Prediction of Sea of Japan (East Sea) acidification over the past 40 years using a multiparameter regression model



[1] A multiparameter linear regression model (MLR) of aragonite saturation state (ΩARG) as a function of temperature, pressure and O2 concentration in the upper 1,000 m of the Sea of Japan (East Sea) was derived with an uncertainty of ±0.020 (1σ). The ΩARG data (n = 1,482) used to derive the basin-wide ΩARG prediction model were collected during a field survey in 1999 and were corrected for anthropogenic CO2. Some biases were resolved by addition of a pressure and O2 concentration interaction term to the proposed model. Correlation between the two predictor terms, caused by addition of this term, was minimized by centering the data for the three variables (thus subtracting the mean from each individual data point). Validation of the model against data sets obtained in 1992 and 2007 yielded correlation coefficients of 0.995 ± 0.013 for 1992 (n = 64, p ≪ 0.001) and 0.995 ± 0.009 for 2007 (n = 137, p ≪ 0.001) and root mean square errors of ±0.064 for 1992 and ±0.050 for 2007. The strong correlation between measurements and predictions suggests that the model can be used to estimate the distribution of ΩARG in the Sea of Japan (East Sea) (including dynamic coastal waters) on varying time scales when basic hydrographic data on temperature, pressure and O2 concentration are available. Application of the model to past measurements for the Sea of Japan (East Sea) indicated that interdecadal variability (2σ from the mean) in ΩARG corrected for anthropogenic CO2 was generally high (0.1–0.7) in the upper water layer (<200 m depth), and decreased (0.05–0.2) with depth for waters deeper than 500 m. The interdecadal variability is largely controlled by variations in the degree of water column ventilation. Superimposed on this natural variability, the input of CO2 derived from fossil fuels has markedly acidified the upper water layers during the anthropocene and thereby moved the aragonite saturation horizon upward by 50–250 m. The impact of CO2 derived from fossil fuels on upper ocean acidification will increase in the future. The present study indicates that, in combination with other easily measurable parameters, a multifunctional model can be a powerful tool for predicting the temporal evolution of ΩARG in the ocean, including coastal waters that are highly likely to be susceptible to ocean acidification in the future.

1. Introduction

[2] The rise in atmospheric CO2 concentration caused by fossil fuel burning, deforestation and cement production is one of the major environmental concerns of the industrial era [Intergovernmental Panel on Climate Change, 2007]. The oceans appear to have mitigated the adverse effects of increased atmospheric CO2 levels by absorbing about one-third of total anthropogenic CO2 emissions [Sabine et al., 2004]. When anthropogenic CO2 dissolves in seawater the protons that are formed from the reaction between CO2 and H2O react with carbonate ions (CO32−) to produce bicarbonate ions (HCO3). The resulting reduction in the CO32− concentration ([CO32−]) causes a decrease in the seawater saturation state with respect to CaCO3 minerals including aragonite (ΩARG), which can be calculated from equation (1):

equation image

where [Ca2+] is the concentration of calcium ions and K*SP-ARG is the stoichiometric solubility product of aragonite. The accumulation of fossil fuel CO2 in the ocean during the anthropocene has decreased the pH of ocean surface waters by 0.1 pH units [Prentice et al., 2001], and a further decrease of as much as 0.3–0.4 pH units is projected by the end of this century [Caldeira and Wickett, 2003]. This pH reduction will cause a 50–60% reduction in [CO32−] [Brewer, 1997] and the seawater ΩARG [Feely et al., 2004].

[3] A reduction in ocean pH of this magnitude will probably have a significant impact on organisms that form CaCO3 skeletons, including coralline algae (high-Mg calcite), pteropods and corals (aragonite), bivalves (calcite, aragonite or mixed), forams (calcite or aragonite) and coccolithophorids (calcite) [e.g., Feely et al., 2004; Gehlen et al., 2007; Doney et al., 2009]. Reduced rates of biogenic calcification caused by a reduction in the seawater saturation state with respect to aragonite and calcite are increasingly being reported in both laboratory and mesocosm [e.g., Riebesell et al., 2000; Kurihara and Shirayama, 2004; Delille et al., 2005], field [e.g., Manzello et al., 2008; Green et al., 2009], and paleo-oceanographic [e.g., Moy et al., 2009] studies. Therefore, it is important to establish the present saturation state of seawater with respect to CaCO3 particles, to assess how this has evolved over time, and to identify the controlling factors. Although relevant data are limited in quantity, there is evidence that in some species biological calcification is primarily controlled by [CO32−] rather than by the saturation state [Trimborn et al., 2007].

[4] In this regard the exceptionally high quality data set derived from the World Ocean Circulation Experiment/Joint Global Ocean Flux Study conducted in the 1990s enabled estimation of the impact of anthropogenic CO2 on the saturation state of seawater with respect to CaCO3 across three ocean basins [Feely et al., 2002, 2004; Sabine et al., 2002; Chung et al., 2003, 2004]. According to this global scale analysis, the penetration of anthropogenic CO2 during the anthropocene has moved the 100% aragonite saturation depth upward by approximately 30–100 m in the Pacific Ocean, 100–150 m in the Atlantic Ocean, and 100–200 m in the Indian Ocean. Moreover, recent modeling studies anticipate that calcifying organisms in the Southern Ocean and in parts of the North Pacific Ocean are likely to be subject to corrosive conditions with respect to aragonite by the end of the 21st century [Orr et al., 2005; McNeil and Matear, 2008; Steinacher et al., 2009]. In addition, some coastal areas that occasionally receive upwelled water from the upper- to mid-thermocline have already been affected by corrosive water, and, in these areas the addition of anthropogenic CO2 has further increased the influence of such water [Feely et al., 2008a].

