Citizen Sailors Provide a Detailed Map of Open‐Ocean Surface Biodiversity and Help Link eDNA Data to Satellite Observations

ABSTRACT

Characterizing open ocean biodiversity at basin scales remains challenging due to sparse sampling and limited opportunities to validate satellite‐derived estimates of plankton community composition. Through the Citizens of the Sea initiative, we deployed standardized environmental DNA (eDNA) sampling kits on 27 sailing vessels during the 2024 Pacific Rally circumnavigating between New Zealand, Tonga, and Fiji, generating 728 surface ocean eDNA samples across 1.5 million km². Using three complementary metabarcoding markers (bacterial 16S rRNA, eukaryotic 18S rRNA and metazoan Cytochrome Oxidase I genes), and following rigorous bioinformatic quality control, we obtained 31,177 unique Amplicon Sequence Variants (ASVs), providing the first high‐resolution plankton biodiversity assessment of this understudied region. After accounting for human‐derived variation, all metabarcoding markers revealed clear latitudinal patterns in alpha diversity and community composition. ASV richness increased significantly with sea‐surface temperature and community turnover was strongly correlated with gradients in temperature and chlorophyll‐a. Comparisons between eDNA‐inferred phytoplankton functional groups and satellite‐based estimates showed moderate agreement for prokaryote‐dominated communities (R ² = 0.27) but a weak correlation for eukaryotic phytoplankton (R ² < 0.15), reflecting functional‐type‐specific mismatches between molecular and remote sensing approaches. Our results reinforce that citizen science fleets can generate high‐quality oceanographic eDNA datasets at scales unattainable by research expeditions, while demonstrating the strong potential of these initiatives to bridge molecular observations with satellite‐derived assessments and improve our understanding of open ocean plankton biodiversity.

1. Introduction

Understanding and safeguarding marine biodiversity is crucial for maintaining life on Earth (Corlett 2020; Ward et al. 2022). Our oceans regulate our climate, supply food and oxygen and support complex ecological processes essential to planetary health (Fleming et al. 2019; Laffoley et al. 2020; Rose et al. 2024). The rapid expansion and increasing sophistication of satellite platforms have dramatically enhanced the global acquisition of physical and biochemical ocean information (Cetinić et al. 2024; Groom et al. 2019). These advancements allow scientists to monitor vast oceanic regions continuously, providing critical data on sea‐surface temperature, ocean colour, phytoplankton pigments (e.g., chlorophyll concentration), sea level and currents, which are essential for climate study, marine resource management and environmental monitoring (Rani et al. 2021; Zhao et al. 2024). With these data, our ability to monitor physical and biochemical changes in our oceans is continually improving and is being used in models to help predict ecological shifts (Follows et al. 2007). However, our understanding of the biological components of our oceans remains limited due to a significant gap in the available data on ocean biodiversity at comparable scales (Canonico et al. 2019).

Most of our oceans remain biologically unexplored, especially in offshore and remote regions, due to the high costs and logistical challenges of traditional research expeditions (de Vargas et al. 2022; Lauro et al. 2014). This gap in biological data, both geographically and temporally, is one of the main limitations to understanding the complex community dynamics in these ecosystems and how they are responding to climate change and human‐related stressors such as pollution, overfishing and other threats (e.g., Boonstra et al. 2015; Halpern et al. 2007, 2015, 2025). Filling this biological data gap is essential, as integrating biodiversity information with physical and biogeochemical observations will improve predictive ocean models and provide the ground‐truthing required to refine satellite‐based assessments of plankton communities (Cetinić et al. 2024).

Environmental DNA (eDNA) technologies are transforming our capacity to monitor biodiversity by providing a non‐invasive and cost‐effective alternative to traditional methods (Deiner et al. 2017; Ruppert et al. 2019; Wood et al. 2013). In marine environments, eDNA is increasingly adopted to characterize biodiversity for a wide range of taxa, from microbes and algae to fishes and marine mammals, spanning ecosystems from the tropics to the poles and from coastal waters to the open ocean and the deep‐sea floor (Capurso et al. 2023; Çevik and Çevik 2025; Miya 2022; Shea et al. 2023; Suarez‐Bregua et al. 2022; Yang et al. 2024). Moreover, eDNA offers a non‐invasive and scalable means to detect biodiversity across multiple trophic levels, capturing both intra‐ and extracellular DNA throughout the water column. While this signal includes both contemporary and relic DNA, which can sometimes blur fine‐scale ecological resolution, it nonetheless enables broad community‐level assessments that would otherwise be infeasible in remote or under‐sampled regions (Cordier et al. 2021; Hoban et al. 2023; Nester et al. 2024; Saenz‐Agudelo et al. 2024; Yang et al. 2024). As a result, eDNA can play a critical role in addressing the biological data gap that persists across much of the global ocean, particularly in offshore and open‐water environments that remain logistically challenging to monitor through traditional means.

For the open ocean, several landmark initiatives have advanced our understanding of plankton diversity and ecosystem functioning at a global scale. Major scientific expeditions such as the Global Ocean Sampling (Rusch et al. 2007), Tara Oceans (Karsenti et al. 2011) and Malaspina (Duarte 2015), as well as citizen‐driven efforts like Plankton Planet (de Vargas et al. 2022), have revealed unprecedented insights into plankton taxonomy, diversity, interactions and community structure. Despite these achievements, substantial scalability challenges remain. Single‐vessel scientific cruises are costly and time‐intensive. For example, Tara Oceans required three years to complete a single global transect, precluding repeated, basin‐scale coverage. Citizen science programmes offer complementary potential, but their success depends strongly on how simple and accessible the sampling protocols are; overly complex procedures can reduce participation and increase the risk of obtaining data of inconsistent or variable quality (Agersnap et al. 2022; de Vargas et al. 2022). These constraints highlight the need for new approaches that combine scientific robustness with operational scalability and ease of use.

The Citizens of the Sea initiative addresses both challenges in ocean science. This charitable programme equips ocean cruisers and offshore yachts with a user‐friendly, towable eDNA sampler known as TorpeDNA (Pochon et al. 2026), which enables the efficient collection of eDNA using in‐water DNA capture technology while the vessel is underway at cruising speeds (von Ammon et al. 2023; Pochon et al. 2024). This innovation eliminates the need for specialized research vessels or laborious on‐board filtration, thereby dramatically reducing the cost and complexity of offshore sampling. By equipping large numbers of vessels, in strategic places, high‐quality biological data can be obtained at scale, especially in data‐poor regions of the open ocean.

Here, we describe the first large‐scale field campaign undertaken by Citizens of the Sea. During the 2024 Island Cruising Pacific Rally, 27 volunteer sailing vessels collected surface‐water eDNA across more than 1.5 million km² of ocean between Aotearoa‐New Zealand, Tonga and Fiji. Participating crews were trained in a standardized sampling protocol and equipped with TorpeDNA samplers, enabling the collection of a high‐density open‐ocean biodiversity dataset across approximately 20 degrees of latitude in the South West Pacific. Samples were analysed using three complementary metabarcoding markers targeting bacterial 16S rRNA, eukaryotic 18S rRNA and metazoan cytochrome c oxidase subunit I genes.

