The UVP5 dataset (Fig. 2) was compiled from observations across all oceans over a 10-year period (2008–2018) through a collaboration of international partners. It includes 3114 profiles from 62 EcoTaxa projects conducted during research cruises by different institutions (Table A1), each aimed at addressing specific local or basin-scale research questions. It comprises vertical profiles validated to more than 99 % which were retained for subsequent analysis. Currently, the different annotators (in total 71, Table A2) of the 62 EcoTaxa projects chosen have sorted the image datasets in more than 250 categories from species level to phyla or by adding morphological attributes and life stages. Such detailed classification is not possible over the whole dataset with high accuracy because recognition and sorting of organisms can be a source of bias depending on the levels of perception and experience of the people who perform them. Several cognitive biases exist, such as boredom, fatigue or a classification biased towards the most used groups (Culverhouse, 2007; Culverhouse et al., 2014).

In the present dataset, to reduce these errors and ensure classification homogeneity, the images were thereafter grouped into only 33 broader taxonomic groups for living organisms by combining all vignettes within the children categories (Table 1), while in EcoTaxa the original categories are maintained. The broad taxonomy list was established following recent published works using UVP5 data (Drago et al., 2022; Panaïotis et al., 2023; Laget et al., 2024) to ensure a minimum of 50 images per category and ecological patterns at regional or global scales for specific groups or community composition. For example, rhizarians were well studied at local (Biard and Ohman, 2020) and global scales (Biard et al., 2016; Laget et al., 2024) while polychaetes were mostly studied in the tropical Atlantic (Christiansen et al., 2018). Community composition analyzes have been conducted at local scales by Forest et al. (2012) and Barth et al. (2020), and at a global scale by Panaïotis et al. (2023). Global biomass was estimated in Drago et al. (2022). Inter-comparison with plankton net data at global scale (Soviadan et al., 2024) allowed to set confidence in the abundance assessments of many of these groups.

Images of unidentifiable objects exhibiting probable biological characteristics (e.g., symmetry, appendages, or tentacles) were classified into the “plankton-like” category. In this dataset, the term “-like” refers to structured objects that resemble organisms but cannot be taxonomically identified due to limitations such as poor focus, low resolution, incomplete capture, or lack of taxonomic knowledge. For example, copepoda-like contains all vignettes of copepods for which the antenna is not visible while the body shape resembles typical prosome and urosome. When recurring objects that were not yet classified appeared frequently but with uneven spatial distribution, temporary categories were created to accommodate them and labeled as temporary (tmpXXX). When contextual information to identify them was used, “contextual” was added to their names. All images of non-living organisms, including poor-quality images (blurred, low grey level, bubbles, artificial nature, badly segmented images) or particles (aggregates, fibers, pellets) were sorted into several categories without consistency criteria across the consortium. Low-quality images (mostly objects not completely in focus) were set in a category named “artefact” that includes bad focus particles, plankton and real artefact, while particles were grouped into a single detritus category.

To ensure data consistency across projects, a subset of the dataset was reviewed to verify the accuracy of taxonomic sorting within each group. A total of 200 images were extracted from each category, or all available images if fewer than 200 were present. For detritus, 10 % of the total images were selected and independently reviewed by a different trained operator, working independently of the initial classification, after the final dataset was assembled. The error or uncertainty rate was under 8 % for taxa classification. Among the detritus, less than 0.1 % were plankton organisms. The resulting global dataset consisted of 7.05 million detritus images and 734 thousand plankton images from 3114 profiles.

Due to the large number of plots required to display the results for all categories, results are shown for a subset of the original categories. From the 33 initial categories derived from taxonomy (Table 1 and Fig. 1), five broader taxonomic groups were defined for further analysis based on shared feeding strategies and ecological traits (e.g. habitat, life mode, and trophic role). These groups can be highly useful as input for biogeochemical or community ecology models. The groups include Crustacea (Copepoda and Malacostraca), gelatinous filter-feeders (Appendicularia, Salpida), gelatinous carnivores (Chaetognatha, Narcomedusae, Siphonophorae, Ctenophora), Rhizaria (Collodaria, colonial Collodaria, other Rhizaria, Foraminifera, Phaeodaria, Acantharea), and Trichodesmium. From these groups, biovolumes were calculated along with concentrations at different depths and for the analysis of the average grey level.