[5] In the work reported here we used objective statistical methods to derive an optimal multiparameter model of ΩARG after removing the effect of anthropogenic CO2. We applied the resulting model to archived data on salinity, temperature and O2 concentration for the period 1960–2000. This enabled reconstruction of the evolution of seawater ΩARG for the Sea of Japan (East Sea) during this period, based on methods similar to those described elsewhere [Feely et al., 2008b; Juranek et al., 2009]. We also evaluated the effect on ΩARG of anthropogenic CO2 accumulated in the Sea of Japan (East Sea) over the period 1800–1999. The Sea of Japan (East Sea) was chosen because it has two unique characteristics. First, as a consequence of changes in water column ventilation the Sea of Japan (East Sea) has undergone considerable variations in O2 concentration throughout the water column [Gamo et al., 2001; Kim et al., 2001, 2002; Kang et al., 2004] and further changes are expected in future years. Second, the Sea of Japan (East Sea) has the highest specific column inventory (80 mol C m−2) of anthropogenic CO2 relative to other marginal seas and the three major ocean basins [Park et al., 2006], and the rate of anthropogenic CO2 uptake into this basin has changed considerably over relatively short periods [Park et al., 2008]. These are both key factors affecting ΩARG in the Sea of Japan (East Sea). As a result, this basin is uniquely suitable for testing the accuracy of our predictive model.

2. Data and Methods

2.1. Data Sources

[6] The CO2 measurement and calibration methods used during surveys in 1992, 1999 and 2007 have been documented elsewhere [Chen et al., 1995; Talley et al., 2004; Park et al., 2008], and the adjustment factors applied to these data sets for consistency were reported by Park et al. [2006, 2008]. As the data set obtained during the 1999 survey provides the densest coverage of the Sea of Japan (East Sea), it was used to derive a multiparameter linear regression model that expresses the relationship between seawater ΩARG values and other predictor variables. The predictive ability of the resulting model was checked against the data sets obtained in different years (1992 and 2007), and the model was then applied to historical measurements (1960–2000; archived at the Japan Oceanographic Data Center, JODC; from the Sea of Japan (East Sea) to predict interdecadal variations in ΩARG.

[7] To avoid seasonal variations in physical and chemical properties of the Sea of Japan (East Sea), we excluded from our analysis all data for waters shallower than 50 m. This cutoff point was based on observations that the depth of the water column directly influenced by local meteorological conditions does not exceed 50–100 m over much of the Sea of Japan (East Sea), with the exception of the northwestern part of the basin [Tishchenko et al., 2003]. As the surface area of the Sea of Japan (East Sea) with bottom depths shallower than 50 m accounts for less than 3% of the total surface area, our predictive model covers nearly all of the Sea of Japan (East Sea). We also excluded data for waters deeper than 1000 m, as interdecadal variability is largely confined to shallower depths [Min and Warner, 2005].

2.2. Data Filtering

[8] As the data archived at the JODC were collected by a variety of investigators in different years it was critical to remove biased data that could lead to erroneous interpretations of the distribution of ΩARG. To ensure data quality for each predictor variable we adopted a traditional mean and standard deviation method to remove erroneous data, whereby values greater than 3 standard deviations from the mean were rejected [Garcia et al., 2006]. In this evaluation process all data for each variable were first interpolated onto a three dimensional grid of 1° latitude × 1° longitude × a 25–100 m depth range, and the mean and standard deviation were then calculated for each grid. However, this data evaluation method did not apply to grids with less than five data points; in these cases all data in the grid were used. As only 5% of the total data fell in this category its inclusion did not significantly bias our results. The application of the data filtration method to the entire data set indicated that only 0.4% of the total data set involved had values greater than 3 standard deviations from the mean; these outliers were not used in predicting seawater ΩARG.

2.3. Calculation Methods

2.3.1. Aragonite Saturation State of Seawater

[9] Seawater ΩARG values were calculated using equation (1), in which [Ca2+] was estimated from salinity using the established [Ca2+] to salinity ratio [Riley and Tongudai, 1967], and [CO32−] was calculated from total alkalinity (AT) and total inorganic carbon (CT) using pressure-corrected thermodynamic constants that are consistent with the calibrated global field data [McElligott et al., 1998; Wanninkhof et al., 1999; Lee et al., 2000; Millero et al., 2002]. These constants include the carbonic acid dissociation constants of Mehrbach et al. [1973], as refitted into a different functional form by Dickson and Millero [1987], and other ancillary constants as suggested by Millero [1995]. The effect of pressure on these dissociation constants was estimated from partial molal volume and compressibility data [Millero, 1995]. The values of K*SP-ARG were taken from Mucci [1983] for a given temperature and salinity at 1 atmosphere. The effect of pressure on K*SP-ARG was also estimated from partial molal volume and compressibility data [Millero, 1995]. The preindustrial 100% saturation horizons with respect to aragonite (ΩARG = 1) were calculated using measured AT and preindustrial CT. The latter was estimated by subtracting anthropogenic CO2 values (estimated by Park et al. [2006] using the ΔC* approach developed by Gruber et al. [1996]) from the measured CT.

