Bjørnsen & Nielsen 1991, Gallager et al. 2004,Genin et al. 2005, Durham et al. 2009). Measure-ments of trophic and demographic rates of phyto-plankton−zooplankton interactions have shown ra -pid rates of plankton growth and grazing, whichcould lead to patch formation and decline on timescales of hours to days (Menden-Deuer & Fredrick-son 2010, Menden-Deuer 2012).

A major consequence of plankton patchiness isthat biomass and related processes are distributedhetero geneously, providing a challenge for accu-rately characterizing features of the water columnand rate processes. For example in East Sound,Washington, USA, vertically restricted phytoplank-ton-rich layers (PRLs) occupied <12% of the watercolumn but represented over 50% of the primaryproduction (Menden-Deuer 2012). Rines et al.(2002) showed that the detection of harmful algalbloom (HAB) species depends on the accuratedetection of plankton thin layers that could harborlocalized concentrations of deleterious species.Accurate identification of patches or layers and theenvironmental factors leading to their formation istherefore an important prerequisite to advances inpassive survey technology and modeling planktonecology.

Prior investigations of PRLs show that they haveunique optical signatures (Donaghay et al. 1992,Cowles & Desiderio 1993). Phytoplankton patchescan be horizontally extensive and persist over timescales from hours to several days (Rines et al. 2002,Menden-Deuer 2008), and early investigations sug-gested that phytoplankton layers are often associatedwith stratified or vertically stable water columns(Dekshenieks et al. 2001, McManus et al. 2003,Cheriton et al. 2009). Patches or layers composed ofphotosynthetic plankton are most commonly de tec -ted using the in situ signal derived from chlorophyll a(chl a) fluorescence. Using CTD mounted fluorome-ters, these already-formed patches can be detectedand sampled rapidly using optical identification of insitu fluorescence profiles (Menden-Deuer 2008, Sul-

livan et al. 2010b). Although accomplished in realtime, this approach to identifying PRLs relies on for-tuitous sampling over extensive time and spacescales or in places of known patch formation (Dek-shenieks et al. 2001, Sullivan et al. 2010a). An alter-nate approach to PRL identification has been to useairborne lidar for remote detection (Churnside &Donaghay 2009). Airborne surveys vastly increasesthe spatial scales over which areas can be sampledand the large-scale, long-term processes to whichthey can be related. Still, subsequent sampling forvalidation and calibration has to be coordinated withship-based operations.

Relying solely on optical identification of phyto-plankton patches through chl a fluorescence signalslimits detection of the diverse types of patches thatcan include non-pigmented particles. More so, patchor PRL detection based on an existing signal limitsdetection to only those features that have alreadyformed. Field studies can miss the early initiationstages when fluorescence characteristics do not devi-ate significantly from background values. Conse-quently, important topics such as patch formationand the evolution of patch or layer characteristics(e.g. optical properties, biological rates) often remain empirically inaccessible. More so, nonphotochemicalquenching (NPQ) in phytoplankton (Niyogi et al.1997, Miloslavina et al. 2009, Milligan et al. 2012),leading to a reduction in measured fluorescence,may alter fluorescence profiles, masking surface lay-ers or falsely indicating a subsurface PRL (Falkowski& Kolber 1995). Photoacclimation to the ambient lightregime and pigment regulation due to nutrientscould also systematically alter the fluorescence sig-nature of PRL-associated phytoplankton (Menden-Deuer 2012), resulting in an apparent decline orincrease in patch intensity. To overcome these obsta-cles, techniques are needed that allow the detectionof diverse PRL types as well as predict the likelihoodof future PRL formation and occurrence to facilitateempirical research of plankton patchiness and itsramifications.

Several parameters, beyond fluorescence, areavailable to characterize plankton patchiness. Theconcentrated phytoplankton biomass within patchesmay be detectable through the use of the particulatebeam attenuation coefficient (cp; m−1). Phytoplanktongrowth and abundance have been correlated with cp

(Siegel et al. 1989, Durand & Olson 1996, Green et al.2003, Gernez et al. 2011), and it may serve as a proxyfor phytoplankton biomass in the open ocean (Beh -renfeld & Boss 2003, 2006). While riverine input andre-suspension of benthic material may alter the par-

2

Graff & Menden-Deuer: Toward strategic sampling of plankton patchiness

ticulate signal in nearshore waters by adding to thetotal particle field, cp is most sensitive in the sizerange ~0.5 to 20 µm (Stramski & Kiefer 1991), whichincludes most phytoplankton and may provide a use-ful optical indicator of plankton biomass in coastalseas. Recently, cp was shown to have a significantrelationship with direct measurements of phyto-plankton biomass while the particulate backscatter-ing coefficient (bbp) showed the strongest correlationwith phytoplankton carbon (Cphyto)(Graff et al. 2015).A proxy such as cp would provide a convenientmeans of assessing the carbon dynamics of PRLs incoastal ecosystems, irrespective of phytoplanktonphotophysiology, when measurements of bbp are notavailable.

In addition to optical indices, water column stabil-ity has been shown to correlate with the presence ofphytoplankton thin layers. These parameters includethe presence of a strong pycnocline (Dekshenieks etal. 2001) and measures indicative of water columnstability, such as a low Brunt-Väisälä frequency,Thorpe scale, and turbulent energy dissipation rate(Cheriton et al. 2009). The specifics of how the rela-tive magnitudes of convergence versus dissipation,by way of shear, buoyancy, motility and turbulenceenhanced diffusion balance to drive the shape of alayer’s profile are being debated (Stacey et al. 2007,2009, Birch et al. 2008, 2009) and need to be resolvedto gain a mechanistic understanding of planktonpatch dynamics.

In the meantime, all of these factors to characterizethe physical properties of the water column broadenthe parameter space of interest when investigatingplankton patchiness and might provideleverage for characterizing the conditionscondu cive to PRL presence. Predictivealgorithms may (1) aid in formulatingtestable hypo theses about PRL formationand dissipation processes, (2) direct auto -no mous vehicles based on environmentalparameters (e.g. Fior elli et al. 2003), and(3) lead to the autonomous collection ofbiological samples and rate measure-ments (e.g. Pennington et al. 2015). Pro-viding gliders with quantitative observer-independent PRL identification criteriawill improve comparability among stud-ies and en able adaptive patch samplingtechniques. Adaptive ap proa ches will benecessary to overcome sample size limi-tations inherent in observation tech-niques that rely on fortuitous observa-tions of events.

