1 Introduction
Particulate organic carbon (POC) is composed of non-living organic matter, i.e., sediments, faecal pellets and detritus, and the living biomass, e.g. phytoplankton, zooplankton, bacteria. It represents one of the most dynamic pools of carbon in the ocean. Knowledge of the spatial and temporal distribution of POC in the global ocean is fundamental to understand the impacts of climate change on the biological carbon pump (Bopp et al., 2001; Sarmiento & Gruber, 2006).
In the last decades, several studies have described the spatial distribution of POC and the processes controlling its temporal variations from the surface to the abyssopelagic zone, leading to relatively extensive coverage of POC (Aumont et al., 2017; Bishop & Wood, 2008; Bishop & Wood, 2009; Cetinić et al., 2012; Gardner et al., 2003; Gardner et al., 2006; Henson et al., 2012; Kiko et al., 2017; Lam et al., 2011; Moran et al., 1999; Rasse et al., 2017; Wangersky, 1976). Given its importance, much effort has been invested in developing methods for determining POC. Nevertheless, we still have a relatively rudimentary understanding of the uncertainties associated with these methods and, therefore, with the estimates of POC. Lack of a consensus protocol (but see IOCCG POC Sampling and Measurement Protocols) and differences in methodologies preclude one from distinguishing between inaccuracies in the analytical methods and the natural heterogeneity of POC in the water column. Thus, improved and standardised methodologies for estimating POC and for assessing and reducing uncertainties associated with those estimates are required (Burd et al., 2016; Gardner et al., 2003; Russell et al., 2015; Stramski et al., 2008).
Different approaches can introduce variability and affect the accuracy of POC concentrations. Possible sources of inconsistency include adsorption of dissolved organic and particulate inorganic carbon on filters, sampling collection systems, filter contamination, water patchiness, pressure differentials, filter loading, artificial formation of particles, and breakage of phytoplankton and other particles during the processing of bottle samples (Gardner et al., 2000; Gardner et al., 2003). However, insufficient work has been done to quantify the uncertainties associated with these potential artefacts. There is no consensus on how to minimise and correct the uncertainty associated with each artefact and standard methods do not evaluate the overall accuracy of POC concentrations. Moreover, it is not known which part of the sample processing and analysis is likely to introduce the largest portion of the uncertainty. Analysing the effects of different artefacts on POC uncertainty would improve the reliability of POC estimates. Also, it would help to develop a consensus protocol which could promote the inter-comparability of POC concentrations across different research groups.
This study aims to contribute to developing a method to quantify experimental uncertainties of POC concentrations using duplicate measurements, to improve the understanding of the uncertainty sources in POC estimates, to promote accurate techniques for measuring and processing POC concentrations, and to contribute to developing a uniform protocol.
2 Data and methods
In an attempt to quantify the uncertainties and to reach higher accuracy in POC concentrations, we modified the sampling, processing and analysing methodologies described in the protocols for the Joint Global Ocean Flux Study (JGOFS) (Knap et al., 1996; Martin et al., 1993). Modifications, described in more detail in Sections 2.1, 2.2 and 2.3, included the sampling of different volumes of water in accordance to expected in-situ concentrations, using different types of blanks to quantify the dissolved organic carbon and any contamination, and using silica gel in the desiccators during the acidification step to minimise any potential contamination. All code used to analyse the data has been made publicly available (see Section 4) (pstrubinger, 2021).
2.1 Sampling
Water samples were collected at 67 stations during the 24th Atlantic Meridional Transect (AMT-24) aboard the RRS James Clark Ross from September 25th to November 1st, 2014 (see Figure 1). Two casts were completed every day (weather permitting): one pre-dawn and the other around solar noon. Different ecological provinces (Longhurst, 2007) were sampled: the North Atlantic Drift Province (NADR), the North Atlantic Subtropical Gyral Province (NAST), the North Atlantic Tropical Gyral Province (NATL), the Western Tropical Atlantic Province (WTRA), the South Atlantic Gyral Province (SATL), and the South Subtropical Convergence Province (SSTC).
Figure 1. The track of AMT-24 cruise with locations of 67 stations where samples for particulate organic carbon concentration were collected.
Colour coding of the stations represents biogeographical provinces that were sampled: the North Atlantic Drift Province (NADR), the North Atlantic Subtropical Gyral Province (NAST), the North Atlantic Tropical Gyral Province (NATL), theWestern Tropical Atlantic Province (WTRA), the South Atlantic Gyral Province (SATL), and the South Subtropical Convergence Province (SSTC).
At each station, water samples were collected from six depths in the upper 500 m using 20-l Niskin bottles. The sampling procedure did not account for dregs, the rare large particles that might have been under sampled by Niskin bottles (Gardner et al., 2006). Water samples were transferred from the Niskin bottles into six 15-l carboys, which were pre-washed with acid and then with sample water prior to use. The carboys were taken to the on-board laboratory, where water samples were transferred into six polycarbonate ~2-l bottles and filtered through pre-combusted (450°C for five hours) 25 mm Whatman glass fiber filters (GF/F) at low vacuum (~125 mm Hg) by inverting the bottles into a standard funnel setup. Each of the polycarbonate bottles was also pre-washed with sample water prior to use. To keep the water samples homogenised, carboys were gently shaken before pouring into the bottles. The volume of sample water filtered for each POC measurement varied between ~1 and ~8 liters, and it was adjusted according to the expected concentration of POC in the respective environment. Thus, each 2-l bottle was typically re-filled (up to four times) during filtration to achieve the total volume for any given sample. No measures were taken to prevent atmospheric contamination during filtration.
We implemented the "double-filter" technique advocated by Banoub & Williams (1972); Feely (1974); Kinney et al. (1971); Loder & Hood (1972); Moran et al. (1999); Smith et al. (1996). These authors suggested using two stacked filters for each estimate of POC concentration. The upper filter was used to collect particles and adsorbed dissolved organic carbon (uncorrected POC sample, henceforth referred to as "uPOC") while the lower filter to quantify the dissolved organic carbon that is adsorbed onto GF/F filters (adsorbed DOC blank, henceforth referred to as "aDOC") (see Table 1 for a list of variable names and their subscripts). After each filtration, uPOC and aDOC filters were removed from the filtration rig, wrapped into separate pre-combusted (450°C for five hours) aluminium foil envelopes, flash-frozen in liquid nitrogen, and then stored in a freezer at −80°C for post-cruise analysis.
