Introduction

Atmospheric carbon dioxide (CO2) levels today are nearly 50% higher than during pre-industrial times and are predicted to increase at similar or accelerating rates over the next hundred years1. The ocean has taken up roughly a quarter of the total anthropogenic emissions2, resulting in decreased ocean pH and associated changes in carbonate system equilibrium due to ocean acidification (OA)3. Globally, the mean surface ocean pH has decreased by ~0.1 since the beginning of the Industrial Revolution and is projected to drop by as much over the next 60 years4,5. These trends have been modeled6 and detected7 in the California Current System (CCS).

The California Cooperative Oceanic Fisheries Investigation (CalCOFI) program was formed in 1949 to study the pelagic ecosystem of the CCS in response to the collapse of an economically important sardine fishery8. The original sampling design included regular cruises making a grid pattern of profiles to 500 m of physically and biologically important parameters such as temperature, salinity, and zooplankton biomass. Observations of carbonate chemistry were incorporated into CalCOFI in 1983 at the same time as Charles David Keeling initiated time series measurements at Hydrostation S near Bermuda9, and Olafsson and Takahashi began a time series in the North Atlantic9,10,11. The most continuous time series in the Southern CCS is surface waters (0–20 m) at CalCOFI Line 90 Station 90 (station 90.90). Observations were made at this location between 1984–present, with a gap from 2002 to 2008. Station 90.90 is the oldest traditional hydrographic time series of inorganic carbon in the Pacific yet, these valuable observations have remained unpublished until this work. The long-term trends established by this work add a direct observation of OA and climate change to the few existing time series of ocean carbonate chemistry9,12.

CalCOFI station 90.90 is located at 31.4°N, 122°W, approximately 450 km from shore, with a water depth of approximately 4000 m. Due to its location in the western California Current, station 90.90 lies near the eastern edge of the North Pacific Subtropical Gyre exhibiting an oligotrophic open-ocean regime13 (the mean phosphate and nitrate concentrations were 0.3 μM and 0.1 μM respectively in sea surface samples with inorganic carbon observations). Samples were analyzed for total alkalinity (AT) and dissolved inorganic carbon (CT). AT and CT were used to calculate other carbonate system variables such as, partial pressure of CO2 in seawater (pCO2), pH, carbonate ion concentration ([\({{CO}}_{3}^{2-}\)]), saturation states of aragonite and calcite (Ωaragonite, Ωcalcite), and Revelle Factor (∂ln[CO2]/∂lnCT). See Methods for details.

Results and discussion

The time series exhibits an unambiguous OA signal in pH of −0.0015 ± 0.0001 yr−1 from 1984 to 2021, in agreement with the global surface ocean average14 of −0.0017 yr−1 (Fig. 1, Table 1), as well as a distinct seasonal cycle (Fig. 1, right column). The sea surface pCO2 at station 90.90 is driven by increasing total inorganic carbon (CT) at a decadal scale and a combination of CT and temperature at a seasonal scale (Fig. 2). There is no significant trend in ocean temperature, salinity, AT or nAT at station 90.90. However, longer-term, near-shore studies in the CCE have shown increasing temperatures shoreward of the California Current15.

Fig. 1: CalCOFI Line 90, Station 90 ocean time series.
figure 1

Seasonally detrended observations (a–f) and average seasonal cycles (g–l). “n” indicates salinity normalization to the mean salinity (33.3). a Temperature and salinity, (b) Total alkalinity (AT) and nAT, (c) Total inorganic carbon (CT) and nCT, (d) pCO2 (μatm) and Revelle factor, (e) pH and CO3, and (f) Ωcalcite and Ωaragonite. There is no significant trend in temperature, salinity or AT (p > 0.35). The ocean acidification trend (shown in c–f) is within the range of observations made at other time series sites9. Regression statistics for time series are shown in Table 1. Descriptive statistics for seasonal cycles are shown in Supplementary Table 2. Error bars in (g–l) represent the standard error of the 3 month sliding bin (Supplementary Fig. 2) with number of observations ranging from 20 to 33.

Full size image
Table 1 Regression statistics of sea surface hydrography and seawater carbon chemistry from 1984 to 2021. (from Fig. 1).
Full size table
Fig. 2: Forcings on the annual pCO2 cycle.
figure 2

The relative contributions of salinity, temperature, AT and CT to the seasonal cycle of sea surface pCO2, computed using CO2SYS40.

