The identification of phytoplankton pigments has a wide range of applications as chemotaxonomic markers to identify community composition (Jeffrey et al., 1997; Mackey, et al., 1996; Wright and Jeffrey, 2005; Roy et al., 2011), species and their biomass in seawater (Vidussi et al., 2001; Uitz et al., 2006; Hirata et al., 2011) and plays a role in the development, refinement, and validation of satellite algorithms for mapping phytoplankton biomass, functional types, and size classes (e.g., Brewin et al., 2010; Mouw et al., 2017).

Ocean color data validation, satellite bio-optical algorithm development, and ocean productivity models within the framework of marine ecosystem studies, also require high-quality in situ measurements of chlorophyll a (TChl a) and primary pigment concentrations. Validation of satellite products of TChl a is particularly relevant as it is listed as an essential climate variable (GCOS, 2011). The phytoplankton pigments of algae can be routinely quantified through liquid chromatographic screening methods to separate known pigments quantitatively. More than 30 high-performance liquid chromatography (HPLC) methods are reported in the literature (Roy et al., 2011) as commonly implemented for phytoplankton pigment determination. The HPLC methods primarily utilized for remote sensing validation activities are based on C8 and C18 chromatographic columns. In the JGOFS Protocols for the Joint Global Ocean Flux Study Core Measurements (1994), the Wright et al. (1991) method was recommended as the HPLC method for determining TChl a. In the oceanographic community, the Zapata et al. (2000) and Van Heukelem and Thomas (2001) methods have gained widespread adoption over the past decade. These two methods offer the advantage of enabling the simultaneous analysis of multiple pigments and unknowns in a single run (Wright and Jeffrey, 2005), while also exhibiting superior resolution for chlorophylls c compared to Wright’s approach (1991).

Moreover, to ensure a continuous high standard in sample analysis, regular participation in comparison exercises involving international laboratories or comparison with a certified reference laboratory are required. The so-called round-robin exercises have been conducted in the past decades to assess the performance of HPLC methods applied for pigment determination, and uncertainties of 15% have been documented for TChl a and other pigments (Latasa et al., 1996; Claustre et al., 2004). It is worth highlighting that the majority of studies tend to focus on TChl a analysis, while comparatively fewer investigations are dedicated to other pigments. This disparity is noteworthy as it limits our understanding of the broader spectrum of pigments present in various samples. Natural samples, such as those obtained from different marine environments, exhibit a diverse range of pigments with unique functional roles. Understanding the full pigment profile of natural samples requires comprehensive HPLC comparisons that encompass not only TChl a but also other pigments such as chlorophyll b, carotenoids, and phycobilins. The SeaWiFS HPLC Analysis Round-Robin Experiment (SH) and the HPLC/DAD Intercomparison on Phytoplankton Pigments (HIP) exercises have compared laboratories on several pigments, relying on both pigment standard and natural samples in their statistics. Such exercises have the great merit of allowing the comparison of methods and performance among the participating institutes, typically 4 to 11 (Table 1) on different water types. In SH and HIP exercises (with the exception of HIP-3), a quality assurance (QA) subset of laboratories was defined that represented the state-of-the-art results based on agreed parameters of performance metrics (Hooker et al., 2005). The TChl a average uncertainties for the QA subset were invariant to water type and lower than 8%, while a result lower than 25% for primary pigments was reached in all the exercise with the exception of SH-4, which was focused on eutrophic waters (Hooker et al., 2010). Although the level of agreement requested for remote sensing applications is demonstrated to be achievable, some participants not part of the QA have shown differences much higher than 15% for the TChl a, and than 25% on average on primary pigments (Hooker et al., 2005; Hooker et al., 2009; Hooker et al., 2010; Hooker et al., 2012; Canuti et al., 2016; Canuti et al., 2022). A known limit of the comparison of marine pigments is the lack of certified reference materials (CRMs); instead, pigment standards certified by analysis from a select few producers (e.g., DHI, DK, and carotenature, CH) were commonly used.

