Coastal ocean acidification and increasing total alkalinity in the northwestern Mediterranean Sea

Coastal time series of ocean carbonate chemistry are critical for understanding how global anthropogenic change manifests in near-shore ecosystems. Yet, they are few and have low temporal resolution. At the time series station Point B in the northwestern Mediterranean Sea, seawater was sampled weekly from 2007 through 2015, at 1 and 50 m, and analyzed for total dissolved inorganic carbon (CT) and total alkalinity (AT). Parameters of the carbonate system such as pH (pHT, total hydrogen ion scale) were calculated and a deconvolution analysis was performed to identify drivers of change. The rate of surface ocean acidification was −0.0028± 0.0003 units pHT yr−1. This rate is larger than previously identified open-ocean trends due to rapid warming that occurred over the study period (0.072± 0.022 C yr−1). The total pHT change over the study period was of similar magnitude as the diel pHT variability at this site. The acidification trend can be attributed to atmospheric carbon dioxide (CO2) forcing (59 %, 2.08± 0.01 ppm CO2 yr−1) and warming (41 %). Similar trends were observed at 50 m but rates were generally slower. At 1 m depth, the increase in atmospheric CO2 accounted for approximately 40 % of the observed increase in CT (2.97± 0.20 μmol kg−1 yr−1). The remaining increase in CT may have been driven by the same unidentified process that caused an increase in AT (2.08± 0.19 μmol kg−1 yr−1). Based on the analysis of monthly trends, synchronous increases in CT and AT were fastest in the spring–summer transition. The driving process of the interannual increase in AT has a seasonal and shallow component, which may indicate riverine or groundwater influence. This study exemplifies the importance of understanding changes in coastal carbonate chemistry through the lens of biogeochemical cycling at the land–sea interface. This is the first coastal acidification time series providing multiyear data at high temporal resolution. The data confirm rapid warming in the Mediterranean Sea and demonstrate coastal acidification with a synchronous increase in total alkalinity.


Introduction
Maintaining time series of oceanographic data is essential for understanding anthropogenic changes in the ocean (Tanhua et al., 2013).On land, fossil fuel burning, cement production, and land use changes have contributed ∼ 600 Gt carbon to the atmosphere during the period 1750-2015(Le Quéré et al., 2016).In the recent decade 2006-2015, an estimated 25 % of this anthropogenic carbon has been absorbed by the ocean in the form of carbon dioxide (CO 2 ; Le Quéré et al., 2016), and is causing global changes to the ocean carbonate system.Absorption of CO 2 by seawater produces carbonic acid, which decreases seawater pH, and is of great concern for biological processes and marine ecosystems (Doney et al., 2009;Gattuso and Hansson, 2011;Pörtner et al., 2014).Since the preindustrial era, global mean ocean pH has declined by 0.1 (Rhein et al., 2013).Due to the declining trend of ocean pH with increasing anthropogenic CO 2 , the process is termed "ocean acidification".This expression represents a suite of chemical changes, including increases in total dissolved inorganic carbon (C T ) and partial pressure of CO 2 (pCO 2 ) and decrease in calcium carbonate saturation states ( , aragonite and calcite; Dickson, 2010).Rates of ocean acidification differ by ocean region and range from Published by Copernicus Publications on behalf of the European Geosciences Union.−0.0026 (Irminger Sea, North Atlantic) to −0.0013 (South Pacific) units pH yr −1 (Bates et al., 2014).Such time series remain spatially limited, especially in coastal regions, which provide valuable ecosystem services (Barbier et al., 2011;Costanza et al., 1997) and are under high anthropogenic impact (Halpern et al., 2008).Here, we present the first coastal ocean acidification time series at high temporal resolution.
Compared to the global ocean, marginal seas play a critical role in anthropogenic CO 2 storage via enhanced CO 2 uptake and export to the ocean interior (Lee et al., 2011).As a marginal sea, the Mediterranean Sea has a naturally high capacity to absorb but also buffer anthropogenic CO 2 (Álvarez et al., 2014;Palmiéri et al., 2015).This is primarily due to the high total alkalinity (A T ) of Mediterranean waters and overturning circulation (Lee et al., 2011;Palmiéri et al., 2015;Schneider et al., 2010).In the Mediterranean Sea, the salinity-A T relationship is driven by the addition of river discharge and Black Sea input, which are generally high in A T (Copin-Montégut, 1993;Schneider et al., 2007).Combined with evaporation, this results in higher A T and salinity in the Mediterranean Sea compared to the Atlantic Mediterranean source water (Jiang et al., 2014).On average, Mediterranean Sea A T is 10 % higher than in the global ocean (Palmiéri et al., 2015).The surface ocean acidification rate, estimated at pH T (total hydrogen ion scale) of −0.08 since 1800, is comparable to that of the global ocean despite a 10 % greater anthropogenic carbon inventory (Palmiéri et al., 2015).Due to its important role in carbon sequestration and ecological sensitivity to anthropogenic change with economic consequences (Lacoue-Labarthe et al., 2016), the Mediterranean Sea could provide insight to global trends (Lejeusne et al., 2010).
Over the last few years, numerous studies have estimated ocean acidification rates across the Mediterranean Sea (Table 1).Together, these studies cover various periods with a range of techniques yielding different results.For example, estimates of change in pH of bottom waters since the preindustrial era range between −0.005 and −0.06 (Palmiéri et al., 2015) and up to as much as −0.14 for full profile estimates (Touratier and Goyet, 2011).Techniques for estimating ocean acidification in the Mediterranean Sea thus far include the following: (1) hind-casting, using highresolution regional circulation models (Palmiéri et al., 2015), the TrOCA approach as applied to cruise-based profile data (Krasakopoulou et al., 2011;Touratier and Goyet, 2011;Touratier et al., 2016), and others (Howes et al., 2015); (2) partially reconstructed time series (Marcellin Yao et al., 2016); (3) comparative study periods (Luchetta et al., 2010;Meier et al., 2014); and (4) sensor-based observations over a short study period (Flecha et al., 2015).Ocean acidification time series of consistent sampling over many years are lacking for the Mediterranean Sea (The MerMex Group, 2011), particularly along the coast where river discharge influences the carbonate system (Ingrosso et al., 2016).
Table 1.Previous estimates or documentation of pH change ( pH) in the Mediterranean Sea."Total" indicates estimates made for the whole Mediterranean Sea.TrOCA is the "Tracer combining Oxygen, inorganic Carbon, and total Alkalinity" method, NR means "not reported", and PI is "preindustrial era".Compared to the open ocean, shallow coastal sites exhibit natural variability in carbonate chemistry over annual time frames (Hofmann et al., 2011;Kapsenberg and Hofmann, 2016;Kapsenberg et al., 2015), complicating the detection and relevance of open ocean acidification in isolation of other processes (Duarte et al., 2013).In the northwestern (NW) Pacific coast, rapid acidification of surface waters ( pH T −0.058 units yr −1 ) at Tatoosh Island was documented in the absence of changes in known drivers of local pH variability (e.g., upwelling, eutrophication, and more; Wootton and Pfister, 2012;Wootton et al., 2008).Further inshore, in the Hood Canal subbasin of the Puget Sound, only 24-49 % of the estimated pH decline from preindustrial values could be attributed to anthropogenic CO 2 (Feely et al., 2010).The excess decrease in pH was attributed to increased remineralization (Feely et al., 2010).Acidification rates documented along the North Sea Dutch coastline and inlets were highly variable in space, with some exceeding the expected anthropogenic CO 2 rate by an order of magnitude while others exhibited an increase in pH (Provoost et al., 2010).