Table 1. List of variables names and their subscripts.
| Subscript | Variable Name | Units |
|---|
| POC | Particulate Organic Carbon | N/A |
| uPOC | Uncorrected Particulate Organic Carbon | N/A |
| aDOC | Adsorbed Dissolved Organic Carbon | N/A |
| MuPOC | Mass of organic carbon on uPOC | µg |
| MaDOC | Mass of organic carbon on aDOC | µg |
| Mcap | Mass of organic carbon on tin capsules | µg |
| Mac | Mass of organic carbon on acidified filters | µg |
| Mnac | Mass of organic carbon on non-acidified filters | µg |
| M | Mass of particulate organic carbon | µg |
| C | Concentration of particulate organic carbon | mg/m3 |
| ∆ | Scaled arithmetic difference between duplicate pairs | mg/m3 |
| ∆r | Scaled relative difference between duplicate pairs | dimensionless |
| σ∆ | Uncertainty of each duplicate difference | mg/m3 |
| σdr | Relative experimental uncertainties of duplicate values | dimensionless |
| s∆r | Standard deviation of duplicate pairs distribution | dimensionless |
| σC | Total experimental uncertainty of POC estimates | mg/m3 |
| σV | Combined uncertainty of volume V | L |
| σCP(V ) | Total uncertainty explained by uncertainty in volume | mg/m3 |
| σM | Uncertainty of carbon mass predicted by the
calibration equation (Prediction interval) | µg |
| σMP | Uncertainty of carbon mass predicted by the
calibration equation (standard law of propagation) | µg |
| σCP(M) | Total uncertainty explained by uncertainty in
regression models | mg/m |
| σCP(η) | Total uncertainty explained by uncertainty in handling | mg/m3 |
At each cast, a duplicate sample from a randomly chosen depth was collected to assess the accuracy of our method. Sample water for each pair of duplicate measurements was taken from the same Niskin bottle. In total, we collected 392 uPOC samples with their corresponding aDOC blanks and 57 duplicate uPOC samples with their related aDOC blanks.
2.2 Laboratory sample handling
All the filters were processed in 16 separate batches or CHN runs. Each run consisted of (1) uPOC filters and their corresponding aDOC blanks from multiple casts, (2) duplicate uPOC and aDOC filters from these casts used to estimate total experimental uncertainties, (3) empty tin capsules, acidified and non-acidified filter blanks (detailed below) used to estimate uncertainties related to the sample handling in the lab and systematic biases, and (4) standards used to calibrate the CHN analyser (see Section 2.3 for details).
Acidification at room temperature for a period of 12 to 16 hours was used to remove the inorganic carbon accumulated on the uPOC and aDOC filters. To do this, a crucible containing a small amount of 37% HCl was located in the middle of a glass desiccator. The uPOC filters and aDOC blanks from each batch were removed from their aluminium envelopes, placed into individual acid-washed glass vials, and positioned around the crucible in the desiccators. In contrast to adding an aliquot of a dilute acid solution directly onto the filters, the technique of exposing them to acid fumes is expected to homogenize the effect of the acid on all the filters within a desiccator and to avoid losses of organic particles (Martin et al., 1993). Silica gel was used to reduce the humidity inside the desiccators and to minimise any potential contamination. To minimise differences in contamination between corresponding uPOC and aDOC samples, paired uPOC and aDOC filters were acidified in the same desiccator. Duplicate filters were acidified in a different desiccator.
To account for any potential contamination during acid fuming, we introduced an additional type of pre-combusted (450°C for five hours) 25 mm Whatman GF/F filter blank. Three of these filter blanks were placed into each desiccator, then acidified and processed along with uPOC filters and aDOC blanks, while three other filter blanks were kept clean and dry outside of the desiccators.
After the acidification phase, all the filters, including the acidified uPOC filters, aDOC blanks, and acidified and non-acidified filter blanks were dried in an oven at 60°C for several hours. Subsequently, the acidified and non-acidified filters and blanks were wrapped into individual tin capsules (Pressed, Standard Clean, 10 × 10 by OEA Labs) and analysed for carbon. An increment on the acidified filter blank masses in comparison with the non-acidified ones would indicate contamination during sample acidification and drying.
2.3 Determination of POC
The mass of carbon contained on filters was determined by high-temperature combustion (Gordon & Sutcliffe, 1974; Menzel & Vaccaro, 1964; Sharp, 1974; Wheeler et al., 1997) using a CHN analyser (FlashEATM 1112 Elemental Analyzer). Filters were processed in accordance with the manufacturer’s manual (ThermoQuest, 1999). In every CHN run, a new reaction tube was used. The extracted CO2 was measured by a thermal conductivity detector.
As samples are analysed, the reaction tube of FlashEATM CHN analyser gradually fills up with the combusted tin capsules and filters. Thus, as the CHN run proceeds and the reaction tube fills up, a variation in the instrument combustion efficiency can occur with time. To stabilise the combustion efficiency of the instrument throughout each run, we performed constant adjustments to sample delay, which represents the time that it takes the CHN analyser to combust each sample, to deliver CO2 released from the sample to the detector, and to run the analysis.
The CHN analyser was calibrated during each run using two sets of pre-weighted acetanilide standards (C = 71.09%, N = 10.36%) contained in tin capsules. The first set of 11 standards covered the entire range of expected masses of carbon on our uPOC and aDOC filters (5 – 300 µg) and it was analysed immediately prior to processing the sample filters. We will refer to this set of standards as the calibration standards. The second set of standards was processed alongside the filters (one standard after every six filters) to validate the initial calibration throughout filter processing. We will refer to the second set of standards as the stability standards.
In the absence of any instrumental drift, we expected the calibration coefficients derived from both sets of standards from the same CHN run to be statistically indistinguishable. However, during seven out of 16 runs the CHN analyser was unstable for unknown reasons and the calibration coefficients derived from the two sets of standards differed significantly. Thus, we decided to use for all CHN runs, both types of standards, calibration and stability, to develop the relationship between the response of the CHN analyser and the mass of carbon on processed filters.
The mass of carbon M on the ith filter, processed during the kth CHN run, was estimated by a linear regression model using a robust fitting method:
where x represents the output signal from the CHN analyser, m and b represent the slope and the intercept of the regression line, respectively. The intercept was removed from the model when it was not statistically significant (p-value > 0.05).