Here, our goal was to develop an em pirically basedalgorithm that would enable the modeling andpotential prediction of plankton patchiness based onautonomously acquirable water column characteris-tics. We used a large data set of water column pro-files collected over 3 years in a shallow fjord to deter-mine if we could identify a unique set of parametersto predict (1) plankton patchiness and (2) patchintensity. We found that patch presence may be mod-eled from water column stability parameters withhigh accuracy and that patch intensity can be estima -ted, though less confidently. This approach appearsto be a step forward towards predicting and samplingplankton patches and may ultimately serve a criticalrole in studies of food webs and carbon cycling in themarine environment.

MATERIALS AND METHODS

Study site and sampling design

East Sound is a temperate fjord within the SanJuan Archipelago in the Northeastern Pacific(48° 38.61’ N, 122° 52.75’ W) (Fig. 1). The fjord ex -tends north−south with a variable width of 1 to 2 kmand a maximum depth of 40 m at the most southernend. A partial sill at the southwestern terminus of thefjord restricts circulation and mixing with nearbywaters. Day-cruises, 8 in 2008, 6 in 2009, and 10 in2010 provided the CTD profile library used for ouranalysis. For each CTD cast, the water column wasprofiled with a SeaBird 19+ CTD mounted with a

3

Fig. 1. East Sound in the San Juan Islands, Washington state, USA, Northeast Pacific Ocean

Mar Ecol Prog Ser 544: 1–14, 2016

WET Labs WETStar fluorometer, a WET Labs C-Starmeasuring beam transmission over 25 cm at 660 nm,and a Biospherical QSP-2300L 4π photosyntheticallyactive radiation (PAR) sensor to collect light profiles.The observations of PRLs were evident in repeatcasts at the same location, and there was excellentagreement between down- and up-casts. All datapresented here were taken from the downcast profiles.

We followed layer identification criteria in Ryan etal. (2008) and Dekshenieks et al. (2001). As such, aprofile with a layer was identified based on its fluo-rescence profile and its maximum had to exceedbackground values by 3-fold. Background valueswere calculated as the mean fluorescence outside of3 m above and below the maximum for the remain-der of the water column. To remove a strict restrictionas to the maximum vertical extent that could be con-sidered a plankton layer, but often employed for de -fining phytoplankton ‘thin layers’, the term ‘phyto-plankton-rich layers’ (PRLs) was used to describepatches of high plankton biomass bordered by steepgradients in plankton concentrations, but criteriaregarding the vertical extent of the layers wererelaxed (Menden-Deuer 2008). The term PRL is usedinterchangeably with ‘patches’ and ‘layers’ through-out this paper. A steep gradient is the key character-istic of a PRL. Profiles exhibiting high fluorescencevalues outside of the 3 m above and below the depthof maximum fluorescence were indicative of broadand not steep gradients and were not identified as apatch.

The PRL profiles used in our analysis were col-lected over multiple years between the months ofApril and July, and the phytoplankton communitycomposition was taxonomically variable, rangingfrom highly diverse assemblages to nearly mono-specific aggregations of motile and non-motile spe-cies (see Menden-Deuer & Fredrickson 2010 formore taxonomic details of patches in East Sound).In the vertical profiles, heterogeneous phytoplank-ton distributions appear as vertically restricted lay-ers. Through repeat profiling, down to 0.1 nauticalmiles (nmile; ~185 m) in distance, we found repeatoccurrences of PRLs, suggesting that at least someof the features may be horizontally extensivelayers, rather than isolated patches. The WETStarfluorometer used in this study was calibrated withdiscrete samples of extracted chlorophyll andyielded a strong positive correlation between the insitu CTD-derived fluorescence and extracted chl a(Men den-Deuer 2008, Menden-Deuer & Fredrick-son 2010).

Data handling and statistical analysis

High-resolution data from 158 downcast profilescollected from the CTD package were processedusing SeaBird SeaSoft (v. 2). The CTD had an aver-age descent rate of 0.13 m s−1 and a sampling rate of4 Hz. Prior to statistical analyses, the processed datawere imported into Matlab (v. 7.4) and depth aver-aged into 20 cm bins to reduce the effect of in -frequent extreme records. This resulted in 11 331binned observations that were included in the analy-sis with ~6 observations per depth bin. Beam trans-mission values were converted to beam attenuationcoefficients (c) using the equation provided in theWetLabs C-Star manual: c = (−1/beam length) ×log(beam transmission). Here we use cp’ instead ofthe traditional cp to refer to the particulate beamattenuation coefficient at 660 nm, where cp’ repre-sents particulate beam attenuation values uncor-rected for the contribution of pure water and coloreddissolved organic matter, with the recognition thatthe variable constituent of colored dissolved organicmatter (CDOM) may influence this analysis (Behren-feld & Boss 2006). However, at 660 nm, the particu-late fraction should dominate the signal, with CDOMbeing a minor constituent of the measurement(Behrenfeld & Boss 2006). The Brunt-Väisälä fre-quency (N 2; s−2), a measure of water column stability,was calculated using the Matlab function sw_bfrqfrom the SEAWATER (v. 3) package provided by theCSIRO (www.cmar.csiro.au). The Thorpe scale (L t;m), a measure of the length of turbulent overturningevents, equals the root mean square displacement ofeach density measurement based on the reorderingof the density profile, where L t = (Δdepth2/2)1/2, tomake the water column gravitationally stable (Thorpe1977, Cheriton et al. 2009). The turbulent energy dis-sipation rate (ε = L t

2 × N 3; W kg−1) was calculatedwith the assumption that the Thorpe scale was equalto the Ozmidov scale following similar assumptionsand lack of turbulence measurements as in Cheritonet al. (2009). These physical parameters, while notwholly independent, were included in this analysisand each has been previously shown to be correlatedwith the presence of PRLs (Dekshenieks et al. 2001,Cheriton et al. 2009).

