1 Introduction

Understanding patterns of marine fish biodiversity is crucial as species richness considerably influences ecosystem functions, such as nutrient cycling and productivity (Thomsen et al. 2012; Gamfeldt et al. 2015). Assessing taxonomic diversity is vital for developing effective conservation and fishery management strategies (Różański et al. 2022), especially in areas with complex oceanic fronts (Saito 2019) and eddies (Xing et al. 2023), which act as biogeochemical hotspots by altering nutrient distributions and primary production (Woodson and Litvin 2015). These dynamics promote high productivity and growth for marine fish and their larvae (Bakun 2006).

Pelagic fish communities in ocean front ecosystems can serve as indicators of biodiversity hotspots (Miller and Christodoulou 2014) and play crucial roles in ecosystem functioning (Watari et al. 2019).

However, monitoring of marine surface biodiversity in pelagic environments remains limited (Martin et al. 2020), and a significant portion of pelagic fish species’ diversity and community composition is unknown (Higgs and Attrill 2015). This gap in sampling effort is largely due to the challenges associated with the diversity in pelagic fish size, behavior (Li et al. 2022), and distribution, which are often linked to environmental parameters (Yu et al. 2023). Marine pollution (Montenegro et al. 2020) and global climate change, which shift the distribution of marine life (Poloczanska et al. 2013), further complicate this issue. Additionally, the negative impacts of human activities on marine biodiversity, such as overfishing and high by-catch rates, have been well-documented (Jackson et al. 2001; Sandoval Gallardo et al. 2021), underscoring the need for ongoing biodiversity monitoring and community observations (Miloslavich et al. 2018).

The traditional sampling methods, such as trawling, or netting, face additional challenges when applied to marine surface fish communities. They are often ineffective at capturing rare species (Schmelzle and Kinziger 2016) thus limiting the taxonomic resolution of entire communities (Jac et al. 2021). Moreover, these traditional methods can be labor-intensive, costly, and invasive (Evans et al. 2017; Zhang et al. 2020), contributing to a significant knowledge gap about the biodiversity of many marine ecosystems (Claudet and Fraschetti 2010). Consequently, large-scale, non-invasive monitoring technologies are urgently needed to assess the biodiversity and community composition of poorly understood marine surface communities (Kennedy et al. 2020).

Environmental DNA (eDNA) has emerged as a valuable tool for ecosystem monitoring, providing a rapid, non-invasive, and cost-efficient approach to biodiversity assessment (He et al. 2023; Gold et al. 2021). eDNA enhances species detection by over 30% compared to traditional sampling methods and offers insights into fine-scale spatial and temporal variations within marine surface communities (Stat et al. 2019). As a comprehensive method, eDNA sampling holds promise for revolutionizing how marine biodiversity is monitored and understood (Afzali et al. 2021; Miya 2022). However, eDNA metabarcoding success depends on factors such as primer selection (Schenekar et al. 2020), methodology (Sakata et al. 2021), choice of bioinformatic pipeline (Brandt et al 2021), and contamination control (Andruszkiewicz et al. 2017; Fujii et al. 2019). The MiFish primer is widely used for fish species and has been shown to outperform other primers (Collins et al. 2018; Zhang et al. 2019), amplifying and identifying fish species’ eDNA by targeting a specific mitochondrial DNA region (Miya et al. 2015). However, the open-ocean application of this primer requires careful consideration of factors such as sampling effort (Kawakami et al. 2023) and the prevailing environmental context (Yu et al. 2023). Additionally, due to the relatively low quantity of DNA quantities found in ocean surface waters, sampling methods need to be optimized to increase the likelihood of the MiFish primer being effective (Yu et al. 2022).

Like all biodiversity sampling techniques, eDNA metabarcoding is subject to various biases and limitations (Pompanon et al. 2012; Fonseca 2018; Jo et al. 2019). In marine environments, the standard eDNA metabarcoding approach typically involves filtered water replicates collected from the water column (Stat et al. 2017). It has been argued that employing multiple sampling methods concurrently is needed to generate the most comprehensive assessment of biodiversity (Alexander et al. 2023). Therefore, MiFish metabarcoding studies usually sample the water column with Niskin bottles—typically deployed at depths starting from 10 m (Yu et al. 2022; Alexander et al. 2023); bucket samples collected from the sea surface (Miyata et al. 2022); and ship-bottom intake samples drawn from the vessel’s laboratory faucet, located several meters below the surface (Kawakami et al. 2023). Each method has advantages and disadvantages: the intake method allows for continuous sampling along the ship tracks, while the Niskin bottle and bucket methods require the vessel to stop (Yu et al. 2022). It is relatively easy to keep the Niskin bottles and buckets clean, whereas maintaining a clean environment with the intake method is more challenging. Additionally, the intake method may face risks from metal ion obstructions (Greco et al. 2022; Dumoulin et al. 2024). That said, all methods are vulnerable to the impacts of calcium ions that—being alkaline earth metals—can interfere with eDNA detection (Dumoulin et al. 2024).

The ocean’s surface layers are relatively well mixed by wind and waves and so eDNA is expected to be homogeneously distributed within this layer if there are no biases in the vertical distribution of fish (an example is Littlefair et al. 2020, although they focused on a lake). All bucket, intake, and 10 m Niskin bottle samples are expected to be within the mixed layer, and thus sampling the same eDNA field. In our previous research (Ahmed et al., 2025, submitted), we investigated species-specific fish detection in the western North Pacific—a region marked by dynamic oceanic fronts where the warm Kuroshio and cold Oyashio currents meet, creating a wide temperature range between the Kuroshio Extension and Oyashio Front (Yasuda 2003; Kida et al. 2015). We found that the bucket, intake, and Niskin sampling methods showed comparable detection performance for frequently detected species, as indicated by relative read abundance (RRA). However, some species such as the Japanese jack mackerel (Trachurus japonicus) and Pacific saury (Cololabis saira) showed better RRA in bucket samples. Nonetheless, the overall observed similarity in detection performance across methods suggests that combining data generated from different water collection methods on different occasions may be a viable strategy to generate larger spatio-temporal scale analyses of marine surface biodiversity. Still, it remains to be formally tested whether the bucket, intake, and Niskin methods actually capture similar types of biodiversity and fish community composition.

A meaningful comparison of eDNA methods requires each protocol to be deployed concurrently to sample the same parcel of water. However, even for such a presumably exact comparison, the sparse distribution of eDNA in the ocean may mean that sample-to-sample differences eventuate by chance (Kawakami et al. 2023). Here, we reduce the influence of single sample anomalies by comparing a large dataset of concurrently collected water samples across three methods compiled from multiple cruises under varying conditions. This approach enabled us to investigate whether using different common surface water sampling methods affect the detection of pelagic fish species and community composition within the well-mixed surface layer, where environmental conditions are generally uniform at a given time point.