Using this dataset, we addressed three main objectives. First, we assessed spatial patterns in taxonomic diversity and community composition across the sampled region and quantified the extent to which variation in the data was attributable to citizen collectors and vessel‐related sampling effects. This allowed us to evaluate the reliability of citizen‐led eDNA sampling for large‐scale open‐ocean biodiversity monitoring. Second, we examined how eDNA‐derived community patterns were related to satellite‐derived environmental variables, including sea‐surface temperature and chlorophyll‐a concentration. Finally, we used the multi‐marker dataset to infer the relative representation of major phytoplankton functional groups and evaluated how these molecular community signatures corresponded with satellite‐derived phytoplankton functional type estimates from the Copernicus Marine Service. Through these analyses, we tested the capacity of citizen‐powered eDNA sampling to generate scientifically robust biodiversity observations at basin scale and to complement remotely sensed assessments of open‐ocean ecosystems.

2. Materials and Methods

2.1. Citizen Science Sampling Campaign

Between 6 May and 3 December 2024, a fleet of 27 sailing vessels participating in the Citizens of the Sea Pacific Rally collected environmental DNA (eDNA) samples across more than 1.5 million km² of the South West Pacific Ocean. This large‐scale citizen science campaign was designed to test and validate a novel approach to open ocean biodiversity monitoring using a standardized, user‐friendly eDNA collection protocol that could be reliably implemented by non‐specialists at sea.

Each vessel was equipped with a TorpeDNA sampler, which enables the in situ capture of eDNA while the boat is underway at speeds of up to 12 knots (Pochon et al. 2026). Prior to deployment, crews received training through a combination of written manuals, video tutorials and live or virtual briefings. These resources covered the four‐step sampling protocol, best practices for contamination control, sample handling procedures and data submission workflows (Supporting Information; https://www.youtube.com/watch?v=9k‐31_44THk). Training emphasized simplicity, reproducibility and data integrity, ensuring broad applicability across a diverse fleet with varying levels of scientific experience.

Sampling was conducted opportunistically by citizen scientists along their sailing routes between Aotearoa New Zealand, Tonga and Fiji, rather than at predesignated offshore stations. Participants were instructed to collect triplicate surface water samples sequentially at suitable locations during daylight hours—between 10 am and 2 pm—when weather and sea conditions permitted the safe and practical deployment of the TorpeDNA sampler. Safety was of paramount concern and crews were explicitly instructed to avoid sampling during rough seas or adverse weather conditions. This flexible yet standardized approach enabled the fleet to gather scientifically valuable data across diverse oceanographic conditions, while ensuring participants' safety and adherence to best practices for sample quality. At each station, crews collected triplicate surface water samples by towing the TorpeDNA sampler approximately 15 m behind the vessel for 5 min per replicate, typically at sailing speeds between 5 knots and 12 knots. The sampling device contained a 20 μm nylon mesh membrane (47 mm diameter; Merck Millipore, Germany) to trap microbes, larvae and free‐floating DNA particles derived from larger organisms, a method that was validated in Pochon et al. (2024).

Immediately following recovery, filters were handled with sterile gloves and forceps, then transferred into 1.7 mL vials prefilled with PowerProtect preservation buffer (QIAGEN, Germany). Each sample was scanned using a custom ArcGIS Survey123 form (Esri, California), which enabled the automatic capture of metadata, including date, time, GPS coordinates and vessel ID. During the first passage from New Zealand to the tropics (May–August 2024), filters were stored on board at 4°C until the vessel reached drop‐box locations in Nukuʻalofa (Tonga), Savusavu and Viseisei (Fiji), where they were consolidated and refrigerated until shipping. During the return passage from the tropics to New Zealand (September–December 2024), samples were inspected by biosecurity officers at the Opua Marina and refrigerated until shipping. All collected samples were transported under chilled conditions to the Cawthron Institute in Nelson, Aotearoa New Zealand, for laboratory analysis. Marine scientific research permits for the collection of seawater filtrates were secured for the Exclusive Economic Zones of Fiji (Collection permit #TPN 546/2025) and Tonga (Collection permit # F3/14/1146/24), which were delivered by the Ministry of Foreign Affairs of the Governments of Fiji and Tonga, respectively, and New Zealand (Collection permit #SP822 delivered by the New Zealand Ministry for Primary Industries).

2.2. eDNA Sample Processing

DNA from each filter was extracted using QIAGEN's PowerSoil Pro Kit on a QIAcube HT system (QIAGEN, Germany), as per the manufacturer's protocols. A blank filter sample was included during extraction protocols for high‐throughput DNA extraction, amplification and sequencing.

DNA extracts from all samples were amplified by Polymerase Chain Reaction (PCR) using three different primer sets targeting bacterial, eukaryotic and metazoan communities (Table 1). All primer sets were modified with Illumina overhang adapters to allow library preparation (forward overhang: TCGTCGGCAGCGTCAGATGTG TATAAGAGACAG; reverse overhang: GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAG). Polymerase chain reactions were performed using 17 μL of MyFi Mix 2× Master Mix (Bioline, UK) with 1 μL of each primer (10 μM), 11 μL of DNA‐free water and 2 μL of purified DNA, in a total reaction volume of 30 μL. Primer sequences and thermocycling conditions for each primer set are detailed in Table 1. Amplification products were normalized using SequalPrep Normalization Plates (Thermo Fisher Scientific, USA), dual‐indexed with Nextera XT adapters and pooled. Paired‐end sequencing was performed at Sequench Ltd. (Nelson, Aotearoa New Zealand) using V2 chemistry on a NextSeq platform (Illumina, USA). Extraction and PCR blanks were included throughout the process.

TABLE 1.

Details of all metabarcoding primer assays used for each targeted community.

Metabarcode	Target community	Primer sequences (5′–3′)	Thermocycling conditions	References
16S	Bacteria	341F CCTACGGGNGGCWGCAG 805R GACTACHVGGGTATCTAATCC	94°C 5 min 35 cycles (94°C 30 s, 53°C 30 s, 72°C 45 s) 72°C 7 min	Klindworth et al. (2013)
18S	Eukaryotes	UNI18SF AGGGCAAKYCTGGTGCCAGC UNI18SR GRCGGTATCTRATCGYCTT	94°C 5 min 37 cycles (94°C 30 s, 54°C 30 s, 72°C 45 s) 72°C 7 min	Zhan et al. (2013)
COI	Metazoans	mlCOIintF GGWACWGGWTGAACWGTWTAYCCYCC jgHCO2198 TANACYTCNGGRTGNCCRAARAAYCA	94°C 5 min 40 cycles (94°C 30 s, 52°C 30 s, 72°C 45 s) 72°C 7 min	Leray et al. (2013)

2.3. Bioinformatics

All fastq files were demultiplexed, and their primers were removed using CUTADAPT (v4.9; Martin 2011), allowing no insertion/deletion and requiring a minimum overlap of 15 bp. Sequences were truncated on their 3′ ends at 215 bp and 190 bp for 16S, and 225 bp and 226 bp for 18S and COI, for the forward and reverse reads, respectively, to remove low‐quality calls. Sequence quality filtering and denoising were performed using the DADA2 R package (v 1.3; Callahan et al. 2016), and paired‐end reads were merged with a minimum overlap of 10 bp. Chimeric sequences were identified and removed using the consensus option of DADA2.

