In the context of global change, the number of studies that include meta-analytical combinations of environmental data increased at a fast pace in the past decades (e.g., Duarte et al., 2009; Harvey et al., 2013; Talarmin et al., 2016; McCrackin et al., 2017; Carstensen and Duarte, 2019; Lheureux et al., 2021; Lheureux et al., 2023). In these studies, it is assumed that the data were of similar quality and reliability, which is not necessarily true: measurement processes are all subject to systematic (i.e., bias) and random (i.e., variations) errors of various types, such as the choice of the method, ambient conditions, or the working practices of analysts (Taverniers et al., 2004; de Boer, 2016). At a time when high-quality scientific journals require a comprehensive documentation of data quality control (QC) procedures for publication (de Boer, 2016), and because erroneous conclusions may arise from studies using inappropriate methods, it is important to increase the confidence levels of data QC, which is often poorly documented in environmental monitoring datasets.

By using the 20 years of experience of the French coastal monitoring SOMLIT network (https://www.SOMLIT.fr/), the goal of our paper is to share a pragmatic approach to QC that focuses on annual inter-laboratory comparison (ILC) exercises as an indication of measurement uncertainty. The SOMLIT network currently includes 12 laboratories and 22 permanent sampling stations distributed along the French littoral ( Figure 1 ) and covers a wide range of environmental and trophic conditions (Goberville et al., 2010; Liénart et al., 2017; Liénart et al., 2018; Lheureux et al., 2021). Implemented by CNRS (Centre National de la Recherche Scientifique) in 1997 to increase our scientific knowledge of the responses of marine coastal ecosystems to natural and human-induced influences, the SOMLIT network is part of the national research infrastructure ILICO (Infrastructure de recherche LIttorale et CÔtière; Cocquempot et al., 2019). Common standard operating protocols are used and all data are continuously routinely quality-controlled since 2006 by means of robust quality assurance procedures guided by the ISO-17025 standard (ISO-17025, 2005, then ISO-17025, 2017).

After an overview of the different QC actions implemented by the SOMLIT to fulfill requirements for quality and competence, we report equipment, accommodation, design of the ILC exercises and statistical methodology to assess individual performance specifically adapted to small (<20 laboratories) environmental networks and multivariate datasets, based on the ISO-13528 standard (ISO-13528, 2005), Héberger and Kollár-Hunek (2011), and procedures for measurement uncertainty assessment. Finally, we provide an expanded uncertainty of measurement for 20 environmental variables routinely measured by SOMLIT from discrete sampling ( Table 1 ), including Essential Ocean Variables (Miloslavich et al., 2018).

Data QC within the SOMLIT network is composed of five actions ( Figure 2 ). Within-laboratory control is continuously performed by each laboratory on the basis of control charts ( Figure 3 ; Table S1 ) and the Nordtest TR 569 principles (Nordtest TR 569, 2007): control charts ensure the quality of produced data according to statistical or reference criteria—when available—and represent an essential step for the calculation of uncertainty. Control charts may also be performed (1) to objectively compare different analytical methods, (2) for new method validation, (3) to qualify an inexperienced analyst, or (4) to ensure stable ambient conditions (e.g., temperature and hygrometry) for measurements. There are two kinds of control charts: X-charts and R-charts for controlling systematic errors and random variations, respectively (Nordtest TR 569, 2007). The setting control limits used to build the control charts used by the SOMLIT network are given in Table S1 .

SOMLIT data are qualified on a trimonthly basis using QC flags (see Table 2 ), which are derived from the World Ocean Circulation Experiment (WOCE; https://exchange-format.readthedocs.io/en/latest/quality.html#woce-bottle-quality-codes). QC flags include two levels. Level 1 ensures interoperability, but the codification is simplified to provide QC flags easily understandable by end-users. Level 2 is detailed in order to answer to the needs of data producers; it corresponds to internal QC according to the ISO 17025 standard (ISO-17025, 2017). Internal and external audits guided by the ISO 17025 standard (ISO-17025, 2017) are made on an annual and triennial basis, respectively, for ensuring laboratory quality management and technical requirements. We provide an example of the report in Figure 4 .

