Potential controls of isoprene in the surface ocean

Isoprene surface ocean concentrations and vertical distribution, atmospheric mixing ratios, and calculated sea‐to‐air fluxes spanning approximately 125° of latitude (80°N–45°S) over the Arctic and Atlantic Oceans are reported. Oceanic isoprene concentrations were associated with a number of concurrently monitored biological variables including chlorophyll a (Chl a), photoprotective pigments, integrated primary production (intPP), and cyanobacterial cell counts, with higher isoprene concentrations relative to all respective variables found at sea surface temperatures greater than 20°C. The correlation between isoprene and the sum of photoprotective carotenoids, which is reported here for the first time, was the most consistent across all cruises. Parameterizations based on linear regression analyses of these relationships perform well for Arctic and Atlantic data, producing a better fit to observations than an existing Chl a‐based parameterization. Global extrapolation of isoprene surface water concentrations using satellite‐derived Chl a and intPP reproduced general trends in the in situ data and absolute values within a factor of 2 between 60% and 85%, depending on the data set and algorithm used.


Introduction
Biogenic volatile organic compounds have profound impacts on atmospheric chemistry, predominantly by forming tropospheric ozone, modifying the oxidative capacity of the atmosphere, and contributing to secondary organic aerosol [e.g., Carlton et al., 2009;Claeys et al., 2004]. Isoprene, emitted mainly from trees and plants, accounts for around half of the global biogenic nonmethane hydrocarbon emissions, with an estimated terrestrial source strength of 500 Tg C yr À1 [Guenther et al., 2012]. More recent research has suggested that an oceanic source of isoprene could significantly impact atmospheric chemistry and cloud microphysical properties in the remote marine boundary layer, despite its much lower emission strength in this region [Meskhidze and Nenes, 2006;Wingenter et al., 2007;Gantt et al., 2009;Arnold et al., 2009;Metzger et al., 2010]. However, estimates of global oceanic emissions vary by up to 2 orders of magnitude between bottom-up (seawater derived;~0.1-1 Tg C yr À1 ) [Arnold et al., 2009;Luo and Yu, 2010;Palmer and Shaw, 2005;Ito and Kawamiya, 2010;Gantt et al., 2009] and top-down approaches (1.68-11.6 Tg C yr À1 ) [Arnold et al., 2009;Luo and Yu, 2010]. The factors controlling marine isoprene emissions are still poorly understood, partly due to a paucity of field measurements, especially those with suitable supporting data.
A review of previous work on marine isoprene and its likely oceanic controls can be found in Shaw et al. [2010], which is briefly repeated and extended here with more recent studies. Bonsang et al. [1992] reported the first observations of isoprene in seawater and suggested a biological source based on the association with concurrently measured chlorophyll a concentrations ([Chl a]). This was further supported by a number of field measurements [Milne et al., 1995;Broadgate et al., 1997;Baker et al., 2000;Wingenter et al., 2004;Moore and Wang, 2006;Kurihara et al., 2010Kurihara et al., , 2012Tran et al., 2013;Kameyama et al., 2014;Zindler et al., 2014;Ooki et al., 2015] and laboratory studies [Moore et al., 1994;McKay et al., 1996;Shaw, 2001;Shaw et al., 2003;Gantt et al., 2009;Bonsang et al., 2010;Exton et al., 2013] that observed correlations with Chl a. Additionally, photochemical production from the sea surface microlayer has recently been proposed as a potential abiotic source [Ciuraru et al., 2015].
The importance or precise nature of a biological sink within the ocean is still unconfirmed. Acuña Alvarez et al. [2009] reported isoprene consumption by various hydrocarbon-degrading bacteria from estuarine sediments in laboratory experiments, while Shaw et al. [2003] did not find an effect of introducing bacteria into cultures.
Based on these observations of isoprene-Chl a relationships, along with Chl a-normalized laboratory monoculture production rates, several authors have attempted a global extrapolation of isoprene fluxes. Palmer and Shaw [2005] assumed steady state oceanic isoprene concentrations, scaling isoprene production by satellite [Chl a] and balancing it with losses due to chemical and biological removal, water column mixing, and air-sea exchange (the latter being the largest loss term by far). They then used the obtained isoprene water concentrations to model global fluxes. Arnold et al. [2009] refined this approach by differentiating isoprene production rates by phytoplankton functional type (PFT), following findings from laboratory studies. Production in the water column and loss to the atmosphere were taken to be in steady state, and both remotely sensed [Chl a] and PFT data based on the PHYSAT algorithm were used for global scaling.
Another recent global emission estimate has been computed using a physically-based method that takes into account known influences of several additional parameters on isoprene production based on laboratory studies [Gantt et al., 2009]. It uses global PFT distributions and [Chl a] from satellite data with PFT-specific production rates and ambient light levels to derive hourly production within the entire euphotic depth of the water column. As in the previous methods, the production is assumed to result in instantaneous emission to the atmosphere, which in combination with the light dependence in this approach leads to a diurnal profile of the flux. However, observations suggest that significant isoprene production takes place at depths of 5 m or deeper [cf. Bonsang et al., 1992;Milne et al., 1995;Moore and Wang, 2006;Tran et al., 2013] (vertical distributions in this work) and so would only gradually be vented to the atmosphere, which would smooth out the diurnal nature of production. In their paper, Gantt and coworkers already reported that field observations [Matsunaga et al., 2002] did not show the strong diurnal pattern and zero nighttime emissions that their model predicted.
Despite a qualitative consensus about the existence of a biological control, published equations to predict isoprene concentrations in water, based on [Chl a] vary widely (this is discussed further in section 3.4 and Table 4). Likely reasons are the apparent dependence of isoprene production on PFTs, or even lower taxonomic groups, as well as growth conditions in laboratory cultures [Shaw et al., 2003;Gantt et al., 2009;Bonsang et al., 2010;Exton et al., 2013], and also variations observed in the field with sea surface temperature (SST) [Ooki et al., 2015].
This work reports a substantial number of new isoprene observations in the surface ocean and marine atmosphere across a large range of latitudes, thereby significantly increasing the available data set and enabling the estimation of sea-to-air fluxes. Supporting measurements of the proposed and potential additional controls provide new field data to improve our understanding of the different parameters influencing marine isoprene. We investigate the suitability of several parameterizations, derived from the new data, for predicting isoprene fluxes on a global scale.
CTD (conductivity-temperature-depth) casts were performed twice daily (pre-dawn and solar noon) during AMT 22 and 23 and once daily (morning) during ACCACIA 1 and 2. Water was subsampled from the Niskin bottles (20 L) for further processing, described in detail separately for each variable. Various measurements were also made from the ships' clean underway seawater supply inlet (nominal depth 5-6 m).