Taxonomic assignation for 16S was performed using the DADA2's implementation of the RDP Naïve Bayesian Classifier algorithm (Wang et al. 2007) with default settings against the SILVA database (v 138.2; Quast et al. 2012). For 18S, sequences were assigned with the RDP Naïve Bayesian Classifier algorithm applied on the SILVA (version 132) and PR2 (v 5.0.0; Guillou et al. 2012) databases. For COI, assignments were made using the classification trees (‘insect’) classifier (v 1.4.0.9; S. P. Wilkinson et al. 2018) trained on the MIDORI UNIQUE 20180221 training set and approximately 14,000 non‐metazoan COI sequences from GenBank (Benson et al. 2008). For 18S and COI, taxonomy was also assigned using blastn and megablast (Camacho et al. 2009) on the GenBank nucleotide database (Benson et al. 2008), using the ‘blastn_taxo_assignment’ function of the biohelper R package (v 0.0.18; Laroche 2024). For each gene, results from different assignment approaches were combined using the ‘taxo_merge’ function of biohelper, which first normalizes the taxonomy across the different outputs using the National Center for Biotechnology Information (NCBI) curated taxonomic database (Schoch et al. 2020), and then uses the highest resolution between approaches whenever there is a majority consensus (> 50%) across all assigned ranks. At each specific rank, taxonomy was attributed only if there was a majority agreement between methods.

Nuclear mitochondrial DNA sequences (NUMTs) in the COI data were identified and removed using MetaMate (Andújar et al. 2021) and the default specification file. The filtering parameters were chosen to maximize the retention of verified authentic sequences (highest recall) while removing as many verified non‐authentic sequences as possible (highest accuracy). Amplicon sequence variants (ASVs) of the COI data were then assigned to operational taxonomic units based on sequence similarity (97%) using the ‘otu’ function of the kmer R package (v 1.1.2; S. Wilkinson 2022).

Potential contamination was investigated using negative field (n = 86) and laboratory controls (n = 5), and removed using default settings of the microDecon R package (v 1.0.2; McKnight et al. 2019). For the bacterial 16S dataset, only ASVs with at least 10 reads in at least three samples were retained to minimize the presence of potentially spurious sequences. After applying these filters, samples with fewer than 10,000 reads were also removed from further analyses. For COI, all ASVs assigned to bacteria were removed. For the COI and 18S datasets, ASVs assigned to terrestrial taxa were manually removed, and only ASVs with more than two reads in at least two samples were retained. After applying these filters, samples with fewer than 5000 reads were removed.

2.4. Satellite Data

Daily sea‐surface temperature (SST) data were obtained from the Multi‐scale Ultra‐high Resolution (MUR) sea‐surface temperature dataset (JPL MUR SST v4.1; Chin et al. 2017) accessed through the NASA Jet Propulsion Laboratory ERDDAP server. Daily chlorophyll‐a concentrations were extracted from the National Oceanic and Atmospheric Administration (NOAA) CoastWatch VIIRS (Visible Infrared Imaging Radiometer Suite) ocean colour product (NOAA NESDIS, 2023), accessed at 2 km spatial resolution. Daily suspended particulate matter (SPM) concentrations were obtained from the NOAA CoastWatch VIIRS‐derived particulate matter product at 2 km spatial resolution. Environmental data extraction was performed using the rerddapXtracto R package (v 1.2; Mendelssohn 2024). For each sampling location, satellite data were extracted with the rxtracto() function using the longitude, latitude and date of each sample location. The bilinear interpolation option, with a search radius of four grid cells, was chosen to account for potential cloud cover or missing data. Surface values (z = 0) were used for the chlorophyll‐a and SPM datasets.

Monthly phytoplankton functional type concentrations were obtained from the Global Ocean Colour (Copernicus‐GlobColour) Bio‐Geo‐Chemical L4 dataset (OCEANCOLOUR_GLO_BGC_L4_MY_009_104; ACRI‐ST, 2023) accessed through the Copernicus Marine Service (https://doi.org/10.48670/moi‐00279; accessed October 2024). We obtained satellite‐derived estimates of total chlorophyll‐a (CHL) and six phytoplankton functional types (PFTs): diatoms (DIATO), dinoflagellates (DINO), green algae (GREEN), haptophytes (HAPTO), Prochlorococcus (PROCHLO) and prokaryotes (PROKAR), along with their respective uncertainty estimates (Xi et al. 2021). Data extraction and downloading were performed using the copernicusmarine Python package (version 2.2.1). For each sampling location, satellite data were accessed using the open_dataset() function and extracted using the isel() function with the longitude, latitude and date of each sample location. A spatial buffer of 0.01° (~1 km) around each point was applied to account for potential data gaps, with nearest‐neighbour interpolation used for the final extraction. Monthly composites corresponding to the mean of available samples in each sampling month were used to ensure data availability and temporal consistency.

2.5. Statistical Analyses

Custom R scripts were used to obtain summary metrics of the data and to generate all the figures. Alpha diversity was explored in different ways. First, sampling effort completeness across all three markers was explored. For this, an ASV accumulation curve analysis was performed using the R package vegan (v 2.6.8; Oksanen et al. 2019), using the function specaccum with the ‘random’ method and 100 permutations. We also investigated whether the number of ASVs, serving as a proxy for alpha diversity, was correlated with sea‐surface temperature (SST), water column depth and chlorophyll‐a (we did not include other colour‐derived satellite variables, as most are highly correlated with chlorophyll‐a). For this, a generalized linear mixed‐effects model was fitted using the glmmTMB R package (v 1.1.11; Brooks et al. 2017). One model was built for each marker set. For each model, ASV richness was used as a response variable. SST, water column depth, vessel speed and chlorophyll‐a concentration were included as fixed effects. To control for the non‐independence in the data structure, the sequencing run ID was included to account for technical variation between sequencing batches. The vessel ID was also included, nested within sequencing run, to account for variation among vessels (vessel ID was nested because different vessels were sequenced in different sequencing runs). This nested structure allowed us to control for these sources of variation that could confound relationships with fixed variables. Finally, because sequencing effort is highly variable among samples and was significantly correlated with ASV richness (16S: Pearson's r = 0.103, p = 0.012; 18S: Pearson's r = 0.689, p < 0.001; COI: Pearson's r = 0.648, p < 0.001), the offset(log(sequence depth per sample)) was included in our models. This approach was chosen instead of rarefying samples to an even depth, as the latter approach discards substantial amounts of data and can introduce biases (McMurdie and Holmes 2014). This approach assumes that ASV richness scales proportionally with sequencing depth, allowing the model to estimate fixed effects while controlling for sequencing effort (Wagner et al. 2011). All environmental variables were standardized prior to model fitting.

Before analysis, samples with missing environmental data and samples from shallow coastal waters (less than 20 m depth) were removed. Model convergence was assessed by examining warning messages and the properties of the Hessian matrix. The statistical significance of environmental predictors was evaluated using Wald tests, with effects considered significant at α ≤ 0.05. Variance components for random effects were extracted to quantify the proportion of variation explained by sequencing run and vessel ID effects.