Statistical analyses were performed using Matlaband included regressions of measured and calculatedvariables, principal component analyses (PCA), andcomparisons of physical parameter distributions forpatch and non-patch samples, using a 2-sample Kol-mogorov-Smirnov (K-S) test. PCA, a method com-monly used for data exploration of a large number of

4

Graff & Menden-Deuer: Toward strategic sampling of plankton patchiness

variables that can be difficult to interpret graphically,is the transformation of normalized data onto dimen-sionally reduced principal component space. Thevariability of the data along each principal compo-nent can be attributed to the contribution of eachvariable to a principal component. Type II linearregressions were performed for the analysis of in situvariables, as each variable was measured with error.Multiple generalized additive models (GAMS)(Hastie & Tibshirani 1986), i.e. linear models com-posed of a sum contribution of multiple variables,were created to test if patch intensity (PI) of a samplecould be estimated from (1) a combination of the vari-ables measured by the CTD package (depth, temper-ature, salinity, PAR, beam attenuation), with andwithout fluorescence, or from (2) the variables meas-ured by the CTD package plus the calculated param-eters relating to physical properties of the water col-umn (Brunt-Väisälä frequency, Thorpe scale, andturbulent energy dissipation rate). The PI is derivedfrom the fluorescence layer identification criteria(Ryan et al. 2008, Menden-Deuer & Fredrickson2010) and is calculated as: PI = (Sample fluores-cence − Mean background fluorescence)/Maximumfluorescence. Thus, observations within a PRL wouldgenerally have PI indices between 0.5 and 1.0,approaching 1.0 in samples with maximum fluores-cence. Samples outside a patch have lower values,typically <0.5, and falling below zero for samples thathave a fluorescence value lower than the mean back-ground fluorescence. The GAMs were created in R(R64.app for Mac OS X GUI 1.36, 5691 Leopard build64-bit, www.r-project.org). We used the gam proce-dure within the mgcv package. This allowed for vari-ables to be non-parametrically estimated smoothfunctions, potentially going to zero. Allowing forvariable functions to approach zero is the equivalentof removing the variable from the model and allowsfor the elimination of variables that do not improvethe fit of the model. We also tested the step.gam function (results not shown) but found the gam func-tion of the mgcv package to provide slightly betterapproximations of PI based on linear regressionanalysis of the calculated PI values and model esti-mated PI values.

Investigations of NPQ were limited to regressionanalyses and did not include direct physiologicalmeasurements on the samples in the field. The analy-sis included depth-binned regressions of fluores-cence (Fl) and cp’ and a regression of Fl normalized tocp’ versus in situ PAR. The linear relationship of Fl/cp’with PAR was fitted with an exponential curve usingthe EzyFit Toolbox (v. 2.41b) for Matlab, by F. Moisy.

This fit was used to correct for NPQ by adding thedifference between the model and the y-intercept toeach data point evaluated at its respective in situPAR. The result was then divided by the correspon-ding cp’ to determine the NPQ-corrected fluorescence.

RESULTS

Of the 158 profiles, 71 (45%) contained PRLs.Observations from within PRLs accounted for 779(nearly 7%) of the 11 331 discrete observations. Ofthe remaining profiles, 68 (43%) contained subsur-face fluorescence maxima that occurred over a broadrange of depths and thus did not meet our criteria tobe identified as PRLs (i.e. mean fluorescence within apatch exceeds background fluorescence by 3-fold).Although not classified as PRLs, the majority of theseprofiles had heterogeneous, but very broad, fluo -rescence distributions lacking steep gradients; i.e.fluorescence here should not be assumed to behomogenous (see Fig. 1 in Menden-Deuer [2012] forexamples). The re maining 19 (12%) of the profileshad no distinct fluorescence maxima due to uniformfluorescence throughout the entirety of the water column. This range of observations provided suitablecontrols to test the validity of our predictive algorithm.

Optical measurements of fluorescence and beamattenuation were collected over a wide range of val-ues. In situ fluorescence ranged from 0.08 to 4.82 V(instrument range 0−5 V) and beam transmissionfrom 38.9 to 95.7% (instrument range 0−100%),equivalent to cp’ values of 0.17 to 3.77 m−1. Regres-sion analysis (Type II) of in situ fluorescence and cp’showed a significant positive correlation (R2 = 0.75,p < 0.0001), indicating a strong correspondence be -tween the fluorescence signal (a proxy for chl a con-centration) and the cp’ signal (a proxy for particulateconcentration) (Fig. 2). Non-linear regression did notimprove the explanatory power of the regression.

A PCA of the factors distinguishing PRL samplesfrom non-PRL samples indicated that 75.6% of thevariability in the data could be explained by a combi-nation of the CTD-measured variables (Fig. 3A). PC1contributed 39.3% of the explained variability, withtemperature and salinity contributing most of the ex-planatory power. PC2 contributed almost equally(36.3%) to the variability explained, with fluorescenceand beam transmission contributing the majority ofthe explanatory power. PC3, comprised largely ofPAR, added a smaller amount of explanatory power(17.3%) and PC4, i.e. exclusively temperature andsalinity, even less at 4.7%. The PCA using only the

5

Mar Ecol Prog Ser 544: 1–14, 2016

variables measured by the CTD package did notresult in particularly coherent groupings of PRL andnon-PRL samples. The inclusion of the physical stabil-ity parameters (N 2, L t, and ε) in the PCA reduced theexplanatory power of the PCA (Fig. 3B) to 54.6% forthe first 2 principal components, with PC1 and PC2contributing 30.0% and 24.6 % of the explained vari-

ability, respectively. The turbulent energy dissipationrate (ε) accounted for the second largest portion of thevariability on PC1. PC3 and PC4 again contributed asmaller portion of explanatory power at 18.5 and12.4%, respectively, and were most highly repre-sented by the physical stability parameters. Althoughthe explanatory power was lower, inclusion of the sta-bility parameters improved the distinction betweenPRL and non-PRL samples (Fig. 3B). This distinction isobserved with the PRL samples (black dots in Fig. 3B)forming a co hesive subset within the ‘boundaries’ ofthe non-PRL samples (grey dots in Fig. 3B).