2 Materials and methods

2.1 Study sites

The samples were collected from diverse habitats across the western North Pacific, including the Kuroshio (subtropical) region, Kuroshio Extension, the mixed water region (transition zone between the Kuroshio Extension and Oyashio Front), the Tsugaru Warm Water Current region, and the Oyashio (subarctic) area (Fig. 1). In addition, a single sample station was located in the Sea of Japan. Sampling was conducted at various stations, yielding a total of 83 sample sets during eight research cruises aboard the R/V Shinsei Maru and one cruise on the R/V Hakuho Maru. Of these, 52 sample sets were obtained from Ahmed et al. (2025, submitted), while 31 additional sets from the KS-22-15 and KH-23-3 cruises were added in this study. Each sample set consisted of samples collected using three different methods: bucket, intake, and Niskin. Table 1 summarizes the cruises, specifying the sampling months and number of stations, and Fig. 1 shows the station locations. The data were also used to analyze species-specific detection performance across sampling methods in the surface layer (Ahmed et al., 2025, submitted), biogeographic boundaries at the Takara Gap (Inoue et al., 2025, submitted), and fish species composition, including subsurface layers (Lin et al., 2025, submitted).

Fig. 1
figure 1

Distribution of sampling sites in the western North Pacific, including a single site in the Sea of Japan. Different colored circles represent sampling sites from various research cruises: KS-22-15 (October 2022), KS-21-11 (June 2021), KS-21-12 (June-July 2021), KS-21-8 (May 2021), KS-21-3 (March 2021), KS-18-5 (May 2018), KS-22-11 (August 2022), KH-23-3 (July 2023), and KS-21-24 (October 2021). To distinguish the overlapping sampling locations of the KS-21-24 and KS-22-11 cruises, KS-21-24 sites are marked with 'X' symbols, while KS-22-11 sites are indicated with circles. Red arrows highlight warm ocean currents, while blue arrows show cold currents. 'TWC' refers to the Tsugaru Warm Current

Full size image
Table 1 Summary of research cruises including the number of stations where samples were collected using all three sampling methods (bucket, intake, and Niskin)
Full size table

Each cruise had distinct research objectives, leading to an uneven distribution of sampling points across various seasons. We joined these cruises to collect samples from a broad range of sites in the western North Pacific, allowing for a comprehensive assessment of marine surface biodiversity and community composition using different eDNA sampling methods under varied environmental conditions.

2.2 Sample collection

At each station, surface water samples were collected using three methods. Niskin bottles (hereafter referred to as “Niskin”) were deployed at depths below 5 or 10 m (depth depending on the season), while surface water was collected at 0 m using a clean bucket (hereafter referred to as “bucket”) dropped directly over the side of the ship. Additionally, water was sampled via the ship’s seawater intake system (hereafter referred to as “intake”), which collects water from approximately 4.5 m below the surface through a designated seawater faucet. Niskin bottles were equipped with a conductivity temperature depth (CTD) system to record various environmental parameters.

At each station, approximately 10 L of water was collected per sample simultaneously using all three methods. The exception was during the KS-18-5 cruise, where intake samples were collected at a different time to Niskin and bucket samples due to underway CTD observation. Here, about 7 L was collected per sample in each method. During the KH-23-3 cruise, samples from all three methods were collected from two stations every 4 h to compensate for the limited number of observation sites.

Prior to sample collection, clean plastic sampling containers were washed three times with the sampled seawater. Each sample was then weighed and filtered using a Sterivex-HV pressure filter unit with a 0.45 μm pore size, following the procedure outlined by Yu et al. (2022). During the KS-18-5 cruise, a 0.22 μm pore-size filter of the same type was used. After filtration, the filter units were filled with 2 mL of RNAlater using a disposable syringe to preserve the DNA and then stored at − 25 °C.

To ensure accurate seawater sampling and assess potential contamination during onboard handling, two negative control samples consisting of 1.5 L of Milli-Q water were collected at the start and end of each cruise. The Milli-Q water was generated on board and filtered following the same procedure used for the seawater samples. The KS-18-5 cruise, conducted in 2018, was among the early applications of eDNA metabarcoding in open ocean studies. At that time, the necessity of including negative controls was not fully recognized, as also highlighted by Sepulveda et al. (2020), and therefore, they were not collected. Filtration took place exclusively in a designated area of the research vessel’s laboratory. All equipment were thoroughly cleaned using a 1% bleach solution and rinsed with Milli-Q water to eliminate any bleach residue before each use, minimizing cross-contamination risk. Fresh gloves and masks were worn during both water collection and filtration to further reduce contamination potential.

2.3 DNA extraction and purification procedures

After samples were transported to land, eDNA extraction and purification for all cruises were performed at the Molecular Marine Biology Laboratory, Atmosphere and Ocean Research Institute, The University of Tokyo, using a slightly modified version of the Miya et al. (2015) protocol, as described by Wong et al. (2020). For the KS-18-5 cruise, extraction followed the original Miya et al. (2015) protocol and was conducted at Bioengineering Lab. Co., Ltd., Kanagawa, Japan. To minimize contamination risk, all equipment, including silicon tubes and connectors, was sterilized with a 1% bleach solution and thoroughly rinsed with Milli-Q water before each use.

The samples were processed in a dedicated eDNA laboratory reserved exclusively for molecular work, with no exposure to other tissue samples or PCR products. Stringent hygiene protocols were observed, including the use of new gloves and masks for each procedural step. The workspace was regularly disinfected with 0.1% bleach spray to maintain a contamination-free environment.

2.4 MiFish eDNA metabarcoding

Following DNA extraction, eDNA metabarcoding was conducted using the MiFish primer, which amplifies a ~ 170 bp fragment of the mitochondrial 12S rRNA gene, widely used for fish species detection due to its broad taxonomic coverage and resolution (Miya et al. 2015). The libraries were prepared via a two-step PCR (Miya et al. 2020), and sequencing was performed using Illumina platforms (HiSeq X, MiSeq, or HiSeq 2500) across cruises. All sequencing data were processed using a standardized pipeline: primers were removed (using DADA2, or Cutadapt for KH-23-3), quality filtering was applied, and a fixed 120 bp segment from the 3′ end of the ~ 170 bp amplicon was retained to reduce sequencing errors and ensure consistent read length across platforms (see Text S1, Supplementary Material 1; Fig. S1, Supplementary Material 2). For all cruises except KS-18-5, amplicon sequence variants (ASVs) were inferred using the DADA2 (or Cutadapt for KH-23-3) plug-in in QIIME2 (Bolyen et al. 2019). For the KS-18-5 cruise, sequencing data were obtained from Yu et al. (2022; Tables S4 and S5 of their manuscript), in which the reads had been clustered into operational taxonomic units (OTUs) based on a 97% sequence similarity threshold.