Method
Variability in coastal carbonate chemistry stems from both physical (e.g., upwelling, river discharge; Feely et al., 2008;Vargas et al., 2016) and biological processes (e.g., primary production, respiration, net calcification).Within watersheds, coastal carbonate chemistry is affected by eutrophication (Borges and Gypens, 2010;Cai et al., 2011), groundwater supply (Cai et al., 2003), and land use and rain influence on river alkalinity (Raymond and Cole, 2003;Stets et al., 2014).Over longer periods, pH can also be influenced by atmospheric deposition (Omstedt et al., 2015).Through primary production and respiration, coastal ecosystems produce pH fluctuations over timescales ranging from hours (e.g., seagrass, kelp) to months (e.g., phytoplankton blooms; Kapsenberg and Hofmann, 2016).Due to existing pH variability in coastal seas, it is necessary to quantify high-frequency trends in order to interpret the pH changes inferred from lowerfrequency sampling.
In this study, we present the first complete time series data quantifying the present-day ocean acidification rate for a coastal site in the Mediterranean Sea, based on weekly measurements of A T and C T sampled from 2007 through 2015.For a subset of this time series, pH variability was documented using a SeaFET ™ Ocean pH Sensor in order to assess hourly pH variability.For comparison and consistency with other ocean acidification time series around the world, we report rates of change based on anomalies (Bates et al., 2014) and identify drivers of change.
2 Materials and methods