The current study first aims to investigate the uncertainties associated with phytoplankton pigment quantification by comparing the analyses performed on duplicate samples by a certified laboratory and an accredited laboratory applying the same method (Van Heukelem and Thomas, 2001): the Joint Research Centre of the European Commission (J) and the DHI Group, Denmark (D). The comparison with one accredited laboratory allows for a focused examination of the same method applied to various water types and is significant as both laboratories have shown good performance in previous intercomparison exercises (e.g., Canuti et al., 2016). The chromatographic method selected here has been successfully applied to a wide range of pigment concentrations, including oligotrophic (Hooker et al., 2009) and eutrophic coastal waters (Hooker et al., 2005). However, the most relevant aspect of the study is the data set available for analysis, made up of batches of natural samples collected at 957 measurement stations during ship-based oceanographic campaigns and analyzed by both laboratories. These data are representative of different trophic conditions and water types of the European Seas, with TChl a concentration in the range of 0.083–27.35 mg/m³, allowing an extensive comparison to be conducted. The phytoplankton pigments considered for the present exercise are the chlorophylls and the carotenoids most commonly used in marine chemotaxonomic and photo-physiological studies, including the major marker pigments used for classification of phytoplankton groups in ecological studies (Roy et al., 2011). Phytoplankton pigment datasets are commonly used to characterize algal species in terms of phytoplankton functional types (PFTs) or phytoplankton size classes (PSCs). Related investigations have been pursued in the context of ocean color remote sensing, as this would allow a synoptic description of PFTs or PSCs (e.g., IOCCG, 2014; Mouw et al., 2017; Xi et al., 2020). The two most widely used PFT methods are CHEMTAX (Mackey et al., 1996) and diagnostic pigment (DP) analyses. The a priori assumption underlying the chemotaxonomy methods is the covariance of pigments. The inherent uncertainties of chemotaxonomy methods are due to the widespread occurrence of most pigments across many different taxa (IOCCG, 2014). Another known limit is the association of a single pigment with a taxon even though it could be, in some case, not representative of the entire group, e.g., some Dinoflagellates spp. have types of “non-canonical” plastids instead of the Peridinin (Matsumoto et al., 2011), and there are known cases (Naik et al., 2011) where the chemotaxonomy alone leads to incomplete interpretation of the taxa distribution. Despite these considerations, most of the validation of remote sensing methods for assessing phytoplankton community composition is based on chemotaxonomy derived from HPLC pigments. Tracing uncertainties and building uncertainty budgets for descriptors of phytoplankton populations are tasks in progress (Bracher et al., 2017), where uncertainties associated with field data of diagnostic pigments are needed (Xi et al., 2021). The present study obviously contributes to that effort. A second contribution is to explore how uncertainties in diagnostic pigment values propagate to phytoplankton indexes that are at the basis of pigment-based algorithms deriving descriptors of phytoplankton populations such as PFTs (Hirata et al., 2011), therefore providing a benchmark value for that step.

The determination coefficients between TAcc and TChl a for the two laboratories were calculated for each of the ship-based campaigns and for all the AAOT natural samples, including those at different depths (Figure 2). The co-variations were overall good (r² > 0.95). The J and D laboratories were then compared through scatter plots with a specific focus on the PPig. The determination coefficient, r², the slope of linear regression (OLS), and the slope for the LAD regression were calculated for all log-transformed PPig (Table 6; Figure 3). In their studies, Bland and Altman (1999) established that the limit of agreement could be determined when the differences between the data from the two laboratories are normally distributed and the standard deviation and the mean are the same across the entire range of measurements. The normal distribution of the difference in our case was verified by observing the histogram plot and by using D’Agostino–Pearson test. If differences are not normally distributed, as in our case, Altman and Bland suggest that a logarithmic transformation of the original data can be tried. The comparison through logarithmic Bland and Altman plots (B&A plots) was considered for each primary pigment of the data set (Figure 4). The LAD slope is between 0.97 and 1.08 for all the PPig (Table 6), indicating a remarkable match. This is confirmed by the determination coefficient, whose value varies from 0.76 for Zea to 0.95 for Fuco (0.91 for TChl a). All pigments exhibited an RSMLD lower than 0.32, with TChl a having a value of 0.16. Based on the assessment of B&A plots, there appears to be a bias between the two laboratories, resulting in an overestimation of J, particularly for xanthophylls and carotenoids (Zeax, Diato, Peri, Allo), with the exception of Caro, where the trend suggests an underestimation of J. The B&A plots evidence that the number of outliers is highest for the pigments with the lowest ranges of concentration, while for TChl a, TChl b, TChl c, Fuco, and Caro, approximately 95% of the natural samples are within the limit of agreement (within 1.96 standard deviation) (Figure 4).