ASVs and OTUs were taxonomically assigned based on BLASTN searches (ver. 2.12.0–2.13.0; Altschul et al. 1990) against the MitoFish mitochondrial genome database (ver. 3.72; Sato et al. 2018b) and MiFish reference sequences (Miya et al. 2015), using thresholds of ≥ 97% sequence identity and query coverage. When multiple candidate species met these thresholds, the species with the longest alignment length (i.e., the highest number of matching bases) was selected. If multiple candidates shared the same alignment length, the sequence was assigned to a higher taxonomic rank (e.g., genus or family). Despite the use of different taxonomic units (OTUs vs. ASVs), station-wise comparisons among sampling methods (bucket, intake, and Niskin) were unaffected, as taxonomic assignments were consistent within stations across methods. The average number of quality-filtered reads per sample ranged from ~ 78,000 (KS-18-5) to ~ 591,000 (KS-21-24) across cruises. Detailed raw and quality-filtered read counts for all individual samples, along with the average values for each cruise, are presented in Table S1 (Supplementary Material 3).

In eDNA metabarcoding, a ceiling effect may occur when a few highly abundant taxa dominate the sequencing output, thereby masking the detection of less abundant species. Furthermore, higher read counts of a species in one sample compared to another do not necessarily reflect greater biological abundance—particularly for non-dominant taxa—due to potential variation in sequencing depth and amplification efficiency (Skelton et al. 2023). Similarly, a sample with more total reads does not inherently contain more species than one with fewer reads (Miya et al. 2015). To account for these biases, raw read counts were normalized to relative read abundance (RRA), expressed as the percentage of total reads within each sample. This normalization enables more meaningful and standardized comparisons of species composition across samples. To further improve detection accuracy, we applied a threshold of 0.5% RRA per sample to confirm species presence/absence. This 0.5% threshold was determined through preliminary visual inspection of species count distributions across various cutoffs (Fig. S2, Supplementary Material 2), which showed a marked decline in detected species at the 0.5% level, while increasing the threshold to 1% led to only a marginal reduction. These patterns indicate that a 0.5% cutoff effectively reduces false positives without substantially discarding true detections. Similar thresholds have been employed in previous eDNA metabarcoding studies to improve detection reliability and minimize erroneous identifications (Drake et al. 2022; Jackman et al. 2021).

The results of the onboard negative controls are summarized in Table S2 (Supplementary Material 3). In most cruises, no fish species were detected in these control samples, with the exception of KS-22-11, KS-21-8, and KS-22-15. During the KS-22-15 cruise, Alburnus alburnus (Common Bleak), a freshwater species, was detected exclusively in the negative control and not in any of the seawater samples, indicating a likely false positive. In contrast, Zacco platypus (Pale Chub), another freshwater species (Froese and Pauly 2024), was detected in the KS-22-11 cruise negative control sample as well as in seawater samples from that and some other cruises. Given its ecological improbability in open marine environments, Zacco platypus was entirely excluded from the dataset. All other species detected in the negative controls were either absent from the seawater samples or appeared at very low read counts. As described above, RRA values were used to account for sequencing bias, and a threshold of > 0.5% per sample was applied to determine species presence. Consequently, species with low read counts in the negative controls were automatically excluded due to their low RRA values.

2.5 Collection of environmental data

Environmental parameters—including seawater temperature, salinity, and chlorophyll-a concentration—were recorded systematically at 1 dbar intervals using a CTD system during seawater sampling. To reduce instrument noise, seawater temperature and salinity data were processed using a Gaussian filter (Gaussian sigma = 2.5 db, width = 11 db) using the “scipy.ndimage.gaussian_filter1d” function in Python 3.10.9 (Python Software Foundation, 2023). Potential density was calculated using the Gibbs Seawater Oceanographic package with “gsw.density.sigma.” Maximum depth was recorded at each sampling site. Mixed layer depth (MLD) was determined using a finite difference method based on potential density at 10 dbar, with a density threshold of Δσθ = 0.125 kg/m3, following Kara et al. (2000). Seasons were defined by sampling months, as outlined by Trusenkova and Ishida (2005). Sample collection time (day or night) was identified using local sunset and sunrise times, calculated via the “datetime” and “astral.sun” functions in Python. A comprehensive summary of all environmental data is provided in Table S3 (Supplementary Material 3).

2.6 Statistical analysis

Taxonomic differences, as proxied by relative frequency and relative family-level composition (proportional representation of species- and family-presence), were visually compared across sampling methods using “ggplot2” (Wickham 2016) in the statistical software R (version 4.3.1) accessed via R Studio (version 2024.09.1 + 394; R Core Team 2023). The relative frequency for each sampling method was determined by dividing the number of occurrences of each species by the total number of occurrences across all species detected within that method. Similarly, relative family composition percentage was calculated as the proportion of occurrences of each family relative to the total occurrences across all families within the same sampling method.

Rarefaction curves helped to evaluate sampling completeness and compare asymptotic species richness and biodiversity among methods. Asymptotic estimates, with 95% confidence intervals, were calculated using R’s “iNEXT” package (Hsieh et al. 2016). Hill's numbers (q = 0, q = 1, and q = 2) were utilized for these comparisons based on incidence frequency data, where the input reflects the number of samples in which each species was detected within each method. Hill’s numbers quantify different aspects of species diversity: q = 0 for species richness, q = 1 for effective species number (exponential of Shannon entropy), and q = 2 for the inverse Simpson index, reflecting species evenness and frequency. Equations for Hill’s numbers are provided in Equations S1, S2, and S3 (Supplementary Material 4).

We compared the differences in each diversity index (q = 0, q = 1, q = 2) among methods using a non-parametric Kruskal–Wallis test, followed by pairwise Wilcoxon rank-sum tests for post hoc comparisons. Statistical significance was set at P < 0.05.

Pairwise comparisons revealed significant differences in the q = 2 diversity index across sampling methods based on pooled species data (see resultes). We, therefore, performed cluster analysis, using the “cluster” package (Maechler et al. 2023), to categorize sampling sites based on their species composition. The optimal number of clusters was determined through elbow and silhouette methods. Clusters were mapped using the Python library “cartopy” (https://github.com/SciTools/cartopy), and species compositions within clusters were visualized with the R package “pheatmap” (Kolde 2019). Indicator species analysis was performed using the multipatt function from the “indicspecies” package in R (De Cáceres et al. 2010) to identify species that are significantly associated with specific clusters. Alpha diversity indices (Shannon and Simpson) were also compared among methods within each cluster using the Kruskal–Wallis test (α = 0.05), followed by Wilcoxon tests for pairwise comparisons.