For each CHN run, we estimated the mass of organic carbon contained on uPOC (MuPOC), aDOC (MaDOC) filters, tin capsules (Mcap), and acidified (Mac) and non-acidified (Mnac) GF/F filters using Equation 1. These MuPOC and MaDOC values, however, do not represent the true load of organic carbon contained in the corresponding water samples as these values may be affected by biases, i.e., contamination during the acidification step, residual organic carbon on the tin capsules, and on the combusted GF/F filters). Therefore, the true mass of organic carbon from the ith uPOC and aDOC filters, which were acidified together in the jth desiccator and processed in the kth CHN run (MuPOCijk and MaDOCijk, respectively), must be written as:
where acjk is the average carbon mass of the three filter blanks acidified in the same desiccator jth as the ith filter and nack and capk are the average carbon mass of the three non-acidified filter blanks and the average carbon mass of three tin capsules, respectively.
For each pair of uPOC and aDOC filters, the mass of POC, M, was determined as
We assumed that uPOC and aDOC filters had adsorbed the same amount of DOC, and their contamination due to sample handling during the CHN analysis was equal to the average mass of the three acidified filter blanks. Hence, the subtraction in Equation 2 removed various systematic biases from the final estimates of the mass of POC. Note that because we processed corresponding pairs of uPOC and aDOC filters in the same desiccator and during the same CHN run, Equation 2 is equivalent to:
To determine the POC concentration, C, for each water sample, we divided M by the volume of water, V, filtered for each sample:
2.4 Uncertainty analysis
The standard law of propagation of uncertainty (JCGM, 2008) was used throughout our uncertainty calculations and we recall it here for the reader:
here is the total (or "combined") variance of the estimate y (POC, in our case), which is determined from the input quantities x1, x2, ..., xN through the functional relationship y = f (x1, x2, ..., xN). The uncertainty of each of the input variables is denoted as σxi and their inter-dependencies are represented by the correlation coefficients r(xi,xj). The total uncertainty is the positive square root of .
2.4.1 Total experimental uncertainty. We first estimated the total experimental uncertainties associated with our POC concentrations by analysing the measurements from the duplicate samples. These total experimental uncertainties are expected to represent the uncertainties arising from all (or at least most of) the steps required to estimate POC concentrations (i.e., from sample collection in the field to sample analysis in the laboratory). Below we show that these uncertainties depend on the POC concentration and therefore we present them as relative uncertainties. These relative uncertainties will then be used to estimate the uncertainty of any sample collected during the cruise. The experimental uncertainties were estimated from the scaled arithmetic differences between the POC concentrations of the duplicates, D1 and D2:
Since both D1 and D2 are uncertain and their uncertainties add in quadrature, to estimate the uncertainty in only one measurement, the difference of the duplicates were divided by (Hyslop & White, 2009).
We found that the duplicate differences ∆ were positively related to POC concentration and that this dependency was different between the productive and mesopelagic zones (Figure 2a). The correlation coefficient between ∆ and the average POC concentration of duplicates was 0.4 in the productive zone and 0.54 in the mesopelagic zone. To remove this dependency on concentration, we expressed the differences in duplicate measurements as a relative difference (∆r = ∆/), where is the average of POC concentration in the two duplicates. Figure 2b confirms that, once normalised, the duplicate differences in the productive and mesopelagic zones did not depend on POC anymore. These findings imply that estimates of POC concentration from a specific zone should have the same relative uncertainty, σr. The total experimental uncertainty was computed by analysing the ∆ values as follows. The uncertainty of each duplicate difference, σ∆, depends on the unknown uncertainties of the individual duplicate estimates, σd1 and σd2. By applying the law of propagation of uncertainty (Equation 5) to Equation 6 and by assuming no correlation between the duplicate estimates, we find:
Figure 2.
Scaled differences in duplicate particulate organic carbon (POC) measurements, expressed as absolute value (a) and as a function of the mean value of each pair of duplicates (). Scaled relative differences in duplicate POC measurements relative to , expressed in absolute value (b). The absolute values were used to more easily demonstrate the dependency of the duplicate differences on POC. Black and white points represent duplicate measurements from the productive and mesopelagic zones, respectively. Points with magenta borders represent samples processed during CHN runs with highly uncertain calibrations. The red dashed and blue dash-dotted lines are the linear fits to the data in the productive and mesopelagic zones, respectively.
here σdr is the relative uncertainty of the duplicate and we assumed that, as for the duplicate differences, also σdi (with i = 1, 2) needs to be expressed in relative terms (i.e., σdi = σdr Di).
By rearranging Equation 7, σdr can be expressed in terms of the uncertainty of the difference, σ∆:
Expressing σ∆ in relative terms yields:
σ∆r was then estimated as the robust standard deviation of the relative duplicates in each zone:
where P84 and P16 are the 84th and 16th percentiles of ∆r (Hyslop & White, 2009).
Finally, the total experimental uncertainty of POC concentration was estimated as:
where σr is the median of σdr, representing 14% and 35% uncertainties in POC concentrations for the productive and the mesopelagic zone, respectively.
2.5 Modeled uncertainty and uncertainty budget
In this section, we used a second method to model an independent estimate of the total uncertainty in POC concentrations. Specifically, the standard law of propagation of uncertainty was used to propagate the uncertainties associated with different steps of the POC determination. This independent model estimate of the POC uncertainty was then compared to the total experimental uncertainty derived from the duplicate measurements to assess the extent to which this latter theoretical calculation could reproduce the experimental uncertainties: the closer these estimates are, the more confident we can be in how we understand the measurement process and its uncertainties.
Equation 5 also allows one to estimate an uncertainty budget, that can be used to partition the total uncertainty into different contributions. Here, the relative contribution of each uncertainty source was computed as the ratio of the variance of each specific modeled uncertainty source divided by the total experimental uncertainty. The sum of the modeled relative variances is an estimate of how well the modeled and total experimental uncertainties agree. The different relative uncertainty contributions can be ranked to identify the most uncertain steps in the methodology and prioritise improvements in the method. Table 2 presents the list of the sources of uncertainty that we were able to quantify.