Full size image

Several features make station 90.90 unique compared to other multi-annual longstanding oceanic inorganic carbon time series9. Station 90.90 is the oldest traditional hydrographic inorganic carbon time series location in the Pacific, it is also the furthest from land globally. The observed CT, Ωaragonite, Revelle Factor, sea surface temperature and salinity at station 90.90 have average seasonal ranges not covered by the other seven time series9. Station 90.90 is one of two located within an eastern boundary current, shown in Supplementary Fig. 1 (alongside European Station for Time series in the Ocean at the Canary Islands i.e., ESTOC)9. Station 90.90 is also one of two time series locations that appear to be annual net sources of CO2 to the atmosphere (alongside CArbon Retention In A Colored Ocean i.e., CARIACO)9,16. Station 90.90 has a mid-range Revelle Factor, and a mean pCO2 close to that of the atmosphere (Fig. 3), resulting in a nominal to low mean ΔpCO2, where ΔpCO2 is the difference in the partial pressure of CO2 between the sea surface and atmosphere. All of the other time series sites are located in areas with either a lower Revelle Factor or a greater ΔpCO2. A lower Revelle Factor allows for greater uptake and storage of anthropogenic CO2 and a greater (negative) ΔpCO2 increases the uptake of both natural and anthropogenic CO2. The unique chemical signature of station 90.90 results in the lowest CO2 uptake rate among all of the timeseries sites9. The direct evidence of a reduced uptake rate of CO2, relative to the global average is presumably due to proximity to a mid-latitude eastern boundary current upwelling zone17. As we discuss below, this finding is significant because models prescribing climate change trends in terms of mean CT uptake rate will differ from those using local Revelle Factor and ΔpCO2.

Fig. 3: Comparison of pCO2.
figure 3

Data products26 for atmospheric boundary layer (SeaFluxatm) and sea surface pCO2 (SeaFluxsea) compared to sea surface pCO2 at station 90.90. SeaFlux data are the average between two points, longitude 121.5°W, 31.5°N and 122.5°W, 31.5°N. From Table 2 in ref. 25, the reported bias ± RMSE (relative to the training data) in the product used to generate SeaFluxsea is 0.1 ± 13.7 μatm.

Full size image

Natural variability at station 90.90 (resulting from, e.g., interannual variability in the proportion of North Pacific Gyre vs California Current water masses) along with the six-year gap in data confound the identification of long period patterns such as El Nino, or the Pacific Decadal Oscillation. Perhaps the most obvious anomaly is a perturbation in temperature and salinity during 2014–2016 (Fig. 1). This anomaly may be a result of the 2014/15 North Pacific marine heatwave, the strong 2015/16 El Nino, or a combination of the two18,19,20.

Power spectral density (PSD) analysis of the detrended time series showed the presence of a strong annual signal in temperature and CT, as well as in all calculated carbonate system variables, but did not resolve interannual features within the time series (Supplementary Fig. 5). While temporal anomalies in this time series may be worthy of further investigation, the goal of this work is to present the 1st order OA trend, mean seasonal cycle, and to finally make the quality-controlled time series publicly available. Follow-on work with this time series may consider implementing gap-filling techniques for the 2002–2008 period21 and alternative modes of trend detection such as simultaneous fitting of harmonics and underlying trends as well as development of a mixed layer carbon budget at station 90.90.

The station 90.90 observations presented here are compared to the global bottle dataset (Global Ocean Data Analysis Project; GLODAP22) and the global underway pCO2 dataset (Surface Ocean CO2 Atlas (SOCAT)23). GLODAP data are used via the widely-accepted open-source empirical function, ESPER24. SOCAT data are the basis of a product used in this study25,26 to estimate the sea surface pCO2 at station 90.90.