A comparison of the frequency distributions of thePRL- and non-PRL-associated stability parameters(N2, L t, and ε) showed that PRLs and non-PRLs couldbe characterized by significantly different ranges foreach parameter. The mean ± SD Brunt-Väisälä fre-quency for PRL samples was an order of magnitudehigher than for non-PRL samples (N 2 = 0.002 ± 0.006and N 2 = 0.0005 ± 0.0016 s−2 in PRL and non-PRLsamples, respectively; p << 0.0001, KS-test). Similarresults were found for mean ε and L t (Table 1); thesebeing an order of magnitude and a factor of 2 greaterin PRL samples, respectively, and coming from sig-nificantly different distributions (K-S test, p <<0.0001). This characteristic difference in the watercolumn stability parameters between PRL and non-PRL samples provided considerable predictive powerfor the likelihood of PRL detection (Fig. 4). The prob-ability of an observation being from within a PRL wasgreatest in stable parts of the water column where

6

Fig. 2. In situ fluorescence (V) versus beam attenuation (cp’,particulate beam attenuation coefficient, m−1) for 11 331,0.2 m depth-binned observations collected from 158 watercolumn profiles. The grey line is a best-fit (Type II) linearregression with the equation shown (R2 = 0.75, p < 0.0001).Black lines are the 95% prediction interval for the obser-

vations

Fig. 3. Principal component analyses of (A) in-water properties: temperature (°C), salinity (psu), photosynthetically active radi-ation (PAR, µmol photons m−2 s−1), beam transmission (%), and fluorescence (V); and (B) in-water properties as in panel (A)and including calculated stability parameters: Brunt-Väisälä frequency (N 2, s−2); Thorpe scale (Lt, m); and turbulent energydissipation rate (ε, W kg−1). Including the stability parameters (in B) resulted in a tighter clustering of patch samples (blackdots) relative to non-patch samples (grey dots) suggesting that the stability parameters are a unifying characteristic of patches

Graff & Menden-Deuer: Toward strategic sampling of plankton patchiness

energy dissipation was high and scales ofoverturning were small. Therefore, watercolumn stability parameters are a necessaryaddition and provide a significant improve-ment to predicting patch occurrence and, toa lesser extent, intensity.

Having established the potential for apatch occurrence predictor, we exploredcharacteristics that may provide predictivepower of patch intensity. Correlation analy-ses of the physical stability parameters indi-cated significant correlations with patchintensity (PI), but the variability explainedby these regressions was low (N 2: R = 0.17,p < 0.001; ε: R = 0.10, p < 0.001; L t: R = −0.04,p < 0.001). GAMs composed of in situ watercolumn properties indicated that a combina-tion of smoothed spline parameters from theCTD provided the best models of PI. Linearregression analysis of the calculated andmodeled PI-values using a GAM includingfluorescence and a GAM excluding fluores-cence produced R2 values of 0.69 and 0.58,respectively (Fig. 5). A smoothed spline ofeach variable’s relative contribution to pre-dicting PI, ranging from < 0 for non-PRLsamples to a maximum of 1 for the peakpatch intensity, is provided in the right handpanels of Fig. 5, and show the varying com-bination of each variable to each GAM.While all parameters were considered to bestatistically significant to the model, somehave a more dynamic contribution over theirrange of measured values (e.g. fluorescenceand beam attenuation) than other variablesthat had relatively constant contributionover their measured range (e.g. depth andsalinity).

The accuracy of the GAMs was assessedby determining the percentage of samples

7

N 2 (s−2) Lt (m) ε (W kg−1)

PRL 0.002 ± 0.006 0.0077 ± 0.016 9.12 × 10−7 ± 0.006Non-PRL 0.0005 ± 0.006 0.0044 ± 0.011 7.45 × 10−8 ± 0.006p <<0.001 <<0.001 <<0.001

Table 1. Mean ± SD values of water column physical stability para -meters (N 2: buoyancy frequency; L t: Thorpe scale; ε: turbulent energydissipation rate) associated with samples taken when a plankton-richlayer was present (PRL) and not present (non-PRL). Mean values of all3 parameters are up to an order of magnitude greater for PRL samplesthan for non-PRL samples. p-values << 0.001 from 2-way Kolmogorov-Smirnov (K-S) tests indicate a significant difference in the distribution

of values between PRL and non-PRL samples

Fig. 4. Normalized histograms of log10-transformed values for (A) the Brunt-Väisälä frequency, (N 2, s−2), (B) turbulent energy dissipation rate (ε, W kg−1), and (C) Thorpe scale (L t, m), for non-patch samples (black)and phytoplankton-rich layer (PRL) only samples(grey), provide the relative frequency of samples in thatrange of values on the left-hand y-axis. The probability(line and circles) of a sample with that particular buoy-ancy frequency to be within a patch or layer is indicatedon the right-hand y-axis. Kolmogorov-Smirnoff (K-S)tests of the distributions of patch and non-patch sam-ples for all parameters were significant (p << 0.05). Note

the axis ranges

Mar Ecol Prog Ser 544: 1–14, 2016

with PI ≥ 0.5 that were accurately modeled to have aPI value of that magnitude. The GAM without fluo-rescence was accurate for 30% of the samples.Including fluorescence in the GAM increased theaccuracy to 33%. A GAM consisting of only theBrunt-Väisälä frequency, Thorpe scale, and the tur-bulent energy dissipation rate (N 2, L t, and ε) did apoor job of reproducing PI values (R2 = 0.06, resultnot shown). Including these parameters in the mod-els along with the CTD parameters did not improvethe predictions of PI. Thus, the magnitude of a PRLmay be determined using only the parameters mostcommonly measured by a CTD package.

A comparison of the depths at which the maximumin situ fluorescence and maximum cp’ occurred inprofiles containing a PRL indicated that these 2depths did not agree. On average, maximum cp’occurred 2.1 m above maximum fluorescence. Someof this discrepancy may be due to instances withinsome profiles of a fresh water lens near the surfacecarrying a high particulate load of apparently non-fluorescent particles (Fig. 6, above 3 m). However,NPQ could be responsible for some of the differencesobserved between the depth of maximum cp’ andmaximum fluorescence, whether this differenceoccurred immediately at the surface or at subsurface

8

Fig. 5. Regressions of calculated patch intensity (PI) values (large panels) with (A) results from a generalized additive model(GAM) using the input variables: depth, temperature, salinity, photosynthetically active radiation (PAR), beam transmissionand fluorescence; R2 = 0.68, and (B) a GAM without fluorescence; R2 = 0.58. A smoothed spline of each variable’s relative con-tribution to predicting PI (black line), ranging from < 0 for non-PRL samples to a maximum of 1 for the peak PI, is provided in

the small right-hand panels. Dashed lines are the 95% CI for each spline

Graff & Menden-Deuer: Toward strategic sampling of plankton patchiness

depths where irradiance was high at the top of alayer. A regression of cp’ normalized fluorescence(Fl/cp’) versus in situ PAR suggests that NPQ wascontributing to the offset between max cp’ and Fl(Fig. 7A). In the absence of NPQ, the regression ofthese parameters would create a horizontal line withslope of zero (dashed line in Fig. 7A). Instead, theresults show a decrease in fluorescence per cp’ athigher in situ PAR (solid line regression in Fig. 7A).Additionally, depth-binned regressions of Fl and cp’also suggest NPQ, with lower slopes, i.e. less fluores-cence per cp’, in samples at the surface (0−5 m:slope = 0.92; 5−10 m: slope = 1.45; 10−15 m: slope =1.54) and greater slopes in samples collected indeeper waters (15−20 m: slope = 1.60; 20−25 m:slope = 1.64; 25−30 m: slope = 1.69) (Fig. 7B). Anincrease of the Fl signal with depth relative to cp’indicates a masking of Fl in near surface samples.