The community composition and structure of the three eDNA datasets (16S, 18S and COI) were explored. First, for each dataset, the average relative abundance at the phylum level was calculated by merging ASVs per phylum and estimating the relative abundance of each per degree of latitude. We used custom R scripts for this and plotted relative abundances as a function of latitude. Second, the community structure was explored by performing a principal components analysis (PCA). For each dataset (ASV abundance), a centred‐log‐ratio (CLR) transformation was applied to account for the compositional nature of sequence data. The function ‘transform’ from the microbiome R package (v 1.24; Shetty and Lahti 2019) was used. This transformation preserves relative abundance information while addressing the unit‐sum constraint inherent in this type of data by transforming each sample's taxonomic abundances relative to their geometric mean, thereby converting the data in a way that allows standard multivariate statistical methods to be applied appropriately (Aitchison 1982; Gloor et al. 2017). Subsequently, a PCA on the CLR‐transformed abundance matrix was performed using the prcomp R function with centre = true, but scale = false as the CLR data already addresses this standardization. The first two axes were used for visualization using ggplot custom scripts. Finally, we evaluated whether the resulting ordination was significantly correlated with three environmental variables: SST, chlorophyll‐a concentration and the depth of the water column. For this, we first used the function envfit from the R package vegan to measure how community composition variation was related to each of these three variables (999 permutations were performed to assess statistical significance). Second, we performed a redundancy analysis (RDA) on the same CLR‐transformed ASV data and used vessel speed, SST, chlorophyll‐a concentration and depth of water column variables as constraining variables and included sequencing run and vessel ID as conditioning variables. The RDA allowed us to quantify the proportion of multivariate composition variance explained by these environmental variables while controlling for the potential effects of sequencing run and vessel ID.

Finally, we examined the relationships between eDNA‐derived phytoplankton abundances and satellite‐derived chlorophyll concentrations. Specifically, we evaluated whether satellite surface total chlorophyll concentration estimates (CHL) correlate with total eukaryotic phytoplankton eDNA relative abundance and alpha‐diversity metrics, and whether the satellite derived chlorophyll concentrations of different phytoplankton functional types (PFT) as defined in (Xi et al. 2021) (DIATO = diatoms, DINO = dinoflagellates, HAPTO = haptophytes, GREEN = green algae, PROKAR = prokaryotes and PROCHLO = Prochlorococcus) obtained from the Copernicus Marine Service, were correlated with estimated relative abundances of these groups from eDNA data. For this, we transformed the data to relative abundance and filtered the 18S and COI datasets to extract taxa that were taxonomically assigned to the phyla Bacillariophyta, Rhodophyta, Chlorophyta, Prasinodermophyta, Haptophyta, and classes Dinophyceae, Cryptophyceae, Pelagophyceae, Raphidophyceae, Dictyochophyceae and Bolidophyceae to create phytoplankton‐specific datasets. From each of these datasets, we created separate sub‐datasets for diatoms (phylum Bacillariophyta), dinoflagellates (class Dinophyceae, excluding the Syndiniales order), haptophytes (phylum Haptophyta) and green algae (phylum Chlorophyta). For prokaryotes, we filtered the 16S dataset using the FAPROTAX database (Louca et al. 2016) to first assign functional groups to each ASV and then selected all ASVs that were assigned to either the photosynthetic_cyanobacteria or photoautotrophy functional groups. For Prochlorococcus, we selected all ASVs that were classified as Prochlorococcus at the genus level. For each dataset and taxonomic group, we estimated relative read abundance, ASV richness (number of ASVs), and relative ASV richness. In addition, we calculated ASV richness scaled by distance (metres). We used distance because our sampling method does not allow for accurately measuring the filtered water volume (which would allow us to standardize by filtered water volume instead). For this, we used vessel speed and deployment time (5 min) to estimate the length of each sampling transect in meters and used this to calculate a standardized ASV richness per meter. We employed a generalized linear mixed model approach (GLMM) to evaluate the relationship between satellite‐derived chlorophyll‐a (total and specific to each taxonomic group), and four eDNA estimates of relative abundances. For each response variable (CHL, DIATO, DINO, HAPTO, GREEN, PROKAR and PROCHLO) and dataset combination, we compared four models with different eDNA predictors: relative read abundance, ASV richness, relative ASV richness and standardized ASV. For ASV richness and standardized richness models (count data) we used Gamma distributions with log links. For relative read abundance and relative ASV richness models we used beta distributions with logit links. All models included Vessel ID as a random effect to account for the observed technical variation among vessels. Model selection was based on AIC, and we calculated marginal and conditional R ² values to partition variance explained by biological predictors versus technical factors. Models were fitted using the glmmTMB package in R (Brooks et al. 2017).

3. Results

3.1. General Patterns

Overall, we successfully recruited a total of 27 vessels, which collectively gathered 831 samples, including 86 field (negative) controls. From these, we excluded 17 samples that failed to meet quality requirements (metadata collection, filter integrity, etc.). A total of 728 samples and 86 negative field controls were processed in the lab and sequenced. Ninety‐seven per cent of samples (n = 708) successfully amplified at least one of the three primers. Amplification success for 16S was 83% (n = 606), for 18S was 85% (n = 618) and for COI was 92% (n = 673). After demultiplexing, fastq‐files quality filtering and contamination removal, we recovered ca. 432 million reads and catalogued 209,478 ASVs across all three metabarcodes (Table 2). Around 381 million reads (88.2%) remained after applying ASV, read depth, and water depth filters. Details on the distribution of the number of samples, reads and ASVs per metabarcode before and after filtering are shown in Table 2. Overall, the estimated total sampled area was around 1,518,200 km². A map showing the geographic distribution of all samples is presented in Figure 1.

TABLE 2.

Distribution of the number of samples, sequences and amplicon sequence variants (ASVs) per vessel during the sampling process, and that passed basic bioinformatic quality filters.

Metabarcode	Before filtering			After filtering			Filters applied
	N	Sequences	ASVs	N	Sequences	ASVs
16S	606	170,493,609	83,484	579	152,212,321	13,842	ASVs > 10 reads, in > 2 samples Samples > 10,000 reads Water depth > 20 m
18S	618	125,980,388	73,607	576	115,339,136	10,081	ASVs > 2 reads, in > 1 samples Terrestrial taxa excluded Samples > 5000 reads Water depth > 20 m
COI	673	135,233,222	52,387	557	113,572,102	7254	ASVs > 2 reads, in > 1 samples Terrestrial taxa excluded Samples > 5000 reads Water depth > 20 m

Abbreviations: 16S = bacterial ribosomal 16S rRNA, 18S = eukaryotic ribosomal 18S rRNA, COI = metazoan mitochondrial Cytochrome c Oxidase I.

FIGURE 1.

Map showing the geographic distribution of samples (red points) and bathymetric profile of the study region.

Taxonomic assignments varied between markers and among different taxonomic ranks. In general, a high proportion of ASVs (> 50%) was assigned to high taxonomic ranks (phylum, class, order and family) for all markers. At the genus and species levels, taxonomic assignments were lower for all markers and ranged between 4% and 50%. Further details on the proportions of taxonomic assignments can be found in Table 3.

TABLE 3.

Distribution of assignments across taxonomic ranks for all markers used.