Using ESPER we estimate CT and AT, where CT and AT are derived from local data collected independent of CalCOFI (see methods)24. While ESPER can also serve as the basis to estimate pCO2, (as well as pH and Ω) products based on SOCAT provide a more robust comparison for pCO2 due to the higher density of surface data in their training set. This was done using two different sets of predictor variables: first, using only temperature and salinity, and second, using all available predictor variables including temperature, salinity, phosphate, nitrate, silicic acid, and oxygen concentrations (in addition to latitude, longitude, depth, and year in both cases). Depending on the “preformed” nature of surface water, additional variables such as oxygen and nutrients may not correlate with carbonate parameters in the surface ocean, hence the use of T and S only as one set of predictors27,28. Using only temperature and salinity, the comparison for CT showed measurements were 6 ± 15 μmol kg–1 (mean ± std) higher than ESPER with less variability (Supplementary Fig. 6). The comparison of AT showed measurements were 4 ± 4 μmol kg–1 (mean ± std) lower than ESPER with similar variability (Supplementary Fig. 7). Both CT and AT are significantly different than the ESPER predictions (p«0.05). When using all available predictor variables, the comparison for CT showed measurements were 0.1 ± 6 μmol kg–1 (mean ± std) higher than ESPER with similar variability (Supplementary Fig. 8). The comparison of AT showed measurements were 3 ± 4 μmol kg–1 (mean ± std) different than ESPER with similar variability (Supplementary Fig. 9). Only AT was significantly different than the ESPER predictions (p«0.05). Increased accuracy in predicted CT and AT when using a greater number of predictor variables matches ESPER’s expected performance24.

There was no significant trend in the AT residuals when using either set of predictor variables. However, there was a significant trend in the CT residuals in both sets of predictor variables (Supplementary Figs. 6 and 8), indicating the predicted CT trend was 40% higher (a residual slope of 0.3 higher than the measured slope of 0.7 μmol kg−1 yr-1) than the observed trend. Both sets of predictor variables resulted in the same trend, albeit with a higher standard deviation when using fewer predictors. The discrepancy between the OA trend observed and predicted may be due to the use of the simplified steady-state assumption used by ESPER to predict anthropogenic CO2 as a function of time via a simple exponential equation with one coefficient applied globally24. We noted a slight decrease in phosphate over the time series (Wolfe, 2022), which could impact the trend in the second ESPER prediction, but not the first. The decrease in phosphate may signal a change in biogeochemistry of the region and is worth further investigation, but the fact a discrepancy of 0.3 μmol kg–1 yr-1 shows up in both versions of measured vs. ESPER, suggests that the discrepancy is due to regional differences relative to the average forcing employed within the ESPER model.

The main drivers of pCO2 seasonality are temperature and CT, with little contribution from AT or salinity (Fig. 2). Each contributes 57%, 34%, 8%, and 1% of the pCO2 seasonality, respectively. The effects of temperature and CT on pCO2 are out of phase, which cancels out much of their seasonal impact on pCO2. In turn, CT is driven by gas exchange, net ecosystem metabolism and mixing (Wolfe, 2022)16. On a decadal timescale, increasing CT is the only significant driver of pCO2, contributing 93% (Supplementary Fig. 3). Similarly, CT is responsible for >90% of the trends in the other carbonate variables shown in Fig. 1 and Table 1. Although it is beyond the scope of this work, a mixed layer carbon budget at station 90.90 is the subject of a separate manuscript in progress.

The measured sea surface pCO2 trend matches the atmospheric CO2 trend over the same timeframe, although the sea surface pCO2 has significantly greater variability than atmospheric pCO2, a feature common to all ocean time series (Fig. 3)29,30. The mean sea surface pCO2 seasonal cycle derived from the CalCOFI bottle data also exhibits a greater amplitude (42 vs. 29 μatm) and a phase shifted maximum (July vs. September) compared to the climatological sea surface pCO2 (Fig. 4, based on the Landschützer climatology25 and SeaFlux product26). Perhaps most importantly, the measured sea surface pCO2 at station 90.90 is, on average, 15 μatm higher than the products derived from the Surface Ocean CO2 Atlas (SOCAT) database23, a discrepancy large enough to lead to characterization of this location as a CO2 source (when using the measured data) rather than a sink (when using e.g., the SeaFlux data products) to the atmosphere (Fig. 3).

Fig. 4: Hydrography vs. climatology.
figure 4

The measured pCO2 cycle at station 90.90 (computed from AT and CT) compared to the Landschützer et al. (2020) climatology at station 90.90. The climatology was adjusted to a mean year of 2002 by applying our observed slope of −1.5 μatm yr-1 (Table 1), and by the scaling of Fay et al. (2021) to generate the SeaFlux data product (sampled for year 2002). Error bars for 90.90 are identical to Fig. 1j. Error bars shown for the Landschützer climatology and SeaFlux product are 13.7 μatm, the RMSE reported in Table 2 in ref. 25.