In order to assess the impact of the potential NPQon layer detection and patch index values, the ex -ponential regression in Fig. 7A was used to removethe potential contribution of NPQ to fluorescenceobservations. This was accomplished by adding thedifference between the exponential fit evaluated ata given PAR and the y-intercept. This correctionresulted in the alteration of some fluorescence pro-files (Fig. 8). In general, as NPQ was most pre -valent in the surface layer, surface fluorescenceprofiles were most significantly affected. Twelve ofthe 158 profiles were significantly altered, resultingin the detection of 9 additional PRLs at the surfaceand altering the profiles such that 3 PRL profileswere reclassified as non-PRL profiles because the

fluorescence profile was extended towards the sur-face. In general, NPQ corrections to the PRLprofiles adjusted the depth of maximum fluores-cence to become shallower (Fig. 8A), broadenedthe feature (Fig. 8B), or generated a new PRLwhere none was previously observed (Fig. 8C).Thus, overall NPQ correction resulted in the detec-tion of additional PRLs. The remaining 92% of allprofiles, including PRL profiles, exhibited onlysmall changes.

9

Fig. 6. Example profile of chlorophyll a, particulate beam at-tenuation (cp’), and density for the same vertical profile con-taining a phytoplankton rich layer in East Sound, WA, USA

Fig. 7. Possible influence of nonphotochemical quenching(NPQ) from (A) a regression of fluorescence (Fl) normalizedto the particulate beam attenuation coefficient (Fl/cp’) versusphotosynthetically active radiation (PAR) and (B) 5 m depth-binned regressions of Fl and cp’. In the absence of NPQ ormajor differences in non-phytoplankton particles, Fl/cp’should not vary as a function of PAR (dashed line in A) andthe slope of Fl versus cp’ would be the same at all depths.The exponential fit y = 1.04e−0.0001x in A was used to correct

for NPQ

Mar Ecol Prog Ser 544: 1–14, 2016

DISCUSSION

PRLs can make significant contributions to the pro-ductivity and carbon dynamics of coastal waters(Menden-Deuer 2012). Particularly with the adventof broader-scale ocean observation capacity, it isimportant to evaluate the environmental parametersassociated with these patches so they can be detec -ted and their ecology and pervasiveness can be bet-ter understood and made accessible to direct sam-pling, autonomous observations, and modeling efforts.Our analysis revealed that PRL occurrence can bereliably predicted from water column stability para -meters despite interannual variability in the waterstability parameters and in phytoplankton commu-nity composition. Patch intensity, on the other hand,was not related to water column stability and re -quired a complex combination of most of the vari-ables measured by the CTD profiler. It is noteworthythat fluorescence, the para meter previously used toidentify PRL presence (e.g. Menden-Deuer 2008) hadlittle predictive power over either patch occurrenceor intensity. This is most likely due to differences inthe maximum fluorescence of one patch compared toanother when analyzed within a large dataset whererelatively low values of fluorescence could representan intense patch due to the even lower backgroundvalues in the surrounding water column. An ap -proach to identifying and describing patches mustinclude such relative terms in order to be applicable

to dynamic datasets collected over large spatial andtemporal scales. More so, localized acclimationstrategies to the light regime (e.g. NPQ) may signifi-cantly alter the detectability and measured intensityof surface patches using in situ measurements and,thus, may be an important factor to consider whenusing remotely sensed data in coastal regions.

The association of PRLs with higher water columnstability, particularly with respect to the Brunt-Väisälä frequency, has been observed in prior fieldstudies in Monterey Bay, CA, USA (Dekshenieks etal. 2001, McManus et al. 2005, Cheriton et al. 2009),indicating that our observations in East Sound arenot specific to this location. This relationship couldpotentially be used to predict where layers will formprior to their optical signatures eclipsing those of thesurrounding waters. Predicting the occurrence ofpat ches before their formation will be an essentialstep in Lagrangian studies of patch dynamics toreveal the time-resolved factors driving patch forma-tion. The strong predictive power of patch occur-rence through water column stability parameters isindicative that, at least for the patches sampled,physical mechanisms were essential to their forma-tion. While motility could theoretically lead to anaggregation of cells, and thus support patch forma-tion anywhere in the water column, it has beenshown that diverse motile plankton respond to phys-ical features through alterations in swimming pat-terns and growth rates, and thus water column stabil-ity may even affect patch formation in motileplankton (Kamykowski 1981, Karp-Boss et al. 2000,Sullivan et al. 2003, Genin et al. 2005, Steinbuck etal. 2009). The majority of pat ches observed here weredominated by non-motile phytoplankton as de -scribed for these patches in Men den-Deuer (2012),suggesting water column stability is even moreimportant to patch formation and persistence bythese organisms.

Interestingly, patch intensity had generally lowcorrelation with parameters describing the physicalstability of the water column (N 2, L t and ε). Here, it isimportant to remember that a patch is distributedthroughout a range of depths and bordered by steepgradients in relation to a particular measure of sta -bility. Thus, it is difficult to pinpoint a narrow rangeof parameter values that correspond to high patchintensities, which may explain the lack of correlation.Phytoplankton are exposed to a range of conver-gence and divergence conditions which influencepatch formation and dissipation (Stacey et al. 2007,Birch et al. 2008). In addition to physical mecha-nisms, biological rates influencing phytoplankton

10

Fig. 8. Example fluorescence profiles before and after apply-ing the correction for non-photochemical quenching (NPQ).(A) The profile of the phytoplankton-rich layer (PRL) re -mains relatively unchanged with the maximum depth of flu-orescence approximately 1.5 m shallower after the correc-tion is applied. (B) The PRL appears as a broader featureafter the NPQ correction. (C) Correction results in a newlyrecognized surface PRL that was not recognized in the

uncorrected profile

Graff & Menden-Deuer: Toward strategic sampling of plankton patchiness

growth and loss can influence the shape and locationof the PRL and its peak (e.g. Benoit-Bird et al. 2009).Thus, it is not necessarily the depth of maximum ver-tical stratification or stability that determines wherethe peak of a layer will occur. Nevertheless, the moststable regions of the water column have a higherprobability of containing a phytoplankton patch, par-ticularly with respect to the Brunt-Väisälä frequency.