The spatial distribution of each cluster was further analyzed with non-metric multidimensional scaling (nMDS) using Jaccard dissimilarity. We visually explored relationships between clusters and underlying environmental variables using the ‘envfit’ function from the “vegan” package (Oksanen et al. 2022). We used a permutational multivariate analysis of variance (PERMANOVA), via the “adonis” function, also from “vegan” package with statistical significance set at P < 0.05 and 999 permutations to formally assess how much variation in fish community composition could be explained by the predictor variables of sampling method, season, cluster (i.e., region), time of sampling (day/night), and environmental variables (chlorophyll-a concentration, MLD, and temperature). Multicollinearity among environmental variables was assessed using the variance inflation factor (VIF) method with “calculate_vif” tool in Python 3.10.9. Potential density was excluded due to high VIF values, indicating multicollinearity with other predictor variables. The remaining variables showed low VIF values (see Table S4, Supplementary Material 3), confirming their suitability for analysis.

3 Results

3.1 Taxonomic richness

Our eDNA metabarcoding analysis from three surface sampling methods (bucket, intake, and Niskin) collectively detected 465 different fish taxa across 83 sampling stations, with each method individually identifying at least 324 species. The relative frequencies of fish taxa across the three methods were consistent (Fig. 2a). The dominant families identified—Carangidae, Clupeidae, Cottidae, Cyprinidae, Exocoetidae, Gobiidae, Gonostomatidae, Labridae, Mullidae, Myctophidae, Salmonidae, Scombridae, Serranidae, Tetraodontidae, Monacanthidae, and others—were detected across all methods, with relative composition percentages being nearly identical (Fig. 2b).

Fig. 2
figure 2

Comparison of taxonomic richness and family composition across three sampling methods (bucket, intake, and Niskin). a Relative frequency of the top 50 fish species detected by each method, with other species grouped as 'Others'. b Relative composition of the top 15 fish families observed using each method, with all other families grouped as 'Others'

Full size image

3.2 Hill’s diversity indices across eDNA sampling methods

Hill’s numbers (q = 0, q = 1, q = 2) were analyzed to characterize marine surface fish diversity across the three eDNA sampling methods (bucket, intake, and Niskin) as a function of sampling effort (Fig. 3A–C). Species richness (q = 0), representing the total number of detected fish species, did not show significant differences among the sampling methods (P > 0.05) (Fig. 3A). Similarly, the Shannon index (q = 1), which accounts for both species detection and evenness using the exponential of Shannon entropy, showed no significant variation across methods (p > 0.05) (Fig. 3B). However, the inverse Simpson index (q = 2), which gives more weight to the proportion of frequently detected species, showed a significantly higher value in the bucket samples (p < 0.05). In contrast, the intake and Niskin samples did not differ significantly (p > 0.05) (Fig. 3C). Details of the statistical tests are provided in Tables S5a and S6a (Supplementary Material 3).

Fig. 3
figure 3

Rarefaction and extrapolation curves for Hill’s diversity indices (q = 0, q = 1, and q = 2) across three eDNA sampling methods (bucket, intake, and Niskin) for marine surface fish diversity. A Richness (q0) represents species richness, indicating the total number of detected species. B Shannon index (q1) accounts for the exponential of Shannon entropy. C The inverse Simpson index (q2) emphasizes species evenness. Solid lines represent rarefaction estimates, and dashed lines represent extrapolation beyond observed data. Shaded areas indicate 95% confidence intervals for each sampling method

Full size image

To address the bioinformatics inconsistency in taxonomic assignment, we excluded data from the KS-18-5 cruise which used a different filter (0.22 µm) and OTU approaches instead of ASVs used elsewhere. Despite excluding this data, we still observed a higher inverse Simpson index (q = 2) in bucket samples (Fig. S3, Supplementary Material 2; Tables S5b and S6b, Supplementary Material 3). We then analyzed Niskin samples collected at the same time at two very close depths (5 m and 10 m) during the KS-18-5 cruise. No significant differences were observed in species richness (q = 0), Shannon index (q = 1), or Inverse Simpson index (q = 2) between the 5 m and 10 m depths (P > 0.05) for this cruise (Fig. S4, Supplementary Material 2). In addition, both depths shared 32 of the 52 detected species. These results suggest that depth within this range did not significantly affect biodiversity indices or species detection in the Niskin samples collected during the KS-18–5 cruise.

3.3 Fish community composition of marine surface

To further investigate the fish community composition in the marine surface environment, we conducted a cluster analysis using Ward’s linkage method and Jaccard distances based on presence/absence data. This analysis revealed eight distinct clusters across all samples collected from the three sampling methods (bucket, intake, and Niskin) (Fig. 4). Mapping the cluster positions for each method showed that seven of the eight clusters were consistently detected across all methods, with similar overall distributions (Fig. 5). However, samples collected using the intake method were not included in Cluster 2. Stations identified as belonging to Cluster 2 in the bucket and Niskin samples were replaced by Cluster 5 in the intake samples for those stations. Additionally, one Niskin sample is also classified as Cluster 5 in place of Cluster 2 (Fig. 5). To further examine method-specific patterns, we conducted cluster analyses independently for each sampling method (bucket, intake, and Niskin) using Jaccard distances. Each method yielded six clusters, named according to their order of appearance in the dendrograms (Fig. S5a–c, Supplementary Material 2). Clusters representing corresponding groups of sampling sites—based on geographical proximity—were generally consistent across methods (Fig. S6, Supplementary Material 2).

Fig. 4
figure 4

Cluster dendrogram of fish community composition in the marine surface environment, based on presence/absence data from bucket, intake, and Niskin sampling methods. The analysis was conducted using Ward’s linkage method with Jaccard distances, revealing eight distinct clusters, each highlighted by a different colored box. Branch colors represent sampling methods: blue for bucket, pink for intake, and green for Niskin, corresponding to the stations sampled with these methods

Full size image
Fig. 5
figure 5

Spatial distribution of identified fish species clusters across the three eDNA sampling methods: a bucket, b intake, and c Niskin. Each map illustrates the geographic positions of the different species clusters (1–8) identified using hierarchical clustering with Ward linkage based on Jaccard similarity. The symbols and colors represent distinct clusters. Notably, Cluster 2 is absent in the intake method. Caution is advised as some stations have very close or almost identical positions, leading to overlapping symbols (e.g., Cluster 1 and Cluster 6)

Full size image

Alpha diversity indices (Shannon and Simpson) were largely similar among methods within each cluster (Kruskal–Wallis test followed by pairwise Wilcoxon test), except for Clusters 5, 6, and 7. In Cluster 5, bucket samples had significantly higher biodiversity than both intake samples (p < 0.01) and Niskin samples (p < 0.05). Similarly, in Cluster 7, bucket samples exhibited slightly higher biodiversity than Niskin samples (p < 0.05) for both indices, with no significant difference between bucket and intake samples. In Cluster 6, bucket samples exhibited slightly higher biodiversity than Niskin samples (p < 0.05) for the Shannon index only. In all of Clusters 5, 6, and 7, biodiversity indices did not significantly differ between Niskin and intake samples (Fig. 6). After excluding the KS-18-5 cruise to minimize inconsistencies related to bioinformatics processing and filter type, we identified eight clusters (see dendrogram in Fig. S7a, Supplementary Material 2). Similar to the results from the full dataset including all cruises, both biodiversity indices did not differ significantly among methods in most clusters (P > 0.05, Kruskal–Wallis test followed by pairwise Wilcoxon test), except in Clusters 2a, 6a, and 8a (Fig. S7b, Supplementary Material 2). These three clusters (2a, 6a, and 8a), which showed significant differences (P < 0.05) in biodiversity indices among sampling methods after excluding the KS-18-5 cruise, were located in the same geographical regions (Fig. S8, Supplementary Material 2) as the previously defined Clusters 5, 7, and 6, respectively, where similar differences in biodiversity indices were observed based on data from all cruises (Fig. 5).