Full size image

The higher mean and more variable pCO2 of the time series relative to the data products highlights the continued need for direct measurements. An important next step should be assimilating the data presented in this work into empirical algorithms and climatology products of the CO2 system at the station 90.90 study site. There is little doubt that future versions of products such as ESPER and SeaFlux will benefit by incorporating new measurements such as the 90.90 data as well as assimilating new data sets (e.g., using SOCAT data in ESPER).

Conclusions

This work establishes station 90.90 as one of very few long-standing marine inorganic carbon time series9, with various unique properties. Over 37 years, the surface ocean at station 90.90 has decreased in pH by 0.0015 yr-1 and increased in pCO2 and CT by 1.6 μatm yr–1 and 0.7 μmol kg–1 yr–1, respectively. While these trends are in line with other open ocean trends documented in the Central North Pacific, and Sargasso Sea9, we found that widely used empirical proxy relationships24 introduced a 40% bias in the rate of CO2 local uptake, underscoring the need for sustained measurements. We also report a strong annual cycle in carbonate system variables, with dominant control of the seasonal cycle by temperature and total inorganic carbon. In contrast, the long-term secular trend in carbonate system variables is directly related to an increase in total inorganic carbon.

Methods

Sampling at station 90.90

CalCOFI station 90.90 is located at 31.4°N, 122°W, approximately 450 km from shore, with a water depth of approximately 4000 m. All observations discussed here were collected near the sea surface (0–20 m), with an average depth of 5.2 m. Although a small subset of these measurements (2009–2015) have been publicly available for several years (see data availability), the remaining 22 years of observations have not been published until this work.

Observations used in this work cover the period 1984–2021 with a gap from 2002 to 2008. Bottle samples were collected on quarterly CalCOFI cruises at station 90.90. Mercuric chloride was added (as a biocide and preservative to final concentration of 50 µM) and the samples were sealed and stored in borosilicate glass bottles following best practices31. Storage times ranged from one month to multiple years before analysis.

Analytical methods

Bottle samples were analyzed for total alkalinity (AT) and dissolved inorganic carbon (CT). AT was measured using a closed cell titration32 until 1992 and an open cell titration33 after 1992. CT was measured using vacuum extraction and manometry34,35 until 1992, coulometry36 from 1992 to 2015 and an infrared (IR) analyzer37,38 after 2015. Throughout the record, CT has been traceable to a Scripps manometer (either directly or via Certified Reference Materials39 after 1992) which, in parallel with NOAA, defines the WMO mole fraction scale for CO2 in air. AT has been measured using the traceable practices (acid standardization, weighing) and data processing that now define the AT of Certified Reference Materials33. The various methods used over the years for both AT and CT differ primarily in their precision and therefore accuracy of AT and CT is estimated to range from 2 to 5 μmol kg-1 and 1–3 μmol kg-1, respectively, over the dataset, reflecting the range of precisions of instruments used with the low end of the range representing the uncertainty of the reference material or method.

Calculating additional carbonate chemistry parameters

The partial pressure of CO2 in seawater (pCO2), pH (total scale), carbonate ion concentration ([\({{CO}}_{3}^{2-}\)]), saturation states of aragonite and calcite (Ωaragonite, Ωcalcite), and Revelle Factor (∂ln[CO2]/∂lnCT), were calculated in MATLAB using CO2SYS40 from AT, CT, temperature and salinity with carbonic acid equilibrium coefficients reported in Lueker et al.35 and other constants as recommended by the Guide to Best Practices31 along with the borate-salinity relationship of Lee et al.41. Based on propagation of error through CO2SYS, the measurement uncertainty in CT and AT may lead to an error in pCO2 in the range 6–10 µatm. The choice of input constants (or choice of uncertainty one assigns to the constants) used in CO2SYS presents a significant source of bias in derived values that would negligibly affect the reported slopes in Table 1, but significantly affect the mean pCO2 and other derived parameters in Figs. 1, 3, and 4. Resolving such biases in derived vs measured CO2 system parameters has long been a central focus of marine carbonate chemists42,43,44 and may indeed prove to be part of the apparent ~15 µatm bias between the pCO2 derived from CalCOFI bottle data and products based on underway pCO2 (Figs. 3 and 4). AT and CT were salinity normalized (indicated by nAT and nCT,) to the average salinity of the time series (33.3, n = 107).