Isoprene Measurements by (autoP&T)-TD-GC-MS
Trace gases were measured in water and air during all four cruises by (automated Purge & Trap)-Thermal Desorption-Gas Chromatography-Mass Spectrometry ((autoP&T)-TD-GC-MS), with details given in the supporting information (Text S1).
Briefly, isoprene in air was measured as discrete samples of 1-2 L air, from a continuously pumped air inlet. Isoprene in seawater was analyzed from the ships' pumped nontoxic seawater supplies as well as from CTD casts from within the photic depth, using the semiautomated P&T system described in Andrews et al. [2015], with modifications detailed in the supporting information. Analysis was performed by GC-MS, with regular calibrations using a premixed gas standard. Detection limits were calculated dynamically to account for changing instrument sensitivity and ranged between 0.1-5 pmol L À1 and 0.1-2.5 pptv (parts per trillion) for water and air, respectively, and the uncertainty of the analysis was typically around 10-20%. Air data were filtered for contamination arising from the ship's exhaust using hydrocarbon concentration thresholds.

Biological and Supporting Data
With the exception of ACCACIA 1, a variety of biological data sets were collected and analyzed during the cruises, with methods described in the supporting information (Text S2). Data included gross biomass (Chl a) and integrated primary production (intPP) using the methods described in Tilstone et al. [2009], as well as flow cytometry and pigment data from high-performance liquid chromatography analysis (HPLC). Furthermore, CHEMTAX analysis was performed for ACCACIA 2 pigment data.
Meteorological data such as wind speed and sea surface temperature (SST) were obtained from the ship systems, provided by the British Oceanographic Data Centre (BODC).