Site description
A carbonate chemistry time series was initiated in 2007 and maintained through 2015 in the NW Mediterranean Sea at the entrance of the Bay of Villefranche-sur-Mer, France (Fig. 1): Point B station (43.686 • N, 7.316 • E, 85 m bottom depth).A second site, the Environment Observable Littoral buoy (EOL, 43.682 • N, 7.319 • E, 80 m bottom depth), was used for pH sensor deployment starting in 2014.These two sites are 435 m apart.Point B has been an oceanographic station since 1957.A full site description and research history has been detailed by De Carlo et al. (2013).Briefly, the Bay is a narrow north-south facing inlet with steep bathymetry and estimated volume of 310 million m 3 .The surrounding region is predominantly composed of limestone with a series of shallow, submarine groundwater karst springs (Gilli, 1995).The north current, a major and structuring counterclockwise current in the Ligurian Sea, can sometimes flow close to Point B. The Bay can also be, on occasion, influenced by local countercurrents.Both of these hydrodynamic movements have signatures of river discharge.Limestone erosion can be observed in the A T of rivers nearest to Point B (Paillon, due 4 km west; Var due 10 km west; and Roya due 26 km east).River A T ranges between 1000 and 2000 µmol kg −1 (data from Agence de l'Eau Rhône-Méditerranée-Corse, http://sierm.eaurmc.fr),and is lower than seawater A T .The Paillon River, whose plume on occasion reaches into the Bay (L.Mousseau, personal observation, 2013Mousseau, personal observation, , 2014)), was sampled on 18 August 2014 and had a A T of 1585 ± 0.1 µmol kg −1 (N = 2, Gattuso, unpublished).Due to low primary productivity, seasonal warming drives the main annual variability in carbonate chemistry at this location (De Carlo et al., 2013).

Point B data collection, processing, and analysis
To document long-term changes in ocean carbonate chemistry at Point B, seawater was sampled weekly from 9 January 2007 to 22 December 2015.Samples were collected at 1 and 50 m, using a 12 L Niskin bottle at 09:00 LT.Seawater was transferred from the Niskin bottle to 500 mL borosilicate glass bottles and fixed within an hour via addition of saturated mercuric chloride for preservation of carbonate parameters, following recommendations by Dickson et al. (2007).Duplicate samples were collected for each depth.For each sampling event, CTD casts were performed either with a Seabird 25 or Seabird 25+ profiler whose sensors are calibrated at least every 2 years.Accuracy of conductivity (SBE4 sensor) and temperature (SBE3 sensor) measurements from CTD casts were 0.0003 S m −1 and 0.001 • C, respectively.
Within 6 months of collection, bottle samples were analyzed for C T and A T via potentiometric titration following methods described by Edmond (1970) andDOE (1994), by Service National d'Analyse des Paramètres Océaniques du CO 2 , at the Université Pierre et Marie Curie in Paris, France.Average accuracy of C T and A T measurements was 2.6 and 3 µmol kg −1 , respectively, compared to seawater certified reference material (CRM) provided by A. Dickson (Scripps Institution of Oceanography).Repeatability of replicate samples was better than 3 µmol kg −1 .
Calculations of the carbonate system parameters were performed using the R package seacarb version 3.1 with C T , A T , in situ temperature, and salinity as inputs (Gattuso et al., 2016).Total concentrations of silicate (SiOH 4 ) and phosphate (PO 3− 4 ) were used, when available from Point B (Mousseau, unpublished, http://somlit.epoc.u-bordeaux1.fr/fr/).Detection limits for nutrients were 0.03 µM for SiOH 4 and 0.003 to 0.02 µM for PO 3− 4 ; relative precision of these analyses is 5-10 % (Aminot and Kérouel, 2007).Total boron concentration was calculated from salinity using the global ratio determined by Lee et al. (2010).The following constants were used: K 1 and K 2 from Lueker et al. (2000), K f from Perez and Fraga (1987), and K s from Dickson (1990).Reported measured parameters are temperature, salinity, A T , and C T , and derived parameters are pH T (total hydrogen ion scale), pH T normalized to 25 • C (pH T25 ), pCO 2 , and aragonite ( a ) and calcite ( c ) saturation states.Salinitynormalized changes in A T (nA T ) and C T (nC T ) were calculated by dividing by in situ salinity and multiplying by 38.Except for pH T25 , all parameters are reported at in situ temperatures.
The average uncertainties of the derived carbonate parameters were calculated according to the Gaussian method (Dickson and Riley, 1978) implemented in the "errors" function of the R package seacarb 3.1 (Gattuso et al., 2016).The uncertainties are ±2.7 × 10 −10 mol H + (about 0.015 units pH T ), ±15 µatm pCO 2 , and ±0.1 unit of the aragonite and calcite saturation states.
To quantify interannual changes in carbonate parameters, the data were detrended for seasonality by subtracting the respective climatological monthly means computed for the period 2009-2015 from the time series (hereafter "monthly means").The resulting residuals were analyzed using a linear regression to compute anomaly trends.This approach follows methods from Bates et al. (2014) to allow for comparisons of trends observed at different time series stations.All analyses were performed in R (R Core Team, 2016).