Coherently with the B&A plots, %PD tends to be higher for lower values of the pigment concentration: the %PD highest values (> 50%) are seen for all pigments at concentrations lower than 0.5 mg/m³ (Figure 5). With measurement stations located mostly in clear or only moderately turbid waters (Figure 1), most of the samples analyzed in support to this comparison had concentrations in the range of 0.05–0.5 mg/m^{3 for} PPig and lower than 0.05 mg/m³ for the secondary pigments (Figure 5; Table 7). Lower correlations between laboratories occurred when pigments were quantified close to the instrumental LOD. This behavior has already been observed in previous intercomparisons and confirmed that the false-positive/negative quantifications performed near LOD thresholds may cause an increase in the bias (Hooker et al., 2005; Canuti et al., 2016). Considering the whole data set, J overestimates TChl a by 10.6% and PPig by 16.9% (Table 8), confirming the presence of a systematic bias as suggested by the B&A plot analysis. The analytical method had an impact on the distribution of differences among pigments. In this regard, we could observe that for all the campaigns, %PD were lower (−5.8%) for Caro and close to 0 for Zea and TChl b (−0.3% and 0.9%, respectively), despite the fact that the concentration of Zea was lower than 1 mg/m³ for all the samples. In the absence of specific tests on extraction on replicates of natural samples and of the analysis of common standards, it is not possible to attribute observed differences to the extraction procedure, to the chromatographic separation, or to a combination of the two.

The closest results between the laboratories were observed for the samples collected on AAOT: for TChl a, the average %PD is 8.2% versus 11.9% for the ship-based oceanographic campaigns and 15.3% versus 17.4% for what concerns the PPig. This could be explained by the large number of campaigns and major confidence in the filtration procedure (i.e., the volume of water to use for preconditioning the filter) and by the fact that, proportionally, the most oligotrophic conditions (associated with high %PD) were mostly found during oceanographic campaigns. The comparison of the surface data (one third of the total AAOT samples) with samples collected at 8 and 14 m depth exhibits comparable %PD for TChl a. Differences in results between AAOT and oceanographic campaigns also suggest that the inhomogeneity of preconditioned samples may affect the assumed equivalence of duplicates and consequently the agreement between independent analyses, regardless of the pigment concentration.

The %PD was further considered in relation to the sampling campaign (all results detailed in Table 8) to evaluate if the water type and a change in the operator, generating an inhomogeneity in the duplicate samples, may also impact the difference distribution. It can be noticed that when the TChl a average concentration is close to 1 mg/m³ (campaigns L04, BL7, I02 and AAOT), the %PD values (in modulus) are lower (from −1% to 8%), independently from the marine basin considered. This suggests that the methods are optimized for analysis of samples with a TChl a pigment content around 1 mg/m³. Other considerations, considering the differences between campaigns, indicate that an increase in %PD for low concentrations is by no means a general rule. The E03 (Eastern Mediterranean Sea) and K09 (Black Sea) campaigns had a comparable number of samples (49 and 53 respectively), were performed within a year, and finally, the time delay between collection and analysis was comparable. E03 TChl a concentration was on average noticeably lower than during K09 (0.098 and 0.678 mg/m³, respectively), but the associated PD% is 15.6%, while it is 24.1% for K09 (Table 8). Similar differences could be appreciated considering the %APD, where the TChl a associated with E03 is 16.9% while itis 26.3% during K09 (Table 9). On the other hand, the BL8 campaign (Black Sea, TChl a of 0.413 mg/m³ on average) shows lower differences (2.6% for TChl a and 8.2% for PPig) than K09. We could conclude that the marine basin and the water type are not influencing the performance of the two laboratories.