Further Analysis
The isoprene sea-to-air flux was calculated following the approach of Johnson [2010], assuming air concentrations of zero due to the large degree of supersaturation of the gas in seawater (see supporting information Text S3.1). [Chl a] and SST were extracted from Moderate Resolution Imaging Spectroradiometer (MODIS)-Aqua (http://oceancolor.gsfc.nasa.gov/cms/); intPP was calculated using MODIS-Aqua Chl a and photosynthetically active radiation (PAR) and mixed layer depth from a climatology (see supporting information Text S3.2). Isoprene concentrations were produced from the satellite data using the SST-binned combined regressions (ALL) detailed in section 3.3 for [Chl a] and intPP (total) .

Isoprene Air and Water Concentrations
A comprehensive, internally consistent set of isoprene atmospheric mixing ratios and surface ocean concentrations is reported here, spanning around 125°of latitude in the Atlantic and Arctic Oceans at comparatively high spatial and temporal resolution. The data substantially increase the number of currently published marine isoprene measurements, especially in oligotrophic oceanic regions [cf. Shaw et al., 2010], as can be seen in Figure 1.
In a typical depth profile (Figure 2), the isoprene concentrations generally followed the shape of the Chl a profile, with the maximum isoprene concentration slightly shallower than the deep chlorophyll maximum when present. This is in agreement with published vertical distributions [Bonsang et al., 1992;Milne et al., 1995;Moore and Wang, 2006;Tran et al., 2013].
As shown in Figure and summarized in Table 1, surface ocean isoprene concentrations varied between 1 and 66 pmol L À1 , with mean concentrations around 20 pmol L À1 for all cruises except the Arctic winter/spring cruise, during which concentrations were much lower (mean of 4 pmol L À1 ). This compares well with previously published mean values of typically 25-30 pmol L À1 in the Atlantic and around 10-80 pmol L À1 across all oceans (Table 2). Some authors have also observed much higher concentrations, mostly where biological productivity in the study region was high (e.g., coastal or spring blooms) [Tran et al., 2013;Kameyama et al., 2014;Ooki et al., 2015].

10.1002/2016GB005531
There was no significant difference between average daytime and nighttime concentrations during any of the cruises (Figure 3), suggesting a lack of diurnal variation for isoprene in surface waters. (Day defined using a threshold of modeled jNO 2 > 5 × 10 À5 s À1 (GEOS-Chem v9.2, 4 × 5°resolution) [Sherwen et al., 2016] during the sampling period, as the photolysis rate of NO 2 (jNO 2 ) is zero at night, threshold allowing for noise in the data.) There were too few ACCACIA 2 nighttime values (due to light summer nights) to establish a meaningful difference to daytime.
Atmospheric mixing ratios were consistently low and often below the detection limit (DL), with mean values of around 1-3 pptv. This lies at the lower end of the range of previous observations; several studies found tens to a couple of hundreds of pptv isoprene above phytoplankton blooms or in coastal areas (Table 2).   (43) a Mean and median are given for only data > DL as well as for all data, substituting a value of half the DL (0.5 × DL) for points that fall below the DL but are otherwise not flagged as bad (or probably bad) data. A dash indicates that no data were < DL. Number of data points n applies to both mean and median (only shown with median). b AMT cruises: North and South Atlantic (October/November); ACCACIA 1: Arctic (March); and ACCACIA 2: Arctic (July/August). c May be compromised due to hydrocarbon contamination. d Dynamic DL. e Surface CTD (<10 m depth).

Global Biogeochemical Cycles
10.1002/2016GB005531 The magnitude of our reported atmospheric concentrations is consistent with past studies in which an insignificant role for isoprene for marine secondary organic aerosol has been inferred [Arnold et al., 2009]. Our air measurements also provided a constraint for flux calculations in this work (see Text S3.1). Similar to water, no diurnal trends were observed in air, even when levels were consistently above the detection limit. However, the generally low biological activity during our cruises is comparable to "background" or "remote" regions for which isoprene near or below the detection limit is often reported ( Table 2). Careful filtering for potential ship influence in this work may also have resulted in lower values; comparisons are limited as the extent of filtering has, to our knowledge, not been described in detail for other studies.
As a substantial part of our air data fell below the DL, mean and median values were also calculated including a value of 0.5 × DL for data < DL. The results should represent a more accurate approximation of the typical isoprene mixing ratios by accounting for low values (and assuming that samples < DL are unlikely to be at zero). A mean or median calculated from only data > DL is necessarily biased toward higher values and omits some of the valid data (cf. Table 1).