Deconvolution of pH T and pCO 2
To identify proportional contributions of various drivers to ocean acidification trends at Point B, deconvolution of time series pH T and pCO 2 was performed following methods from García-Ibáñez et al. (2016) for observations at 1 and 50 m.The equation is described below for pH T , where changes in pH T are driven by changes in temperature (T ), salinity (S), A T , and C T , over time (t).
Here, , where var is either temperature (T ), salinity (S), A T , or C T .The sensitivity of pH to var ∂pH T ∂var was estimated by calculating pH T using the true observations of var and holding the other three variables constant (mean value of the time series) and regressing it to var.Sensitivity ∂pH T ∂var was then multiplied by the anomaly rate of var (Table 2).The calculation was repeated for pCO 2 dpCO 2 dt in order to compare the rate of increase with that of atmospheric CO 2 .dt , the rate of anthropogenic CO 2 increase was estimated from atmospheric CO 2 concentrations nearest to Point B (Plateau Rosa, Italy, courtesy of the World Data Center for Greenhouse Gases, http:// ds.data.jma.go.jp/gmd/wdcgg/).For these data, missing daily values were linearly interpolated, climatological monthly means were calculated and subtracted from the time series to generate an anomaly time series.A linear regression was performed on anomalies where the slope represents the rate of anthropogenic CO 2 increase in the atmosphere.Finally, to help identify different processes that might have contributed to the observed trends, linear regressions were performed on changes in A T and C T per month (mean value of observations within 1 month) from 2009 through 2015 and on the salinity-A T relationship by year.

SeaFET data collection, processing, and analysis
To capture pH variability at higher-than-weekly sampling frequencies, a SeaFET ™ Ocean pH sensor (Satlantic) was deployed on the EOL buoy (435 m from the Point B sampling site) starting in June 2014, at 2 m depth.Autonomous sampling was hourly and deployment periods averaged 58 ± 25 days with 5 ± 2 calibration samples per deployment.Field calibration samples for pH were collected using a Niskin bottle next to SeaFET within 15 min of measurement.This sampling scheme was sufficient for this site as there is no large high-frequency pH variability.Unlike Point B sampling, SeaFET calibration samples were processed for pH using the spectrophotometric method (Dickson et al., 2007) with purified m-cresol purple (purchased from Robert H. Byrne's laboratory, University of South Florida).In situ temperature, salinity, and A T measured at Point B, within 30 min of the SeaFET sampling, were used to calculate in situ pH T of the calibration samples.
SeaFET voltage was converted to pH T using the respective calibration samples for each deployment period, following the methods and code described in Bresnahan et al. (2014) but adapted for use in R.Only 5 % of the data was removed during quality control, due to biofouling in one deployment and battery exhaustion in another, yielding 610 days of data.The mean offset between calibration samples and the calibrated SeaFET pH time series was ±0.007, indicating a highquality pH dataset (data shown in Fig. 7c).
The estimated standard uncertainty in SeaFET pH T is ±0.01 and was calculated as the square root of the sum of each error squared.The sources of errors are measurement error of spectrophotometric pH (±0.004, N = 68 mean SD of 5 replicate measurements per calibration sample for samples   2. collected between 16 July 2014 and 3 May 2016), spatiotemporal mismatch sampling at EOL (±0.007, mean offset of pH T of the calibration samples from calibrated time series), and variability in purified m-cresol dye batch accuracy compared to tris buffer CRM pH (±0.006, mean offset of pH T of the spectrophotometric measurement of tris buffer from the CRM value).

Time series trends
At Point B from January 2007 to December 2015, more than 400 samples were collected for carbonate chemistry at both 1 and 50 m.Anomaly trends detected at 1 m (Fig. 2) were also significant at 50 m (Fig. 3, Table 2), with the exception that salinity only increased at 50 m (0.0063 ± 0.0020 units yr −1 ).At 1 m, trends were significant for pH T (−0.0028 units yr −1 ), A T (2.08 µmol kg −1 yr −1 ), C T (2.97 µmol kg −1 yr −1 ), pCO 2 (3.53 µatm yr −1 ), and a (−0.0064 units yr −1 ).At the same time, temperature anomaly increased (0.072 • C yr −1 ).Changes in carbonate chemistry parameters were greater at 1 compared to 50 m, with the exception of salinity and temperature.The warming rate at 50 m was slightly greater compared to 1 m, mostly due to increasing summer temperatures since 2007.Time series data are available online (Gattuso et al., 2014).
Strong seasonal cycles of carbonate chemistry parameters were present at Point B at 1 m (Fig. 4).Climatological monthly means (2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) are described briefly and listed in Supplement Table S1.Mean temperature range was 11.2 • C with a maximum of 24.77 ± 1.35 • C in August and minimum of 13.58 ± 0.41 • C in February.The range in A T was 19 µmol kg −1 from June to September.The C T range was 33 µmol kg −1 with a peak in late winter and minimum values in August and October.Due to summer warming co-inciding with the period of peak primary productivity (De Carlo et al., 2013), warming countered the influence of low C T on pH.As a result, pH T reached minimum values in summer (8.02 ± 0.03, July and August) and peaked in late winter (8.14 ± 0.01, February and March), for an overall annual pH range of 0.12.The corresponding pCO 2 range was 128 µatm from February to August.Seasonal cycles were smaller at 50 compared to 1 m (Table S1).

