Separating Faces in ARMS Metabarcoding Improves Marine Biodiversity Monitoring: A Comparison Across Protocols, Experimental Designs and Photographic Surveys Anne Chenuil, Elyna Bouchereau, Térence Legrand, Virgile Calvert, Cécile Chemin, Sandrine Chenesseau, Dorian Guillemain, José Miguel Gutiérrez Ortega, Anne Haguenauer, Michèle Leduc, Frédéric Legendre, Florent Marschal, Christian Marschal, Fatma Mirleau, Marjorie Selva, Laurent Vanbostal, Frédéric Zuberer, Pascal Mirleau, Laetitia Plaisance, Vincent Rossi, Sandrine Ruitton, Emese Meglécz, Vincent Dubut Abstract Monitoring marine biodiversity requires approaches capable of capturing its spatial and temporal complexity. DNA metabarcoding coupled with Autonomous Reef Monitoring Structures (ARMS) is increasingly used for this purpose, yet most applications still pool all sessile fractions and rarely benchmark molecular outputs against photographic observations. Here, we combined photographic analysis with cytochrome c oxidase I (COI) metabarcoding across 10 north‐western Mediterranean sites to compare and refine ARMS‐based monitoring protocols. We first optimized laboratory procedures (DNA extraction and polymerase choice) and applied the control‐driven, replicate‐aware VTAM pipeline to minimize false positives and ensure traceability. We then conducted the first face‐by‐face comparison of α‐ and β‐diversity between imaging and eDNA, metabarcoding each ARMS face separately rather than pooling samples. Metabarcoding detected ~15× higher site‐level richness and revealed stronger correlations with geographic distance and environmental gradients—which stemmed from its finer taxonomic resolution—whereas photography provided complementary information on macro‐taxa and surface cover. For metabarcoding, processing each face separately yielded much higher richness and stronger β‐diversity–distance correlations than with the NOAA pooling protocol, demonstrating that pooling inflates sampling variance, weakening ecological signal. Grouping the 17 faces into five structural categories offered a more operational alternative while further increasing α‐diversity and strengthening β‐diversity correlations. Overall, our results show that retaining ARMS microhabitat structure is critical for maximizing metabarcoding performance. Using five structural sessile fractions per ARMS combined with a control‐driven bioinformatic workflow provides a reproducible, scalable framework for long‐term eDNA monitoring and early detection of biodiversity change. Introduction Biodiversity protection is increasingly recognized as a global priority (Cowie et al. ). Global assessments warn that current biodiversity loss is unprecedented and threatens vital ecosystem services (Díaz et al. , ). Marine ecosystems face intensifying pressures from climate change, ocean acidification, deoxygenation, overfishing, pollution and habitat destruction (Halpern et al. ; Duarte et al. ). Effective protection therefore depends on timely, evidence‐based monitoring of species composition and ecological change (Estes Jr et al. ; Miloslavich et al. ). Meeting this need requires standardized, sensitive, and scalable methods that capture taxonomic and genetic diversity and their associated spatio‐temporal dynamics across broad geographic ranges (Laikre et al. ). For hard‐bottom benthic communities, two major challenges remain: (i) developing standardized sampling schemes and (ii) ensuring consistent, taxonomically reliable assessments (Danovaro et al. , ). Historically, benthic assemblages were identified morphologically by taxonomic experts, which is a labour‐intensive process that depends on a diminishing pool of experts (Appeltans et al. ; Costello et al. ) and is prone to inconsistencies among experts, limiting reproducibility (Aylagas et al. ). These limitations hinder the long‐term and large‐scale monitoring plans required to track biodiversity change. The challenge is particularly acute in regions of exceptionally high species richness, such as the Mediterranean and many tropical seas, where small, morphologically similar or cryptic taxa dominate hard‐bottom assemblages (Coll et al. ; Bianchi et al. ; Obura ). In these biodiversity hotspots, the sheer number of species and the prevalence of cryptic or early‐life stages make comprehensive morphological inventories impractical. Autonomous Reef Monitoring Structures (ARMS) offer a promising solution for standardized hard‐bottom sampling. Initially developed in tropical seas (Plaisance et al. ), they are now also deployed in temperate and polar environments (Couëdel et al. ; David et al. ; Obst et al. ; Pearman et al. ; Daraghmeh et al. ; Pagnier et al. ). ARMS provide a uniform, replicated habitat that becomes colonized by diverse sessile and mobile organisms. ARMS community composition has typically been characterized either through photography‐based analyses of (some of) the 17 exposed plate faces (David et al. ) or through DNA metabarcoding of the organisms scraped from the structure after retrieval (Leray and Knowlton ) and from their mobile fractions. The global ARMS programme provides a standardized protocol, from deployment to data analysis, particularly suitable for large‐scale biodiversity observatories such as ARMS‐MBON (Obst et al. ; Pearman et al. ). This programme integrates photography analyses and metabarcoding across European seas and polar regions, providing long‐term, FAIR (Findable, Accessible, Interoperable and Reusable) biodiversity data (Klymus et al. ; Takahashi et al. ) and delivering Essential Biodiversity Variables (EBVs; Miloslavich et al. ) as well as early detection of non‐indigenous species. Both photography‐based and metabarcoding approaches have advantages for biodiversity assessment and can be partly automated thanks to advances in artificial intelligence and next‐generation sequencing, respectively (Beijbom et al. ; Pearman et al. ; Deiner et al. ; Taberlet et al. ). Furthermore, beyond community inventories, ARMS metabarcoding can also capture intraspecific genetic diversity and reveal cryptic population structure (Thomasdotter et al. ). On one hand, metabarcoding typically identifies taxa from bulk DNA extracts using universal primers and high‐throughput sequencing (e.g., Andújar et al. ; Brandt, Pradillon, et al. ; Pearman et al. ). However, false positives and false negatives can be generated without careful quality control (Zinger et al. ; Alberdi et al. ; Brandt, Trouche, et al. ). Here, we apply a stringent, control‐driven pipeline (VTAM; González et al. ) that minimizes such errors and ensures full traceability. On the other hand, photography‐based analyses can reach species‐level resolution, for some taxa, when performed by trained experts, but large monitoring networks often require coarser, higher‐taxonomic classifications (David et al. ; Pearman et al. ; Miloslavich et al. ). The standard NOAA protocol for metabarcoding uses the two smallest mobile fractions (100–500 μm and 500–2000 μm) separately and a single composite sessile fraction scraped from all nine ARMS plates (Leray and Knowlton ; Pearman et al. ), thereby sacrificing