Sea-to-Air Fluxes
Fluxes determined in this study are of comparable magnitude to previously published fluxes for the Atlantic and Arctic Oceans (Figure 4 and values in Table S1, supporting information), taking into account the overall large spatial and temporal variability. The error bars shown on the previously published fluxes reflect the large range in isoprene concentrations observed within each study, with an additional distinction between different water masses in the Tran et al. [2013] data. The Tran et al. [2013] data shown are the two water masses with the lowest and highest isoprene concentrations and fluxes observed during that study, to highlight the extent of the variation in the North Atlantic and Arctic Oceans.
There is a further distinction for different approaches to the flux calculations in two studies (values in Table S1). The calculated isoprene fluxes from Meskhidze and Nenes [2006] highlight that the choice of empirical relationship strongly influences results from remotely sensed Chl a: mean fluxes calculated according to Palmer and Shaw [2005] (based on laboratory production rates) were 3 times higher than the SOFeX method (calculated from in situ Southern Ocean isoprene concentrations), even though both approaches scaled surface isoprene by the same satellite [Chl a]. Furthermore, the choice of flux calculation can affect the results by more than 50%, as evident from the two methods employed by Tran et al. [2013]. Owing to this combination of large variations in concentrations and calculations, it is difficult to obtain close agreement between different studies. Considering this, the mean values and ranges for the cruises from this work generally compare very well (within a factor of 2) with the corresponding literature, such as Atlantic values from various studies, or the in-ice ACCACIA 2 and all ACCACIA 1 data with Tran et al. [2013] polar waters.
From equation (1) (supporting information Text S3.1), it can be seen that the main controls of the calculated bottom-up fluxes are wind speed and isoprene surface ocean concentration, as k w is strongly dependent on wind speed and flux scales linearly with seawater concentration. The former is most obvious in Figure 4 for Global Biogeochemical Cycles 10.1002/2016GB005531 the AMT cruises, with distinctly higher fluxes for the periods with higher wind in one year compared to the other year (around 30-45°S), while isoprene concentrations matched closely between both years (see Figure 1a).
The latter is evident from the larger isoprene concentrations for ACCACIA 2 out-of-ice compared to in-ice, with higher fluxes despite similar wind speeds. As wind speed can be obtained on a global scale from a number of sources including climatologies, comprehensive databases, and satellite products (e.g., from http://giovanni.gsfc.nasa.gov/giovanni/), a better understanding of the factors affecting isoprene concentrations would be a large step toward a more accurate global extrapolation of isoprene fluxes. We will therefore now focus on potential controls of isoprene in the surface water.