The dataset presented here revealed a heterogeneous global distribution. The UVP5 deployments predominantly occurred in the northern hemisphere (66.3 % of profiles, Fig. 2), especially in tropical and temperate latitudes where seasonal variations are well observed. In contrast, the southern hemisphere had fewer observations (33.7 % of profiles), resulting in a less comprehensive dataset. The region between 30° S–30° N represented the highest proportion of UVP5 profiles (47.25 %). The vertical profiles distribution revealed a predominant concentration of measurements in the upper 1000 m of the water column, representing 40.6 % and 46.8 % for the SD and HD models respectively, with a peak occurring within the upper 500 m. Thus, most deployments targeted shallower depths. The SD version of the UVP5 demonstrated the most extensive coverage, with profiles spanning the entire latitudinal range. Meanwhile, the HD profiles were prevalent in the Northern Hemisphere. The profiles also presented a remarkable annual and interannual variation observed primarily in the northern hemisphere and the under-sampling of the southern hemisphere and the deep sea.

Detritus was the most abundant category (90.5 % of total images), indicating a substantial presence of non-living particulate organic material (Fig. 3). Copepoda dominated zooplankton (2.2 %), followed by Trichodesmium (1.3 %), Trichodesmium-contextual (1.1 %), the plankton-like category (1 %), artefacts (0.9 %), Phaeodaria (0.8 %), and Bacillariophyta-contextual (0.6 %). Each of the remaining groups represent <0.16 % of the total number of objects.

The spatial patterns of plankton and detritus concentrations in the global ocean vary significantly with depth and location. Across all depth layers, plankton and detritus abundances increased (up to 5 plankton per m³ and 60 detritus per m³ in the 0–100 m depth layer and up 2 plankton per m³ and 20 detritus per m³ in the 500–1000 m depth layer, Fig. 4) in eastern boundary current systems (especially in the Californian and Senegal upwelling systems) or in coastal seas. Minimum values (up to 2 plankton per m³and 20 detritus per m³ in the 0–100 m depth layer and up 1 plankton per m³ and 1 detritus per m³ in the 500–1000 m depth layer) were usually found in the center of large ocean gyres (notably in the South Pacific and the Indian Ocean).

The distribution of detritus followed a dome-like pattern (criteria from Drago et al., 2022), with UVP5 showing a lower detection threshold for particles below 1 mm for UVP5 SD and 600 µm for UVP5 HD (Fig. 5). In contrast, plankton categories exhibited a much flatter normalized biomass/biovolume size spectra (NBSS) distribution. Detritus dominated the smaller size range (<1 mm) while the contribution of plankton increased relative to detritus in the larger size range. Except for Rhizaria, smaller planktonic organisms were more abundant in the upper 100 m in the UVP5HD dataset, emphasizing the detection limits of the two camera versions in the small size range for both plankton and detritus.

For both UVP5s, the proportion of plankton images on all images (detritus and plankton) increased with size from 0.01 at 500 µm to about 0.8 for objects larger than 3 mm in ESD. The two instruments displayed a similar pattern with a plateau in this ratio starting at about 3 mm. In general, the ratio did not vary with depth and remained at 0.8 for objects larger than 3 mm.

When analyzed across vertical layers (Fig. 7), detritus showed the highest concentrations in the upper 100 m, with biovolume patterns mirroring concentrations and exceeding those of living plankton by several orders of magnitude. Trichodesmium (excluding Trichodesmium-contextual) was the most abundant group in surface waters, but contributed relatively little to biovolume due to its small size (Fig. 4). Crustacea displayed comparable patterns in concentration and biovolume, with high surface abundance and decreasing values with depth. Gelatinous plankton, including filter feeders (Appendicularia, Salpida) and carnivores (Chaetognatha, Narcomedusae, Siphonophorae, Ctenophora), were characterized by low concentrations but disproportionately high biovolume, particularly in the upper 100 m, with both metrics declining with depth. Rhizaria also peaked in concentration and biovolume in surface waters and decreased gradually down to 1000 m.