microhabitat‐scale heterogeneity among individual plate faces as observed by David et al. (). To overcome this limitation and enhance ARMS‐based biodiversity assessments, we refined existing protocols through several methodological improvements. First, we evaluate laboratory procedures by comparing two DNA extraction kits and two polymerases to identify the most efficient combination for diverse marine assemblages. Second, we implement a face‐by‐face metabarcoding design in which the sessile community of each of the 17 colonizable ARMS faces is analysed individually rather than as a single pooled sample. This enables the first direct face‐by‐face comparison of hard‐bottom community composition using both photography‐based annotation and COI metabarcoding, providing unprecedented cross‐validation of α‐ and β‐diversity estimates. Third, we use two additional ARMS for each site where metabarcoding used pooled sessile fractions to test whether retaining face‐level resolution improves richness detection and strengthens correlations of community dissimilarity with environmental variables (geographic distance and environmental gradients) relative to the standard pooled NOAA protocol. Finally, we tested a more operational scheme in which five sessile fractions per ARMS (corresponding to distinct microhabitats) were analysed separately while including three ARMS units per site. Here, we evaluate whether separating ARMS faces for metabarcoding improves biodiversity assessments relative to pooled designs. Specifically, we compare four monitoring protocols—one based on photography (images taken on the 17 individual plate faces of each ARMS) and three based on metabarcoding (full face‐by‐face processing of the same 17 faces, pooling all faces per ARMS and grouping faces into five structural categories)—to quantify their ability to recover α‐, β‐ and γ‐diversity and to detect geographic and environmental structuring of communities. Because ARMS provide contrasting microhabitats (upward‐ vs. downward‐facing surfaces, open vs. closed compartments, varying exposure to light), we hypothesized that retaining face‐level resolution would enhance the ecological signal recovered by metabarcoding. Accordingly, we predicted that: (i) metabarcoding of separated faces would reduce the impact of sampling variance and detect higher richness than pooled protocols, and (ii) geographic and environmental gradients would be more strongly recovered when ARMS faces are analysed individually or by structural category. Material and Methods Field Sampling and ARMS Sampling Design Autonomous Reef Monitoring Structures (ARMS) were deployed by scuba divers at 12 sites distributed across three regions of the French Mediterranean coast (Figure ) at depths of 16–22 m (Table ). Units were immersed between 27 February and 7 April 2017 and retrieved between 5 March and 9 May 2018 (Table ). Before dismantling and photographic documentation, each ARMS was kept in aerated seawater for 1 h to overnight. To remove vagile fauna, each plate was gently shaken prior to photography. After imaging, the sessile benthos was scraped from each face, blended for 15 s with sea water or ethanol, placed in a 40 μm mesh bag, slightly rinsed, then placed in a Falcon 50 mL tube filled with ethanol, and stored at +4°C until DNA extraction. For some samples processed in 2019, homogenization was performed after a period of storage in ethanol rather than immediately after collection. Detailed processing information for each sample is provided in the File (MIEM sheet). Each ARMS consisted of nine superimposed plates (Figure ), with an upper face (oriented upward) and a lower face (oriented toward the seafloor). In alternating inter‐plate spaces, crossbars subdivided the gap into four compartments (see Pearman et al. ). Plate faces were designated by plate number and orientation (e.g., 1T, 2B, 2T ... 9T, where ‘T’ denotes the upward‐facing surface and ‘B’ the downward‐facing surface). For subsequent statistical analyses, we grouped the 17 faces into five structural categories: 9T, the uppermost face;BC (bottom closed): faces 2B, 4B, 6B, 8B;BO (bottom open): faces 3B, 5B, 7B, 9B;TC (top closed): faces 1T, 3T, 5T, 7T;TO (top open): faces 2T, 4T, 6T, 8T. This hierarchical face‐type classification was used to test how fine‐scale microhabitat influences community composition, and to evaluate whether grouping faces by structural categories before DNA extraction is an efficient approach for metabarcoding‐based monitoring. To assess the benefits of face‐by‐face sampling and compare imaging with metabarcoding at the finest level, we analysed at each site: (i) one ARMS unit with all 17 individual faces processed separately, and (ii) two ARMS units with all sessile material pooled into a single DNA extraction for each ARMS. ARMS retrieved in 2018 were replaced by uncolonized ARMS during the same dive. This new set of ARMS was retrieved and dismantled in 2019 (Table ). In the present study, these samples were exclusively analysed with metabarcoding under the new five structural categories scheme, with one DNA extraction and metabarcoding analysis for each category in each of the three ARMS per site. Photo Analyses We used a SONY DRC‐RX10M3 camera mounted on a fixed arm to photograph each face of every ARMS unit. Images were taken with a focal length of 35 mm, aperture F/6.3, exposure time 1/100 s and ISO 80 sensitivity. Photographs were captured under optimal natural light conditions. Each image was 4864 × 3648 pixels (350 ppi, 24‐bit colour depth, sRGB colour space). Images were subsequently cropped in IrfanView without quality loss and uploaded to the CORALNET website (https://coralnet.ucsd.edu/; Beijbom et al. ). Final cropped image files ranged from a few to about 10 MB, depending primarily on the proportion of uncolonized substrate. Each image was then subdivided in 16 sub‐squares with 4 randomly drawn points, to a total of 64 points (stratified random sampling, Figure ). For the annotation process, a list of labels was prepared (i.e., names of taxa or names of non‐taxonomic categories to which a point could be assigned, fully described in File ). During the annotation process, this list has been modified to add new taxa and to refine categories. For labels added to distinguish two categories which were previously merged in a single label, experts had to update already assigned images. The CORALNET search engine helped finding the corresponding images. Annotations were done manually until the automatic classifier was trained enough to propose labels, which were then either validated or corrected by experts (Beijbom et al. ; Chen et al. ; Williams et al. ). An initial model was generated after 50 annotated images, and a new model was trained every 50 additional annotated images. Over the course of 595 annotated images, 11 models were generated, with a maximum average accuracy of 69%. However, all label assignments were reviewed by experts, and the AI system served only as a tool to assist in the annotation process. Metabarcoding Protocol: From Benchtop to Desktop For DNA extraction, 1.5 mL of homogenate (see above: field sampling) was used. To minimize cross‐contamination, extractions were performed in individual tubes rather than in the 96‐well plate format of the extraction kit (see Corse et al. ). DNA extractions and COI metabarcoding followed the protocol described by Thomasdotter et al. (), with the following key steps: (i) DNA was extracted using the NucleoSpin Soil kit (Macherey‐Nagel, Germany); (ii) the cytochrome c oxidase subunit I (COI) gene was amplified in triplicate PCRs with primers IIICRrevN (3′‐GGNTGAACNGTNTAYCCNCC‐5′) and HBR2d (5′‐TAWACTTCDGGRTGNCCRAARAAYCA‐3′), which include a heterogeneity spacer and 11–13 nucleotide sample‐identifying tags, and target a broad spectrum of eukaryotic and algal phyla; (iii) a two‐step tailed PCR generated paired‐end, ready‐to‐load libraries; (iv) libraries were sequenced on an Illumina MiSeq using v2 chemistry (2 × 250 bp); and (v) a full series of negative and positive controls (two distinct mock communities) was incorporated. Before the workflow was published in Thomasdotter et al. (), we conducted optimization trials (File ) to assess how the choice of DNA extraction kit and polymerase affected sequencing yield and taxonomic breadth. Six ARMS‐derived samples (four sessile, two mobile) were each extracted at least twice using two DNA extraction kits, and all extracts were amplified in triplicate with two polymerases, generating 72 uniquely tagged amplicons for direct comparison. Optimization trials showed that the NucleoSpin Soil kit recovered broader diversity and that the Qiagen Multiplex polymerase produced higher read and ASV yields; this combination (described by Thomasdotter et al. ) was therefore adopted throughout the study. For bioinformatic processing we used VTAM v0.2.0 (Validation and Taxonomic Assignment of Metabarcoding data; González et al. ), a pipeline specifically designed to leverage the presence of technical replicates and control samples (including negative controls and mock communities). The core principle of VTAM is to optimize parameter values across multiple filtering steps to maximize the removal of false‐positive occurrences in control samples while retaining all expected variants (i.e., no false negatives in mock samples and the minimal possible number of false positives). Once optimized, these parameters are applied uniformly to all samples within the same sequencing library and run, under the rationale that parameter values validated on control samples are also appropriate for the corresponding experimental samples. Briefly, paired‐end reads were merged, quality filtered, demultiplexed, and trimmed to remove sequencing tags and primers. Default parameters were used, except that the minimum and maximum lengths of merged and trimmed reads were set to 295 and 331 bp, respectively. Amplicon Sequence Variants (ASVs) were initially filtered to remove low‐frequency noise using low‐stringency default thresholds, excluding occurrences with low absolute or relative read counts within samples or across variants. Such occurrences likely originate from sequencing or PCR errors, tag‐jumps, or minor cross‐sample contamination. Optimal parameter values for these filters were then inferred from the control samples, after which the filtering procedure was repeated using the optimized parameters. Final parameter settings are provided in File . To ensure reproducibility among technical replicates, only ASV occurrences detected in at least two of the three replicates per sample were retained, and replicates exhibiting composition profiles strongly divergent from the others were discarded. Subsequent filtering steps included the removal of potential chimeras and pseudogenes. Taxonomic assignment was performed in VTAM against the COInr database (version 2022‐05‐06; https://zenodo.org/records/6555985), after removing Insecta sequences to avoid spurious assignment of marine ASVs within Arthropoda (Meglécz ). Assignments were made by the algorithm of mkLTG (Meglécz ), using a stepwise identity threshold decreasing from 100% to 85%, ensuring high‐resolution identifications (genus or species) when close reference sequences were available and lower‐rank assignments when references were missing. The Minimum information for reporting of environmental metabarcoding data (MIEM checklist; Klymus et al. ) is reported in File . Statistical Analyses of Diversity Assemblage Datasets For the photography analyses, CORALNET output tables give, for each photography, the number of points assigned to each label. For most comparisons with metabarcoding, non‐biological labels such as not colonizable or bare substrate were removed, as was the label unknown, which generally represents undetermined organisms and could bias biodiversity estimates by conflating distinct taxa or by duplicating categories. For each remaining biological label, we calculated a standardized abundance by dividing the number of annotated points by the total number of labelled points on the image. From these data, we constructed three nested datasets: (i) PhoLab0, comprising all biological labels (38 in total); (ii) PhoLab2, in which several labels were merged to yield 34 broader biological categories; and (iii) PhoPhy, in which labels were pooled into nine higher‐level taxa corresponding to six animal phyla (Annelida, Bryozoa, Chordata, Cnidaria, Mollusca, Porifera), two photosynthetic groups (Ochrophyta and Rhodophyta) and a single label representing Foraminifera (File ). For metabarcoding, the filtered dataset was first analysed at the level of molecular operational taxonomic units (mOTUs)—clusters of amplicon sequence variants (ASVs) generated with Swarm v3 (Mahé et al. ) at d = 7. It was then aggregated into higher‐rank taxa (34 classes or 19 phyla or equivalent ranks) after removing mOTUs without any class or phylum‐level assignment (File ) to enable comparisons with photographic data with more similar numbers of taxa. Taxonomic richness and relative abundance were calculated for each ARMS plate face and for both methods. Coverage‐based rarefaction and extrapolation curves were computed from incidence data (presence–absence across samples) with the package iNEXT, version 3.0.1 (Chao et al. ; Hsieh et al. ) for both photography and metabarcoding datasets. These curves estimate how diversity accumulates with sample number and provide asymptotic estimates of total richness. Taxonomic richness was expressed as the Hill numbers with parameter q = 0 (D0), the exponential of Shannon–Wiener entropy (Hill number with q = 1, D1), and the inverse of the Simpson concentration index (Hill number with q = 2, D2, i.e., the inverse of the sum of squared frequencies of each taxon). α‐Diversity For each of the four datasets, α‐diversity (species richness) was quantified as the number of distinct photographic labels or mOTUs using PRIMER 7 (Clarke and Gorley ) with the PERMANOVA+ add‐on (Anderson et al. ). To verify that the choice of richness metric did not affect cross‐method comparisons, we also calculated Shannon and Simpson indices and the related (not identical) Hill numbers (q = 1 and q = 2, respectively); because