Potential Biological Controls on Seawater Isoprene Concentrations
Based on previously published associations of marine isoprene with phytoplankton biomass (Chl a), PFTs, and general biological productivity, relationships of surface ocean isoprene concentrations with concurrently monitored biological variables were investigated. Spatial and temporal variation of selected biological variables is shown alongside isoprene concentrations in Figure 5. These analyses were not possible for ACCACIA 1 due to the lack of supporting biological data; therefore, the cruise was excluded from the following investigations. Due to the nature of the cruise tracks, AMT cruises are shown as latitudinal transects, while ACCACIA data are shown as time series.
In order to explore controls on concentrations of isoprene in seawater, we assume that concentrations are directly correlated with production rates, i.e., that oceanic losses do not vary significantly across the Global Biogeochemical Cycles 10.1002/2016GB005531 different water bodies sampled. Chemical and physical loss rates due to reaction in the water column and downward mixing are believed to be slow, with estimated lifetimes of 19, >1000, and >250 days with respect to reaction with OH, singlet oxygen, and losses due to mixing, respectively [Palmer and Shaw, 2005]; thus, to a first approximation they can be assumed to be constant. (Using the values of k 1O2 = 10 6 M À1 s À1 and [ 1 O 2 ] = 10 À14 M cited in Palmer and Shaw [2005] results in a calculated lifetime of 1150 rather than 115 days as stated by the authors. Using the value of k 1O2 = 3.7 × 10 4 M À1 s À1 (in chloroform solution) given in the original reference [Monroe, 1981], the lifetime with respect to singlet oxygen becomes >30 000 days.) Biological losses as a result of bacterial consumption were similarly assumed to be very small by these authors, based on evidence from laboratory experiments by Shaw et al. [2003]. The main loss from the surface ocean is, in fact, due to sea-air gas exchange, which is strongly dependent on wind speed (see section 3.2 above). This means that our assumption breaks down at high wind speeds especially, as the concentrations are controlled by losses to a larger extent and therefore no longer proportional to production (also observed by Moore and Wang [2006] outside a fertilized patch).
Several core measurements are collected on AMT voyages each year, including pigments, primary production (PP), and phytoplankton abundance and community composition. Typical results are described in detail in Robinson et al. [2002], Tilstone et al. [2009Tilstone et al. [ , 2015aTilstone et al. [ , 2015b, Tarran et al. [2006], Heywood et al. [2006], and Aiken et al. [2009]. Significant linear relationships were found (for each cruise and for combined data sets) between isoprene and Chl a, other pigments including the sum of photoprotective carotenoids (PPC), some PFTs, and PP (only AMT data for the latter two). Correlations generally improved when binned by sea surface temperature (SST) with a threshold of 20°C (Table 3 and Figure 6). The threshold value was chosen based on inspection of the correlation plots for each variable, which generally exhibited a change in slope at that value ( Figure 6; also cf. Figures S2-S5 in the supporting information). For variables where binning produced an insignificant correlation (p > 0.05) for at least one bin, only relationships for the complete data set are shown and used for further calculations.
Regression equations were calculated as the Robust Line of Organic Correlation (RLOC), which is affected less than the LOC approach by outliers and points below the detection limit [Khalil and Adamowksi, 2012]. Equations were determined for each data set internally (A22 for AMT 22, A23 for AMT 23, and Ac2 for Figure 5. Transect/time series plots of surface ocean isoprene concentrations alongside selected biological measurements: total Chl a, sum of photoprotective carotenoids (PPC; sum of aeaxanthin, alloxanthin, diadinoxanthin, αand β-Carotenes, and diatoxanthin), integrated primary production (intPP total), and Synechococcus and Prochlorococcus cell counts (Syn, Pro; Syn in Syn + Pro are scaled by a factor of 3.5 to account for cell-normalized isoprene production rates given in Shaw et al. [2003]).

10.1002/2016GB005531
ACCACIA 2) and also for all available data sets for that variable combined (ALL; may refer to all three cruises or only AMT data; Table 3).

Isoprene-Pigment Relationships