2.5.1 Uncertainty in volume. Errors in measuring the volume of sample seawater translate into uncertainties in POC concentration. Since each POC sample required from one to five bottles of sample seawater, each with a volume Vn, the combined uncertainty of the total volume, V, of seawater used for a sample depended on the number of bottles, n, used and the uncertainty in volumetric measurements, σVn, of each bottle. σVn was set equal to half of a graduation mark of the measuring cylinder (uncertainty of 10 ml). The volume of each bottle (∼2.2 l) was measured with a measuring cylinder multiple times. Thus, the combined uncertainty of volume V can be expressed as
Table 2. Sources of uncertainty that contribute to the total uncertainty of particulate organic carbon (POC) estimates and that we could quantify.
More detailed explanation of calculations are given in Section 2.5. The median values of each uncertainty is given for the productive zone (PZ) and for the mesopelagic zone (MZ).
| No | Source | Method of calculations | Values | Units |
|---|
| PZ | MZ |
|---|
| 1 | Volume | Volume uncertainty of each sample is equal to half graduation
mark of the measuring cylinder used to estimate the volume of
each bottle, multiplied by the number of bottles used during a
given filtration | 0.01 | 0.02 | L |
| 2 | Calibration | Prediction intervals (PI) of the calibration curves of each CHN run | 2.14 | 2.15 | µg |
| 3 | Sample handling
during CHN analysis | Quantify the uncertainty that could arise from acidification and
processing of sample and blank | 0.37 | 0.34 | µg |
| 4 | Total | Combined uncertainty calculated from POC duplicates | 2.29 | 2.60 | mg/m3 |
We estimated the uncertainty in POC concentration that could be explained by the uncertainty in volume, σC (V), by applying the propagation of uncertainty to Equation 4 and obtained:
2.5.2 Uncertainty in mass predicted by the calibration equation. The uncertainty of carbon mass Mik predicted using the calibration equation (Equation 1) was expected to be one of the largest contributors to the uncertainty budget because it determined the quality and reliability of the prediction of the masses of carbon in all components of the analysis. To estimate the uncertainties of the carbon masses predicted by our calibration equations, σM, we used 68% prediction intervals (PIs) as they serve in a normal distribution as one standard deviation. A PI is defined similarly to the better known confidence interval. However, the PI is more appropriate for quantifying the uncertainty of the calibration process because it represents the expected uncertainty of an individual future observation taking into account the uncertainties arising from all the regression parameters into the total uncertainty (Rawlings et al., 1998).
For each CHN run, we estimated σM as stated in Altman (2000).
where σres represents the robust standard deviation of the residuals of carbon mass about the regression line, xik is the instrument response (Equation 1), and sx are the mean and the standard deviation of xik, nS is the number of standards used to fit the model, and t1−α/2, is the value from the t distribution with nS − 2 degrees of freedom and α equal to 0.32 corresponding to a 68% prediction interval. The uncertainty of the regression model and the estimated masses depended on both the accuracy of the weight of the standards and the sensitivity of the instrument to the carbon at the time when filters were processed. Therefore, σres and σM varied among CHN runs. Two CHN runs had significantly higher σres and therefore higher uncertainties of the regression models and mass estimates, compared to the rest of the runs. We believe that the cause of the poorer performance of the regression analysis in these instances was poor weighting of the standards, and not an instability or low sensitivity of the CHN analyser. The differences of the duplicate POC estimates derived from these regression curves, on average, were not larger than those derived from regression curves with lower uncertainties (Figure 2). This means that there was no bias in the "poorly-made" standards - they resulted in accurate calibration coefficients, even though the random uncertainties around these coefficients were high. Nevertheless, to avoid skewing our results, we did not include data derived from these two CHN runs when we determined our total experimental uncertainty (Equation 11).
The uncertainties of carbon masses estimated for uPOC and aDOC using the regression models were propagated into the uncertainties of POC masses by applying Equation 5 to Equation 2:
where r (MuPOC, MaDOC) is the correlation coefficient between MuPOC and MaDOC.
Then, for each estimate of POC concentration, we calculated the contribution of this source of uncertainty to the total uncertainty of POC concentration, σC (M ):
2.5.3 Uncertainty due to sample handling during CHN analysis. The stages of acidification and drying of the filters may contaminate the samples, thus adding positive biases to the estimates of POC concentration and a higher variability between the duplicates (King et al., 1998).
Considering that the uPOC samples, their corresponding aDOC blanks, and the filters blanks were acidified in the same desiccator and dried together in the oven, the amount of contamination that they received should have been approximately equal. Thus, the effect of the possible contamination should have been removed from the estimates of POC during the subtraction of MaDOC(true) from MuPOC(true) in Equation 2. However, there was a degree of random uncertainty associated with the estimation of contamination as the average mass of carbon of the three acidified filter blanks for each desiccator, varied slightly.
The contribution of the uncertainty due to the handling of the filters to the total uncertainty, σC (η), can be calculated by applying the standard law of propagation of uncertainty to Equation 4:
where ση1 and ση2 represents the uncertainties of contamination due to handling of uPOC and aDOC filters, respectively. These uncertainties were estimated as the standard error of the mean of the three corresponding estimates of acidified filter blanks,
3 Results and discussion
3.1 Distribution of POC
The POC concentrations from the AMT-24 cruise were highly variable, ranging between 2 and 76 mg/m3 (Figure 3). The productive zone was characterised by relatively high POC concentrations, while low POC concentrations defined the mesopelagic zone. The overall median POC concentration from the productive zone was 18(±14) mg/m3, whereas that from the mesopelagic zone was 7(±3) mg/m3. We also observed latitudinal patterns of POC driven by the seasonality and differences between the biogeographical regimes sampled during the cruise (Figure 1 and Figure 3). In the productive zone, the highest POC concentrations were found at temperate latitudes and around the equator, whereas the most oligotrophic provinces were characterised by lower POC concentrations (Table 3). Such distribution in the POC concentration matched the typical latitudinal patterns encounter for the upper layer of the Atlantic ocean (Poulton et al., 2006; Rasse et al., 2017; Wangersky, 1976). Particularly, high POC concentrations were found in the sub-surface (50 m) of the South Subtropical Convergence Zone. In this province, Poulton et al. (2006) found relatively high POC concentrations deeper than 200 m as a result of sinking particles from the euphotic zone of the SSTC. Overall, POC concentration below the productive zone was less variable throughout the transect, but the latitudinal pattern remained (see Table 3). Even though our POC concentrations are comparable to those published in the literature, a direct comparison is complicated due to differences in methodologies, sampling times, and regions. The range of POC concentration estimated from bottle samples in the Atlantic ocean by numerous researchers spans from 5 to 350 mg/m3 (Banoub & Williams, 1972; Balch et al., 2010; Cetinić et al., 2012; Gardner et al., 2006; Gardner et al., 2003; Gardner et al., 1993; Graff et al., 2015; Marra et al., 1995; Menzel & Goering, 1966; Mishonov et al., 2003; Poulton et al., 2006; Stramska & Stramski, 2005; Stramski et al., 2008; Wangersky, 1974; Wangersky, 1976). A limited number of studies present POC concentrations measured in the Atlantic ocean between 100 and 500 m and spanning from ~0 to 35 mg/m3 (Carlson et al., 2000; Cetinić et al., 2012; Menzel, 1967; Poulton et al., 2006; Wangersky, 1974; Wangersky, 1976).