these metrics were strongly correlated with sample richness, they were not discussed further. Concordance between methods was evaluated by computing Pearson and Spearman correlation coefficients for paired photographic and metabarcoding α‐diversity values from the same plate‐face samples. Relative‐abundance barplots were generated with the ggplot2 package, version 3.5.2 (Wickham et al. ) using all photographic labels and metabarcoding mOTUs aggregated to higher‐rank taxa. Linear regressions confirmed that correlations were consistent across the different diversity indices (not shown). Correlation matrices and visualizations were generated with the corrplot package, version 0.95 (Wei et al. ). These analyses were performed at two spatial scales: the sample level (166 individual plate faces) and the site level (10 sites represented by the summed abundances of the 17 faces of their analysed ARMS). β‐Diversity, Geography and Environment For each dataset, β‐diversity was assessed from pairwise community dissimilarities calculated with the Bray–Curtis index (after a square root transformation of standardized abundances) and, for comparison, the Sørensen index, which relies solely on presence–absence data and is sometimes preferred for metabarcoding as read numbers may imperfectly reflect biomass or abundance (Deiner et al. ; Hajibabaei et al. ). Mantel tests were used to evaluate correlations in community dissimilarity between photographic and metabarcoding data (using the package ape, version 5.8.1; Paradis et al. ) at both sample level (all individual plate faces) and site level (45 non‐redundant site pairs). To examine spatial and environmental structuring, least‐cost sea distances (km) were computed among sites and correlated with Bray–Curtis dissimilarities using Mantel tests. Environmental variables used in this study were derived from E.U. Copernicus Marine Service Information products. Physical time series (1987–2019) were obtained from the MEDSEA_MULTIYEAR_PHY_006_004 reanalysis (Escudier et al. ), and biogeochemical time series (1999–2018) from the MEDSEA_ANALYSIS_FORECAST_BIO_006_008 reanalysis (Teruzzi et al. ). For the 12 ARMS anchor points (latitude, longitude), 15 variables were extracted from 180 monthly CSV files, including temperature (thetao, bottomT), salinity (so), eastward and northward current velocity (uo, vo), sea surface height (zos), mixed layer thickness (mlost), nitrate (nit), phosphate (pho), phytoplankton carbon biomass (pcb), chlorophyll (chl), primary production (ppn), dissolved oxygen (dox), partial pressure of CO2 (pco) and pH (ph). For each anchor point, basic statistics (mean, median, maximum, minimum, range, and first and third quartiles) were calculated for all variables using data averaged vertically between 1 and 50 m, on 4 × 4 km grids, from the full temporal coverage of the physical (1987–2019) and biogeochemical (1999–2018) models. All computations were performed in GRASS GIS version 8 (GRASS Development Team et al. ). Euclidean distances based on scaled environmental variables (Latitude, Longitude, mean_thetao, range_thetao, mean_bottomT, range_bottomT, min_BottomT, max_bottomT, so, uo, vo, zos, mlost, nit, pho, pcb, chl, ppn, dox, pco, ph) were used in Mantel tests and partial Mantel tests (i.e., after sea least cost distance was accounted for) to disentangle geographic and environmental effects. Permutational multivariate analyses of variance (PERMANOVA; PRIMER 7 with PERMANOVA+) were performed on all photographic and metabarcoding datasets (across three taxonomic aggregation levels each) with models including combinations of factors such as region, site (nested within region or alone), face type (9T, BO, BC, TO, TC), defined by the combination of orientation (top vs. bottom), and compartmentalization (open vs. closed) (Table ). We tested for homogeneity of dispersion in community composition across site, face type and region using the PERMDISP routine in PRIMER7 PERMANOVA+. For metabarcoding data only, we compared the ARMS units analysed face‐by‐face (one per site) against the two companion ARMS per site whose sessile fractions were pooled by means of barplots of relative abundance and richness by phylum and region. We also regressed site‐level β‐diversity with the logarithm of ‘geographic distance + 1’ and with Euclidean environmental distances using Mantel tests as above. R analyses used the version 4.2.0 (R Core Development Team ) and all scripts and input files are in the github repository indicated in the data availability statement. Results Taxonomic Composition, Rarefaction Patterns and Face‐Type Contrasts Photography and metabarcoding revealed distinct yet partially overlapping characterization of ARMS sessile communities. To ensure robust cross‐method comparisons, we removed four low‐diversity samples—due to one photographic sample lacking any biological label (BAN3‐3_T5) and three metabarcoding samples with very few reads and mOTUs (BAN2‐3_T9, BAN2‐3_B8 and VIL2‐1_B8). In the photographic dataset, non‐biological categories dominated: ‘bare substrate’ accounted for ~43% of the surface of all 598 images, ‘not colonizable’ for ~12% (reflecting ARMS structural elements such as compartments and screws) and ‘detritus’ for ~5%. Comparisons were therefore restricted to the 38 biological labels (algae, animals and Foraminifera) distributed across nine to 10 phyla (File ). Metabarcoding of the remaining 166 paired samples detected 1493 mOTUs, an order of magnitude higher than the 38 photographic labels. Out of these ~1500 mOTUs, 150 were assigned to 139 species, some also observed visually. Twenty‐six of the 38 labels found in the 166 photographs were species (13) or genus (13) names. Only two of them were identified from the 166 metabarcoding samples (the bryozoans Idmidronea atlantica , Schizomavella mamillata) but seven additional shared species (metazoans and algae) were found when considering all SEAMoBB samples. Some appeared or became identifiable visually only in ARMS immersed for longer durations; others were only identified via close‐ups or a hand lens. Eight species ( Amphipholis squamata , Antedon bifida , Bittium latreillii, B. reticulatum , Limaria hians , Mimachlamys varia, Paracentrotus lividus and Pseudoceros velutinus ) detected via metabarcoding (from sessile fractions) had been visually identified but excluded from the photography dataset, as they were considered part of the mobile fraction. Of the 12 species identified in the 166 photographs but absent from the metabarcoding, two had been detected in other SeAMoBB French ARMS metabarcoding samples ( Caberea boryi and Ciona intestinalis ) and one, Anomia ephippium —known for COI primer site polymorphism (Chenuil, pers. comm.)