Slopes of the isoprene-Chl a correlations for waters with SST <20°C in the current study (Table 3) are of the same order of magnitude as in published relationships for similarly low-SST regions, listed in Table 4. The threshold SST of around 20°C for these relationships agrees in principle with findings by Ooki et al. [2015] who calculated different linear regressions for different SST bins. Their data were, however, divided into four SST bins (<3.3, 3.3-17, 17-27, and >27°C) as opposed to only two in this work (see section 3.3). It is likely that the SST is an indicator of changes in biological variables rather than a direct cause of the change in isoprene emissions. The differences between Ooki et al. [2015] and our work (cf. Figure 7) could stem from differences in the phytoplankton community between the Atlantic and Pacific/Indian Oceans, different sampling seasons, and frequencies and different [Chl a] measurement techniques (see below). The discrepancies between the studies indicate that either data set is likely not directly applicable to all parts of the world's oceans and further refinement is needed in order to obtain a more accurate global algorithm, especially for Arctic regions. Exton et al. [2013] also reported different isoprene production rates (normalized to [Chl a]) from phytoplankton monocultures grown at different temperatures in accordance with their origins (polar, temperate, or tropical). However, the relative difference in slope between the temperate and tropical regimes was considerably smaller than between slopes for low-and high SST isoprene-[Chl a] correlations in the current study (factors of <2 and 6-8, respectively, even larger for Ooki et al. [2015]; Table 4), which could point to limited comparability of laboratory and field data. This might include additional controls on the production in the field such as the proposed photochemical surface source [Ciuraru et al., 2015] and/or currently unknown loss processes. It highlights that global emission estimates based on laboratory data  may not accurately reflect those in the wider marine environment and validation against field measurements is required.
A significant limitation to comparability between studies, unrelated to potential differences in isoprene measurements, may be Chl a measurement techniques. HPLC pigment analysis can be considered the most accurate method but is more expensive and labor intensive than fluorometric Chl a analysis. The latter is faster and easier, allowing shipboard analysis to be completed typically in less than 2 days (including 24 h for pigment extraction). The different techniques can provide results that vary by up to a factor of 2, with HPLC Chl a concentrations typically lower than those reported from fluorometric analysis [Jacobsen and Rai, 1990]. Fluorometric analysis does not physically separate Chl a from the suite of pigments present; rather, it analyzes the pigment extract as a whole. It has been shown that Chl b and Chl c can interfere with the estimation of Chl a [Coveney, 1982;Trees et al., 1985;Welschmeyer, 1994], and HPLC analysis indicated that both Chl b and Chl c were present in samples from this study. Additionally, the pigment extraction method should be considered as a source of discrepancy between the two results. While both methods used 90% acetone as a solvent, extraction for fluorometric analysis relied on this solution over time to fully extract the pigments. This has been suggested to be less efficient than additionally using physical disruption through sonication [Neveux and Panouse, 1987;Jacobsen and Rai, 1990], which was the extraction method used for HPLC analysis.
Associations of isoprene with specific pigments other than Chl a are consistent with the literature, as the isoprene-pigment correlations found in this study included both fucoxanthin and zeaxanthin (Figures S2-S3 in the supporting information). Diatoms, which contain fucoxanthin as a dominant carotenoid, have been shown to produce isoprene [Moore et al., 1994;Milne et al., 1995;McKay et al., 1996;Colomb et al., 2009;Bonsang et al., 2010;Exton et al., 2013]. Prochlorococcus spp. have also been shown to produce isoprene [Shaw et al., 2003] and contain zeaxanthin. However, a significant correlation between isoprene and DV-Chl a, which is specific to Prochlorococcus, was only present in the high SST bin ( Figure S2), so although Prochlorococcus may at least be partly responsible for the isoprene signal at high SST, other phytoplankton containing zeaxanthin are also likely to have contributed throughout.