Deconvolution of pH T and pCO 2
Deconvolutions of pH and pCO 2 are presented in Tables 3 and 4, respectively.The estimated anomaly trends dpH T dt , dpCO 2 dt from the deconvolution fall within the error of the observed anomaly trends (Table 2).The contribution of warming to the pH T anomaly (−0.0011 units yr −1 , at 1 m) matched the difference between the trends of pH T and pH T25C (Table 2), which verifies that the deconvolution reproduced influences of temperature sensitivity well.Overall, these results indicate that the deconvolution analyses represent the observed trends well.
At both depths, the predominant driver of (2 %).The atmospheric CO 2 anomaly at Plateau Rosa increased by 2.08 ± 0.01 ppm yr −1 (F 1,3285 = 4664, P 0.001, R 2 0.93) during the study period 2007-2015, and represents the anthropogenic CO 2 forcing on seawater pH.To estimate the influence of anthropogenic CO 2 forcing at Point B, we assume air-sea CO 2 equilibrium (e.g., an increase in atmospheric CO 2 causes an equal increase in seawater pCO 2 ) for the water mass at 1 m.This assumption is based on evidence that Point B is a weak sink for atmospheric CO 2 with near-balanced air-sea CO 2 flux on an annual time frame (De Carlo et al., 2013).Considering the error associated with deconvolution of pCO 2 at 1 m, the atmospheric CO 2 increase can, at most, represent 38-43 % of the total C T contribution to dpCO 2 dt .This leaves 57-62 % of the total C T contribution to pCO 2 trends unaccounted for.
As A T is not influenced by addition of anthropogenic CO 2 to seawater, but it did increase, the next question was whether or not the changes in A T and C T were process-linked.At 1 m, regressions of annual monthly trends of A T and C T revealed similar seasonal cycles for both parameters (Fig. 5, Table S2).The fastest increases in A T and C T occurred from May through July.The smallest (non-significant) changes occurred in January.The synchronicity between monthly trends of A T and C T was also observed at 50 m, but the rates were slower (analysis not shown).

Salinity and A T relationships
Over an annual observation period at 1 m, salinity was a poor predictor of A T , with the exception of 2007 (Fig. 6).The R 2 value for each annual salinity-A T relationship at 1 m ranged from 0.00 (in 2013) to 0.87 (in 2007) with y intercepts (A T0 , total alkalinity of the freshwater end-member) ranging between −176 µmol kg −1 (in 2007) and 2586 µmol kg −1 (in 2013).The interannual variability of the salinity-A T relationship was driven by the variability in A T observed at salinity < 38.0 that was present from November through July.
Changes in salinity (based on monthly means) at Point B were small and ranged from 37.64 ± 0.26 to 38.21 ± 0.11 from May to September, following freshwater input in winter and spring and evaporation throughout summer and fall (Fig. 4).The highest (> 38.0) and most stable salinity observations were made in August through October and coincided with the period of maximum A T (2562 and 2561 ± 9 µmol kg −1 in September and October, respectively).Minimum A T (2543 ± 14 µmol kg −1 ) was observed in June, 1 month after minimum salinity.To capture this seasonality without the interannual variation of A T , the salinity-A T relationship at 1 m was estimated from climatological monthly means (cA T and cS, N = 12) where cA T units are measured in micromoles per kilogram (µmol kg −1 ) and error terms are standard errors (R 2 = 0.74): cA T = 1554.9(±185.9)+ 26.3 (±4.9) × cS. (2) At 50 m, monthly salinity and A T were less correlated over an annual cycle (analysis not shown).Salinity remained stable at 38.0 from January through September while A T declined by 13 µmol kg −1 .In general, seasonal changes were dampened at 50 compared to 1 m.

High-frequency pH data
To verify the weekly sampling scheme at Point B, a continuous record of high-frequency pH observations was obtained via SeaFET deployments from June 2014 to April 2016 (Fig. 7).Sensor data corroborated the seasonal pH and temperature cycle observed at Point B (Fig. 7a-b).Eventscale effects (e.g., pH T change ≥ 0.1 for days to weeks, sensu Kapsenberg and Hofmann, 2016) were absent at this site, suggesting that weekly sampling was sufficient to describe seasonal and interannual changes in carbonate chemistry at Point B. The magnitude of diel pH T variability was small: the 2.5th to 97.5th percentiles ranged between 0.01 and 0.05 units pH T (Fig. 7d-e).Variability increased from winter to spring with the greatest variations in April, May, and June ( pH T > 0.035).The magnitude of diel pH variability was not correlated with temperature or the concentration of chlorophyll a (data not shown).