2.3.2. Multiparameter Linear Regression Model

[10] The multiparameter linear regression (MLR) method has often been used to fill data gaps in data-sparse basins via interpolation and extrapolation, or to detect anthropogenic CO2 in the ocean by determining the rates of change of CT and 13C/12C over time [e.g., Brewer et al., 1995; Wallace, 1995; Sonnerup et al., 2000; Ono et al., 2000; Peng et al., 2003; Tanhua et al., 2007; Park et al., 2008]. The MLR model used here takes the form equation (2):

equation image

where Y is a dependent variable, Xp is the pth independent variable, and mp is the mean of the Xp values. I is an intercept and α, β and γ are partial regression coefficients of additive, polynomial and interaction terms, respectively. The final term, ɛ, is an unexplained random error. Not all terms shown in equation (2) were used in our model. Three key tests (see below) enabled selection of the optimum form of the model with the minimum number of predictor variables. The final MLR model chosen yielded high predictability while maintaining simplicity of form.

[11] The first test evaluated whether there was correlation (also referred to as colinearity and intercorrelation) between all possible pairs of predictor variables. The purpose of this test was to avoid the use of correlated variables in our functional model, because they make the estimation of regression coefficients unreliable. For example, the addition of interaction (e.g., X1X2) or polynomial (e.g., X12, X22) terms to a model of additive terms (e.g., X1, X2) is likely to cause correlation between the additive terms, and their interaction and the polynomial terms. To minimize these adverse effects of correlation, all data for each predictor variable were centered by subtracting their mean from each individual data point [Quinn and Keough, 2002]. To assess the possibility of correlation between some pairs of predictor variables in our model we used the variance inflation factor (VIF) [Snee, 1977], which is defined as 1/(1-Rj2), where Rj2 is the coefficient of determination of regression between each predictor variable (e.g., Xj) and the remaining variables (X1, X2,…, Xj−1, Xj+1,…) [Quinn and Keough, 2002]. This test assumes that there is no colinearity in our model if the VIF value is less than 5. For example, without centering the data our final MLR model yielded a value of VIF much greater than 5 (model V in Table 1, see more details in section 3.2).

Table 1. Coefficients for Various MLR Models and the Errors Associated With Eacha
ModelPredictor VariablesCoefficientsVIFbModel Errors
  • a

    Model VI was chosen as the optimal model in this study.

  • b

    Variance Influence Factor.

  • c

    Ordinary Least Square Fit.

  • d.

    Robust Least Square Fit.

  • e

    Cross Validation.

 S−1.50 × 10−12.3“”“”“”
 T1.61 × 10−196“”“”“”
 P−2.57 × 10−46.0“”“”“”
 O25.91 × 10−319“”“”“”
 PO43−−4.12 × 10−284“”“”“”
 Si3.89 × 10−326“”“”“”
 T8.00 × 10−23.0“”“”“”
 P−1.62 × 10−42.7“”“”“”
 PO43−−6.42 × 10−16.1“”“”“”
 S−1.78 × 10−11.2“”“”“”
 T1.57 × 10−14.1“”“”“”
 P−1.64 × 10−42.6“”“”“”
 O25.33 × 10−32.2“”“”“”
IVConstant−3.22 × 10−1-±0.058±0.028+0.0017
 T1.55 × 10−12.0“”“”“”
 P−1.64 × 10−42.7“”“”“”
 O25.40 × 10−31.9“”“”“”
VConstant−5.13 × 10−1-±0.048±0.020+0.003
 T1.54 × 10−12.0“”“”“”
 P5.03 × 10−4100“”“”“”
 O26.20 × 10−33.5“”“”“”
 P × O2−2.88 × 10−692“”“”“”
 (T − TMEAN)1.54 × 10−12.0“”“”“”
 (P − PMEAN)−2.24 × 10−43.2“”“”“”
 (O2 − O2MEAN)5.29 × 10−31.8“”“”“”
 (P − PMEAN) × (O2 − O2MEAN)−2.88 × 10−62.1“”“”“”
 (T − TMEAN)1.55 × 10−12.1“”“”“”
 (P − PMEAN)−2.17 × 10−43.1“”“”“”
 (O2 − O2MEAN)5.13 × 10−31.8“”“”“”
 (P − PMEAN) × (O2 − O2MEAN)−2.02 × 10−62.3“”“”“”
 (O2 − O2MEAN)29.15 × 10−61.5“”“”“”

[12] The second test enabled selection of the optimum predictor variables in the final MLR model. In this test we compared values of the root mean square error (RMSE) for all possible function forms that included only some or all of the additive terms of predictor variables. The functional model (including additive terms only) that yielded a minimum RMSE value was considered optimal. We then used the 10-fold cross validation method [Picard and Cook, 1984] to evaluate whether the addition of univariate polynomial and/or product terms to the model yielded better predictions. In this method the total data set used to derive the model was divided into 10 subsets. The first step of this analysis was to remove 1 subset from the 10 data sets. A regression model was then derived without this subset, the output values for this subset were predicted using the derived regression model, and the residuals were computed. This calculation routine was repeated for each subset, and the resulting residuals were averaged. Smaller mean residuals indicated better prediction.

[13] The third test was used to assess whether a functional model that passed the first two tests contained unnecessary terms. In deriving a multiparameter linear regression model from a given data set that is considerably larger than the number of variables (in our study there were 6 predictor variables and 1,482 data points for each variable), the presence of unnecessary polynomial or interaction terms generally lowers prediction errors, and can occasionally reduce them to below the analytical error. To avoid the inclusion of unnecessary terms in the final functional form we choose the simplest functional model from among all those that gave prediction errors smaller than the analytical errors. The error in ΩARG in this study was approximately ±0.03, which was primarily due to the measurement uncertainties of ±2 μmol kg−1 for CT and ±2 μmol kg−1 for AT [Talley et al., 2004; Park et al., 2006].