Sample storage was considered as well. In a few cases, the samples were analyzed at different times by the two laboratories. J analyzed the WM campaign’s samples 5 years after their collection. It was expected that for these samples the MVChl a could be partially degraded in DVChl a and Chlide a. Jeffrey et al. (1997) recovered 98% and 83% of the original TChl a concentration in microalgae after storage at −196°C for 60 and 328 days, respectively. WM was compared with E03 and I02 campaigns, where E03 and I02 samples were analyzed within a year from collection and no degradation of MVChl a is expected (Jeffrey et al., 1997). The TChl a results for the three campaigns are comparable to what is observed over all the stations: J overestimated D by 12.8% for WM, 15.6% and 7.3%, respectively, for E03 and I02 (Table 8). The MVChl a is on average 9.9% higher for J with respect to D. We should thus expect a lower overestimation of MVChl a for WM with respect to E03 and I02 due to an increasing contribution of Chlide a and DVChl a to TChl a. In the WM campaign, MVChl a quantified by J was 13.8% higher than D’s value, while this overestimation is 17.8% in the case of E03, and is not observed for I02, with J and D values of MVChl a very close (J< D by 0.9%). In the good analysis practice, 1 year is considered the maximum interval between sample collection and analysis (Jeffrey et al., 1997) The case described here indicates that time longer than 1 year between collection and analyses (5 years for WM) did not appear to affect the final quantification of TChl a.

It was expected that for higher-level variables (i.e., pigment sum, see Table 2) the differences between J and D should decrease as the concentration increases (Tables 8, 9). The TChl and DP generally follow the expected trend (differences between laboratories over all samples of 10.8% and 15.2%, respectively). TPig and other sums show differences around 15%, similar to or lower than those encountered for the single PPig. As for the pigment indexes PIndex used to evaluate the proportion for micro- (mPF), nano- (nPF), and picoplankton (pPF), the difference is low for the mPF and nPF fractions (2.5% and 5.7%, respectively, and of 10.0% and 10.6% APD), and higher for pPF (−14.2% PD and 17.8%APD). Looking at results across campaigns, it can be observed that pPF was high (unbiased %PD of −35.2% and −33.7% and %APD of 37.9% and 35.3%, respectively) for the K09 and K10 campaigns (Black Sea, late summer) but not so for BL7 and BL8 (Black Sea, early summer), where the values are −6.6% and −3.6% for pPF, respectively (12.0% and 6.2%, if we considered the %APD), thus suggesting that the basin itself is not a source of differences. The %PD values of mPF and nPF are usually found below 10% (in modulus).

The average %PD can be interpreted in terms of accuracy for the concentration of the various pigments. Accuracy estimates can be fed into error propagation techniques applied to algorithms dedicated to the calculation of PFT’s abundance or PSCs (e.g., Mouw et al., 2017). Additional information required for error propagation is the possible error correlation existing among pairs of pigments. This point has been explored here by using the 611 stations of the oceanographic campaigns and correlating the difference J–D for pairs of diagnostic pigments (Table 10). In general, the correlation (i.e., Pearson correlation coefficient) between differences is fairly low, but there are cases where the correlation is close to or exceeds 0.5. The highest being between TChl a and Fuco (> 0.9), other remarkable correlations were found between TChl a and Caro and TChl c (> 0. 7), and similarly between Fuco and TCl c and Caro (> 0.7).