Discussion
High-resolution time series are necessary to document coastal ocean acidification.Based on weekly measurements from 2007 through 2015, near-surface pH T decreased by 0.0028 ± 0.0003 units yr −1 at Point B in the NW coastal Mediterranean Sea.Temperature increased by 0.072 ± 0.022 • C yr −1 .In addition, A T increased by 2.08 ± 0.19 µmol kg −1 yr −1 , a change that is unrelated to direct effects of CO 2 absorption by seawater.In less than a decade, the total change in pH at Point B ( pH T : 0.0252, Table 1) was of the same magnitude as the diel pH variability ( pH T : 0.01-0.05)and one-fifth of the magnitude of the seasonal pH cycle ( pH T : 0.12) found at this site.We can identify the drivers of ocean acidification at Point B using the deconvolution of pH and pCO 2 trends and by assuming that the increase in A T was due to increases in its carbon constituents, bicarbonate (HCO − 3 ) and carbonate (CO 2− 3 ) ions.This assumption is based on the fact that (1) HCO − 3 and CO 2− 3 ions together make up 96 % of A T at Point B, (2) increases in both HCO − 3 and CO 2− 3 would contribute to A T and C T and thereby explain the observed synchronicity of monthly trends in A T and C T (Fig. 5).Us-ing the pCO 2 deconvolution, we can then sum the contributions to pCO 2 from A T (−3.08 µatm pCO 2 yr −1 ) and C T (5.14 µatm pCO 2 yr −1 ) to determine the proportional contribution of C T to pCO 2 that is unrelated to changes in C T brought on by A T .This remaining 2.06 µatm pCO 2 yr −1 increase due to increasing C T closely matched the magnitude of atmospheric CO 2 increase during the observation period (2.08 ± 0.01 ppm yr −1 at Plateau Rosa, Italy).The closeness of these numbers does not imply causation.However, given that surface waters at Point B exhibit a near-zero air-sea CO 2 flux over annual cycles (De Carlo et al., 2013), the evidence supports the conclusion that the ocean acidification trend at Point B closely follows the rate of atmospheric CO 2 increase.The influence of atmospheric CO 2 can also be observed in the monthly changes in C T .Monthly C T trends are positive and statistically significant over more months than A T trends (8 vs. 3 months), which are more seasonally restricted (Fig. 5).
Applying this simple model to pH T , assuming changes in C T are due to both increasing A T (a neutralizing effect) and atmospheric CO 2 forcing, the surface ocean acidification trend at Point B can be attributed primarily to atmospheric CO 2 forcing (59 %) and secondarily to warming (41 %).  2.
Using this same approach for observations at 50 m, the C T contribution to pCO 2 trends that was unrelated to A T change was 1.41 µatm pCO 2 yr −1 , which is 68 % of the rate of atmospheric CO 2 increase, in contrast to 99 % at 1 m.Changes in pH T can then be attributed to atmospheric CO 2 forcing (42 %), warming (54 %), and increasing salinity (3 %).Due to slightly enhanced warming and reduced CO 2 penetration observed at 50 m, warming had a greater effect on the ocean acidification trend than atmospheric CO 2 forcing at depth compared to the surface.
The acidification rate at Point B (−0.0028 units pH T yr −1 ) is larger than those reported for other ocean time series sites (−0.0026 to −0.0013 units pH yr −1 , Bates et al., 2014), likely due to differences in warming rates.The observed warming from 2007 through 2015 (0.72 to 0.88 ± 0.2 • C decade −1 ) is extremely rapid relative to global trends in the upper 75 m from 1971 to 2010 (0.11 • C decade −1 , Rhein et al., 2013).The coastal region of Point B has warmed steadily since 1980 and with periods of rapid warming (Parravicini et al., 2015).Exacerbated warming may be related to the positive phases of the Atlantic Multidecadal Oscillation (AMO) and North Atlantic Oscillation (NAO), both of which are associated with episodic warming of the Mediterranean Sea (Lejeusne et al., 2010).The AMO has been positive since the 1990s 1 and positive NAO phases were prevalent during the second half of our study2 .
Nearest to Point B, the acidification rate at DYFAMED, an open-sea site about 50 km offshore from Point B (Fig. 1), was estimated at −0.003 ± 0.001 units pH SW yr −1 from 1995 to 2011 (Marcellin Yao et al., 2016).At DYFAMED, warming contributed approximately 30 % to the acidification rate and the remaining 70 % was attributed to anthropogenic CO 2 (Marcellin Yao et al., 2016).While the uncertainty for DY-L.Kapsenberg et al.: Coastal acidification in the Mediterranean Sea FAMED pH data is large, the trends are comparable to those observed at Point B. However, A T at the DYFAMED did not change significantly from 2007 through 2014 (F 1,51 = 3.204, P = 0.0794, R 2 0.08; data from Coppola et al., 2016).This probably indicates that the processes driving changes in A T at Point B are unique to the coastal environment.Similar changes in coastal carbonate chemistry were observed elsewhere in the near-shore Mediterranean Sea.In the northern Adriatic Sea, Luchetta et al. (2010) determined an acidification rate of −0.0025 units pH T yr −1 and an increase in A T of 2.98 µmol kg −1 yr −1 at depths shallower than 75 m, by comparing cruise data between the winters of 1983 and 2008.Point B and Adriatic Sea observations are independent but reflect changes in seawater chemistry that may be occurring in more coastal regions of the Mediterranean.Additional time series would help resolve the spatial extent of these observations.