Figure 3. Depth-resolved distribution of particulate organic carbon (POC) concentration along the AMT-24 cruise.
Borders of the sampled biogeographical provinces are marked by blue vertical lines.
Table 3. Particulate organic carbon distribution presented for productive and mesopelagic zones across the sampled biogeographical provinces: the NorthAtlantic Drift Province (NADR), the North Atlantic Subtropical Gyral Province (NAST), the North Atlantic TropicalGyral Province (NATL), the Western Tropical Atlantic Province (WTRA), the South Atlantic Gyral Province (SATL), and the South Subtropical Convergence Province (SSTC).
"Std" are robust standard deviations.
| No | Province | Productive zone | Mesopelagic zone |
|---|
| Samples | Median | Std | Samples | Median | Std |
|---|
| 1 | NADR | 25 | 33 | 31 | 16 | 10 | 9 |
| 2 | NAST | 36 | 17 | 7 | 33 | 7 | 3 |
| 3 | NATL | 36 | 15 | 8 | 18 | 8 | 2 |
| 4 | WTRA | 32 | 19 | 10 | 16 | 9 | 6 |
| 5 | SATL | 57 | 14 | 5 | 44 | 7 | 2 |
| 6 | SSTC | 39 | 46 | 16 | 39 | 9 | 3 |
| 7 | ALL | 225 | 17 | 14 | 167 | 7 | 4 |
3.2 Detection limits
To determine the limit of detection of our technique we used the approach recommended by The International Union of Pure and Applied Chemistry (IUPAC) and The International Organization for Standardization (ISO) (Analytical Methods Committee AMCTB No. 92, 2020). We calculated the detection limit, (LD ), by first computing the critical value, (LC, Equation 18), which establishes the presence of the analyte (carbon in our case), and is defined as the minimum significant estimated value of an analytical result, which is used as to discriminate against background noise (Currie, 1995):
where and s0 are the mean and the standard deviation of a blank material free from carbon, in our case, the tin capsules. t0.95;df is the one-tailed 95% quantile for Student’s t with degrees of freedom d f , according to the number of values used to estimate and s0.
Given LC, we estimated LD , which is defined as the lowest carbon mass that our analytical method is reliably capable of detecting 95% of the time:
The median of the carbon mass from all the tin capsules used in the 16 CHN runs was 1.7(±1.3) µg. Thus, the estimated LC and LD were 3.2 and 4.9 µg C, respectively (Figure 4). The vast majority of mesopelagic aDOC filters collected and analysed during this study were above LD, with just four filters fell between the estimated LD.
Figure 4. Distribution of carbon mass determined on tin capsules, non-acidified and acidified blank filters (GF/F) used for sample processing, and adsorbed dissolved organic carbon (aDOC) and uncorrected particulate organic carbon (uPOC) filters.
In each element of the box plot, the central rectangle spans from the first quartile (25th percentile) to the third quartile (75th percentile). The green line inside each rectangle shows the median value and whiskers below and above the box show the locations of the 5th and the 95th percentile, respectively. Red circles represent outliers. Red and blank horizontal lines represent the critical value (LC) and detection limit (LD), respectively.
3.3 Correction for biases
We corrected additional sources of bias in the estimates of POC concentration by subtracting aDOC blank measurements from the corresponding uPOC measurements (see Section 2.3 and Equation 2). Since the carbon mass determined on an aDOC blank includes the mass of the adsorbed DOC, the carbon masses detected on empty tin capsules, clean GF/F filters and any contamination occurring during filter acidification and handling, it represents the cumulative bias for which the carbon mass on a corresponding uPOC filter needs to be corrected. Hence, to minimise biases introduced by any potential contamination and mass predicted by the calibration equation, we processed pairs of uPOC and aDOC filters together during sampling, acidification, handling, and processing stages, i.e., acidified in the same desiccator and analysed during the same CHN run.
The magnitude of the masses on these additional components in comparison with the magnitude of the masses of aDOC and uPOC filters are presented in Figure 4. The median of the carbon mass from all the tin capsules used in the 16 CHN runs was 1.7(±1.3) μg, whilst the medians of the mass corresponding to non-acidified and acidified filter blanks were 3.2(±0.8) μg and 4.1(±1.1) μg, respectively, indicating that our method minimised contamination during acidification. On average, the acidification and handling of the filters resulted in contamination of 1(±1) μg. Nevertheless, when comparing corresponding sets of acidified and non-acidified filter blanks, the carbon masses from the acidified filters could be up to twice as large the amount of those from the non-acidified filters. Finally, the cumulative effect of all the biases and potential contamination that composed our aDOC blanks were 12(±4) μg in the productive zone and 9(±3) μg in the mesopelagic zone.
The range in carbon masses of our aDOC is comparable with those from the Atlantic ocean reported by Cetinić et al. (2012): their average mass of DOC adsorption was 10.9 μg with 95% of the masses of their aDOC blanks within the range of 8.5 to 40.5 μg. Also, our aDOC values are within the findings of Abdel-Moati (1990), who reported varying amounts of DOC adsorption between 9.2 and 15.0 μg and 3.5 and 6.5 μg for eutrophic and oligotrophic waters, respectively.