Metabarcode	ASVs	Phylum	Class	Order	Family	Genus	Species
16S	13,842	42 (99%)	66 (94%)	158 (90%)	238 (76%)	504 (50%)	425 (4%)
18S	10,081	28 (38%)	62 (93%)	134 (83%)	257 (54%)	349 (40%)	287 (10%)
COI	7254	22 (80%)	50 (72%)	97 (64%)	122 (50%)	129 (23%)	153 (14%)

Note: The number of reported amplicon sequence variants (ASVs) corresponds to the ASVs left after applying the following filters: ASV abundance (n ASV per sample > 10 for 16S, and 2 for 18S and COI), ASV prevalence (ASVs present in at least three samples for 16S, and two samples in 18S and COI) and sample‐sequencing depth (samples with at least 10,000 sequences for 16S and 5000 for 18S and COI). The numbers in parentheses indicate the percentage of ASVs successfully assigned to that particular taxonomic category.

Abbreviations: 16S = bacterial ribosomal 16S rRNA, 18S = eukaryotic ribosomal 18S rRNA, COI = metazoan mitochondrial Cytochrome c Oxidase I.

3.2. Alpha Diversity

Rarefaction curves indicated that approximately 413 samples (16S), 502 samples (18S) and 487 samples (COI) were needed to recover 99% of ASV richness, with each marker's curve approaching an asymptote at these sampling efforts (Figure 2).

FIGURE 2.

Amplicon sequence variants (ASVs) accumulation curves for each of the three markers used with this dataset. Shaded areas around each curve indicate standard deviation estimates derived from 100 permutations. Vertical dashed lines indicate the number of samples at which 99% of the ASV richness is achieved. 16S = bacterial ribosomal 16S rRNA, 18S = eukaryotic ribosomal 18S rRNA, COI = metazoan mitochondrial Cytochrome c Oxidase I.

The results of the mixed‐effects models revealed significant drivers of ASV richness across all three marker datasets, with some consistent patterns among markers (Table 4). Random effects analysis revealed that the vessel ID consistently explained most of the random variation across all datasets (bacterial 16S: 99.9%; eukaryotes 18S: 69.8%; metazoan COI: 98.4% of total random variance, respectively), while sequencing run effects were minimal (less than 30% for all), indicating a higher source of uncontrolled variation in sampling (vessel) compared to laboratory procedures. All three datasets showed a significant and positive relationship between ASV richness and SST and chlorophyll‐a concentrations. The estimated slopes for all markers and both parameters were similar (range: 0.10–0.17) (Table 4). All datasets displayed a positive correlation between ASV richness and vessel speed, but it was only significant for the 16S (estimate = 0.183, p = < 0.001) and COI markers (estimate = 0.103, p = 0.001). All datasets showed a significant relationship with water column depth. This relationship was positive for the COI marker (estimate = 0.084, p = 0.022), but negative for both the 16S and 18S (16S: estimate = −0.1, p < 0.001; 18S: estimate = −0.098, p < 0.001). Plots depicting the correlations from this analysis are shown in the Supporting Information, Figure S1.

TABLE 4.

Summary of generalized linear mixed‐effects models examining different drivers of amplicon sequence variant (ASV) richness across three metabarcoding datasets.

Marker	Parameter	Estimate	SE	p
Bacterial 16S AIC: 8463.4 BIC: 8498.1 N = 569	Random effects
	Seq_run (variance)	0.006
	Vessel (variance)	0.323
	Fixed effects
	Vessel speed	0.183	0.032	< 0.001
	SST	0.123	0.035	0.003
	Chlorophyll‐a	0.124	0.036	< 0.001
	Water column depth	−0.138	0.035	< 0.001
Eukaryotic 18S AIC: 6702.4 BIC: 6737.7 N = 546	Random effects
	Seq_run (variance)	0.017
	Vessel (variance)	0.042
	Fixed effects
	Vessel speed	0.048	0.025	0.058
	SST	0.162	0.030	< 0.001
	Chlorophyll‐a	0.115	0.030	< 0.001
	Water column depth	−0.097	0.028	< 0.001
Metazoan COI AIC: 6076.3 BIC: 6110.6 N = 555	Random effects
	Seq_run (variance)	0.008
	Vessel (variance)	0.151
	Fixed effects
	Vessel speed	0.103	0.032	0.001
	SST	0.145	0.034	< 0.001
	Chlorophyll‐a	0.122	0.034	< 0.001
	Water column depth	0.083	0.037	0.023

Note: Models were fitted using negative binomial distributions with log(seq_count) as an offset term to control for sequencing depth variation. Random effects included sequencing run and vessel ID (nested within sequencing run) to account for technical and methodological variation. Fixed effects included vessel speed (knots), sea surface temperature (SST), water column depth and log‐transformed chlorophyll‐a concentration (μg/L). Akaike Information Criterion (AIC), Bayesian Information Criterium and number of observations (N) are shown for each model. Standard errors (SE) and p‐values (p) are shown for each of the fixed parameters. All fixed effects were scaled, thus the reported parameter estimates are a measure of their relative importance.

3.3. Community Composition and Structure

The relative abundance (proportion of reads) and diversity (ASV richness) of different phyla assigned to the 16S, 18S and COI metabarcoding datasets as a function of latitude are depicted in Figure 3. Briefly, for the bacterial 16S, the most abundant phyla included Proteobacteria, Bacteroidota and Cyanobacteria, while the most diverse phyla were Pseudomonata, Bacteroidota and Bdellovibrionata. For the eukaryotic 18S, the most abundant phylum was Arthropoda, followed by ASVs that could not be assigned to a phylum and were displayed as the Eukaryota domain (including all eukaryotes unidentified at lower taxonomic levels and eukaryotes identified at lower taxonomic levels but without a recognized phylum category in the NCBI taxonomy, such as brown algae, dinoflagellates, etc.), and subsequently by Chordata and Cnidaria. In contrast, the most diverse group was composed of all ASVs assigned to the Eukaryota domain, followed by Arthropoda and Ciliophora. For the metazoan COI, the most abundant phylum was Arthropoda, followed by ASVs assigned to the Eukaryota domain, Cnidaria and Mollusca. In terms of ASV diversity, Arthropoda had the highest ASV richness, followed by the Eukaryota domain, Bacillariophyta and Cnidaria.

FIGURE 3.

Relative abundance of sequence reads (top row) and amplicon sequence variants (bottom row) for each of the three markers (columns) of different phyla as a function of latitude. For COI and 18S, all ASVs that could not be assigned to a phylum were grouped at the higher taxonomic level Eukaryota domain.