While the trends in atmospheric CO 2 forcing and temperature account for the ocean acidification trend at Point B, the increase in A T and C T beyond what can be attributed to changes in atmospheric CO 2 was unexpected.The fastest increases in A T and C T occurred from May through July (Fig. 5), when the thermal stratification settled.In the NW Mediterranean, the main processes governing seasonal variability in A T are evaporation, which increases A T , in summer (i.e., June through September at Point B) and, to a lesser extent, phytoplankton uptake of nitrate (NO − 3 ) and phosphate (PO 3− 4 ), which increase A T from January through March (Cossarini et al., 2015).During the transition of these processes, salinity decreases to a minimum in May, reflecting freshwater input that dilutes A T to minimum values at the start of summer.For C T , peak values occur in winter when the water column is fully mixed.For reference, at DY-FAMED, mixing occurs down to more than 2000 m depth and C T is up to 100 µmol kg −1 higher in deep waters (Copin-Montégut and Bégovic, 2002).Notably, monthly trends of C T at Point B were not statistically significant from November through January for the period 2007-2015.Following winter, C T declines due to a combination of phytoplankton bloom carbon uptake and freshwater dilution (assuming river C T < seawater C T ), until the onset of summer stratification.Summer warming leads to pCO 2 outgassing (De Carlo et al., 2013), thereby further decreasing C T .In addition, the increases in A T and C T from 2007 through 2015 were more pronounced at 1 compared to 50 m.Thus, the processes driving A T and C T trends are stronger at the surface and affect carbonate chemistry primarily during the springsummer transition from May through July.
Identifying the processes causing an increase in A T and C T beyond what can be attributed to changes in atmospheric CO 2 requires some speculation.Some biogeochemical processes can be ruled out as drivers.For example, changes in benthic processes are very unlikely.Reduced calcium carbonate (CaCO 3 ) precipitation rates would increase A T but would increase C T .Even so, the dominant ecosystem in the Bay of Villefranche-sur-Mer is seagrass meadows, which harbor relatively few calcifying organisms.Dissolution of calcium carbonate sediment would contribute to A T and C T increase in the water column.However, as the water was supersaturated with respect to both aragonite and calcite, this could only be mediated by biological processes.Carbonate dissolution following CO 2 production via oxic degradation of organic matter releasesA T and C T in a 1 : 1 ratio (Moulin et al., 1985).Likewise, anaerobic remineralization produces alkalinity (Cai et al., 2011).In the sediment of the Bay of Villefranche, sulfate reduction coupled with precipitation of sulfide minerals is the dominant anoxic mineralization pathway (Gaillard et al., 1989).An increase in these processes would explain the observed increase in A T and C T , but as trends were slower at 50 compared to 1 m, this would suggest the dominance of a process taking place in surface waters.
In the upper water column at Point B, changes in biological processes are unlikely to explain the observed trends in A T and C T .For example, the concentration of chlorophyll a (a proxy of primary production) has decreased since 1995, nutrients have increased, and phytoplankton blooms have shifted towards earlier dates in the year (Irisson et al., 2012).While a decrease in net primary production could drive C T trends, the observed increase in NO − 3 and PO 3− 4 would cause a small decrease in A T (Wolf-Gladrow et al., 2007).Stimulated community respiration could result from warming waters but enhanced remineralization would cause a decrease in A T (Wolf-Gladrow et al., 2007).
The lack of salinity change at the surface excludes additional processes as drivers of the A T and C T increase at Point B. For example, increased summertime evaporation (concentration effect) and reduced freshwater input (decreased dilution effect) would both be expected to cause an increase in salinity, which was not observed.Increased input of eastern Mediterranean Sea waters could increase A T , but this is unlikely as this water mass flows much deeper than Point B (Millot and Taupier-Letage, 2005).
Instead, the observed changes in A T and C T could be due to increased limestone weathering, which would increase the input of A T from land to the sea via rivers and groundwater.Rivers contribute both A T and C T to the Mediterranean Sea (Copin-Montégut, 1993;Tamše et al., 2015).River A T originates from erosion and is correlated with bedrock composition (e.g., McGrath et al., 2016).Positive trends in river A T have been documented in North America and occur via a number of processes including (1) the interplay of rainfall and land use (Raymond and Cole, 2003), ( 2) anthropogenic limestone addition used to enhance agricultural soil pH (Oh and Raymond, 2006;Stets et al., 2014) and freshwater pH (Clair and Hindar, 2005), and (3) potentially indirect effects of anthropogenic CO 2 on groundwater CO 2 acidification and weathering (Macpherson et al., 2008).These and other processes were hypothesized to have driven A T changes in the Baltic Sea (Müller et al., 2016).There, an increase in A T of 3.4 µmol kg −1 yr −1 was observed from 1995 to 2014 (mean salinity = 7).In contrast to Point B, the increase in Baltic Sea A T was not noticeable at salinity > 30 (Müller et al., 2016).
Given the above speculations, the simplest plausible mechanisms causing the unexpected A T and C T trends would be through increasing A T of the freshwater end-member of Point B. Local precipitation, however, did not have an influential effect and was not correlated with salinity or A T (Fig. S1 in the Supplement).While submarine groundwater springs can be a significant source of nutrients, A T , and C T for the ocean (Cai et al., 2003;Slomp and Van Cappellen, 2004), carbonate chemistry contributions of local submarine springs are currently unknown (Gilli, 1995).Signatures of limestone erosion can be observed in A T of nearby rivers (Var, Paillon, and Roya) but detailed time series are not available.Likewise, riverine influence at Point B has not been quantified.If river runoff exerts a dominant control on Point B carbonate chemistry, there is a lag effect, as freshwater influence peaked in May but A T and C T increased fastest from May through July.Consequently, this hypothesis needs further investigation.Until the source of A T increase is properly identified, use of this observation in modeling should be implemented with caution.