[14] Partial regression coefficients were derived by minimizing the sum of squared deviations between measured values and those predicted from the model (referring to an ordinary least squares model). In the event of outliers the ‘robust fit’ method is better because at each iteration step a certain weight value is assigned to each data point, depending on its quality. In the first iteration step the robust fit was comparable to the ordinary least squares method. In subsequent iterations a new weight value was calculated for each data point and assigned to it. In this procedure lower weight values were given to data points that were distant from those predicted by the model derived from the preceding step. This statistical procedure continued until a new iteration yielded no significant change in weight values for the data.

[15] Prior to deriving a functional model of ΩARG for the Sea of Japan (East Sea) using the 1999 data set, we subtracted the total anthropogenic CO2 that had accumulated during the period from 1800–1999 from the 1999 CT values. The anthropogenic CO2 concentration as of 1999 in the Sea of Japan (East Sea) was estimated by Park et al. [2006] using a modified ΔC* method [Gruber et al., 1996]. We then calculated the seawater ΩARG values using AT and the anthropogenic CO2-corrected CT data. Finally, the calculated ΩARG values were related to other predictor variables to derive the optimal functional model. Without this correction the functional model derived using the 1999 CT data could not adequately predict seawater ΩARG values for years other than 1999 because of the influx of anthropogenic CO2 over time.

3. Results and Discussion

[16] In this section we first present the distribution of seawater ΩARG values for the Sea of Japan (East Sea) derived from the 1999 data set, and describe the procedure by which we arrived at the final functional model for predicting seawater ΩARG values corrected for the effect of anthropogenic CO2 We next present the predicted interdecadal variability of seawater ΩARG values using the chosen model in conjunction with measured temperature, pressure and O2 concentration data for the period 1960–2000. Finally, we evaluate the effect of anthropogenic CO2 (accumulated between 1800 and 1999) on the upward shift of the 100% aragonite saturation depth (ΩARG = 1) in the Sea of Japan (East Sea).

3.1. Distribution of Saturation States of Seawater With Respect to Aragonite in the Sea of Japan (East Sea)

[17] Values of ΩARG for near-surface waters (shallower than 100 m depth) of the Sea of Japan (East Sea) increased dramatically from north to south (Figure 1a). In the northern basin (nominally north of 40°N) ΩARG values ranged from 1.2 to 1.6 at 100 m depth, which is considerably lower than the values (1.6–2.4) found at the same depth in the southern basin. At 200 m depth the large meridional difference across the basin disappeared, and, at greater depths the opposite pattern was observed, with lower values to the south (Figures 1c1e). The ΩARG values for waters below 300 m depth were close to or below saturation with respect to aragonite (Figures 1c1e and 2a).

Figure 1.

Distributions of the aragonite saturation state of seawater at the (a) 100, (b) 200, (c) 300, (d) 400, and (e) 500 m depth levels. Aragonite seawater saturation values were calculated from measurements of CT and AT from the 1999 survey. Dots indicate sampling locations. (f) The three major basins are indicated, and the surface current directions are shown as red (warm currents) and blue (cold current) arrows.

Figure 2.

Distribution of (a) the aragonite saturation states of seawater and (b) salinity for the area 130.8°E–138.4°E and 37.7°N–42.6°N. The contour lines indicate the aragonite seawater saturation values and salinity. High salinity in the upper 100 m is a key characteristic of Tsushima Warm Current. Black dots indicate the sampling locations for the 1999 cruise.

[18] The meridional difference at the 100 m level is attributable to the interactions between two contrasting water masses; warm and salty water masses in the south, and colder and less salty water masses in the north [Talley et al., 2006]. The water mass prevailing in the south comprises the Tsushima Warm Current, which has a higher AT/CT ratios (∼1.13) than its counterpart (∼1.08) in the north (Figure 2b), and thus has higher ΩARG values (Figure 2a). The ΩARG values changed markedly across the front between these two water masses, which generally forms along the 40°N line, although its position can vary considerably [Talley et al., 2006] (Figures 1a and 2). Contrary to the large meridional difference at 100 m depth, the ΩARG values for waters deeper than 100 m gradually decreased from north to south, and the meridional difference in ΩARG became smaller with increasing depth. This meridional pattern was more apparent at two isopycnal surfaces: the East Sea Intermediate Water (σθ ∼ 27.1) and the Upper Japan Sea Proper Water (σθ ∼ 27.31) [Talley et al., 2006]. At the core of both water masses ΩARG gradually decreased in a southward direction, which is consistent with the southward decrease in the O2 saturation level reported by Talley et al. [2006] (Figures 3a, 3b, 3d, and 3e). The strong correlation between the ΩARG values and the O2 concentration occurs because oxidation of organic matter by bacteria continues to decrease both O2 and CO32− concentrations as the water moves along the isopycnal surface from north to south (Figures 3c and 3f). This supports our use of O2 concentration as a key proxy for predicting seawater ΩARG values (see more details in section 3.2). These two water masses dominate the depth ranges 100–250 m and 200–500 m, respectively.

Figure 3.

Aragonite seawater saturation values and O2 concentration for the two isopycnal surfaces of (a, b) 27.1 and (d, e) 27.31 kg m−3, which are the core densities of East Sea Intermediate Water and Upper Japan Sea Proper Water, respectively [Talley et al., 2006]. Plots of aragonite seawater saturation values as a function of O2 concentrations for the two isopycnal surfaces of (c) 27.1 and (f) 27.31 kg m−3.