Fig. 6
figure 6

Boxplots of alpha diversity indices (a. Shannon and b. Simpson) for each cluster, comparing the bucket, intake, and Niskin sampling methods. Kruskal–Wallis and post-hoc Wilcoxon tests were applied to determine significant differences between methods. NS: not significant. Significance levels: *P < 0.05; **P < 0.01

Full size image

Cluster-wise rarefaction and extrapolation curves of Hill numbers (q = 0, 1, and 2) also indicated no significant differences among the three sampling methods—bucket, intake, and Niskin—in most clusters (Fig. S9a and S9b, Supplementary Material 2; Table S7, Supplementary Material 3). However, some clusters exhibited method-specific differences. For example, in Cluster 5, bucket samples showed higher diversity estimates than both intake and Niskin samples for q = 1 and q = 2 (P < 0.05), while in Cluster 7, bucket samples showed higher diversity than Niskin samples for q = 1 only (P < 0.05). These results are highly consistent with the alpha biodiversity findings, where Shannon and Simpson indices also showed significantly higher values in Bucket samples within the same clusters. Species richness (q = 0) varied significantly in Cluster 1 (intake > Niskin), Cluster 6 (bucket and intake > Niskin), and Cluster 7, where all three methods differed significantly from each other. No significant differences in Hill numbers were observed among methods in the remaining clusters (P > 0.05; Table S7, Supplementary Material 3).

Each cluster was characterized by the presence of one or more indicator species, which help define and distinguish the clusters from one another (Table 2). A heatmap showing the detection frequencies of the top 10 most frequently detected species in each cluster is presented in Fig. 7. The geographical distribution of the clusters revealed by eDNA metabarcoding data from the bucket, intake, and Niskin methods showed that the clusters were associated with different regions. Cluster 1 was located in the Oyashio region (subarctic gyre). Cluster 2 was along the Kuroshio Extension. Clusters 3 and 4 were in the coastal region of the northern part of the mixed water region. Cluster 5 was widespread along the coastal region of the southern part of the mixed water region, as well as the Pacific coast of the Kuroshio region. Cluster 8 was associated with the Kuroshio Current and Tsushima Current, and Cluster 7 was located in the East China Sea (Fig. 5). Clusters 2 and 5 are geographically close, with the latter replacing the former when considering biodiversity patterns as revealed by just the intake method (Fig. 5).

Table 2 Indicator species for different marine clusters based on Indicator Species Analysis (ISA)
Full size table
Fig. 7
figure 7

Heatmap showing the species composition across the eight clusters, illustrating the detection frequency of the top 10 most detected species in each cluster. The color gradient represents the number of times each species was detected within each cluster, with blue indicating low detection frequency and red indicating high detection frequency. Species not in the top 10 for each cluster are grouped under "Others."

Full size image

The PERMANOVA analysis revealed that multiple factors contributed to variation in fish community composition, with the highest variation explained by cluster (region) at 25.4%, followed by season at 11.3%. In contrast, the effect of method accounted for only 1.4% of the variation, followed by the influence of temperature (1.08%) (Table 3). However, as the cluster (region) variable was derived a priori from the same species matrix, we also tested a model excluding this factor. Notably, the variation explained by method remained unchanged at 1.4% (Table S8, Supplementary Material 3).

Table 3  Results of PERMANOVA based on Jaccard dissimilarity (presence/absence data) testing the effects of sampling method (bucket, intake, Niskin), season, region (cluster 1 to 8), time of sampling (day/night), and environmental variables (chlorophyll-a concentration, mixed layer depth, and temperature) on fish community composition
Full size table

3.4 Influence of environmental parameters

Non-metric multidimensional scaling (nMDS) was conducted to visualize patterns in species composition across different sampling methods (bucket, intake, and Niskin) using the Jaccard dissimilarity index. The nMDS analysis produced a two-dimensional solution with a stress value of 0.164, effectively distinguishing the ordination of the eight identified clusters (Fig. 8). When ordinations were performed independently for each method, the overall patterns were broadly consistent, with minor differences likely reflecting variations in stress values: Bucket (0.159), Intake (0.162), and Niskin (0.116) (Fig. S10, Supplementary Material 2).

Fig. 8
figure 8

nMDS ordination plot (Stress = 0.164) illustrating cluster patterns across three sampling methods (bucket, intake, and Niskin) using the Jaccard dissimilarity index. Points represent samples, colored by method (blue for bucket, pink for intake, and green for Niskin), and shaped according to the identified clusters (1–8). Ellipses denote the dispersion of each cluster at a 95% confidence interval. Significant continuous environmental parameters (temperature, chlorophyll-a (Chl_a), and salinity) are displayed as vectors, and the centroid positions of significant seasonal categories (winter, spring, summer, and fall) are shown as orange points with labels. These vectors and centroids illustrate the influence of environmental variables on the clusters and their separation in the ordination space

Full size image

To evaluate the influence of environmental variables on community composition, we applied the “envfit” function to fit both continuous variables (temperature, salinity, chlorophyll-a, mixed layer depth, and maximum depth at sampling sites) and categorical variables (season and time of day for sample collection) onto the nMDS ordination.

The continuous variables that showed significant correlations with the ordination were temperature (R2 = 0.59, P < 0.001), salinity (R2 = 0.58, P < 0.001), and chlorophyll-a (R2 = 0.26, P < 0.001) (Table S9, Supplementary Material 3). Among the categorical variables, “season” demonstrated a significant association (R2 = 0.26, p < 0.001). To enhance clarity, only significant parameters are displayed on the nMDS plot (Fig. 8). These findings suggest that the clusters may be influenced by both environmental factors (continuous and seasonal) as well as geographical position. Specifically, Cluster 1 was associated with lower salinity; Cluster 2 with higher salinity and the spring season; Clusters 3 and 4 with elevated chlorophyll-a levels; Cluster 5 with a combination of multiple variables (cooler temperature, higher Chlorophyll-a and salinity during winter and spring) distributed across a larger area on the ordination plot; Cluster 6 with higher temperatures; and Clusters 7 and 8 with favorable temperatures, differentiated by the fall and summer seasons, respectively (Fig. 8).