ILCs are undertaken annually at the end of summer in one of the SOMLIT laboratories ( Figure 1 ). To strengthen the robustness of comparisons, we alternate among different hosting laboratories, with sampling stations in the oligo- to ultra-oligotrophic Mediterranean Sea, oligo- to mesotrophic Atlantic waters, and the meso-eutrophic English Channel ( Table S2 ). ILC allows best practices to be shared among participants from different laboratories, to evaluate data reliability (uncertainty calculation), and to assess the performance of individual operators from each laboratory. If issues are identified, ILC provides a forum to help participants to improve their practices and to correct their errors. In the following sections, we describe the methodology used during ILC and show how efforts to fulfill quality requirements have helped to improve the trueness and reproducibility of SOMLIT data over time.

Although the SOMLIT program routinely measures 25 environmental variables, only the 22 variables subjected to discrete sampling ( Table 1 ) were considered each year. Figure 5 describes the different steps involved in the ILC exercises. Given that stable reference material is lacking for most of the environmental variables measured by the SOMLIT network, performance in total measurement was evaluated by robust analysis, using measurement data from real natural seawater (see below).

Two approaches exist for the evaluation of measurement uncertainty. The first is based on a physical model relating the measuring process at all steps, using the propagation law of uncertainty (GUM, 1993). The second relies on a statistical model that uses data from within-laboratory control and/or ILC exercises or PT exercises (ISO-13528, 2005; Ferretti, 2011; Magnusson et al., 2012; de Boer, 2016). The statistical model is often favored, however, and was chosen by the SOMLIT network both for its less apparent mathematical complexity and for the lack of knowledge about the processes that need to be measured for the physical model (e.g., firm statistical background). For that purpose, we used the data from ILC exercises from the last 6 years. Note that the values we obtained are an approximation of the total measurement uncertainty, given that sampling uncertainty was never investigated.

Total measurement uncertainty of the SOMLIT network (U, in ± Unit and/or in %) was calculated following Equation 4 (Magnusson et al., 2012):

where u_RW is the precision uncertainty and b is the bias uncertainty obtained from Equations 5 and 6, respectively:

with σ_ι and x_i being the median and mean, respectively, of the repeatability values obtained by the different laboratories during one ILC exercise for one measure, s_bias being the standard deviation of the different median values of b for at least six different ILC exercises, n being the number of ILC exercises used for uncertainty calculation, and u_x being the median value of the uncertainty of the assigned values calculated for at least six ILC exercises. U is the expanded uncertainty at a confidence interval of 95% (k = 2).

The resulting expanded uncertainty values of the total measurement of the 22 environmental variables monitored by the SOMLIT network given in Table 1 (see also Table S3 ) are either below or very close to those found in the literature, when available.

In this paper, we presented the “road map” and the statistical methodology for the ILC exercises performed by the SOMLIT network, a framework especially adapted to small (<20 participating laboratories) and multivariate environmental datasets. By applying this framework to data acquired during 21 years, we showed that the SOMLIT program provided reliable data, and that the data quality increased over time thanks to the implementation of ILCs. Given the paucity of uncertainty values of the measurement of environmental variables—such as measured by the SOMLIT network—in the literature, our study stresses the importance of implementing ILC exercises in order to improve the quality of data representative of long-term environmental monitoring. We also highlight general acceptance limits in the measurement of these environmental variables.

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

Conceptualization and methodology: EBr, NS, PR-M, BS, TC, SF, ACai, VD, YA, GGr, EGr, AG, OJ, VL, EMac, DM, FM, SM, DP, PP, FR-J, IS, P-GS, and NG. Data collection and curation: EBr, NS, PR-M, TC, SF, SAl, FA, SAu, EBe, HB, LB, SB, ACau, J-BC, LCh, RC’h, M-AC, LCo, OC, MC, HL, GD, JD, AD, MD, CE, SG, EGr, AG, JG, KG, GGu, VG, MH, OJ, PL, NLab, ML, VL, CLa, JL, NLac, BL, EL, DL, CLe, CLi, EMac, EMar, BMa, SM, LMor, LMou, AN, SN, RP, DP, FP, PP, FR-J, PRIM, CS, ES, RVa, FVi, FVo, LZ, and NG. Formal analysis: EBR and ALh. Project administration, funding acquisition, and resources: EBr, NS, PR-M, BS, TC, HA, CB, YB, SB, PCl, PCo, OC, ACai, YA, GGr, EGr, AG, MC, JF, MG, OJ, JL, LL, ALe, SL’H, YL, SM, FM, BMo, LMou, SN, PP, PRIM, ES, IS, LS, VV, RVu, P-GS, and NG. Writing—original draft preparation and writing—review and editing: EBr, NS, PR-M, EGo, ALh, VD, BL, BMo, PRIM, P-GS, FV, and NG. All authors contributed to the article and approved the submitted version.