The results of the PCA revealed clear, similar patterns for all three datasets (Figure 4). The first two principal components captured between 18% (bacterial 16S) and 10.2% (eukaryotic 18S) of the overall variation in the dataset. The PCA ordination showed a clear biogeographic pattern in community structure for all three datasets. While there were no clear clusters of samples in any of the three datasets, a clear sea‐surface temperature (SST) gradient was observed along the first axis of variation for the 16S and COI datasets, and along the second axis for the 18S dataset (Figure 4). There was no sign of sample clustering by vessel ID (based on the visual inspection of the PCA ordination plot when samples were colored according to vessel ID). Environmental fitting analyses revealed that vessel speed, SST, chlorophyll‐a and water column depth were significantly correlated with community variation for all three datasets. Specifically, SST exhibited the highest correlation coefficients, ranging between 0.53 and 0.73 across datasets, while the correlations with chlorophyll‐a and water column depth were more modest, ranging from 0.17 to 0.321 for the former, and from 0.08 to 0.14 for the latter. Correlations with vessel speed were minimal, ranging between 0.01 and 0.06. Details of the environmental fitting analyses are provided in Table 5. The results of the RDA indicated that, after removing the effects of sequencing run and vessel ID, the proportion of community composition variance explained by the three environmental variables was modest for all metabarcodes (adjusted R ² 16S = 11.5%, 18S = 4.01%, COI = 6.37%), with SST explaining most of the constrained variation (adjusted R ²: 2.8%–9%), and vessel speed explaining the least amount of constrained variation (adjusted R ²: 0.1%–0.3%). The details of individual adjusted R ² values and statistical significances are provided in Table 5. Similar patterns were observed when PCAs were run separately for phytoplankton and copepod (class Hexanauplia) datasets (Supporting Information, Figures S2, S3).

FIGURE 4.

Principal component analyses (PCAs) for three different metabarcoding assays. Each PCA was performed on centered log‐ratio–transformed amplicon sequence variant (ASV) abundance data. Each point represents a sample, coloured according to sea‐surface temperature (SST) (top row), chlorophyll‐a concentration (bottom row). Both SST and chlorophyll‐a concentrations were obtained from satellite data. The total variance explained by each of the first two principal components (PC1 and PC2) is indicated in parentheses.

TABLE 5.

Results of the environmental fitting analyses for each of the three datasets.

Marker	Variable	PC1	PC2	r 2	p	Adj‐R 2	p
Bacterial 16S	SST	−0.999	−0.026	0.738	< 0.001	0.090	< 0.001
	Chlorophyll‐a	0.482	−0.875	0.321	< 0.001	0.030	< 0.001
	Water column depth	−0.751	−0.659	0.111	< 0.001	0.003	< 0.001
	Vessel speed	0.083	0.996	0.010	0.042	0.001	< 0.001
Eukaryotes 18S	SST	−0.188	0.982	0.532	< 0.001	0.028	< 0.001
	Chlorophyll‐a	−0.401	−0.916	0.169	< 0.001	0.013	< 0.001
	Water column depth	−0.460	0.888	0.082	< 0.001	0.002	< 0.001
	Vessel speed	0.987	−0.155	0.011	0.036	0.001	< 0.001
Metazoan COI	SST	−0.59	0.807	0.727	< 0.001	0.046	< 0.001
	Chlorophyll‐a	−0.099	−0.995	0.210	< 0.001	0.018	< 0.001
	Water column depth	−0.977	0.213	0.137	< 0.001	0.002	< 0.001
	Vessel speed	0.880	0.474	0.065	< 0.001	0.003	< 0.001

Note: Columns PC1 and PC2 indicate the coordinates of the direction for each variable. The r ² column shows the correlation coefficient of each variable with that particular component. SST denotes sea‐surface temperature. R ² indicates the explanatory power fraction of the total variance explained by each variable after removing the effects of sequencing run and vessel ID obtained from the redundancy analysis. The p‐value columns indicate permutation‐based statistical significance of r ² (envfit) and adjusted R ² (Redundancy Analysis; RDA).

3.4. Phytoplankton‐Satellite Correlations

The relative read abundance and ASV richness of different phytoplankton classes assigned to the datasets for the markers 18S and COI and the genera assigned to the dataset for the bacterial 16S marker as a function of latitude are depicted in Figure 5. Briefly, the most abundant and diverse genera included Prochlorococcus, Trichodesmium and Synecococcus. For the eukaryotic 18S, the class Dinophyceae represented over 95% of the total phytoplankton reads and total phytoplankton ASVs. For the metazoan COI, the most abundant class was Dinophyceae, followed by ASVs assigned to the class Bacillariophyceae and phylum Bacillariophyta. In terms of ASV diversity, Dinophyceae had the highest ASV richness, followed by the class Bacillariophyceae and phylum Bacillariophyta. For both COI and 18S markers, haptophytes and green algae had minimal relative read abundance and ASV richness.

FIGURE 5.

Relative abundance of reads (top row) and amplicon sequence variants (bottom row) for each of the three markers (columns) of different phytoplankton groups as a function of latitude.

We explored the relationship between satellite‐derived chlorophyll concentrations (total and specific phytoplankton functional types) and four metrics of abundance of corresponding phytoplankton groups derived from eDNA. A summary of the generalized linear mixed modelling (GLMM) analyses is presented in Figure 6A. For all comparisons involving eukaryotic groups, the best correlations were achieved with ASV richness metrics. In contrast, for prokaryotes, the best correlation was achieved with relative read counts (Figure 6A). For total chlorophyll (CHL), all eDNA metrics from the 18S marker displayed a negative relationship with satellite chlorophyll concentrations. In contrast, COI metrics were positively correlated, with relative read abundance showing a weak correlation (R ² = 0.018, p‐value < 0.001), compared to all three ASV metrics which produced similar correlations: ASV relative richness marginal R ² = 0.068, (Figure 6B); ASV richness marginal R ² = 0.093; ASV richness per metre marginal R ² = 0.102, all with p‐value < 0.001. Of the six phytoplankton functional types, four displayed positive and significant relationships with at least one of the eDNA‐derived metrics (diatoms, green algae, haptophytes and photosynthetic prokaryotes). The remaining two (dinoflagellates and Prochlorococcus) showed only negative relationships (Figure 6A). For diatoms (DIATO), the highest marginal R ² (0.03, p‐value < 0.001) was achieved with the 18S dataset and relative ASV richness. For green algae (GREEN), the highest marginal R ² (0.122, p‐value < 0.001) was achieved with the COI marker and ASV richness. For Haptophytes (HAPTO), the best marginal R ² (0.04, p‐value < 0.001) was achieved with ASV richness per metre and the COI marker (Figure 6B). Interestingly, the relationship between relative read abundance of photosynthetic prokaryotes (PROKAR) and prokaryote chlorophyll satellite concentrations had the best fit (marginal R ² = 0.272, p‐value < 0.001) (Figure 6B). For all models except those including CHL, GREEN and PROKAR, conditional R ² values (variance explained by fixed and random effects) were substantially higher than marginal R ² (Figure 6A), indicating that technical factors (sequencing run and vessel ID) explained considerable variance beyond the biological predictors.

FIGURE 6.

(A) Summary of GLMMs analyses results. Different shades for each bar indicate the marginal (dark) and conditional (light) R ² estimates for each model. y‐axis labels correspond to: Relative read abundance (Rel_Reads), Relative ASV richness (Rel_ASVs), ASVs per meter (ASVs_stand) and ASV richness (ASVs). Marginal R ² refers to the variance explained by fixed effects only. Conditional R ² refers to the combined variance of fixed and random effects. Red bars indicate negative relationships between satellite and eDNA‐derived data. Blue bars indicate a positive relationship between satellite and eDNA variables. (B) Relationships between satellite‐derived data (y‐axis) and eDNA data (x‐axis) and the model regression prediction (blue line) for total chlorophyll (CHL) and four functional types (DIATO, Diatoms; DINO, dinoflagellates; GREEN, green algae; HAPTO, haptophytes; PROCHLO, prochlorococcus; PROKAR, photosynthetic prokaryotes) that showed a positive relationship. For simplicity, only the highest R ² positive regressions for each functional type are shown. 16S, 18S and COI refer to metabarcoding assays (see main text for details).