To quantify the impact that uncertainties on pigment concentrations might have on PFTs and PSCs, the pigment distributions from the two laboratories were independently used as input to the model of Hirata et al. (2011) (summarized in Table 3), with results elaborated with the Data-Interpolating Variational Analysis (DIVA) software tool of the Ocean Data View software (https://odv.awi.de) to visualize the observations spatially interpolated (Figure 6). These maps should not be interpreted as actual PFT or PSC distributions, as oceanographic campaigns in a given basin took place in different periods, but they offer an efficient overview of the agreement between the distributions obtained from the two data sets. In general, the J and D distributions appear fairly coherent, with some exceptions. PSCs show a good agreement, confirming the results obtained for the pigment index (Tables 8, 9), which is not surprising as the formula are comparable. The picoplankton fraction observed in the Black Sea shows some differences between J and D in line with those observed for pPF in this specific basin (particularly for K09 and K10). For PFTs, the largest differences are seen for diatoms (higher values for J) and, to a much lesser extent, for dinoflagellates.

The present exercise relied on a unique data set of 957 measurement stations covering several European basins, for which samples could be analyzed for pigment concentration by two quality-certified laboratories. An average difference of 10.6% was observed for TChl a for concentrations across a wide range of values (0.083 to 27.35 mg/m³). If interpreted in terms of accuracy, this result fully satisfies the requirement of a 15% accuracy threshold associated with TChl a, measurement recommended for the validation of satellite data products. The average differences for PPig is 16.9%, with systematic differences (biases) observed for the J values of PPig with respect to those from D for chlorophylls and several xanthophylls. However, for some pigments (i.e., Zea and Caro), the differences are close to 0 or very low, regardless of the pigment concentration or the basin considered. This indicates that the extraction methods employed by the laboratories may diversely influence the quantification results in relation to the pigment’s chemical structure. Overall, differences between J and D were lowest for natural samples containing approximately 1 mg/m³ of TChl a, suggesting that both methods are optimized for a target analysis around this concentration.

Other aspects observed concern the time of preservation and the homogeneity of the samples. In terms of preservation, for the Western Mediterranean Sea oceanographic campaign, the long time between collection and analyses (5 years) does not appear to affect the final quantification of TChl a. On the other hand, the analysis of the data set by single campaign suggested that in several cases, the difference between laboratories could result from the in-homogeneity of duplicate samples, regardless of the pigments concentration or the basin characteristics, but in the absence of series of replicates, it was not possible to support this correlation statistically. In the intercomparison exercises, where the participants are usually four laboratories or more, the distribution of triplicates of each natural sample together with the analysis of standards provided to all the participants, allows to clearly address if the statistical differences among participants are due to the methods or to the samples inhomogeneity.

As for the higher-level variables, the differences between laboratories are, of course, in line with the results given for the individual pigments. The average difference for TChl a and DP is 10.8% and 15.2%, respectively. Pigment indexes, which are simple indicators for PSCs, showed a good agreement, with larger differences for the picoplankton fraction pPF (average of 14%) that can be explained by larger differences observed for Zea. This result for pPF can also be traced to the impact of two Black Sea campaigns (K09 and K10) that usually show higher differences for all quantities. In view of developing error propagation models for PFT or PSC algorithms, the analysis correlating differences between the two laboratories indicates a modest degree of error correlation, but exceptions suggest the need for further investigation. It is nevertheless stressed that this result is valid for in situ pigment data and would not necessarily apply to pigment concentration obtained by dedicated algorithms.

In general, it was observed that, regardless of the pigment concentration, the inhomogeneity in the water sample preparation may affect the assumed equivalence of duplicates and consequently the agreement between independent analyses. In conclusion, the comparison with an accredited laboratory is a good option for developing a new method or continuously assessing an existing one. To take one step further and properly address the differences between laboratories, the comparison should include certified standards analyzed at regular time intervals and, in the case of natural sample collection, a series of replicates (minimum three to be analyzed by each laboratory) to build a proper statistic on the natural samples collected. As illustrated by this study and previous intercomparison exercises, defining uncertainty values associated with single in situ pigment measurement is no easy task, but it is required to promote pigment data as fiducial reference measurements. It would also be a good practice before ingesting pigment data into global data bases of pigment data.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

EC has made contributions to this paper in the areas of conception, study design, analytic performance, and calculation.