In this study, the intensity (PI) of patches was bestmodeled by a GAM composed of passively sampledvariables including fluorescence, but were modeledalmost as well without the use of fluorescence. Themodel accurately predicted a patch index ≥0.5, 33%of the time when using fluorescence and 30% of thetime without this identifying metric. An a prioriexpectation that the inclusion of fluorescence in theGAM, since it was used to calculate the PI, wouldincrease the accuracy of the model more than it didbelies the fact that the index used to identify a patchsample represents the ratio of mean water columnand peak fluorescence for a particular profile. Inother words, the patch index is a relative measure offluorescence and not an absolute value. A PRL canexist where fluorescence for an entire profile is gen-erally low compared to other profiles but where apeak is present. Conversely, a PRL may not existwhere the fluorescence values for the entire watercolumn are extremely high (see Fig. 1 in Menden-Deuer 2012) but lack a significantly distinguishablestructure with steep gradients resulting in a peak.Both of these fluorescence patterns were observedand included in the analysis in this study. Thestrength of the GAM approach, thus, is not due to aninherently circular nature of including fluorescenceas a parameter. The complex nature of the fluores-cence spline for the GAM using this parameter, infact, shows the complexity of its usage in the model.Thus, the GAM helps to identify the contribution ofmultiple parameters to the observation of patchiness,taking into account that their relative contributionmight change over the range of values measured. Inthis case, it showed that while salinity and depth hada constant contribution to the GAM models, fluores-cence and beam attenuation were more variable,indicative of a non-linear relationship between patchintensity and the latter 2 variables. Thus, this non- linearity has to be taken into account when interpret-ing patchiness. In a practical sense this means thatour prior approach of focusing on the fluorescencesignal alone is not consistently a reliable contributingvariable to identifying patches. This may be subjec-tive to the definition of a patch and may seem intu-itive to an observer looking at data in real-time.

Reliance on fluorescence must be dampened to somedegree as it is not always a strong parameter for mod-eling purposes, and additional factors will provideadditional strength for autonomous platforms toidentify where a patch is and, better yet, to predictwhere a patch might form.

Patches have long been proposed to be trophic hot -spots that represent essential resources in an other-wise dilute ocean (Lasker 1975, Mullin & Brooks1976) and could be an important source of carbon todeeper waters via grazing and sinking mechanisms.This trophic hotspot hypothesis has been empiricallyverified in East Sound, where measured grazingrates by heterotrophic protists on phytoplanktonwithin patches was significantly higher than outsideof patches, with grazers consuming 65% of primaryproduction within patches in comparison to 26% inthe remainder of the water column (Menden-Deuer& Fredrickson 2010). Moreover, plankton patcheshave been identified as production hotspots, wherethe majority of primary production occurred withinpatches, which only occupied a small fraction of thewater column (Menden-Deuer 2012). Due to the non-linear dynamics of production and grazing rates, it ischallenging to extrapolate bulk rates and planktondynamics from these heterogeneous measurements.A proxy independent of cellular physiology andrelating to phytoplankton carbon, such as the partic-ulate beam attenuation coefficient (cp) or the particu-late backscattering coefficient (bbp), as demonstratedby Graff et al. (2015) and Behrenfeld & Boss (2003,2006), may provide a convenient measure of the car-bon dynamics of patches over a broad range of spa-tial and temporal scales. Despite the potentially largevariability in the contribution of different particletypes to the total particle field in coastal areas (e.g.species composition, freshwater inputs, re-suspen-sion), we found a strong relationship between cp’ andin situ fluorescence. The systematic discrepancybetween the depth of maximum beam attenuationand maximum in situ fluorescence for profiles withpatches, however, is a concern for using cp as a proxyfor phytoplankton biomass when investigating sur-face PRLs. Application of a new method to measurephytoplankton biomass (Graff et al. 2012) has led toadvances in relating in situ optical properties, specif-ically the bbp coefficient, to phytoplankton standingstocks (Graff et al. 2015) and could be applied infuture studies of patch dynamics and alleviate con-siderations of phytoplankton photophysiology and itsimpacts on pigments and fluorescence that are oftenused in patch identification. However in field studieswhere bbp is not an available metric, cp may be a

11

Mar Ecol Prog Ser 544: 1–14, 2016

good alternative (Graff et al. 2015). Combining in situmeasurement approaches for deteriming phyto-plankton biomass with the predictive power of patchoccurrence may make studies of plankton populationdynamics more accessible and provide much neededdata on the relative importance of (1) physical versusbiological mechanisms of patch formation, persist-ence, and decline and (2) overcome obstacles indetermining the contribution of plankton patchinessto ecosystem functions.

The divergence in depths of peak values for beamattenuation and fluorescence optical proxies could bedue to non-photochemical quenching (NPQ), e.g.photophysiology, depressing fluorescence at shal-lower depths, or the input of non-chlorophyll-con-taining particles from nearby coastal sources. Wehighlight the potential for NPQ in this study and thesubsequent effects this has on the interpretation offluorescence profiles, that is, increasing the overallmagnitude of a profile and/or altering the depth ofmaximum fluorescence, creating a broader structure,or hiding a surface patch completely. Physiologicalmeasurements were not collected during these fieldstudies, and as such, the extent to which NPQ wasaccounted for is our best approximation. From thesecorrections, only 12 of the 158 PRLs were signifi-cantly affected by NPQ, and thus in East Sound itappears to be of minor concern. This is also evident inthe robust relationship between extracted chl a andfluorescence in East Sound (Menden-Deuer &Fredrickson 2010) which included the extraction ofchl a from non-PRL samples collected above PRLswhere NPQ would be most prevalent. However,more rigorous physiological studies in relation tolayer identification and sampling should be per-formed in the future for more accurate identificationof, correction to, and interpretations of patch dynam-ics when using fluorescence as the optical index forPRLs. This would be particularly helpful during stud-ies following the life history of PRLs that can undergofluctuations in depth and thus in light exposure andphotoacclimation.