4 Discussion

4.1 Assessment of marine surface fish diversity in the western north pacific

eDNA metabarcoding detects more species than traditional methods like trawling, offering a broader, more comprehensive view of fish diversity (He et al. 2023; Fraija-Fernández et al. 2020). This technique has proven to be a reliable tool for continuous biodiversity monitoring (Stoeckle et al. 2021). It also holds promise in assessing threatened species’ distributions and population trends over time, which is essential for effective conservation and management (Zou et al. 2020). In this study, we identified 465 species across three sampling methods. Most species’ distributions were successfully cross-referenced with FishBase (Froese and Pauly 2024; Boettiger et al. 2012) (Table S10, Supplementary Material 3), arranged by Nelson’s Fishes of the World (5th edition) (Nelson et al. 2016). The dominant families identified align with previous studies (Orlov and Tokranov 2019; Kawakami et al. 2023), with Myctophidae as the most abundant. Our study also detected rare species—those found at very low frequencies across samples—many of which are typically challenging to capture with traditional methods (Stat et al. 2019). This is supported by the observation that the top 50 most frequently detected taxa accounted for more than 50% of total detections across all three methods, despite a total of 465 taxa being identified (Fig. S11a, b, Supplementary Material 2), highlighting eDNA’s potential to uncover hidden biodiversity.

A key finding of our study is that each of the three sampling methods (bucket, intake, and Niskin) yielded similar taxonomic identifications, with each method detecting at least 324 taxa from 83 sampled stations and over 480 asymptotic counts based on rarefaction and exploration analyses (Table S6a, Supplementary Material 3). The consistency in taxa and families across the sampling methods is likely due to the filtration of a substantial volume of water (~ 10 L). This volume may be sufficient to detect most fish species, but for more accurate and comprehensive species detection, larger volumes (e.g., > 100 L) may be required (Kawakami et al. 2023). While larger volumes generally improve species detection, the benefit plateaus beyond a certain threshold (Cantera et al. 2019). Ahn et al. (2022) observed that, in their winter trial, a single 10-L sample detected more OTUs than five combined 2-L replicates, highlighting the potential benefit of using sufficient volume in a single sampling event. However, they cautioned that this result may reflect the sparse distribution of eDNA rather than a consistent advantage of larger volume. Similarly, Kawakami et al. (2023) reported that smaller volumes, such as 3L, require tens of replicates to reliably capture the fish community in surface waters, or a greater total filtration volume is needed for more accurate results. For deep ocean metabarcoding, Yoshida et al. (2023) found that 10–20L is optimal for biodiversity monitoring, but more than 10 replicates are necessary for robust species detection. Although we did not directly compare variation arising from filtration volume and sampling method, the observed consistency in taxa across methods suggests that filtering a sufficiently large volume (e.g., ~ 10 L) may help minimize variability in species detection among methods. Therefore, filtering as much water as possible—ideally with multiple replicates—remains essential for enhancing detection reliability within logistical constraints.

Since eDNA can decay, disperse, and diffuse in marine environments, a minimum species abundance is required to shed enough DNA to enable reliable detection (Barnes and Turner 2016). Although many of the detected species in this study are challenging to identify using traditional methods, their widespread distribution in the study area (Fujikura et al. 2010) likely contributed to the minimal differences in taxonomic identification among the sampling methods. Overall, our results—particularly the similar taxonomic profiles and consistent rarefaction trends—suggest that each of the three sampling methods (bucket, intake, and Niskin) is effective for detecting surface marine fish taxa and can be reliably employed in marine surface ecosystems.

4.2 Geographical distribution and composition of marine surface community

The geographical distribution of fish species clusters, as identified through eDNA metabarcoding using bucket, intake, and Niskin sampling methods, revealed highly consistent clustering patterns across all three methods, reflecting distinct biogeographic regions. Based on our results, the sampling stations were broadly divided into two major regions, as indicated by the clustering structure in the dendrogram (Fig. 4), with the separation occurring around 34°N, as observed in the latitudinal gradient of the spatial distribution map (Fig. 5).

South of this boundary, three clusters—Clusters 6, 7, and 8—were consistently identified by all methods, demonstrating the robustness of eDNA metabarcoding in capturing spatial biodiversity patterns (West et al. 2020). Cluster 8, situated along the western coast of Honshu and influenced by the Kuroshio Current, likely supports fish communities adapted to warm, saline waters (Yu et al. 2023; Zhang et al. 2020). Cluster 7, associated with the East China Sea region, is also under the influence of the Kuroshio Current, highlighting the role of warm, subtropical waters in shaping regional biodiversity (Guo et al. 2006). Cluster 6, located in the offshore open ocean, represents a pelagic environment supporting diverse marine fish species (Fujikura et al. 2010).

North of the 34°N boundary, five clusters were identified: Cluster 1 in the sub-Arctic region and Clusters 2, 3, 4, and 5 along the Sanriku coast. These clusters were also consistently detected across all three sampling methods, with the exception of Cluster 2, which was not included in the intake samples. Cluster 1, shaped by the cold, nutrient-rich Oyashio Current, reflects a sub-Arctic environment that supports species adapted to low temperatures and high productivity (Ishak et al. 2020). Clusters 3 and 4, influenced by the convergence of the Kuroshio and Oyashio currents, are characterized by intermediate temperature and salinity conditions that promote nutrient availability and support species adapted to dynamic environments (Yasuda 2003; Yamamoto et al. 2017). Cluster 2, influenced by the warm, saline waters of the Kuroshio Current, may harbor fish species specific to Kuroshio-dominated systems. Cluster 5, which spans a wide area along the eastern coast, exhibits similar hydrographic conditions and likewise reflects the Kuroshio’s influence on fish community composition (Yu et al. 2022) (Fig. 5).

These regional and cluster-specific biodiversity patterns are further supported by nMDS analyses incorporating environmental variables (Fig. 8), reinforcing the ecological relevance and robustness of the clustering derived from eDNA metabarcoding (Jeunen et al. 2019). With only minor discrepancies in a few assignments across the three methods, the convergence of biogeographic patterns underscores the effectiveness of each of the bucket, intake, and Niskin sampling methods in capturing spatial biodiversity via eDNA metabarcoding. Furthermore, the correspondence between cluster distribution and major ocean currents suggests that physical and oceanographic features surrounding Japan play a key role in shaping distinct ecological zones (Ashjian et al. 2006).

The eight distinct clusters of fish communities, each characterized by indicator species and environmental factors, highlight eDNA metabarcoding’s ability to distinguish localized fish assemblages (Yamamoto et al. 2017; Jeunen et al. 2019). Key factors such as temperature, chlorophyll-a, salinity, and seasonality likely contribute to these clusters. As ectotherms, fish need to remain within their optimal temperature range; chlorophyll-a indicates areas of high fish biomass (Sato et al. 2021); salinity correlates with eDNA concentrations and species’ habitat preferences (Takahara et al. 2019); and seasonal sampling enhances eDNA detection during breeding or migration (Hayami et al. 2020; Inui et al. 2021).