[19] As fishing and aquaculture activities are largely concentrated in coastal regions, we examined the distribution of ΩARG values in near-bottom coastal waters off Korea, Russia and Japan. In this analysis we used the 1999 data collected from these coastal regions (Figure 4). The near-bottom waters shallower than 200 m depth in all three coastal areas are mostly supersaturated with respect to aragonite. Although not illustrated in Figure 4, seasonal coastal upwelling brings corrosive waters to the surface, and thereby makes coastal waters undersaturated or close to saturation with respect to aragonite. For example, the Tsushima Warm Current and the complex shelf bathymetry induce corrosive water (located at approximately 300 m depth; ΩARG = 0.75–1) to seasonally rise to the Japanese shelf slope [Nakada and Hirose, 2009], whereas seasonal winds push coastal waters off the coasts of Korea and Russia, and thereby bring corrosive interior waters to the surface [Lee and Kim, 2003; Vilyanskaya and Yurasov, 2008]. Along the Russian coast, strong monsoonal winds blowing from the northwest during the winter favor upwelling in areas close to Peter the Great Bay. Strong northwesterly winds weaken when they encounter Russian coastal waters and turn southwesterly, blowing parallel to the shoreline; this induces upwelling of corrosive waters (ΩARG ∼ 1) by pushing coastal waters away from the Russian continent [Vilyanskaya and Yurasov, 2008]. A similar mechanism operates along the Korean coast. Northwesterly winds prevail over the Korean peninsula for extended periods during summer, and this pushes coastal waters away from the land and induces upwelling of cold and corrosive waters (ΩARG ∼ 1) [Lee and Kim, 2003]. In particular, direct temperature observations made by the National Fisheries Research and Development Institute of Korea (NFRDI) highlight that a wind-driven upwelling event occurred along the Korean coast in 2007. This affected the surface temperature distribution as the core of the upwelled water was approximately 10°C cooler than the ambient water temperature (Figure 5). According to our analysis of water properties in the interior of the Sea of Japan (East Sea), this cold and corrosive water is likely to have been derived from the isopycnal surfaces of 27.2–27.25 kg m−3, which largely occupy the 200–300 m depth range and have ΩARG values of 1 or lower, and pH values of 7.8–7.9.

Figure 4.

Distribution of aragonite saturation values (colors) for coastal bottom waters along the (a) Korean, (b) Russian and (c, d) Japanese coasts. The color scales indicate the degree of seawater saturation, whereas contour lines and numbers represent bottom depths of 0.2, 0.5, 1, 1.5 and 2 km, respectively. The cross symbols indicate data collection locations for the 1999 cruise. The boxes in the central map represent the location of Figures 4a–4d.

Figure 5.

(a) Map of sea surface temperature (SST, °C) for August 2007 along the Korean coast (35.2°N–38°N) and (b) hourly mean wind directions and the number of observations in August 2007. The data shown here were collected at Pohang, Korea (36.10°N and 129.55°E). The height of each histogram represents the proportion of observations for a given wind direction.

[20] In summary, as most upwelled waters along the three coastal regions are derived from relatively shallow depths (∼300 m) of the Sea of Japan (East Sea), they are corrosive because of both the production of CO2 from organic matter oxidation and the accumulation of anthropogenic CO2 This source water for upwelling will continue to absorb anthropogenic CO2 in the future. Although coastal upwelling is a key mechanism affecting acidification in these three coastal areas, the continuous accumulation of anthropogenic CO2 in upwelling source water will intensify and increase coastal acidification. In section 3.4 we discuss the effect of anthropogenic CO2 on the aragonite saturation state of seawater in the Sea of Japan (East Sea).

3.2. Derivation of the Multiparameter Linear Regression Model

[21] The predictor variables that are available to be related to ΩARG values include salinity (S), pressure (P), temperature (T), and the concentrations of O2 (O2), phosphate (PO43−) and silicate (Si). With an optimal combination of the predictor variables T, P and O2, the three key tests described in section 2.3 yielded the lowest RMSE (±0.028; model IV in Table 1). Inclusion of the variable S in our functional model did not improve predictive ability (model III in Table 1), as the S variability in the Sea of Japan (East Sea) is only one-tenth of that found in the major ocean basins (i.e., the magnitude of variability was too small to influence the model). Therefore, our functional model included only the variables T, P and O2 (model IV in Table 1). Although the fitting error (±0.028) associated with our functional model was small, the model-data residuals exhibit a positive bias for O2 concentrations lower than 230 μmol kg−1 or higher than 300 μmol kg−1 (Figure 6a). These biases were successfully removed by centering the data for the three variables, and adding the interaction term between centered P and centered O2 (model VI in Table 1; see also Figure 6b). The final functional model is shown in equation (3):

equation image

where numeric values in parentheses indicate the mean values for the predictor variables. This MLR model yielded a RMSE of ±0.020 and a mean cross-validation error of 0.003. Values of VIF were less than 5 for all terms in the MLR model, indicating that there was no correlation between the predictor variables. We also assessed whether the chosen MLR model led to any spatial biases in residuals, by dividing the 1999 data set into 16 subsets and estimating the mean residuals for each subset. The residuals for all subsets were smaller than the RMSE of the entire data set, suggesting no discernable biases.

Figure 6.

Distributions of residuals (open circle), measured ΩARG values: those predicted using equation (3) (a) without or (b) with the interaction term, as a function of dissolved O2 concentration (μmol kg−1).