The aDOC concentrations (carbon mass on aDOC filters normalized as a function of volume and corrected for biases) in the productive zone had a median value of 2.2(±1.3) mg/m3, and 1.0(±0.5) mg/m3 in the mesopelagic zone. The distribution of these concentrations was spatially correlated with POC concentrations (see Figure 5, r = 0.7). This correlation was particularly strong in regions characterized by high surface concentrations of POC: NADR (r = 0.8) and SSTC (r = 0.9).
Figure 5. Depth-resolved distribution of adsorbed dissolved organic carbon (aDOC) estimates (mass of carbon on aDOC filters normalized as a function of volume and corrected for biases) along the AMT-24 cruise.
Borders of the sampled biogeographical provinces are marked by blue vertical lines.
Since the filtered volumes were adjusted according to the expected POC concentration, and all GF/F filters were from a single manufacturer and treated identically during the cruise, the aDOC concentration should be relatively constant across the samples and not correlated with POC concentration. However, POC could be present on the aDOC filters, increasing the aDOC concentration due to fragmentation of particles through uPOC filters (Bishop & Edmond, 1976). This fragmentation of larger particles from the uPOC filters into smaller particles has been observed (Wangersky, 1974) to contaminate aDOC filters and, perhaps, causing the higher aDOC values found in our study. For instance, Banoub & Williams (1972) filtered multiple samples of seawater through four stacked GF/C filters and found particles as evidence of contamination on the second filter from top to bottom. Abdel-Moati (1990) carried out a similar experiment with similar conclusions, suggesting the sum of masses from the first two filters should be used as a POC mass, while using the third filter, from top to bottom, as a true blank for aDOC. Cetinić et al. (2012) also pointed to the high variability of their aDOC values and potential contamination from the overlying uPOC filters.
If a higher than usual aDOC value is due to the contamination of an aDOC filter with particles filtered through the uPOC filter, particle loss from the uPOC filters should be higher in productive areas, thus explaining the observed stronger positive correlation between POC and aDOC concentrations (Zhou et al., 2016). Assuming that particles contaminated the aDOC filters, we expect that the typical mass of organic carbon adsorbed onto our GF/F filters should be better represented by the aDOC measurements at depth, where particles are less abundant. Thus, we estimated the loss of particles from uPOC to aDOC filters by subtracting from the true carbon mass of all aDOC blanks (corrected for biases), the median true carbon mass of aDOC filters collected in the mesopelagic zone (≥200 m). Then, we added this difference to our POC masses and found that POC concentrations increased by 3(±4)% in the productive zone and by 0(±5)% in the mesopelagic zone.
We cannot prove which mechanism determined the correlation between aDOC and uPOC concentrations. Nonetheless, the carbon mass ratio MaDOC(true) / MuPOC(true), corrected for biases, ranged from 3.5 to 58.2% with medians of 9.4(±2.2)% and 12.0(±3.4)% in the productive zone and the mesopelagic zone, respectively, indicating that DOC adsorption was important in the open ocean.
An additional reason that aDOC and POC concentrations showed a positive correlation pattern could be because DOC and POC have a similar decreasing pattern as a function of depth in the open ocean (Dai et al., 2009).
3.4 Relative and total experimental uncertainties
The relative experimental uncertainty of POC concentrations, , (Equation 9), was on average ~14% and ~35% in the productive zone and the mesopelagic zone, respectively. Higher were estimated for the mesopelagic zone where POC concentrations were lower and biases might have had a greater effect on the estimates. The resulting total experimental uncertainties of the estimates of POC concentration (σC) are shown in Figure 6. The median of σC for the productive and mesopelagic zones were 2.5(±1.2) mg/m3 and 2.6(±0.6) mg/m3, respectively.
Figure 6. Depth-resolved distribution of the total uncertainty associated with the estimates of particulate organic carbon (POC) concentration derived during the AMT-24 cruise.
Borders of the sampled biogeographical provinces are marked by blue vertical lines.
3.5 Uncertainty budget
The modeled uncertainty for POC is presented in Table 4 for each component, and compared to the estimated total experimental uncertainty of POC derived from the duplicate measurements. The uncertainty in volumetric measurements σC(V) typically contributed about 1% to the total uncertainty of POC concentrations in the mesopelagic zone. However, since the volume of water filtered in the productive zone was smaller, the contribution of volumetric uncertainties was greater (~2%). Overall, the contribution of this source of uncertainty was insignificant, except when POC concentration was particularly high and, as a consequence, the volume of the water sample was ~2 L.
Table 4. Uncertainty budget presenting the contributions of each source of uncertainty that we could quantify, σC(xi), relative to the total experimental uncertainty of particulate organic carbon, σC.
Median values are given for the productive and mesopelagic zones.
| No | Source | σC (xi) /σC |
|---|
Productive
zone | Mesopelagic
zone |
|---|
| 1 | Volume | 0.02 | 0.01 |
| 2 | Calibration | 0.17 | 0.12 |
| 3 | Laboratory sample handling | 0.03 | 0.02 |
| 4 | Unquantified | 0.78 | 0.85 |
The uncertainty due to the calibration equation σC(MP ), after excluding the unstable CHN runs, explained a median of 16(±20)% of the total experimental uncertainty of POC concentrations (Figure 7). Since unstable CHN runs were characterised by greater residual errors in the regression analysis, their contribution to the total experimental uncertainty was significantly higher with a median of 65(±36)%.
Figure 7. Portion of the total uncertainty of particulate organic carbon concentration that can be explained by the uncertainty of regression analysis.
Borders of the sampled biogeographical provinces are marked by blue vertical lines.
Our acidification method and handling during the CHN analyses allowed us to minimise the effect of contamination of POC estimates. The median contribution of this source of uncertainty, σCP(η), in productive waters was 3(±3)%, while in the mesopelagic zone attained 2(±2)%.
Overall, the three sources of uncertainty described above explained only 21(±22)% of the total experimental uncertainty in POC. Thus, other sources of uncertainty must be responsible for the relatively large and unexplained part of the estimated experimental uncertainty (Figure 8).
Figure 8. Portion of the total uncertainty of particulate organic carbon concentration that can be explained by the uncertainty budget.
Borders of the sampled biogeographical provinces are marked by blue vertical lines.
3.6 Missing sources of uncertainty
In this section we discuss potential sources of uncertainty that could explain the missing part of the uncertainty budget.