—had all specimens removed for population genetics. Of the nine remaining species, four lacked COInr entries ( Miniacina miniacea , Platonea stoechas, Schizobrachiella sanguinea, Tubulipora plumosa ). Two had entries but were assigned at higher ranks because: (i) the barcode sets of two species in COInr (Ciona roulei and Ciona intestinalis ) were too similar, so some variants were assigned only at the genus rank, or (ii) a mislabelled COInr sequence caused Alcyonium coralloides to be assigned at the Eumetazoa rank. These deductions came from reassigning mOTUs against a COInr subset of the 12 ‘missing’ species. Therefore, three species, Ascandra contorta, Ascidia mentula and Ascidiella scabra were likely affected by amplification failure. Sample richness ranged from 1 to 13 biological labels for photography and from 7 to 76 mOTUs for metabarcoding, underscoring the much finer taxonomic resolution of DNA data. Coverage‐based rarefaction and extrapolation (Figure ) confirmed this pattern: photographic richness approached saturation, whereas metabarcoding curves remained far from asymptotic, especially for q = 0 (rare taxa). Asymptotic Hill numbers of order 0 (richness), order 1 (abundant taxa, related to Shannon's entropy) and order 2 (dominant taxa, related to Simpson's index) were 40 (s.e. = 6.9), 20.4 (s.e. = 0.5) and 16 (s.e. = 0.44) labels for photography, compared with 2391 (s.e. = 75), 902 (s.e. = 16) and 455 (s.e. = 10) mOTUs for metabarcoding. Taxonomic distributions also differed between methods and among plate faces (Figure ). Both approaches revealed higher Rhodophyta abundance on top faces, but Bryozoa and Annelida were proportionally far richer in photographs, whereas Arthropoda and Echinodermata were detected only by metabarcoding. Pooling all sessile fractions per ARMS drastically reduced diversity in metabarcoding: 22 pooled samples (two ARMS per site plus one biological replicate for MAR3 and MAR4) yielded 559 mOTUs—less than half of the 1493 mOTUs obtained from 166 separated‐face samples from 10 ARMS (Figure ). Phylum abundances were more even in the face‐by‐face design, whereas pooled samples over‐represented Ochrophyta and, to a lesser extent, Porifera (Figure ). In the 2019 metabarcoding dataset, based on five structural face categories across three ARMS per site, 137 samples were retained after quality filtering (13 removed from 150). These yielded 1541 distinct mOTUs, slightly exceeding the 1493 recovered from the 166 individual‐face samples analysed in 2018. Relationships Between Photographic and Metabarcoding Diversity Sample‐level α‐diversity values showed a slight but significant negative correlation between photographic and metabarcoding datasets (Pearson R = −0.17, p = 0.0245; Figure ). After pooling all plate‐face samples within each site, the relationship became positive but non‐significant (R = 0.24, p = 0.51), insignificance being likely driven by an outlier site from the Villefranche‐sur‐Mer region. Site richness was roughly 15‐fold higher for metabarcoding than for photography (Figure ). When taxa were aggregated to higher ranks, positive site‐level correlations became clearer. The strongest correspondence (Figure ) occurred when metabarcoding data collapsed to 34 classes were compared either with the full photographic label set (38 labels; R = 0.71, p = 0.021), or with the intermediate labelset 2 (34 labels; R = 0.69, p = 0.026; Figure ). In contrast, β‐diversity (Bray–Curtis dissimilarity) exhibited consistently positive and stronger correlations between methods, both at the sample level (R = 0.39, Mantel p = 0.0001; Figure ) and at the site level (R = 0.72, Mantel p < 0.0001; Figure ). These positive correlations persisted at all taxonomic aggregation levels, but coefficients declined steadily with coarser taxonomic ranks, reaching non‐significant values (R ≈ 0.20) when both datasets were reduced to phylum level (Figure ). The contrasting signs of α‐diversity correlations at sample versus site level can largely be attributed to strong differences among ARMS face types: for instance, the superior 9T face was the most species‐rich in metabarcoding but among the least diverse in photographic data, whereas bottom‐open (BO) faces showed the opposite trend (Figure ). Geographical and Ecological Effects on β‐Diversity Both photographic and metabarcoding datasets revealed strong spatial and ecological structuring of community composition. PERMDISP detected differences of dispersion among regions (p = 0.0001, p = 0.0006 for metabarcoding and photography respectively), sites (p = 0.0256 and p = 0.0001) and face types (p = 0.0165 for metabarcoding only). PERMANOVA models (Table ) detected highly significant effects of region, site and plate‐face type on β‐diversity. When region and site (nested within region) were included jointly, regional effects weakened (higher p‐values), indicating that much of the regional signal was captured at the site level. Face orientation (top vs. bottom) was consistently highly significant, whereas compartmentation (open vs. closed) was strongly significant for photographic data and moderately significant for metabarcoding. Non‐metric multidimensional scaling (nMDS) ordinations (Figure ) corroborated these patterns: samples separated clearly by face orientation, open–closed differences were more diffuse, and regional structure appeared as an east–west gradient along a diagonal axis. This geographical gradient was slightly more pronounced for metabarcoding, whereas fine‐scale ecological differentiation within‐ARMS (related to light and microhabitat) was clearer in the photographic ordination. Although dispersion in community composition varies between regions, sites or face types, nMDS plots confirm that PERMANOVA results reflect actual shifts in community composition, not just differences in dispersion. Analyses based on the Sørensen index, which is based on presence–absence rather than abundance, yielded very similar results (data not shown). Community similarity declined significantly with the logarithm of sea least‐cost geographic distance for both methods, but more tightly for metabarcoding (Mantel R = 0.70, p = 0.00003) than for photography (R = 0.55, p = 0.00038; Figure ). The weaker relationship for photography likely reflects unusually high similarity between the most distant regions (Banyuls‐sur‐Mer and Villefranche‐sur‐Mer) and low similarity among some Marseille sites. Taxonomic aggregation reduced distance–decay correlations for both datasets, though correlations remained significant. When metabarcoding mOTUs were collapsed to class or phylum, Mantel coefficients decreased to R = 0.43 (p = 0.00183) and R = 0.34 (p = 0.00482), respectively (File ). Similarly, aggregating photographic data to labelset 2 or phylum yielded R = 0.47 (p = 0.00138) and R = 0.27 (p = 0.01593) (File ). By contrast, pooled‐faces metabarcoding performed worst, producing the lowest and non‐significant correlations with geographic distance (R = 0.18, p = 0.1359; Figure ). Metabarcoding based on five structural face categories (three ARMS per site, 2019 dataset) yielded the strongest distance–decay relationship (R = 0.84, p = 0.00002; Figure ), outperforming the 17‐face design despite using fewer samples (137 vs. 166). Regression intercepts indicated that similarity at 0 km averaged 64% for photography, 27% for face‐by‐face metabarcoding and only 16% for pooled metabarcoding, highlighting the reduced consistency of pooled samples. The greater distance of metabarcoding