[22] To evaluate the ability of the chosen MLR model to predict temporal variability in ΩARG for the Sea of Japan (East Sea), values of ΩARG predicted using the model were compared to those measured in other years (1992 and 2007). Prior to comparison, the anthropogenic CO2 values were subtracted from the 1992 and 2007 CT values. Omega values were then calculated using AT and anthropogenic CO2-corrected CT data. Unlike the 1999 data set, the 1992 and 2007 data sets do not include CFC data and thus it was not possible to apply the ΔC* method [Gruber et al., 1996] to the 1992 and 2007 data sets to estimate the anthropogenic CO2 concentration. Therefore, we used an indirect approach to estimate such CO2 values for 1992 and 2007. In these calculations we utilized anthropogenic CO2 values for 1999 based on the ΔC* method [Park et al., 2006] and the accumulation rates of anthropogenic CO2 for the periods 1992–1999 and 1999–2007, based on isopycnal and extended multiparameter linear regression methods [Park et al., 2008]. To estimate anthropogenic CO2 values for 1992 we subtracted the amount of anthropogenic CO2 absorbed for the period 1992–1999 from the 1999 anthropogenic CO2 values, and the resulting anthropogenic CO2 values were then subtracted from the 1992 CT values. To estimate anthropogenic CO2 values for 2007 we added the amount of anthropogenic CO2 absorbed for the period 1999–2007 to the 1999 anthropogenic CO2 values, and the resulting anthropogenic CO2 values were then subtracted from the 2007 CT values. This method of estimating anthropogenic CO2 yielded a CT error of ∼5 μmol kg−1 for 1992 and 2007, which in turn resulted in an error of ±0.034 in ΩARG. Validation of the model against measurements (1992 and 2007 data sets) yielded Pearson product–moment correlation coefficient of 0.995 ± 0.013 for 1992 (n = 64, p ≪ 0.001) and 0.995 ± 0.009 for 2007 (n = 137, p ≪ 0.001), and RMSE of ±0.064 for 1992 and ±0.050 for 2007 (Figure 7); these RMSE values are broadly consistent with those derived from the 1999 data set. The mean residuals were acceptable within the error ranges of ±0.030 and ±0.034 associated with derivation of the MLR equation, and estimation of anthropogenic CO2 values, respectively. This comparison indicates that our MLR model can be used to predict seawater ΩARG values in the Sea of Japan (East Sea) for years other than 1999.

Figure 7.

Profiles of measured (closed circles) and predicted (open circles) ΩARG values for (a) 1992 and (b) 2007. Predicted values were derived using equation (3) and measurements of temperature, pressure and O2 concentration. The insets are residual plots. The vertical dashed lines in the insets indicate the mean residuals.

[23] We also evaluated the importance of each of the three predictor variables in determining ΩARG values. This was achieved by subtracting the mean from individual data points for each predictor variable, and then dividing the resulting data by one standard deviation. This data transformation converted values for the three variables to the same scale, thereby making possible direct comparison of the three data sets. In this analysis we first fitted the three variables for each 100 m layer of water, and then compared regression coefficients for the three variables. The greater the coefficient of a given variable, the higher its importance in determining seawater ΩARG value. Our analysis indicates that the variable temperature is particularly important for depths less than 300 m, where temperature variation is considerable, whereas the variable pressure becomes more important at depths greater than 300 m. The O2 concentration is generally significant throughout the water column.

[24] An important assumption in our approach was that the empirical relationship derived from the contemporary spatial variations in seawater ΩARG and the three predictor variables can be used to infer temporal variations in ΩARG. However, the relationship of ΩARG to O2 concentration can be modified by sudden enhancement of convective mixing [Levine et al., 2008]. For example, in all locations, the increase in O2 concentration for the transition period between the 1960s and 1970s indicates that water column ventilation intensified in this interval (Figure 8), and as a result, the application of the derived model to this transition period may be problematic. In contrast, in all locations but one of the northern locations (indicated by a blue square in the inset of Figure 8h) the decreasing trend of O2 concentration during the period after the 1970s indicates a weakening of water column ventilation (Figure 8), so the derived relationship is likely to hold true for this period because the relationship between ΩARG and O2 concentration was largely determined by the oxidation of organic matter. The good correlation between predictions using the model and the 1992 and 2007 measurements supports the latter case. However, full validation of our derived model for the period prior to 1992 was not possible because carbon parameter data were not available.

Figure 8.

Plots of (a–d) temperature (°C), (e–h) O2 concentration (μmol kg−1) and (i–l) ΩARG value at depths of 500 m (dotted lines with triangles) and 600 m (dashed lines with squares) at four locations (area of each = 2° latitude × 2° longitude) centered at 40.5°N and 132°E (blue); 43°N and 138°E (red); 37°N and 131°E (black), and 38°N and 136°E (green), respectively. The inset in Figure 8h shows four selected locations. ΩARG values were predicted using equation (3) and data for O2 concentration, temperature and pressure.

3.3. Interdecadal Variability in the Aragonite Saturation State of the Sea of Japan (East Sea) (Without Anthropogenic CO2)

[25] The predicted interdecadal variability of ΩARG values (corrected for the effect of anthropogenic CO2) in the Sea of Japan (East Sea) exhibited considerable spatial variation. The variability (2σ values from the mean) ranged between ±0.1 (north of 41°N) and ±0.7 (39–41°N and 131–132°E) in the upper ocean (at 100 m depth), but was ±0.05 (in most areas) and ±0.2 (41–42°N and 134–136°E) in deeper water (at 500 m depth) (not shown). The decadal variability was considerably greater than the fit error (±0.020), indicating that the chosen model can be used to predict interdecadal changes in ΩARG values.