Seasonal cycle

Monthly binning is a common approach used to extract the seasonal cycle45,46. In this work the 12-month climatology was computed from quarterly observations using a 3-month sliding bin. Where, for example, April is represented as the average of all observations from March, April, and May. Due to variability in the scheduling of CalCOFI cruises, there are some observations in each month before binning (Supplementary Fig. 2). The resulting 12-month climatology was used to seasonally detrend the observations as removing periodic signals is a best practice for assessing OA trends47. The climatology was used to calculate the relative contributions of salinity, temperature, AT, and CT to the seasonal cycle of pCO2. The climatology of a single parameter and the average of the remaining three were used with CO2SYS to calculate individual contributions, at time t, as follows,

$${\triangle p{{{{{\rm{CO}}}}}}}_{2,{{{{{\rm{t}}}}}}}^{{{{{{\rm{temperature}}}}}}}={p{{{{{\rm{CO}}}}}}}_{2}\left(\bar{{A}_{{{{{{\rm{T}}}}}}}},\bar{{C}_{{{{{{\rm{T}}}}}}}},\bar{{{{{{\rm{S}}}}}}},{{{{{{\rm{T}}}}}}}_{{{{{{\rm{t}}}}}}}\right)-\bar{{p{{{{{\rm{CO}}}}}}}_{2}}$$
(1)
$${\triangle p{{{{{\rm{CO}}}}}}}_{2,{{{{{\rm{t}}}}}}}^{{{{{{\rm{salinity}}}}}}}={p{{{{{\rm{CO}}}}}}}_{2}\left(\bar{{A}_{{{{{{\rm{T}}}}}}}},\bar{{C}_{{{{{{\rm{T}}}}}}}},{{{{{{\rm{S}}}}}}}_{{{{{{\rm{t}}}}}}},\bar{{{{{{\rm{T}}}}}}}\right)-\bar{{p{{{{{\rm{CO}}}}}}}_{2}}$$
(2)
$${\triangle p{{{{{\rm{CO}}}}}}}_{2,{{{{{\rm{t}}}}}}}^{{C}_{{{{{{\rm{T}}}}}}}}={p{{{{{\rm{CO}}}}}}}_{2}\left(\bar{{A}_{{{{{{\rm{T}}}}}}}},{{C}_{{{{{{\rm{T}}}}}}}}_{{{{{{\rm{t}}}}}}},\bar{{{{{{\rm{S}}}}}}},\bar{{{{{{\rm{T}}}}}}}\right)-\bar{{p{{{{{\rm{CO}}}}}}}_{2}}$$
(3)
$${\triangle p{{{{{\rm{CO}}}}}}}_{2,{{{{{\rm{t}}}}}}}^{{A}_{{{{{{\rm{T}}}}}}}}={p{{{{{\rm{CO}}}}}}}_{2}\left({{A}_{{{{{{\rm{T}}}}}}}}_{{{{{{\rm{t}}}}}}},\bar{{C}_{{{{{{\rm{T}}}}}}}},\bar{{{{{{\rm{S}}}}}}},\bar{{{{{{\rm{T}}}}}}}\right)-\bar{{p{{{{{\rm{CO}}}}}}}_{2}}$$
(4)

The same procedure was used to calculate the contribution to the long-term trends in pCO2 (Supplementary Fig. 3).

Data processing

Model I linear regression (function fitlm in MATLAB) was performed on observations (Supplementary Fig. 4, Supplementary Table 1) and seasonally detrended data (Table 1). The slope, error, r2, p-value and n values are reported. PSD analysis was performed using Lomb-Scargle periodograms with the ‘plomb’ function in MATLAB (Supplementary Fig. 4)48. Observations were also compared to empirically derived proxy estimates “ESPER_MIXED”24,40,49 and to a climatology25 derived from data sources (Global Ocean Data Analysis Project; GLODAP22, The Surface Ocean CO2 Atlas; SOCAT23) including local observations form the Eastern Pacific, independent from the CalCOFI CO2 record.