The clusters varied in their responses to environmental variables. For example, Cluster 1, which included several Salmonidae species, reflected the geographical distribution characteristics of salmonids and their preference for regions with lower salinity (Sato et al. 2018a) and lower temperatures (Stehfest et al. 2017), although these environmental preferences can vary depending on developmental stage. Clusters 3 and 4 included species such as Repomucenus planus, Sardinella zunasi, Laemonema longipes, Lateolabrax japonicus, and Leuroglossus schmidti, which generally feed on zooplankton and crustaceans. These food sources are directly linked to primary productivity, as zooplankton and crustacean abundance increases in environments with high chlorophyll-a levels. Consequently, these fish species thrive in such conditions (Lowerre-Barbieri et al. 2017; Sato et al. 2021). Cluster 5 is characterized by Myctophum spinosum, a species commonly found in tropical and subtropical waters and often appearing in the bycatch of deep-sea trawlers. These fish are opportunistic carnivores, feeding on crustaceans and small fishes, and constitute a significant portion of the mesopelagic zone's biomass (Vipin et al. 2015). Additionally, cluster 5 is dominated by sardines, anchovies, and mackerel, highlighting the ecological plasticity of these species. Their ability to thrive in diverse environments is likely driven by factors such as temperature and chlorophyll-a availability (Froese and Pauly 2024). Cluster 6, characterized by Katsuwonus pelamis and Oxyporhamphus micropterus, represents species highly adapted to pelagic, warmer waters and specific salinity gradients, reflecting their reliance on specific oceanographic conditions (Collette and Nauen 1983; Froese and Pauly 2024).

Clusters 7 and 8, composed of various lanternfish and other mesopelagic fish, differ not only in their geographical locations but also in their seasonal associations. Cluster 7 aligns with fall, while Cluster 8 aligns with summer, reflecting seasonal, life-history-dependent differences in exploiting periods of high biological activity (Hayami et al. 2020; Inui et al. 2021).

Overall, the spatial and seasonal environmental associations of each cluster are consistent with previous knowledge of species’ ecology and life history and emphasize the importance of accounting for both spatial and temporal variation in marine biodiversity assessments. In our study, spatial structuring (Cluster) and seasonal variation emerged as the dominant contributors to fish community dissimilarity, while only minor differences were attributed to sampling method. These method-specific differences likely stemmed from very specific presence/absence cases involving certain species, such as Pacific Saury (Cololabis saira) (Ahmed et al., 2025, submitted). Thus, our observed biodiversity patterns mostly likely reflect true spatio-temporal patterns rather than methodological artifacts. Further supporting this assertion was that all our species clusters were consistently detected across the bucket, intake, and Niskin sampling methods, except for Cluster 2, which was absent in the intake sampling. Likewise, negligible differences were generally observed in Shannon and Simpson diversity indices estimated using data from different sampling methods within clusters.

In shallow, well-mixed marine areas, eDNA concentrations and estimated community composition can be similar between surface and bottom samples. This pattern has been attributed to the likely minimal vertical stratification of eDNA in such environments (Andruszkiewicz et al. 2017). Since all our collected samples were taken within the mixed layer (Table S3, Supplementary Material 3), any observed community differences for specific clusters or methods are unlikely the result of differences in sampling depth. This interpretation is supported by our nMDS analysis (Fig. 8) and PERMANOVA results (Table 3). In addition, the comparison between 5 and 10 m Niskin samples also supports the interpretation (Fig. S4, Supplementary Material 2).

4.3 Differences among methods

While all three sampling methods demonstrated comparable taxonomic richness and community composition, the bucket samples showed higher inverse Simpson index values and biodiversity indices, particularly in Clusters 5 and 7, than Niskin and intake samples. This difference reflects a greater evenness in bucket samples, where the proportion of frequently detected species was lower than in intake and Niskin samples (Fig. S11a, b, Supplementary Material 2). Two factors may explain this increased evenness in bucket samples.

First, transparent exopolymer particles (TEPs), which are abundant in the sea surface microlayer, may influence eDNA capture and detection. Some TEPs are lighter than seawater and tend to accumulate at the surface, potentially trapping eDNA into a microlayer (Wurl et al. 2011). This accumulation is influenced by factors such as wind speed, stratification, and regional productivity (Zamanillo et al. 2019; Engel and Galgani 2016). Due to their high stickiness, TEPs aggregate with other organic and inorganic particles. Furthermore, their lighter density allows TEP-rich aggregates to remain suspended or even ascend, enabling them to remain in the surface microlayer unless ballasted by denser particles (Mari et al. 2017; Wurl et al. 2009). This TEP-enriched surface microlayer enhances light scattering, traps and aggregates particulates, and could thus facilitate the accumulation of eDNA in the water column, potentially increasing its detectability (Stramski et al. 2019). As a result, the TEP-enriched surface microlayer may serve as a reservoir for genetic material from various sources, including rare and less abundant species. Consequently, the aggregation and retention of eDNA in the surface microlayer—driven by TEPs (Wurl and Holmes 2008; Thornton 2014) and less accessible to deeper sampling methods like intake and Niskin—may lead to the detection of more species in bucket samples, which likely explains the higher evenness observed in those samples. However, compared to the vast volume of the global ocean, the surface microlayer is extremely limited in total volume (Cunliffe et al. 2013) and is primarily formed under weak to moderate wind conditions and the presence of surfactants (both natural and anthropogenic) (Wurl and Holmes 2008; Wurl et al. 2011). Therefore, despite our research being conducted in the dynamic Kuroshio-Oyashio system, intake sampling can be applied globally due to the limited presence of the microlayer in the ocean, which leads to better evenness and biodiversity in bucket samples.

Second, a ceiling effect related to high eDNA read counts may influence metabarcoding results (Skelton et al. 2023). UV radiation degrades eDNA more rapidly in surface layers, reducing the eDNA pool density, especially for species that release genetic material in these upper layers (Strickler et al. 2015). This degradation may mitigate the ceiling effect by lowering the read counts for frequently detected species, thereby allowing rarer species to be detected in bucket samples (Fig. S11a and S11b, Supplementary Material 2). In contrast, samples collected using intake and Niskin methods are less exposed to UV, which may lead to a more pronounced ceiling effect, favoring the prevalence of frequently detected species (Lamb et al. 2019). Consequently, bucket samples tend to exhibit higher evenness and potentially a higher inverse Simpson index.

Filter pore size is also an important consideration in fish eDNA metabarcoding studies, as it can influence the number and diversity of species detected (Li et al. 2018). Filters with a 0.45 μm pore size are consistently used in fish eDNA metabarcoding studies (Roblet et al. 2024; Li et al. 2018). Larger filter pore sizes (e.g., 0.45 μm) do not necessarily decrease the yields of eDNA (Jo et al. 2020), while 0.22 μm filters may improve the completeness of fish species identification in situations when the DNA is degraded (Kawakami et al. 2023). In our study, we conducted separate analyses by excluding data from the KS-18–5 cruise, which used a different filter size (0.22 μm) compared to the other cruises (0.45 μm). Our results revealed similar trends in fish biodiversity and community composition estimates, regardless of whether we included all cruise data or excluded the KS-18–5 data.