3.6.1 Rare particles. Patchiness is ubiquitous in the ocean, ranging from microscale thin layers in the water column, to sub-mesoscale and mesoscale fronts, and could have contributed to our total experimental uncertainties in POC. For instance, Bochdansky et al. (2016) analysed the dynamics and abundances of particles at depth using a custom-made digital inline holographic microscope. They found that the concentration of patchy marine snow (large particles >500 µm) was 100 times higher than expected in comparison with the concentration of smaller particles.
Additionally, Ohman et al. (2012) analysed the concentration and vertical distribution of suspended particulate matter and mesozooplankton at a deep-water front in the California Current System using a high resolution digital camera system. They observed that the front had a different composition of particulate matter, and it was a zone of higher marine snow particles where the volume of all size fractions of suspended particulate matter, especially organic aggregates, increased several times in comparison with the surrounding seawater.
Therefore, despite mixing samples in carboys before dispensing them to the filtration bottles, some of our duplicate filters might have captured different types of particles, i.e. rare large aggregates or (invisible) zooplankton might have appeared only on one of the two duplicate filters.
Furthermore, Wangersky (1974); Wangersky (1976) found that a water mass in the open ocean had a homogeneous background of POC concentration, upon which occasional small patches with up to five times POC concentration were superimposed. He reported uncertainty of a single POC estimate derived from replicate 5-litre samples to be equal to 2.5(±1) mg/m3. Our median total uncertainty for the mesopelagic region of 3(±1) mg/m3 is comparable to these findings, suggesting that samples might have been drawn from water masses of similarly spatially variable POC concentrations.
Further experiments are required to better understand the natural heterogeneity and small-scale patchiness of seawater using Niskin bottles. Since the collection of intact and flocculent marine snow particles using bottle samplers is highly challenging and time-consuming (Bochdansky et al., 2016), the oceanographic community would benefit from analysing the natural heterogeneity of suspended particles and the effect of patchiness on estimates of POC concentration and their uncertainties using optical observations from instruments such as the Underwater Vision Profiler or holographic cameras which have the advantage of displaying details of the in-situ vertical distribution of particles across a range of sizes (Bochdansky et al., 2016; Ohman et al., 2012).
3.6.2 Contamination during filtration. The filtration system employed during this study was an open-funnel filtration set which might have increased the risk of contamination (IOCCG POC Sampling and Measurement Protocols). Samples might have been contaminated with carbon-containing particles from the ambient air, i.e., dust, clothing fibres and ashes or soot associated with residual oil burning in the shipping emissions. We consider that mainly uPOC filters would have been contaminated in such a way because they were exposed to the laboratory atmosphere for longer than the underlying aDOC filters. Assuming that the uPOC filters were contaminated during filtration, their duplicate filters might have received a similar amount of contamination. Thus, atmospheric contamination of uPOC filters would result in greater differences between uPOC duplicates compared to the differences between less contaminated aDOC duplicates. In fact, the median of the absolute differences between duplicate uPOC concentrations was 2(±2) mg/m3, which is higher than the median of the absolute difference between duplicate aDOC concentrations of 0.2(±0.4) mg/m3. This result is also consistent with our previous finding that the differences between duplicate concentrations depend on POC concentrations (see Figure 2). As a consequence, we do not have enough grounds to state that differences between duplicate uPOC concentrations are higher than the differences between duplicate aDOC concentrations because of contamination of uPOC filters.
The longer a filter is exposed to the laboratory atmosphere, the more contamination it should receive. The duration of filter exposure to the laboratory atmosphere depends on the volume of seawater filtered through the filter. To further investigate contamination of uPOC filters, we analysed how differences between duplicate uPOC concentrations depended on differences between volumes of seawater filtered through two duplicate filters and found no significant correlation.
Thus, there was no evidence to suggest that the time of exposure of filters to the atmosphere was the cause of disagreement between duplicate estimates. However, analysing duplicates to determine this source of contamination might not be sufficient as duplicate estimates could have been affected by other biases or contaminants that mask a single source of contamination. Consequently, with the available data, we cannot quantify this source of contamination and its uncertainty.
For future experiments, especially when using an open-funnel setup, filtering Milli-Q water through additional blank GF/F filters may be used to quantify contamination from the laboratory atmosphere. Even better, to minimise contamination from airborne particles, it would be advisable to filter samples under a laminar flow hood or by employing a closed filtration system (Cetinić et al., 2012).
3.6.3 Storage of samples. Freezing and storing of sample filters might have also introduced some contamination. Published values suggest that the average mass of unused GF/F filters from a cruise may range from 3(±10) µg to 10(±5) µg (Cetinić et al., 2012; Menzel, 1966; Stramski et al., 2008; Wangersky, 1974). The difference between our non-acidified (3±1 µg) filter blanks and these published values might indicate that the contamination during filter storage could amount to between 0 and 7 µg. However, assuming homogeneous contamination of uPOC and aDOC filters, we would expect that this contamination would be accounted for when aDOC is subtracted from uPOC (Equation 2). To quantify uncertainties due to filter storage, we recommend preserving multiple unused GF/F filter blanks along with the samples for post-cruise analysis.
3.6.4 Collection of samples. Uncertainties can also be introduced by different operators. In our case, samples were collected by two operators, whereby one operator systematically collected samples from pre-dawn casts, while the other from noon casts. Thus, we thought that analysing duplicates pairs collected during pre-dawn and noon casts separately might give us an insight into this source of uncertainty. Due to constraints in the water budget, pre-dawn duplicates were collected from quasi-random depths, while 25 out of 28 duplicate pairs collected during noon time represented deep waters (≥400 m). For pre-dawn duplicates, the median POC concentration and the median absolute differences of duplicate POC concentrations were 9(±9) and 1(±2)mg/m3, respectively, while for noon duplicates were 9(±5) and 2(±2) mg/m3, respectively, suggesting that there was no statistical difference between these medians in the two groups of duplicates. If we take into consideration that the majority of noon duplicates were collected from deep waters (≥400 m), pre-dawn duplicates seem to be slightly more precise than noon duplicates with medians of the relative differences of duplicates collected from deep waters of 19(±10)% and 19(±31)% for pre-dawn and noon duplicates, respectively. However, there is no evidence that samples collected by different operators are biased by the operators themselves rather than by varying composition of particles at different depths. Even though the higher uncertainties that we found in the mesopelagic zone might be partially explained by the varying precision of duplicates collected by the two operators from pre-dawn and noon casts, we cannot quantify the bias and the uncertainty introduced by each operator.