rarefaction curves from their asymptotes likely contributes to this pattern, indicating higher sampling variance relative to photography. Environmental variables also helped explain β‐diversity patterns (File ). Mantel tests showed that scaled Euclidean distances in several environmental descriptors were significantly correlated with community dissimilarity, although geographic distance consistently explained more variance. Partial Mantel tests controlling for geography identified four significant predictors of species composition for metabarcoding (eastward current velocity, nitrate, phosphate and primary production) and two for photography (nitrate and chlorophyll concentration). In addition, chlorophyll was nearly significant for metabarcoding, and eastward current velocity, phosphate, phytoplankton carbon biomass, and primary production approached significance for photography. Metabarcoding using five structural face categories (three ARMS per site) identified six significant environmental variables, the highest number across all datasets. Discussion Our key methodological advance—separating ARMS faces for metabarcoding—proved decisive for improving biodiversity assessments. By increasing detected richness, reducing sampling variance, and strengthening biogeographic and environmental correlations, this approach markedly outperformed the standard NOAA pooling protocol and provides a new benchmark for standardized eDNA monitoring. In the following discussion, we examine (i) the complementarity between photography and metabarcoding, (ii) the causes of their differing performance and (iii) the implications of our results for sampling design and for routine monitoring frameworks—with a focus on the benefits of separating ARMS structural face categories (or all 17 faces) compared with pooled designs for metabarcoding. Finally, we consider the broader implications of these findings for long‐term marine biodiversity monitoring and management. Complementarity of Photography and Metabarcoding α‐ and γ‐diversity were markedly higher in metabarcoding than in photographic analyses, even under a stringent bioinformatic pipeline (González et al. ). This pattern reflects the finer taxonomic resolution of COI metabarcoding—which can distinguish cryptic species lacking morphological differentiation (Cahill et al. ; Chenuil et al. )—as well as its ability to detect a broader size range of organisms. The two methods also differed in how they resolved variations among ARMS structural face categories. Photography recorded the lowest diversity on superior faces (9T), whereas these yielded the highest richness by metabarcoding. A likely explanation is that dense macroalgal cover (e.g., Ochrophyta such as Dictyota spp.) obscures underlying microhabitats from visual detection, while providing more colonizable surface for cryptic taxa that are well captured by DNA. Conversely, photography revealed the lowest richness on top‐closed faces (a pattern not observed in metabarcoding) likely reflecting differences in light exposure and the availability of bare‐substrate. These contrasting face‐type effects explain the negative correlation of α‐diversity between methods at the sample level, whereas pooled site‐level richness showed a positive—but non‐significant—correlation, possibly constrained by the limited number of sites (n = 10). The taxonomic spectra recovered by the two methods overlapped but differed in detail. Every phylum detected by photography was also identified via metabarcoding. However, groups like Bryozoa and Annelida were abundant in photographs but rare in metabarcoding, unlike Cnidarians. Arthropods, absent in photographs, were abundant in metabarcoding, likely due to poor separation of sessile and mobile compartments and the high sensitivity of PCR (Pearman et al. ). Of the nine species identified in photographs but never detected in metabarcoding, three likely resulted from PCR failure or inefficiency, four from missing entries in the reference database, one from a database error, and one from a barcode marker resolution issue. Many photography–metabarcoding discrepancies could therefore be reduced through improved reference databases and taxonomic curation, whereas others reflect intrinsic limitations of single‐marker metabarcoding. Metabarcoding's failure to detect a taxon may also result from low DNA abundance or stochastic sampling effects (Lamb et al. ; van der Loos and Nijland ). Most metabarcoding diversity nevertheless consisted of mOTUs lacking phylum‐level assignments, either because reference sequences remain unavailable for many lineages or because the amplified COI fragments originated from poorly characterized non‐metazoan organisms, including unicellular eukaryotes and even prokaryotes (Collins et al. ; Hintikka et al. ). Although photographic analysis detected far fewer taxa than metabarcoding, both approaches effectively revealed significant effects of all tested spatial and ecological factors on community composition. These factors included region (three regions separated by up to ~395 km), site within region (0–20 km apart), and ARMS face type (9T, top open, top closed, bottom open, and bottom closed). However, subtle but consistent differences emerged in the drivers of community composition detected by each method; p‐values differed slightly between them, indicating nuanced contrasts in sensitivity. Photography detected more significantly than metabarcoding differences in microhabitat within ARMS (plate compartmentation). By contrast, metabarcoding appeared more tightly correlated with geographical distance and environmental variables (see below). Together, these comparisons show that photography and metabarcoding provide complementary insights into marine hard‐bottom biodiversity. Photography offers direct visual estimates of macroscopic cover (e.g., macroalgae), while metabarcoding captures cryptic, microscopic or early‐stage organisms and preserves fine‐scale heterogeneity. Future monitoring programmes may combine both approaches according to management objectives: photography for dominant taxa and surface cover, and metabarcoding for cryptic diversity, early colonizers and compositional change in turbid, complex or species‐rich environments. The broad convergence between methods supports their reliability and robustness, and the stronger geographic and environmental signal in metabarcoding highlights its greater sensitivity to subtle compositional change. Taxonomic Resolution Explains the High Sensitivity of Metabarcoding to Environmental and Spatial Gradients In contrast with α‐diversity patterns at sample levels, β‐diversity patterns were strongly congruent between methods, with significant Mantel correlations at both sample and site levels (Figure , Figure ). This indicates that both approaches reliably captured community structure. However, metabarcoding revealed stronger correlation of community similarity with geographic distance (R = 0.70 vs. 0.55 for photography, with a single ARMS per site). Likewise, both methods revealed similar environmental variables correlated with community dissimilarity, but