Conclusion
Predictions of coastal ocean acidification remain challenging due the complexity of biogeochemical processes occurring at the ocean-land boundary.At the Point B coastal oceanographic station in the NW Mediterranean Sea, surface ocean acidification was due to atmospheric CO 2 forcing and rapid warming over the observation period 2007-2015.However, additional trends in A T and C T were observed and remain unexplained, but these trends could relate to riverine and groundwater input.The influence of coastal boundary processes influencing seawater A T and C T presents a potentially major difference between coastal and offshore changes in ocean chemistry.This study highlights the importance of considering other anthropogenic influences in the greater land-sea region that may contribute to coastal biogeochemical cycles (sensu Duarte et al., 2013) and alter projections of anthropogenic change in near-shore waters.

Figure 1 .
Figure 1.Map of study region in the NW Mediterranean Sea (a), along the north current (b) in the Bay of Villefranche-sur-Mer, France (c).Point B, the EOL buoy, and the offshore time series station DYFAMED are marked.Bathymetric line units are measured in meters (m) (c).
slope contribution of changing "var" to the estimated change in pH T dpH T dt

Figure 2 .
Figure2.Time series observations (a-g) and anomaly trends (h-n) for temperature, salinity, and seawater carbonate chemistry at Point B, 1 m.Regression slopes are drawn ±SE (standard error, in grey) and noted with a star for significance at α = 0.05.Variable abbreviations are the same as in Table2.
in C T .Increasing A T countered 66-69 and 60 % of the influence of increasing C

Figure 3 .
Figure3.Time series observations (a-g) and anomaly trends (h-n) for temperature, salinity, and seawater carbonate chemistry at Point B, 50 m.Regression slopes are drawn ±SE (standard error, in grey) and noted with a star for significance at α = 0.05.Variable abbreviations are the same as in Table2.

Figure 4 .
Figure 4. Monthly distribution of seawater carbonate chemistry at Point B, 1 m, using a combination of a violin plot showing the relative frequency of the observations (shaded blue area) and a box plot showing the median, first, and third quartiles, as well as outliers (blue).

Figure 6 .
Figure 6.Salinity and total alkalinity (A T ) relationships at Point B for the period 2007-2015, by year, at 1 m.Data points are colored by month.

Figure 7 .
Figure 7. Time series pH T (a) and temperature (b) from SeaFET pH sensor deployments at EOL, 2 m.Discrete calibration samples are noted in green, and grey vertical lines bracket deployment periods (a).Calibration sample offsets from processed pH time series are shown in (c).Violin and box plots (see Fig. 4) show diel pH range by month (d), and an example of this pH variability is shown for May 2015 (e).

Table 2 .
Time series anomaly regression analyses on seawater carbonate chemistry at Point B for salinity (S), temperature (T ), dissolved inorganic carbon (C T ), total alkalinity (A T ), pH T , pH T normalized to 25 • C (pH T25 ), pCO 2 , calcite ( c ) and aragonite ( a ) saturation state, and salinity-normalized A T (nA T ) and C T (nC T ), at 1 and 50 m.Slopes represent the change in the variable unit per year.P 0.001 indicates p values far smaller than 0.001.SE stands for standard error.Degrees of freedom are denoted by df.

Table 3 .
Deconvolution of pH T anomaliesdpH T dt , units pH T yr −1 at 1 and 50 m.Sensitivity of pH T with respect to variables ∂pH T ∂var , where the variable var is either temperature (T ), salinity (S), total alkalinity (A T ), or dissolved inorganic carbon (C T ), was multiplied by the anomaly of var dvar dt , Table2.SE is standard error and RMSE is root mean square error.Rounding was performed at the end of the calculations, prior to the estimation of the percentage of contributions.

Table 4 .
Deconvolution of pCO 2 anomalies dpCO 2 dt , µatm yr −1 at 1 and 50 m.Details are the same as in Table3.