4. Discussion

Ocean biodiversity is under increasing threat from accelerating climate and human pressures (Duarte 2014; Dybas 2006; Sala and Knowlton 2006; Talukder et al. 2022). Yet the capacity to monitor these changes at relevant spatial and temporal scales remains a critical bottleneck. Our findings show that the Citizens of the Sea initiative can overcome key limitations of traditional marine biodiversity surveys by leveraging a trained fleet of citizen sailors, a standardized and scalable sampling protocol and high‐throughput molecular tools. The Citizens of the Sea partnership achieved extensive spatial coverage compared to traditional scientific expeditions, which are typically constrained by complex logistics and high costs. Similarly, we show that the data gathered by seafarers meet high standards of scientific quality and, to our knowledge, provide the highest‐resolution ‘picture’ at a basin scale of this part of the Pacific Ocean to date. This dataset therefore represents one of the first comprehensive open ocean eDNA baseline assessments of marine biodiversity in this region and, as such, will be a cornerstone for monitoring biodiversity changes in the South West Pacific Ocean in the coming years. Below, we discuss the patterns of our eDNA results on the spatial distribution of biodiversity, possible citizen‐science effects and the potential and limitations of this approach for satellite ground‐truthing. We then conclude by outlining a roadmap for future sampling strategies, as well as data management and sharing options to maximize the impact of this endeavour.

With more than 700 usable samples collected over a 6 months period, this dataset exceeds previous endeavours such as Plankton Planet (214 samples in 1 year; de Vargas et al. 2022) or the Tara Expeditions (147 sampling sites in 3 years, although the number of samples per site was much higher as they also sampled across several depths; Bork et al. 2015). To date, it is also one of the most comprehensive sampling eDNA campaign in terms of taxonomic coverage at the ocean basin level. Our analyses revealed that for three markers (16S, 18S and COI), ASV rarefaction curves plateaued between 500 and 600 samples. This suggests that our sampling strategy and the number of sailing boats involved (n = 26, post filtering) were sufficient to obtain a comprehensive characterization of the biodiversity of surface waters that is detectable using these markers and sampling technology across this oceanic region. In general, the results reinforce the notion that only citizen‐science approaches can yield sampling designs that are comprehensive and affordable at open ocean spatial scales.

We found significant variation among samples that can be attributed to human procedures. Specifically, a significant proportion of the variation in ASV diversity was linked to the vessel identity and the vessel speed. Significant differences in ASV richness among vessels were expected and can be attributed to differences in on‐board processing behaviours (de Vargas et al. 2022). The positive correlation between ASV richness and vessel speed can, in turn, be explained by larger volumes of water filtered at higher vessel speeds and longer distances of sampling transects. Similar large proportions of variation explained by vessel differences (human manipulation) were found in the eDNA satellite correlations. In the future, it is important to take this source of variation into consideration. However, our statistical modelling was able to accommodate this variation, and after controlling for it, we were able to detect significant environmental patterns. The most predominant patterns included a positive correlation between ASV richness and sea‐surface temperature, and with chlorophyll‐a across all three markers. This pattern is consistent with previous reports that described latitudinal gradients of plankton diversity using molecular (de Vargas et al. 2022; Ibarbalz et al. 2019; Pommier et al. 2007; von Ammon et al. 2020) and morphological data (Chust et al. 2013), with diversity increasing from temperate to tropical waters in the southern hemisphere and thus highlights the quality of the data obtained despite the noise associated with human procedures.

Within the molecular phytoplankton community data, we found marked latitudinal differences in the relative abundance of the two most abundant photosynthetic bacterial genera (Trichodesmium and Prochlorococcus). The observed higher relative abundance of Trichodesmium towards the tropics, and of Prochlorococcus towards 25°–30°S are consistent with known thermal tolerance ranges previously reported for the genera (Smith et al. 2021; Stal 2009). However, niche models based on temperature and photosynthetically active radiation predict monotonically increasing cell abundance of Prochlorococcus towards the tropics (Flombaum et al. 2013), while our results indicate a relative decay towards the tropics. This discrepancy may reflect several factors not captured by temperature‐light models alone, including nutrient co‐limitation in oligotrophic tropical waters (Moore et al. 2013), competitive exclusion by nitrogen‐fixing Trichodesmium spp. (Snow et al. 2015) or ecotype‐specific distributions of high‐light versus low‐light adapted Prochlorococcus strains (Thompson et al. 2021). Additionally, eDNA‐based relative abundances may not directly correlate with cell counts due to the inherent nature of metabarcoding (Deagle et al. 2019). We anticipate that this dataset will provide a valuable resource to help fine‐tune such modelling approaches by incorporating eDNA observations alongside traditional oceanographic parameters.

While previous studies using both traditional and molecular analyses have reported weak or absent latitudinal structuring of plankton communities at the basin scale (Chust et al. 2013; Pommier et al. 2007), our results reveal a clear and consistent latitudinal gradient in community composition across all three markers, which remained evident even when specific taxonomic groups such as copepods and phytoplankton were analysed separately. To our knowledge, only one previous study has characterized open ocean planktonic community structure across this ocean basin using 18S metabarcoding (v9 region), reporting three discrete community clusters and relatively weak biogeographical structuring across five sampling vessels and three legs of the voyage (de Vargas et al. 2022). Taken together, these findings suggest that increased sampling density reveals a more continuous biogeographical turnover in plankton communities, structured along environmental gradients such as latitude and chlorophyll‐a, rather than discrete assemblages. If sustained over time, this unprecedented citizen‐powered sampling framework holds immense potential for detecting biodiversity shifts at spatio‐temporal scales, capturing signals of seasonal plankton dynamics, species range expansions, invasive introductions, or climate‐driven community reorganization across ocean basins.

Correlations between satellite Phytoplankton Functional Types (PFTs) and eDNA‐based metrics were generally weak (all R ² < 0.3). Only three relationships exceeded 10% of explained variance by the eDNA metrics (fixed factors): the correlation between photosynthetic prokaryote relative read abundance from 16S data and satellite‐derived prokaryote chlorophyll concentrations, and the correlations between photosynthetic eukaryote relative ASV richness and satellite total chlorophyll concentrations, and between green algae ASV richness and satellite green algae chlorophyll concentrations (both from COI data). These findings suggest that photosynthetic prokaryotes, particularly cyanobacteria, exhibit a better correspondence between eDNA and satellite data than eukaryotic groups. This pattern is consistent with a recent controlled mesocosm study that reported positive correlations between fluorescence‐based biomass and both DNA‐normalized read counts and ASV richness for cyanobacteria, but not for eukaryotic phytoplankton (Salis and Hansson 2025).