The key morphological properties for detritus and the five plankton categories presented similar patterns between the two UVP5 versions (SD and HD) (Fig. 8). In some cases, the range differed between versions for the groups depending on the property examined. Gelatinous groups were the biggest categories in terms of area measured, while they presented the lowest circularity, compared to the remaining groups. Crustacea was the darkest group, whilst gelatinous were the lightest ones.

The dataset provides decade-long global coverage with standardized methodologies across UVP5 versions, spanning all major ocean basins. However, the southern hemisphere and deep sea remain undersampled, and profiles are mostly concentrated in mid-high latitude of the northern hemisphere (Fig. 2). Well sampled regions correspond to long-term monitoring or time-series programs, such as the California Current, the Western Mediterranean, and the Equatorial region, through initiatives like the CCELTER program (e.g., Biard and Ohman, 2020), ongoing observational efforts of MOOSE (e.g., Llopis Monferrer et al., 2022), and long-term studies in the Equatorial region (e.g., Kiko et al., 2017, Fernández-Carrera et al., 2023). The even distribution across ocean basins is mostly the result of the Tara Ocean Expeditions (2009–2013). Despite broad coverage, small sampled volumes can limit detection of rare taxa, affect image based identification, and cause “zeros” that may reflect detection limits rather than true absence.

At larger size ranges, rare taxa require aggregation counts across thick vertical ocean layers to achieve statistical robustness (Hauss et al., 2016; Panaïotis et al., 2023). At the surface, abundant taxa may permit 10 m binning intervals (Dupouy et al., 2018; Biard and Ohman, 2020), though 20 m bins are typically more appropriate (Picheral et al., 2010). In this study, data were grouped into five depth layers (0–100, 100–500, 500–1000, 1000–2000 and 2000–4000 m) balancing resolution for abundant surface taxa with sufficient counts for rare deeper organisms. While previous publications have adopted fixed depth ranges of 0–200 and 200–500 m (Drago et al., 2022), or a dynamic layer, set as the deepest value among the mixed layer depth and the euphotic depth (median 88 m, Panaïotis et al., 2023), the choice of the vertical layering ultimately depends on the scientific question.

The two UVP5 versions are inter-calibrated based on the size spectra from repeated joint profiles (Picheral et al., 2010; Kiko et al., 2022). Users should also account for unit-specific settings: sn000 (used, for example, during Tara Oceans 2009–2012) shows a different distribution of features and may require a quantile-to-quantile transformation, or exclusion, in cross-unit analyses. Detection efficiency varies with object size (Fig. 5), where HD version captures detritus < 1 mm more effectively than the SD version, while both perform similarly above 1 mm. Size spectra generally show a peak at the detection limit, a steep decline for intermediate sizes, and a flatter distribution for large objects (Stemmann and Boss, 2012).

Variations in the depth of taxonomic classification among annotators necessitated the merging of certain groups, resulting in a final list of 33 taxa at the lowest common taxonomic denominator (Table 1). To validate classification accuracy, a random subset of images (200 from each taxon and 10 000 for detritus) was independently reviewed by a single operator. Error rates were consistently below 10 % across all groups, with most groups exhibiting error rates under 2.5 %. A source of error in the case of detritus is that many annotators had moved un-focused particles into the category artefact. This category includes images of objects arising from external or procedural factors. Examples of such artifacts include reflections or shadows from a poorly tied rope, presence of seals near the camera, cables, or other non-biological objects inadvertently captured during image acquisition, as well as badly segmented objects in nepheloid layers that modify the background. On the other hand, it may also consist of effects like bubbles caused by the descent through the air/water interface where bubbles are captured on the hardware, as well as artifacts resulting from image blurriness where no actual object or non-recognizable object is present (real artifact). When analyzing the artefact category which contains some subjectivity, considering for example that bad-focused could be very transparent objects, we found that it consists of 85 % of real detritus shaped objects, 5 % of plankton and 10 % of artefact. Sorting them in the EcoTaxa database system would be too time-consuming, so users must decide whether to include or exclude them during the analysis. Here, we merged them with detritus as they would become a minority (less than few percent). However, it is important to note that objects classified as artefacts (not shown in Fig. 7) represent less than 1 % of the total dataset when the concentration is calculated, but contribute to more than 200 % when volume is considered. Their impact is high mostly on large objects because many artefacts are observed in very turbid water where particles are badly segmented. This highlights the significance of their removal.