Predicting Isoprene Concentrations From In Situ Observations
The regression equations (Table 3) were used to predict isoprene concentrations from the different biological measurements made during each cruise and were found to replicate the observed surface isoprene reasonably well (Figure 7; Table S2 in the supporting information). The empirical formulae published by Ooki et al. [2015] (referred to as O15; details in Table 4) were also applied to the current data sets and gave generally good agreement except for Arctic data and between 20 and 27°C SST (both underpredicted by O15; Figure 7).

Predicting Isoprene Concentrations From Remotely Sensed Data
Satellite data could provide a basis for global extrapolation using the empirical relationships presented here, if they can be shown to reproduce observations well. Ocean surface Chl a concentration is a standard satellite product, the accuracy of which has been rigorously assessed for the Atlantic Ocean [Brewin et al., 2016]. IntPP has become routinely derived from remote measurements and is well validated [Tilstone et al., 2009[Tilstone et al., , 2015a[Tilstone et al., , 2015bCarr et al., 2006;Friedrichs et al., 2009;Saba et al., 2010;Brewin et al., 2017]. Both were used here to evaluate their suitability for predicting marine isoprene concentrations, and hence ultimately sea-to-air fluxes, globally. The best match with in situ data would be expected for daily satellite products; however, we also investigate the performance of composite images compiled over longer time periods, which are relevant for stable areas of the ocean such as Atlantic oligotrophic gyres, as daily images provide less comprehensive data due to cloud and reduced spatial coverage. Only paired values for isoprene observations and satellite-derived values were used in comparisons. Only a few pairs were available at low SST due to comparatively poor satellite coverage and increased cloud cover over high latitudes. The satellite validation was performed using the full data set, but as a control was also calculated for only independent data (Text S3.2 in the supporting information).
Predicted isoprene from optically derived in situ [Chl a] produced a close match with observations (Figures 8a and S6a in the supporting information), showing that the algorithm works well with underway measurements of Chl a derived from the absorptive properties of phytoplankton particles, which were validated and corrected for bias against HPLC pigment measurements on the same cruise [Brewin et al., 2016].
There was also overall good agreement between observations and isoprene estimated from both daily and monthly satellite [Chl a] (Figures 8a and 8b; Table S1 and Figure S6a in the supporting information), with trends being captured well for the AMT cruises and predicted values within a factor of 3 of the measurements 90% of the time. No obvious difference could be seen in goodness of fit between the daily and monthly data (Figure 8b; Figure S6b in the supporting information; R 2 = 0.33 and 0.23, respectively). A slight underestimation in the oligotrophic gyres (higher SST) could be partially due to degradation of the sensor (slightly lower satellite [Chl a] than in situ techniques in recent years) [Brewin et al., 2016]. For coastal data (>50°N), satellite [Chl a] resulted in significant overestimation (Figure 8a; Figure S6a in the supporting information), which could be related to the use of an open ocean algorithm for optically complex waters.
Overestimation of isoprene concentrations in some coastal locations can also be observed in a comparison of O15 with satellite-derived values (Figure 9a). However, the inverse is even more common for that data set: in situ data near the Japanese coast (around 138-148°E, 34-44°N; especially April/May 2009) show high isoprene concentrations of >100 pmol L À1 (cf. Figure 1), which are often underestimated by a factor of 2-4 by the satellite algorithm. Even though it does not capture the magnitude of the variations ( Figure S8 in the supporting information), the [Chl a]-based algorithm is able to reproduce general trends in the in situ data and 90% of the absolute measured values within a factor of 3 (Table S2 in the supporting information).
Monthly intPP predicts isoprene with varying degrees of success: AMT data are generally good (91% within a factor of 2), with the exception of low latitudes (high SST; Figure 8c; Figure S6 in the supporting information). Figure 8. In situ and satellite-derived predicted isoprene concentrations from this work; (a) AMT 22, Chl-based algorithm, with satellite-derived monthly data extracted every 20 min along the cruise track and "5d"-satellite data from matchups ±2 days of the observation, alongside isoprene based on in situ AC-s [Chl a]; (b) all data (AMT, ACCACIA) matched with "5 day" satellite-derived data, for Chl-based algorithm, with 1:1 line; (c) as Figure 8a but for intPP-based algorithm (monthly satellite data only); (d) as Figure 8c but for monthly satellite data and intPP-based algorithm. Some outlier points (circled; coastal data) were excluded from the correlation. Further data in Figure S6 (supporting information).

10.1002/2016GB005531
Arctic values are within a factor of 10 (96% within factor of 5) of the measurements, but trends are not always reproduced and additionally suffer from sparse coverage particularly in spring ( Figure S6 in the supporting information). A comparison of intPP-predicted isoprene with O15 data appears reasonable (Figure 9b inset graph; Table S3 in the supporting information), but there is significant underprediction in the tropical North Pacific (Figure 9b). A latitudinal plot shows that performance is also poor at high latitudes and trends are not reflected very well overall ( Figure S8 in the supporting information).
These difficulties can be partly attributed to the satellite product being less well validated than Chl a, especially outside the Atlantic, and can vary significantly depending on the satellite model used [Carr et al., 2006]. Lee et al. [2015] suggest that PP models need to be carefully tuned for the Arctic in order to perform well for that region and note that the variability of PP is underestimated by most models. Furthermore, PP measured in situ using 14 C on-deck incubations has been found to differ significantly from other methods [Quay et al., 2010]. The satellite production model used here was not parameterized using on-deck incubations, instead using photosynthesis-irradiance experiments [Brewin et al., 2017]. IntPP is controlled by a variety of factors; its relationship with isoprene may be less consistent on a global scale than for Chl a, and relationships for the open ocean cannot necessarily be applied to coastal regions. In addition, despite a similar correlation coefficient and bias for the intPP-based algorithm compared to that based on Chl a, the regression was performed for far fewer points and should therefore be considered less reliable. Despite these issues, the algorithm still predicts observed values within a factor of 3 for all oceanic regions evaluated here at least 84% of the time (and 98% within a factor of 5; Table S3 in the supporting information).