In coastal and open ocean systems, eDNA decay rates are key factors in interpreting eDNA signals (Pastor Rollan et al. 2024; Andruszkiewicz et al. 2019). These decay processes are influenced by several environmental factors, including temperature and water chemistry (Jo and Minamoto 2020; Sassoubre et al. 2016), and potentially sunlight—although some studies have reported no significant impact of light exposure on eDNA degradation (Andruszkiewicz et al. 2017). Additionally, the ecology of eDNA—its state and characteristics—in the open ocean remains largely unknown. Therefore, the differences observed among sampling methods (for example, higher biodiversity in bucket samples in certain regions) may be attributed to the physical characteristics (e.g., DNA buoyancy, particle size) and state of eDNA in the water column (Jo and Minamoto 2020).

Another notable difference was the absence of Cluster 2 in the intake samples, despite its presence in bucket and Niskin methods. A closer examination of the sampling stations revealed that intake samples replaced Cluster 2 with Cluster 5 (and one Niskin sample also showed this replacement) (Fig. 5). All Cluster 2 samples were collected during the KS-18–5 cruise where the intake samples were collected at a different time of day compared to bucket and Niskin methods (which were collected simultaneously) (see Fig. S12, Supplementary Material 2). Temporal differences in eDNA sampling may result in biases associated with inter-species differences in diurnal activity (Suzuki et al. 2022) or observational biases due to ocean advection (Yu et al. 2022). All intake samples, as well as the single Niskin sample identified as Cluster 5, were collected during the daytime.

Further examination of the KS-18-5 intake samples and the Niskin sample classified as Cluster 5 (hereafter referred to as Cluster 5_I_N) revealed that their species composition closely resembled the overall species profile of Cluster 5 (Fig. S13a, b; Supplementary Material 2). In contrast, Cluster 2 was characterized by Chlorurus sordidus (see Fig. S13a, Supplementary Material 2); Table S11 (Supplementary Material 3)), a coastal species likely transported offshore by the Kuroshio Current (Yoo et al. 2004) and not detected in all samples.

Myctophum spinosum—the indicator species for Cluster 5—was likewise identified as the indicator species for Cluster 5_I_N (Table S11, Supplementary Material 3), accompanied by other Myctophidae species (Fig. S13a, b; Supplementary Material 2). These mesopelagic fishes are known for their nocturnal vertical migrations to surface waters for feeding (Watanabe et al. 2002). However, this study revealed higher occurrences of detection of these species in daytime samples compared to nighttime samples during the KS-18-5 cruise which may result from eDNA persistence from their nocturnal activity (Barnes and Turner 2016) or local behavioral variations (Giske et al. 1998). When comparing day and night eDNA detections in the KS-18-5 cruise, we did not observe strong variation between the two, except for Myctophum spinosum and Sardinops melanostictus/S. sagax, which were better detected (P < 0.05) during the day (Fig. S14, Supplementary Material 2). To fully understand potential diel vertical migrations in fish behavior, samples should be collected at regular intervals (e.g., every 4 h). Future research should be conducted to investigate this phenomenon in greater detail. Owing to the time difference in sampling between the KS-18-5 intake and other methods, we excluded these KS-18-5 samples from biodiversity rarefaction analysis for q = 2, as bucket samples had significantly higher q = 2 values. Nonetheless, bucket samples maintained higher diversity than intake and Niskin (P < 0.05) (see Fig. S3, Supplementary Material 2), indicating that the discrepancies in cluster classification and the higher biodiversity index in bucket samples are attributable to different underlying factors, as discussed above. These differences mainly stem from environmental conditions and sampling timing, rather than sampling efficiency.

A key limitation of this study is the lack of replicates within each sampling method, which prevents a full assessment of the variability within and between methods. As eDNA results can vary significantly even among concurrent samples from the same location, replicates are crucial for capturing a comprehensive range of species. While studies recommend using 7–8 replicates and filtering more than 100 L of water to adequately capture fish diversity (Kawakami et al. 2023; Stauffer et al. 2021), implementing this ideal design was not feasible during the cruises due to logistical constraints (e.g., limited space, restricted ship stationary time at each station, and limitations in storage and filtration capacity). Instead, we analyzed a large dataset of concurrently collected water samples across three methods, compiled from multiple cruises conducted under varying oceanographic conditions to enable robust comparisons among sampling methods. This study design reflects a pragmatic balance between methodological rigor and the logistical realities of large-scale marine surveys, contributing to future efforts aimed at optimizing open ocean eDNA sampling protocols. Additionally, variations in the plumbing systems used for seawater intake across ships may introduce unintended methodological variation.

4.4 Practical recommendations

Although the three sampling methods demonstrated comparable performance in assessing marine surface fish community composition and biodiversity, it is important to compare their practical trade-offs when selecting a method. Bucket sampling may provide a better representation of surface communities in the surface microlayer, but it requires the vessel to stop and involves manual, labor-intensive water collection (Miyata et al. 2022). Intake sampling, in contrast, is less disruptive to ship operations (Kawakami et al. 2023) and is widely accessible across various global ocean vessels worldwide (Hoshina et al. 2018; Montoto-Martínez et al. 2020). However, it poses a higher risk of contamination, particularly from heavy metals in ship plumbing systems (Greco et al. 2022). The Niskin method, while also requiring the ship to stop (Yu et al. 2022), remains the most reliable option for depth-specific water collection (Govindarajan et al. 2021). A detailed summary of these trade-offs is provided in Table 4.

Table 4 Summary of practical characteristics, benefits, and limitations associated with three surface eDNA sampling methods: Intake, Niskin, and Bucket sampling
Full size table

Therefore, the selection of a sampling method should be guided by the research objectives, logistical constraints, and the required depth resolution. Careful consideration of these factors will help ensure effective and efficient application of eDNA metabarcoding for monitoring marine surface fish communities. Since each method can broadly capture and distinguish distinct fish community structures on its own, integrating data from all methods offers a more comprehensive approach to biodiversity assessments in marine surface environments.

5 Conclusion

This study demonstrates the effectiveness of three surface sampling methods—bucket, intake, and Niskin—for assessing fish biodiversity and community composition using eDNA metabarcoding. Each method reliably captured a broad range of species and community structures, offering complementary insights into the complex dynamics of marine surface ecosystems. While minor differences between methods exist, understanding these differences allows for more informed interpretation of eDNA data. Although the combined use of all three methods provides a more comprehensive perspective, it is often impractical in large-scale ocean surveys due to logistical constraints. Therefore, in situations where all three methods cannot be utilized, any one of them can be prioritized based on resource availability and ease of sampling.