metabarcoding consistently produced higher statistical power, with several variables that were marginally non‐significant in photographic data becoming significant in metabarcoding. Aggregating mOTUs into higher ranks (i.e., classes and phyla) progressively weakened these correlations, ultimately falling below those from photography, demonstrating that the stronger biogeographic and environmental signal with metabarcoding stems from its finer taxonomic resolution. We expected mOTU‐level metabarcoding to reflect geographic distances more clearly than photography because it captures intraspecific diversity (e.g., Thomasdotter et al. ; Turon et al. ) that is frequently structured by isolation‐by‐distance or oceanographic connectivity (Legrand et al. ). Marine species frequently exhibit spatial variation in mitochondrial COI haplotype frequencies, including along the French Mediterranean coast where the present study takes place (Boissin et al. ; Cahill et al. ; Penant et al. ). Moreover, numerous cryptic species complexes, genetically distinct but morphologically indistinguishable, exhibit clear COI divergence (Cahill et al. ; Chenuil et al. ; Egea et al. ). Photographic analyses cannot resolve such variation, and most visually distinguishable taxa using our photographic protocol are widespread throughout the north‐western Mediterranean and therefore less likely to reflect geographic distances. Consequently, metabarcoding was expected to reflect geographic distances more clearly than photographic analyses, consistent with the stronger distance–decay relationships observed. For environmental factors, expectations are less obvious. Because metabarcoding detects small, fast‐responding organisms, including unicellular taxa amplified by COI, community changes might be captured more rapidly by DNA than by imagery. This aligns with the classical view that microorganisms respond more strongly to environmental conditions than to geographic barriers, a concept captured by the phrase ‘everything is everywhere, but the environment selects’ (De Wit and Bouvier ; Van der Gast ). Under this paradigm, metabarcoding could be especially sensitive to environmental variation. However, photography better reflects dominant taxa that are likely well adapted to local conditions. They should, in principle, have abundances in photography that mirror biomass more faithfully than with metabarcoding. Read counts are imperfect indicators of abundance, owing to multiple sources of technical biases: (i) subsampling of the homogenized community may under‐represent rare taxa; (ii) DNA extraction efficiencies vary among taxa; and (iii) PCR amplification is taxon‐specific, with some lineages weakly or not amplified (Lamb et al. ; van der Loos and Nijland ). All else being equal, such biases should reduce the sensitivity of metabarcoding to environmental gradients. Yet our results contradict this expectation: metabarcoding exhibited stronger correlations with environmental variables than photography. This indicates that the far greater taxonomic richness recovered by metabarcoding compensates for its lower quantitative fidelity, enhancing its ability to detect environmentally driven changes in community composition (see also Deiner et al. ). Implications and Perspectives for Routine Monitoring Standardized ARMS protocols have become a cornerstone of global marine biodiversity observation, enabling comparable DNA‐ and photography‐based data across studies, regions and time (Ransome et al. ; Obst et al. ; Daraghmeh et al. ). However, substantial heterogeneity remains in sample treatment, molecular workflows and bioinformatic processing, potentially limiting interoperability across monitoring programmes and observatory networks (Jessop et al. ). Recent initiatives therefore emphasize standardized metadata, transparent workflows and FAIR practices to support data reuse and cross‐study integration (Klymus et al. ; Takahashi et al. ). These requirements are particularly critical for eDNA metabarcoding, because even minor changes in sampling, storage, sequencing or bioinformatic processing can cascade through analytical workflows and compromise inter‐study comparability (Zinger et al. ; van der Loos and Nijland ; Alberdi et al. ). Building on these challenges, our study refines key steps of ARMS‐based monitoring—from field deployment to molecular and photographic processing—to improve the consistency and long‐term comparability, providing the baseline on which more detailed sampling strategies can build. Sampling Design and Sampling Variance: Why Pooling All ARMS Faces Before Metabarcoding Should Be Avoided A major difference between the four methods implemented in this study concerns sampling variance. Metabarcoding was more prone to stochasticity than photography, missing a larger fraction of taxa that are potentially detectable. In our dataset, this was evident from the markedly lower community similarity between adjacent sites and from the rarefaction curves further from their asymptotes compared with photographic data. Such variability is consistent with previous reports of eDNA detection heterogeneity (Zinger et al. ; Alberdi et al. ; Lamb et al. ). Because sequencing depth was not limiting (data not shown), this variance most likely reflects technical artefacts introduced during DNA extraction, PCR amplification and data filtering (Lamb et al. ; van der Loos and Nijland ; Zinger et al. ). By contrast, photographic analysis—although based on 64 randomly drawn points per image and thus theoretically subject to some sampling errors—missed far fewer taxa and yielded more consistent estimates of local similarity. These results strongly argue against pooling all sessile fractions from a single ARMS prior to metabarcoding. Indeed, the similarity of adjacent sites reached values as low as 5% in pooled faces metabarcoding, compared with 20% for separated‐faces metabarcoding and 50% for photographic analyses. Analysing the 17 faces of a single ARMS per site separately increased the number of detected mOTUs by more than 2.4‐fold compared with pooling all sessile fractions from 22 ARMS (across 10 sites). The gain in diversity was even higher with the design separating five structural face categories across three ARMS per site. In addition, separating structural face categories may also reduce diversity loss caused by DNA degradation or PCR inhibition from bioactive compounds unevenly distributed among sessile taxa (Leray and Knowlton ). Shifts in phylum representation between pooled and separated designs further suggest that DNA extraction and amplification efficiencies vary with sample composition (see also Lamb et al. ; van der Loos and Nijland ). This hypothesis—namely that separating samples into structural face categories improves site characterization beyond the mere reduction of sampling variance—is also supported by the observation that community composition based on the five structural face categories captured geographic and environmental variation more strongly than the 17‐face design, even though it was based on fewer samples (137 vs. 166 across 10 