Conclusions
We found strong relationships between surface ocean isoprene concentrations and concurrently monitored biological variables such as Chl a, PFTs, and productivity, with better correlations for two separate SST bins with a threshold at 20°C. This is qualitatively consistent with previous studies [Bonsang et al., 1992;Moore et al., 1994;Milne et al., 1995;McKay et al., 1996;Broadgate et al., 1997;Baker et al., 2000;Shaw et al., 2003;Wingenter et al., 2004;Moore and Wang, 2006;Gantt et al., 2009;Bonsang et al., 2010;Exton et al., 2013;Kurihara et al., 2010Kurihara et al., , 2012Kameyama et al., 2014;Ooki et al., 2015]. Our current lack of understanding of isoprene production processes and biological functions prevents an assignment of factors responsible for this threshold. We also observed associations of isoprene with specific pigments other than Chl a, which to our knowledge have not been previously reported. A strong and relatively consistent relationship with the sum of photoprotective carotenoids across all three cruises, could point to a photoprotective function of isoprene or at least to isoprene being a by-product of a Figure 9. Ratio (log 10 scale) of predicted isoprene from satellite data to the Ooki et al. [2015] data set; (a) using Chl a-based algorithm (one outlier at 1.5) and (b) using intPP-based algorithm (full-scale version of inset graph in supporting information Figure S7). Maps created with Ocean Data View [Schlitzer, 2016].
Global Biogeochemical Cycles 10.1002/2016GB005531 photoprotective response of the organisms present. An alternative explanation could be the effect of a physical driver that affects both PPC and isoprene abundances: photoprotective pigments are abundant in surface samples in highly stratified marine environments due to the high light exposure of the cells [Babin et al., 1996]. This could also correlate with other light-driven reactions or processes that occur under these conditions and may increase isoprene concentrations. Emission has been linked to light (and light stress) by various laboratory studies [Shaw et al., 2003;Gantt et al., 2009;Bonsang et al., 2010;Meskhidze et al., 2015] but not yet directly to a photoprotective mechanism of the phytoplankton such as has been proposed for terrestrial plants [Sharkey and Yeh, 2001].
Arguably, individual regression analyses have limited diagnostic potential as each assumes by definition that the investigated variable is solely responsible for the total isoprene, so that the resulting equation is only useful for predicting isoprene if that assumption is true and the variables are independent of each other. Following the same argument, a steeper slope by no means automatically implies that the entity in question (e.g., microplankton; cf. Figure S5 in the supporting information) is a strong emitter: the phenomenon can also be caused by the presence of other organisms or processes responsible for isoprene emission in regions where the examined variable has low concentrations.
For a global emission estimate, the investigated variables must have good global coverage so that an extrapolation using the equations presented here is possible. Currently, surface [Chl a] is available as a standard satellite product, while dominant PFTs can be obtained from ocean color satellite products using the PHYSAT method [Alvain et al., 2008;Brewin et al., 2011;Ben Mustapha et al., 2014;IOCCG, 2014] and intPP can also be derived from satellite data [Carr et al., 2006;Brewin et al., 2017]. Information on PPC distributions can be obtained from data sets such as MAREDAT [Peloquin et al., 2013] and, potentially, from biogeochemical and ecosystem models that are beginning to resolve PPC concentrations on regional and global scales [Bissett et al., 1999, Dutkiewicz et al., 2015. If all parameterizations found here can be extrapolated, the degree to which they converge on a global value should give an indication of how well they reflect the actual emissions.
The extrapolations based on satellite Chl a and intPP performed as part of this study, using the following equations, show that the proposed algorithms are able to reproduce in situ data reasonably well, including data sets acquired independently and in different ocean basins. Chl a generally captured the shape and trends of the data better than intPP but could still not closely reproduce the magnitude of the variations.
This study has contributed a large number of observations to the existing data set of marine isoprene, covering areas of the Arctic and the Atlantic Ocean at basin scale and providing representative values of isoprene in surface water for a large range of latitudes (80°N-45°S). Using a variety of supporting data, relationships with different biological variables have been confirmed and extended. When applied to the different cruises, equations derived from the combined data sets were able to predict isoprene concentrations close to the measured values, including Arctic data. However, no single variable has been identified as a more suitable proxy for predicting isoprene concentrations in water than Chl a, which is currently the most widely used, except perhaps integrated PP and photoprotective carotenoids (PPC). It was nevertheless shown that a separation of ocean regions by temperature ranges is crucial for most proxies in order to obtain representative predictions with empirical relationships. The uncertainty in setting global threshold values for that separation, as seen when contrasting Atlantic and Pacific/Indian Ocean data sets, highlights the need for further field measurements spanning different oceans and seasons and comparison of empirical relationships between studies. Ideally, trace gas measurements should be accompanied by as wide a range of biological measurements as possible, in order to validate the correlations found in this work and examine the suitability of the suggested additional proxies for predicting isoprene concentrations on a global scale.