Among eukaryotic markers, COI consistently outperformed 18S in alignment with satellite‐based chlorophyll estimates. Correlations between 18S‐derived metrics and satellite chlorophyll were weak and often negative, likely reflecting multiple confounding factors. First, high variability in 18S rRNA gene copy number across taxa, especially among dinoflagellates, can bias relative abundance estimates and obscure biomass signals (J. Liu and Zhang 2021; Y. Liu et al. 2021; Stuart et al. 2024). Second, although the V4 region provides better taxonomic resolution than V9, it performs poorly at capturing relative abundance patterns when compared with traditional pigment‐based or microscopy‐derived assessments (Stuart et al. 2024; Zimmermann et al. 2024). In this context, the 18S‐V9 region may be better suited for abundance estimation (though it has the same copy‐number variability limitation as the V4) and warrants testing in future analyses of this dataset. Third, dinoflagellates, the dominant group in both 18S and COI phytoplankton reads, showed no positive correlations with satellite observations, potentially due to their variable pigment expression (Zapata et al. 2012), complex mixotrophic behaviour (Stoecker et al. 2017), or other traits that decouple biomass from chlorophyll signals (Dierssen 2010). Further research to improve our understanding of the physiology and ecology of these phytoplankton groups is needed to refine eDNA estimates of their photosynthetic contribution and help improve marine ecosystem monitoring using ocean colour remote sensing.

We recognize several methodological limitations that may have contributed to the weak eDNA‐Satellite correlations. At the biological level, the spatial scales of comparison are vastly different, with molecular analysis occurring on a small fraction of the filtered water collected and being compared to a large swath of ocean (representative of thousands of cubic metres) observed in the 5 km × 5 km satellite pixels. Methodologically, standardizing ASV richness by transect length (metres) may be suboptimal, as suggested by the minimal improvements in model fit; more accurate estimation of filtered water volume would likely improve model performance. Additionally, a significant portion of unexplained variance may stem from uncertainty in the satellite‐derived PFT products themselves. Many of the functional groups we examined, particularly green algae, haptophytes, and dinoflagellates, are associated with high uncertainty in remote sensing models (see Figure S4) (Xi et al. 2021). Notably, the PFTs with the most robust eDNA correlations (total chlorophyll and prokaryotes) also had the lowest satellite uncertainty estimates in this region, suggesting that improvements in both molecular sampling protocols and satellite algorithm refinement are needed for better integration.

Despite these challenges, our findings demonstrate that citizen‐driven eDNA sampling can contribute meaningfully to the refinement of satellite‐based phytoplankton models. Spectral differentiation of PFTs is a rapidly evolving field, and with the launch of the NASA PACE (Plankton, Aerosol, Cloud, ocean Ecosystem) mission, regular hyperspectral measurements of the global oceans have become available (Dierssen et al. 2023) and represent a major advance over limited multi‐band spectral data. Associated with these new measurements are new approaches for differentiating phytoplankton groups (e.g., Bracher et al. 2022; Kramer et al. 2022, 2024). As eDNA ocean‐monitoring datasets become larger, and better hyperspectral measurements and algorithms to infer PFTs are developed, future satellite observations may help estimate the representativeness of smaller‐scale molecular sampling. While the present correlations are modest, they highlight the importance of improving molecular reference data and reducing satellite model uncertainty (Cetinić et al. 2024). With sustained sampling and targeted methodological refinements, this approach has strong potential to support the calibration and validation of large‐scale coupled ocean–ecosystem models and advance the integration of molecular and remote sensing tools for ocean observation.

5. Conclusion

The physical and biochemical properties of the ocean are changing at an unprecedented pace, reshaping marine ecosystems in ways we are only beginning to understand. Yet, our capacity to monitor these changes remains limited, especially across the vast expanse of the open ocean. Just as medicine relies on physiological metrics to assess patient health, biodiversity observations are vital for tracking the ocean's ecological condition. This study demonstrates that with the support of trained citizen scientists, it is now feasible to collect high‐resolution, basin‐scale biodiversity data using standardized eDNA protocols. While we found only moderate correlations between eDNA and satellite‐derived phytoplankton functional types, this highlights both the limitations of current satellite products and the promise of molecular tools to ground‐truth remote observations. With continued investment and refinement, citizen‐powered molecular sampling can help calibrate ocean‐colour models, improve ecosystem forecasting, and inform conservation policy. Crucially, this approach is scalable in both space and time, offering a path towards globally distributed, repeatable ocean biodiversity monitoring. Sustained, long‐term efforts will be essential to track ecological baselines, detect regime shifts and guide adaptive management as ocean conditions continue to evolve.

Author Contributions

X.P. and E.B. designed the research, all authors performed the research, and X.P. secured funding for the research. E.B. led the logistics and coordination of citizen sailors and sampling. P.S.‐A., O.L. and B.K. performed the remote sensing extractions, bioinformatics and/or statistical analyses. P.S.‐A. wrote the first draft of the manuscript, with assistance from X.P., and all authors contributed revisions to produce the final version.

Funding

This work was funded by Minderoo Foundation Ltd. (collaborative grant# CLB‐3024), KiwiNet (grant # CW002641), with additional in‐kind contributions from Illumina Inc., the Cawthron Institute, and the Citizens of the Sea charitable trust.

Disclosure

Benefit‐Sharing Statement: This research is the product of a collaboration among citizen sailors, scientists, and government authorities from three different countries. The results of this research have been shared with the participating sailing community, the government authorities, and all biodiversity data generated have been made available to the public by sharing them in the GBIF eDNA archive. This research is the first step towards a long‐term citizen science‐driven open ocean biodiversity monitoring framework.

Ethics Statement

Sampling complied with local regulations and institutional guidelines. The collection of seawater filtrates was undertaken under research permits from Fiji (Collection permit #TPN 546/2025), Tonga (Collection permit # F3/14/1146/24) and New Zealand (Collection permit #SP822).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Plots illustrating the significant correlations between ASV richness for each marker (columns) and different environmental variables (rows). Blue dots correspond to empirical data, with values in the x‐axis standardized to mean = 0 and variance = 1. Black lines and grey shades show the GLMMs regression predictions. See the main text for details on the model.

Figure S2: Principal component analyses (PCAs) including only phytoplankton taxa for each of three different metabarcoding assays. Each PCA was performed on centred‐log‐ratio‐transformed amplicon sequence variant (ASV) abundance data. Each point represents a sample, coloured according to sea‐surface temperature (SST) (top row), chlorophyll‐a concentrations (bottom row). Both SST and chlorophyll‐a concentrations were obtained from satellite data. The total variance explained by each of the first two principal components (PC1 and PC2) is indicated in parentheses.

Figure S3: Principal component analyses (PCAs) including only copepod taxa for 18S and COI metabarcoding assays. Each PCA was performed on centred‐log‐ratio–transformed amplicon sequence variant (ASV) abundance data. Each point represents a sample, coloured according to sea‐surface temperature (SST) (top row), chlorophyll‐a concentrations (bottom row). Both SST and chlorophyll‐a concentrations were obtained from satellite data. The total variance explained by each of the first two principal components (PC1 and PC2) is indicated in parentheses.

Figure S4: Bar plot showing the distribution of uncertainty estimates for satellite observations of each of the phytoplankton functional groups. Individual data points were overlaid on top of boxplots for reference.

Data Availability Statement

Raw sequence reads have been deposited in the NCBI SRA archive under project: PRJNA1456174. All related sample metadata and associated ASV curated data have been deposited in GBIF (Citizens of the Sea: Pacific Rally 2024 eDNA Metabarcoding Dataset, https://doi.org/10.15468/d7wkp9).