sites). These results show that sampling design strongly influences both detected diversity and the identification of ecological drivers, making it a critical component of monitoring programmes. Our findings challenge the standard NOAA protocol, which recommends three ARMS units per site with pooled sessile fractions. Given the very low correlations we observed between community differentiation and geographic or environmental distances in pooled designs (Figure ), it is unlikely that NOAA protocol (based on three pooled ARMS per site) would perform substantially better than the two pooled ARMS per site analysed here. Because separating 17 sessile samples per ARMS entails substantial extra work, and because the five‐category scheme provided an even clearer ecological signal, we subsequently adopted the intermediate solution for the SEAMoBB project: pooling faces into five structural categories based on orientation and compartmentalization, while retaining three ARMS per site. We considered including plate number (i.e., distance from the seafloor) as an additional structural factor. However, preliminary analyses showed no consistent pattern (data not shown), and plate number cannot be fully crossed—with respect to statistical design—with orientation and compartmentalization. Overall, this modified protocol offers a balanced compromise between logistical effort and ecological resolution. Refining Photographic Analysis: Greater Precision With Minimal Additional Effort Photography‐based analyses also offer room for methodological refinement. Most ARMS image studies have relied on 50 random points per plate face, following a comparative analysis indicating that this density is generally sufficient. By increasing the density to 64 points, we improved precision without a meaningful increase in processing time, and found that photographic data were much less prone to sampling variance than metabarcoding, because adjacent sites harboured more similar communities (see also David et al. ). Although not formally quantified, increasing point density from 50 to 64 required little additional effort while improving abundance estimates. Nevertheless, further gains are likely to come from improving image quality and resolution, which would require higher‐grade cameras and greater digital‐storage capacity, rather than from further increasing point density. Expanding the number of taxonomic labels can enhance detection of changes in community structure, biomass and anthropogenic impacts (PERMANOVA tests, not shown), but it also increases training time and inter‐expert variability (see also Beijbom et al. ; Williams et al. ). Looking ahead, artificial intelligence–assisted image recognition is expected to accelerate annotation, improve consistency and reduce expert bias (Bravo et al. ; Raphael et al. ; Trotter et al. ). Traceable, Quality‐Controlled Bioinformatics Long‐term eDNA monitoring requires datasets that remain comparable across sites, seasons and sequencing runs through rigorous quality control and standardization (Pawlowski et al. ; Petit‐Marty et al. ; Rimet et al. ). International standards and recent reviews emphasize bioinformatic workflows that minimize false positives and false negatives, ensure traceability, and support reproducible large‐scale comparisons (Pawlowski et al. ; Leese et al. ; Hakimzadeh et al. ; Rimet et al. ). The bioinformatic considerations raised by Turon et al. () reinforce these requirements, showing that temporal trends can be confounded by technical discrepancies across sequencing runs or workflows and thus emphasizing the need for control‐driven, fully traceable pipelines. In this context, the VTAM pipeline proved crucial as it explicitly integrates negative controls, mock communities and technical replicates to optimize filtering parameters, thereby reducing false positives and false negatives and ensuring full traceability at every filtering step (González et al. ). By adopting this validated pipeline, we ensured that the strong biogeographic and environmental patterns revealed by our face‐by‐face metabarcoding reflect true biological signals rather than analytical artefacts. This conclusion is further supported by recent studies demonstrating that VTAM‐filtered metabarcoding can generate highly validated data for fine‐scale, hypothesis‐driven analyses of community assembly and ecological interactions across aquatic ecosystems (e.g., Esposito et al. ; Ruiz et al. ; Villsen et al. ). VTAM provides a robust framework for ARMS metabarcoding by integrating negative controls, mock communities, technical replicates and transparent parameter optimization. This approach improves data comparability across sites and sequencing runs while aligning with recent FAIR data and reporting recommendations for environmental metabarcoding (Klymus et al. ; Takahashi et al. ), thereby supporting interoperability among large‐scale biodiversity observatories. Costs and Scalability Cost and time considerations reinforce these conclusions. Assigning 64 points per image required only about 4–8 min for a trained analyst, so processing our 166 images amounted to less than 1 week of full‐time work after an initial 10‐day training period. By contrast, metabarcoding required roughly 1 month of benchtop labour for ~160 samples and about €6000 for reagents, consumables and sequencing, half of which was for high‐throughput sequencing. The adoption of Illumina NovaSeq sequencing has reduced per‐sample sequencing costs more than tenfold, but exploiting these economies of scale requires pooling sufficient libraries to fill sequencing runs, which may introduce delays that must be balanced against monitoring frequency. Similar trade‐offs between detection sensitivity and economic efficiency have been demonstrated in other eDNA studies, where laboratory costs can offset field savings and strongly influence optimal survey design (Smart et al. ). Broader reviews likewise stress that the choice of sampling and processing methods remains a major determinant of cost, throughput and inter‐study comparability in benthic metabarcoding (Pawlowski et al. ), underlining the need to balance detection sensitivity with cost efficiency when designing long‐term observatories. Overall, field effort and processing time are likely to remain the main constraints for large‐scale monitoring. Sustainable implementation will therefore depend on balancing ecological resolution with logistical effort and on shared infrastructures that ensure cost‐efficiency and data comparability across regions and monitoring programmes. Conclusion By coupling high‐resolution photographic validation with a control‐driven metabarcoding workflow, this study demonstrates that retaining ARMS microhabitat structure substantially improves biodiversity assessment. Analysing structural face categories increases sensitivity while maintaining manageable effort, providing a practical and scalable design for marine biodiversity observatories. These methodological refinements align with initiatives such as ARMS‐MBON and OBON and support standardized biodiversity monitoring across space and time.