CONCLUSIONS

The ramifications of plankton patchiness for char-acterizing the optical and acoustical properties of thewater column, as well as the quantification of trophicand demographic rates, makes it pertinent thatpatches are sampled in a deliberate and predictivemanner, rather than based on fortuitous observa-tions. The combination of factors identified here as

strong predictors of patch occurrence (water columnstability) and intensity (CTD-derived measures)makes that goal feasible. The incorporation of physi-ological measurements to determine the extent ofNPQ would also provide evidence for patches in sur-face waters that may otherwise be missed or under-sampled. It is conceivable that water column datafrom remotely operated or autonomous vehicles,moored profilers, or CTD casts could be input into apredictive model based on methods similar to thosepresented here. This could include the blending ofPCA and regression techniques similar to that pro-vided by partial least squares regression (PLSR).Integration of models like these in autonomous,mobile sampling platforms (e.g. gliders) could ulti-mately inform the observation and adaptive sam-pling of patches and direct efforts to locations wherepatches are likely to form. Future studies that incor-porate data from additional regions, seasons, andplankton communities not represented here can testthe applicability and reliability of these predictivealgorithms across diverse conditions, potentiallyleading to a robust universal model for identifyingand predicting plankton patchiness.

Acknowledgements. We thank the faculty and staff at theShannon Point Marine Center, Western Washington Univer-sity for their support and hospitality. RV ‘Zoea’ skippersG. McKeen and N. Schwarck are gratefully acknowledgedas well as the many graduate and undergraduate studentswho have helped with this research. This research was sup-ported through grants from the Office of Naval Research(N000140710136 and N000141010124).

LITERATURE CITED

Alldredge AL, Cowles TJ, MacIntyre S, Rines JEB and others (2002) Occurrence and mechanisms of formationof a dramatic thin layer of marine snow in a shallowPacific fjord. Mar Ecol Prog Ser 233: 1−12

Behrenfeld MJ, Boss E (2003) The beam attenuation tochlorophyll ratio: an optical index of phytoplanktonphysiology in the surface ocean? Deep-Sea Res I 50: 1537−1549

Behrenfeld MJ, Boss E (2006) Beam attenuation and chloro-phyll concentration as alternative optical indices ofphytoplankton biomass. J Mar Res 64: 431−451

Benoit-Bird KJ, Cowles TJ, Wingard CE (2009) Edge gradi-ents provide evidence of ecological interactions in plank-tonic thin layers. Limnol Oceanogr 54: 1382−1392

Birch DA, Young WR, Franks PJS (2008) Thin layers ofplankton: formation by shear and death by diffusion.Deep-Sea Res I 55: 277−295

Birch DA, Young WR, Franks PJS (2009) Plankton layer pro-files as determined by shearing, sinking, and swimming.Limnol Oceanogr 54: 397−399

Bjørnsen PK, Nielsen TG (1991) Decimeter scale hetero-geneity in the plankton during a pycnocline bloom ofGyrodinium aureolum. Mar Ecol Prog Ser 73: 263−267

12

Graff & Menden-Deuer: Toward strategic sampling of plankton patchiness

Cheriton OM, McManus MA, Stacey MT, Steinbuck JV(2009) Physical and biological controls on the mainte-nance and dissipation of a thin phytoplankton layer. MarEcol Prog Ser 378: 55−69

Churnside JH, Donaghay PL (2009) Thin scattering layersobserved by airborne lidar. ICES J Mar Sci 66: 778−789

Cowles TJ, Desiderio RA (1993) Resolution of biologicalmicrostructure through in-situ fluorescence emissionspectra. Oceanography 6: 105−111

Dekshenieks MM, Donaghay PL, Sullivan JM, Rines JEB,Osborn TR, Twardowski MS (2001) Temporal and spatialoccurrence of thin phytoplankton layers in relation tophysical processes. Mar Ecol Prog Ser 223: 61−71

Donaghay P, Rines H, Sieburth J (1992) Simultaneous sam-pling of fine scale biological, chemical, and physicalstructure in stratified waters. Arch Hydrobiol Beihefte36: 97−108

Durand MD, Olson RJ (1996) Contributions of phytoplank-ton light scattering and cell concentration changes to dielvariations in beam attenuation in the equatorial Pacificfrom flow cytometric measurements of pico-, ultra-, andnanoplankton. Deep-Sea Res II 43: 891−906

Durham WM, Stocker R (2012) Thin phytoplankton layers: characteristics, mechanisms, and consequences. AnnuRev Mar Sci 4: 177−207

Durham WM, Kessler JO, Stocker R (2009) Disruption of ver-tical motility by shear triggers formation of thin phyto-plankton layers. Science 323: 1067−1070

Falkowski P, Kolber Z (1995) Variations in chlorophyll fluo-rescence yields in phytoplankton in the world oceans.Funct Plant Biol 22: 341−355

Fiorelli E, Bhatta P, Leonard NE, Shulman I (2003). Adaptivesampling using feedback control of an autonomousunderwater glider fleet. In: Proc 13th Int Symp onUnmanned Untethered Submersible Technology(UUST). www.princeton.edu/~naomi/uust03FBLSp.pdf

Franks PJS (1995) Thin layers of phytoplankton: a model offormation by near-inertial wave shear. Deep-Sea Res I42: 75−91

Franks P, Anderson D (1992) Alongshore transport of a toxicphytoplankton bloom in a buoyancy current: Alexan-drium tamarense in the Gulf of Maine. Mar Biol 112: 153−164

Gallager SM, Yamazaki H, Davis CS (2004) Contribution offine-scale vertical structure and swimming behavior toformation of plankton layers on Georges Bank. Mar EcolProg Ser 267: 27−43

Genin A, Jaffe JS, Reef R, Richter C, Franks PJS (2005)Swimming against the flow: a mechanism of zooplanktonaggregation. Science 308: 860−862

Gernez P, Antoine D, Huot Y (2011) Diel cycles of the partic-ulate beam attenuation coefficient under varying trophicconditions in the northwestern Mediterranean Sea: observations and modeling. Limnol Oceanogr 56: 17−36

Graff JR, Milligan AJ, Behrenfeld MJ (2012) The measure-ment of phytoplankton biomass using flow-cytometricsorting and elemental analysis of carbon. LimnolOceanogr Methods 10: 910−920