3.6.5 Uncertainty model. The uncertainty model we employed in this study (Equation 11) was based on the empirical relationship we observed between duplicate differences and POC concentrations (Figure 2). Admittedly, this model is likely an approximation of the experimental uncertainty of our POC measurements. To improve the model (e.g., by adding a constant offset to it), we would need to better understand the role that each source of uncertainty plays in the total uncertainty of POC concentrations, to reconcile experimental and theoretical estimates of the POC uncertainty, which should ultimately allow us to then understand the extent to which each source of uncertainty is either a multiplicative or an additive term to the total uncertainty as POC, as the POC concentration in the ocean varies. Further dedicated experiments and analyses would be needed to achieve this deeper understanding.
3.7 aDOC blanks
Even though the double-filter technique employed in this study significantly increases filtration times, this procedure allowed us to collect a sample blank, i.e., aDOC filter, for each uPOC filter. Therefore, it is interesting to investigate how the uncertainty in the final POC concentrations would vary if fewer or no aDOC blanks were collected.
Some researchers avoid collecting aDOC blanks, under the assumption that by maximising the filtered volumes of seawater, uncertainties related to aDOC could be minimised (Stramski et al., 2008). Since we collected multiple types of blanks, i.e., aDOC, non-acidified, and acidified, we can test this assumption under different scenarios.
First, by comparing uPOC concentrations to our POC concentrations, the former are higher than the latter by about 13(±7)% in the productive zone and 19(±11)% in the mesopelagic zone.
Second, by subtracting the median carbon mass of non-acidified filters, i.e., clean GF/F filter blank, from the uPOC carbon mass, we obtained uPOC concentrations (corrected from clean GF/F filter blank) that are greater than our original POC concentrations by 9(±6)% in the productive zone and 12(±9)% in the mesopelagic zone. Finally, by subtracting the median carbon mass of the acidified filters from the uPOC carbon mass, the uPOC concentrations (corrected from acidified GF/F filter blank) are larger than our original POC concentrations by 8(±6)% in the productive zone and 10(±8)% in the mesopelagic zone. Thus, by not correcting for aDOC blanks, we would have introduced positive biases in POC concentrations of the order of 10–20%, even if we filtered up to 8 liters of water.
To reduce filtration time, one could collect fewer aDOC blanks. To quantify the potential uncertainty introduced by this method, we corrected our uPOC estimates using a single value of aDOC blank, which was determined from the median of the aDOC blanks from deep (≥200 m) stations. The resulting POC concentrations were 3(±4)% and 0(±6)% higher than the original concentrations in the productive and the mesopelagic zones, respectively. Thus, by using fewer aDOC blanks, one could significantly reduce the bias generated when not using an aDOC blank.
Overall, the above exercises could be used as quantitative guidance regarding how many (if at all) aDOC blanks to collect, based on the level of uncertainty that one is willing to accept.
4 Conclusions
In this study, we modified the standard protocol used to determine POC concentration (Knap et al., 1996; Martin et al., 1993) by: sampling different volumes of water in accordance to the expected in-situ POC concentrations; adopting different types of blanks, i.e., aDOC filter blanks and acidified and non acidified GF/F filter blanks to quantify the dissolved organic carbon and any contamination (biases); employing silica gel in the desiccators during the acidification step to reduce the humidity and minimise any potential contamination; and collecting and analysing duplicate samples. This methodology enabled us to quantify experimental uncertainties of our POC concentrations. Moreover, the various filters used as blanks allowed us to discuss several sources of uncertainty that contributed to the discrepancies between the duplicate estimates.
We found that the total experimental uncertainty of our POC estimates varied with depth and with POC concentration and was ∼14% and ∼35% in the productive and in the mesopelagic zones, respectively. However, the quantification of all the different sources of uncertainty associated with POC concentrations was impossible with our measurements, and we could explain only ∼21% of the total experimental POC uncertainty we observed. Further work is required to identify the unexplained portion of the uncertainty.
Nevertheless, we moved one step forward in understanding uncertainties of bottle-derived POC estimates. The analysis of duplicates and of the various blanks presented in this paper improved our understanding of the limitations of the methods and of some of the stages of sample collection and processing that contributed to the variability of our results. Further experiments would be required to fully understand the uncertainty budget of POC estimates. This understanding would allow us to concentrate our effort on those parts of the methodology that are more prone to produce errors and to develop a better-informed protocol to foster comparability of POC estimates across different studies.
Data availability
Underlying data
Published Data Library (PDL), British Oceanographic Data Centre (BODC): AMT24 (JR20140922/JR303) Particulate organic carbon (POC) measurements from CTD bottles. http://doi.org/10/fzw5 (Dall’Olmo et al., 2021).
This project contains the following underlying data:
README.txt. (contains important information that is commonly required to understand the following files or spreadsheets deposited in CSV format.)
Standards.csv. (Contains information regarding standards used to calibrate the CHN analyser, see README.TXT)
Capsules.csv. (Contains information regarding empty tin capsules used to estimate uncertainties related to the sample handling in the lab, see README.TXT)
NonAcidifiedFilters.csv. (Contains information regarding non-acidified filter blanks used to estimate uncertainties related to the sample handling in the lab, see README.TXT)
AcidifiedFilters.csv. (Contains information regarding acidified filter blanks used to estimate uncertainties related to the sample handling in the lab, see README.TXT)
aDOCFilters.csv. (Contains information regarding aDOC filters used to determine POC concentrations, see README.TXT)
uPOCFilters.csv. (Contains information regarding uPOC filters used to determine POC concentrations, see README.TXT)
DuplicateaDOC.csv. (Contains information regarding duplicate aDOC filters used to estimate total experimental uncertainties, see README.TXT)
DuplicateuPOC.csv. (Contains information regarding duplicate uPOC filters used to estimate total experimental uncertainties, see README.TXT)
POC.csv. (Contains overall information regarding POC concentrations, including nominal depth, amount of seawater filtered for each filter, geographic coordinates, date and time of collection, POC mass and POC concentration of each sample, see README.TXT)
This dataset is available under the terms of the UK Open Government Licence version 1.0 for public sector information. This licence governs access to and use of Open Data supplied by the Natural Environment Research Council (NERC).