[26] Our discussion primarily focuses on the variability in the 100% aragonite saturation (ΩARG = 1) depth in the Sea of Japan (East Sea), because, from a thermodynamic perspective, ΩARG = 1 represents a transition from CaCO3 precipitation to CaCO3 dissolution [Orr et al., 2005; Fabry et al., 2008], even though calcification and dissolution are subject to the so-called “vital effect,” including organic coating of shells and species-specific calcification mechanisms [Langdon et al., 2003; Tunnicliffe et al., 2009]. The most noteworthy feature of the distribution of the ΩARG = 1 depth in the Sea of Japan (East Sea) was the large interdecadal variability in the northern basin relative to the southern basin. In most parts of the northern basin the ΩARG = 1 depth was 400–450 m in the 1960s, but this deepened by another 100–200 m in the 1970s (Figure 9). During the 1970s and 1980s, two locations (centered at 42.5°N and 139°E; south of Peter the Great Bay) in the northern basin showed considerable variation in the ΩARG = 1 depth, with the shift in this depth at one location often being in the opposite direction to that observed at the other location. For example, during the transition period (the 1970s and 1980s) the ΩARG = 1 depth moved up by more than 200 m in the eastern part of the northern basin, whereas it moved downward by a similar depth in the western part of the northern basin (Figure 10). Similar displacements in the ΩARG = 1 depth re-occurred during the transition period between the 1980s and 1990s, with some variation in location.

Figure 9.

Distribution of the aragonite saturation depth for the (a) 1960s, (b) 1970s, (c) 1980s and (d) 1990s, predicted using equation (3) in conjunction with historical data on temperature, pressure and O2 concentration. The dots indicate data points. The contour lines and numbers represent 50 and 100 m intervals, respectively.

Figure 10.

Anomaly (ΩARG = 1 depths for each time period - 40-year means) plots for the (a) 1960s, (b) 1970s, (c) 1980s and (d) 1990s. Positive and negative values denote the mean ΩARG = 1 depths as deeper or shallower, respectively. Sampling locations are not shown but are identical to those shown in Figure 9.

[27] The sudden and substantial shift in the ΩARG = 1 depth between the 1960s and 1990s is probably attributable to the waxing and waning of water column ventilation, as this directly affects both temperature, and O2 and anthropogenic CO2 concentrations, in the interior waters of the Sea of Japan (East Sea) [Gamo et al., 2001; Park et al., 2008], and variations in these parameters are closely related to variations in seawater ΩARG. To test this hypothesis, we assessed the temporal evolution of O2 concentration and temperature in waters centered at 500 and 600 m depths at four locations (two in the northern basin and two in the southern basin) of the Sea of Japan (East Sea), as much of the interdecadal variation in water column ventilation is largely confined to waters shallower than 600 m [Min and Warner, 2005], and the ΩARG = 1 is generally found at these depths. Our analysis indicated that, in all locations but one of the northern locations, the O2 concentration briefly increased during the transition period between the 1960s and 1970s, and gradually fell thereafter. Contrary to this general trend, the exceptional northern location showed a slightly different trend, with an increased O2 concentration from the 1960s to the 1980s, followed by a decrease in the 1990s. This trend (an inverse relationship between O2 concentration and temperature) was also evident for seawater temperature. The consistency of results for these two parameters indicates that water column ventilation intensified during the transition period between the 1960s and 1970s (or between the 1960s and 1980s in the northern location), and weakened during more recent periods (Figure 8). This finding is broadly in agreement with earlier studies concluding that reductions in O2 concentration in the open ocean have resulted from changes in ocean circulation [Joos et al., 2003; McDonagh et al., 2005; Deutsch et al., 2006]. The strong correlation between variations in ventilation proxies (i.e., temperature and O2 concentration) and changes in ΩARG = 1 depth supports our hypothesis.

3.4. Effect of Anthropogenic CO2 on the Aragonite Saturation State of the Sea of Japan (East Sea)

[28] In addition to variations in the degree of water column ventilation and coastal upwelling, the input of anthropogenic CO2 is another important and emerging factor in determining the seawater ΩARG of the Sea of Japan (East Sea). The input of anthropogenic CO2 decreases seawater ΩARG value, unlike the effects of other natural factors that can cause seawater ΩARG value to either decrease or increase. In the Sea of Japan (East Sea), the effect of anthropogenic CO2 on the seawater ΩARG value may not be linear over time, because varying accumulation rates of anthropogenic CO2 have recently been reported [Park et al., 2008]. In addition, anthropogenic CO2 absorbed by the Sea of Japan (East Sea) in the period 1992–2007 has moved the ΩARG = 1 horizon up by only ∼40 m, which is less than predicted by the errors associated with estimation of ΩARG and anthropogenic CO2 More strikingly, there has been little change in the ΩARG = 1 depth caused by anthropogenic CO2 for the period 1999–2007, during which only negligible anthropogenic CO2 was found in waters below 300 m [Park et al., 2008]. With respect to the effect of anthropogenic CO2 on the ΩARG = 1 depth in the Sea of Japan (East Sea), in the present study we assessed only the impact of the total amount of anthropogenic CO2 that had accumulated in the period between 1800 and 1999.