Graff JR, Westberry TK, Milligan AJ, Brown MB and others(2015) Analytical phytoplankton carbon measurementsspanning diverse ecosystems. Deep-Sea Res I 102: 16−25

Green RE, Sosik HM, Olson RJ, DuRand MD (2003) Flowcytometric determination of size and complex refractiveindex for marine particles: comparison with independentand bulk estimates. Appl Opt 42: 526−541

Hastie T, Tibshirani R (1986) Generalized additive models.Stat Sci 1: 297−310

Holliday DV, Greenlaw CF, Donaghay PL (2010) Acousticscattering in the coastal ocean at Monterey Bay, CA,USA: fine-scale vertical structures. Cont Shelf Res 30: 81−103

Kamykowski D (1981) Laboratory experiments on the diur-nal vertical migration of marine dinoflagellates throughtemperature gradients. Mar Biol 62: 57−64

Karp-Boss L, Boss E, Jumars PA (2000) Motion of dinoflagel-lates in simple shear flow. Limnol Oceanogr 45: 1594−1602

Lasker R (1975) Field criteria for survival of anchovy lar-vae — relation between inshore chlorophyll maximumlayers and successful first feeding. Fish Bull 73: 453−462

Martin A (2005) The kaleidoscope ocean. Philos Trans R SocA 363: 2873−2890

McManus MA, Alldredge AL, Barnard AH, Boss E and others (2003) Characteristics, distribution and persist-ence of thin layers over a 48 hour period. Mar Ecol ProgSer 261: 1−19

McManus MA, Cheriton OM, Drake PJ, Holliday DV, Stor-lazzi CD, Donaghay PL, Greenlaw CF (2005) Effects ofphysical processes on structure and transport of thin zoo-plankton layers in the coastal ocean. Mar Ecol Prog Ser301: 199−215

Menden-Deuer S (2008) Spatial and temporal characteristicsof plankton-rich layers in a shallow, temperate fjord. MarEcol Prog Ser 355: 21−30

Menden-Deuer S (2012) Structure-dependent phytoplank-ton photosynthesis and production rates: implications forthe formation, maintenance, and decline of planktonpatches. Mar Ecol Prog Ser 468: 15−30

Menden-Deuer S, Fredrickson K (2010) Structure-depen-dent, protistan grazing and its implication for the forma-tion, maintenance and decline of plankton patches. MarEcol Prog Ser 420: 57−71

Milligan AJ, Aparicio UA, Behrenfeld MJ (2012) Fluores-cence and nonphotochemical quenching responses tosimulated vertical mixing in the marine diatom Thalas-siosira weissflogii. Mar Ecol Prog Ser 448: 67−78

Miloslavina Y, Grouneva I, Lambrev PH, Lepetit B, Goss R,Wilhelm C, Holzwarth AR (2009) Ultrafast fluorescencestudy on the location and mechanism of non-photochem-ical quenching in diatoms. Biochim Biophys Acta 1787: 1189−1197

Mullin MM, Brooks ER (1976) Some consequences of distri-butional heterogeneity of phytoplankton and zooplank-ton. Limnol Oceanogr 21: 784−796

Niyogi KK, Bjorkman O, Grossman AR (1997) Chlamy-domonas xanthophyll cycle mutants identified by videoimaging of chlorophyll fluorescence quenching. PlantCell 9: 1369−1380

Pennington JT, Blum M, Chavez FP (2016) Seawater sam-pling by an autonomous underwater vehicle: ‘Gulper’sample validation for nitrate, chlorophyll, phytoplankton,and primary production. Limnol Oceanogr Methods 14:14–23

Rines JEB, Donaghay PL, Dekshenieks MM, Sullivan JM,Twardowski MS (2002) Thin layers and camouflage: hid-den Pseudo-nitzschia spp. (Bacillariophyceae) popula-tions in a fjord in the San Juan Islands, Washington,USA. Mar Ecol Prog Ser 225: 123−137

Rines JEB, McFarland M, Donaghay P, Sullivan J (2010)Thin layers and species-specific characterization of the

13

Mar Ecol Prog Ser 544: 1–14, 2016

phytoplankton community in Monterey Bay, California,USA. Cont Shelf Res 30: 66−80

Ryan JP, McManus MA, Paduan JD, Chavez FP (2008)Phytoplankton thin layers caused by shear in frontalzones of a coastal upwelling system. Mar Ecol Prog Ser354: 21−34

Siegel DA, Dickey TD, Washburn L, Hamilton MK, MitchellBG (1989) Optical determination of particulate abun-dance and production variations in the oligotrophicocean. Deep-Sea Res 36: 211−222

Stacey MT, McManus MA, Steinbuck JV (2007) Conver-gences and divergences and thin layer formation andmaintenance. Limnol Oceanogr 52: 1523−1532

Stacey M, McManus M, Steinbuck J (2009) Response to‘Plankton layer profiles as determined by shearing, sink-ing, and swimming’ by D. A. Birch, W. R. Young, and P. J.S. Franks. Limnol Oceanogr 54: 400

Steinbuck JV, Stacey MT, McManus MA, Cheriton OM,

Ryan JP (2009) Observations of turbulent mixing in aphytoplankton thin layer: implications for formation,maintenance, and breakdown. Limnol Oceanogr 54: 1353−1368

Stramski D, Kiefer DA (1991) Light scattering by microor-ganisms in the open ocean. Prog Oceanogr 28: 343−383

Sullivan JM, Swift E, Donaghay PL, Rines JE (2003) Small-scale turbulence affects the division rate and morphologyof two red-tide dinoflagellates. Harmful Algae 2: 183−199

Sullivan JM, Donaghay PL, Rines JEB (2010a) Coastal thinlayer dynamics: consequences to biology and optics.Cont Shelf Res 30: 50−65

Sullivan JM, Van Holliday D, McFarland M, McManus MAand others (2010b) Layered organization in the coastalocean: an introduction to planktonic thin layers and theLOCO project. Cont Shelf Res 30: 1−6

Thorpe S (1977) Turbulence and mixing in a Scottish loch.Philos Trans R Soc A 286: 125−181

14

Editorial responsibility: Steven Lohrenz, New Bedford, Massachusetts, USA

Submitted: May 8, 2015; Accepted: December 30, 2015Proofs received from author(s): February 15, 2016

➤

➤

➤

➤

➤

➤

➤

➤

➤

➤