Overall, this dataset provides a unique and comprehensive tool for the scientific community, enabling researchers to examine the intricate relationships between plankton, detritus, and ecosystem function. The detailed classification of plankton and detritus provides a valuable resource for understanding the complexity between their abundance distribution and the environmental processes happening at global (Drago et al., 2022; Panaïotis et al., 2023; Laget et al., 2024) and local scales (see references in Table 1) within marine systems. By analyzing the abundance, diversity, and distribution of planktonic organisms, researchers can gain insights into food web structures, energy flow, and the interactions between different trophic levels (e.g., D'Alelio et al., 2016; Perhirin et al., 2025). The data allows ecologists the monitoring of plankton communities, which are essential indicators of environmental health and can reflect changes in water quality, nutrient levels, and ecosystem productivity (Muller-Karger et al., 2018). The dataset is equally valuable in biogeochemical studies, so researchers can better predict how changes in plankton community composition and detritus abundance influence nutrient cycling and the overall functioning of marine ecosystems. Additionally, biogeochemical models that integrate this dataset can help refine predictions of carbon sequestration in the ocean, an essential process for regulating atmospheric CO₂ levels and mitigating global warming. As plankton and detritus play pivotal roles in the ocean's carbon pump, understanding their distribution in different environmental conditions is crucial for assessing the resilience of marine ecosystems and their ability to mitigate or exacerbate climate change effects (e.g., Le Quéré et al., 2016; Rohr et al., 2023). Through these integrated models, the dataset offers a robust framework for exploring the complex feedback mechanisms between marine biological communities and biogeochemical processes, advancing our ability to predict future changes in marine ecosystems under varying environmental scenarios.

Future classification should focus its efforts on sorting the detritus category, as a recent study highlights that sinking speed estimated for various particle categories varies with particle morphology (Trudnowska et al., 2022; Soviadan et al., 2025) or that mesoscale activities mix them vertically according to their morphology (Accardo et al., 2025). In addition, they represent a very high proportion in abundance and probably of organic biomass compared to the plankton. Additionally, the specific classification of detritus particles, such as fecal pellets and their associated zooplankton communities, will enhance our understanding of their dynamics, improve the accuracy of carbon flux calculations, and refine their representation in biogeochemical models (Perhirin et al., 2025). Therefore, classifying particles based on their traits could provide valuable new insights into their origin and role in the carbon cycle through the gravitational pump (Boyd et al., 2019). When it comes to zooplankton, recent studies revealed that analyzing their traits, such as body size, morphology, and metabolic rates, which significantly influence their ecological roles and interactions within marine ecosystems, provides a framework for linking organismal characteristics to ecosystem functions, offering deeper insights into zooplankton behavior, community structure, and their contributions to biogeochemical cycles (Titocci et al., 2025).

Transfer learning has shown promising potential in leveraging existing datasets to enhance training for improved recognition tasks. By utilizing the present dataset as a model training for machine learning and artificial intelligence (AI) techniques can be employed to better classify and extract relevant features (Lumini and Nanni, 2019; Irisson et al., 2022). Specifically, AI could play a key role in sorting and categorizing the 7 000 000 detritus images in the dataset, enhancing the processing of the remaining 5598 UVP5 profiles and their corresponding 70 million images. Furthermore, integrating existent profiles with those acquired in the future, such as the anticipated 500 annually for UVP5 and 5000 annually for the new version of the UVP (UVP6LP and HD; Picheral et al., 2022), into the observation system will enhance the robustness of the database, ensuring its adaptability and scalability for future data handling and feature extraction. This observation effort is supported by GOOS (Dexter and Summerhayes, 2010) and already in operation in large-scale observing programs, like the Bio-GO-SHIP (Clayton et al., 2022) or possibly by the BGC-Argo (Claustre et al., 2020) program. Having better embedded classifiers is particularly important for BGC Argo for which the floats and the cameras are not recovered. The new UVP6, which can be attached or associated with different platforms, is now widespread distributed and has allowed the acquisition of innumerable profiles and 87 millions of images in the last 4 years since its commercialization in 2020. These advancements will ultimately contribute to more efficient data analysis and support the continuous improvement of environmental monitoring systems.