[29] The anthropogenic CO2 that accumulated in the Sea of Japan (East Sea) from 1800 to 1999 has variously displaced the ΩARG = 1 depth upward by 50–250 m across the basin (Figure 11), with the most substantial upward displacements occurring in the northern basin relative to the southern basin. The magnitudes of displacement of the ΩARG = 1 depth in the northern basin were considerably more than observed for the same latitude in the North Pacific (i.e., ∼50 m) [Feely et al., 2002]. The greater upward displacement of the ΩARG = 1 depth in the northern basin of the Sea of Japan (East Sea) relative to the North Pacific can be attributed to the deeper penetration of anthropogenic CO2 in the former.

Figure 11.

Distribution of basin wide shoaling of the aragonite saturation depth (ΩARG = 1) due to the input of anthropogenic CO2. Open circles indicate sampling locations during the 1999 cruise. The contour lines and numbers indicate the aragonite saturation depth at 25 and 50 m intervals, respectively.

[30] The differences in the vertical gradient of ΩARG caused significant differences (as much as 100 m) in the upward shift of the ΩARG = 1 depth within the Sea of Japan (East Sea). The greatest upward depth displacement (>200 m) occurred in the northern and western basins, whereas in the southern basin the displacement was less than 100 m (Figure 11). These contrasting regional differences are largely attributable to regional variation in the vertical gradient of ΩARG. The areas where the effect of anthropogenic CO2 on ΩARG was less significant tended to show larger gradients in ΩARG (−0.1 to −0.3 ΩARG/100 m) over the depth range between 125% and 75% of the ΩARG = 1 horizon, whereas in the Japan and western basins, where the effect on ΩARG was significant, there were smaller gradients (−0.05 to −0.1 ΩARG/100 m) over the same ΩARG interval.

[31] The impact of anthropogenic CO2 on the ΩARG = 1 depth in coastal waters of three countries was also examined at sampling stations located within 50 km of the coast (Figure 12). Anthropogenic CO2 has significantly affected the coastal waters of Korea, where the ΩARG = 1 depth has risen by as much as 200 m (north of ∼37°N). Its impact has been less significant in the coastal waters of Russia and Japan, where the ΩARG = 1 depth has risen by only approximately 100 m. Overall, shelf waters (less than 100 m depth) in all three coastal regions were oversaturated with respect to aragonite. However, seasonal upwelling makes shelf waters undersaturated with aragonite by inducing upwelling of corrosive slope waters to the surface. Future upwelling events will make shelf waters more corrosive by bringing increasingly acidic waters to the shelf slope over time, because of the accumulation of anthropogenic CO2 in slope waters.

Figure 12.

Alongshore sections of the aragonite saturation state of seawater (ΩARG; solid and dashed lines) and anthropogenic CO2 concentration (colors) for coastal waters of (a) Korea, (b) Russia and (c) Japan. Values of ΩARG and anthropogenic CO2 were calculated using data collected during the 1999 survey. The dashed and solid lines represent the ΩARG = 1 horizon for the pre-industrial era and the year 1999, respectively. Open circles indicate sampling locations.

[32] Because the upper coastal waters (shallower than 100 m) of the Sea of Japan (East Sea) are generally supersaturated with respect to all phases of CaCO3, seawater carbonate chemistry has not previously been considered to be a limiting factor in the growth of calcifying organisms. However, recent studies indicate that the degree of supersaturation affects the calcification rates of species in both planktonic and benthic habitats [e.g., Riebesell et al., 2000; Kurihara and Shirayama, 2004; Delille et al., 2005; Manzello et al., 2008; Green et al., 2009; Moy et al., 2009]. Acidification induced by both natural and anthropogenic factors could have profound consequences for both benthic and pelagic calcifying organisms including commercially important organisms endemic to the coastal waters of the Sea of Japan (East Sea), such as oysters, mussels, echinoderms, and pteropods [Fabry et al., 2008; Doney et al., 2009]. Acidification may also affect the reproductive success of many benthic invertebrates because juvenile larvae and gametes are particularly vulnerable to low ΩARG [Kurihara and Shirayama, 2004].

4. Conclusion

[33] In this study we established an objective statistical procedure through which a multiparameter linear regression model relating seawater ΩARG values to other parameters was derived. The procedure described should prove useful as a guide for future similar studies. Three key variables (temperature, pressure and O2 concentration) are intimately linked to changes in seawater ΩARG in the Sea of Japan (East Sea), including coastal waters. The functional model relating seawater ΩARG to these variables in the Sea of Japan (East Sea) enabled prediction of interdecadal changes in the distribution of ΩARG, corrected for anthropogenic CO2. When the optimum functional model was run using past measurements of temperature, pressure and O2 concentration, the model identified the waxing and waning of water column ventilation over time as the primary factor determining the interdecadal variability in ΩARG for the neritic Sea of Japan (East Sea), whereas seasonal upwelling events were found to be of primary importance in coastal waters. In addition to these natural factors, the impact of anthropogenic CO2 on the ΩARG value has become significant in both neritic and coastal waters, and its influence is expected to increase in the future. The continued acidification of coastal waters of the Sea of Japan (East Sea) will put numerous commercially valuable organisms in jeopardy.


[34] This paper benefited a great deal from the numerous constructive suggestions by Scott Doney of Woods Hole Oceanographic Institution and an anonymous reviewer. This work would not have been possible without the efforts of many scientists who contributed data to the Japanese Ocean Data Center (JODC). We extend our thanks to scientists of the JODC who compiled and enabled access to the data. This work was primarily supported by the National Research Laboratory program of the Korean Science and Engineering Foundation and the MOEHRD (KRF-2005-070-C00143). We also thank Mike Johnson, Joel Levy, and Ken Mooney of the NOAA Climate Program Office for their support of the NOAA studies.