| 02 Sep 2026
A FAIR annotated set of 60k+ in situ zooplankton images from the Greater North Sea and Greenland
Abstract. The North Sea is a productive shelf sea with a wide variety of hydrographic and biological conditions. It is subject to high anthropogenic pressures, including climate change, terrestrial nutrient loading, and the development of offshore wind farms, and the associated environmental changes have major consequences for zooplankton abundance and community composition. In situ imaging provides new opportunities to monitor meso- and macrozooplankton, as it enables sampling at meter-scale resolution and improved assessments of fragile gelatinous taxa. The vast amounts of data that can be collected with in situ imaging necessitate the use of automated data processing pipelines. However, developing these requires a substantial amount of expert-annotated data. In this paper, we share a large annotated set of plankton and particle images from an in situ shadowgraph imager (ISIIS-DPI, 46,331 images) and the Continuous Particle Imaging and Classification Sensor (CPICS, 14,496 images) among, respectively, 64 and 42 taxonomic and trait-based classes that were collected in the Greater North Sea and complemented with data from Greenland. Based on the geographical coverage, size distribution, and environmental data collected simultaneously with the images, we show that this set is taxonomically representative of the Greater North Sea. In addition, we trained Convolutional Neural Network classifiers, one for the ISIIS-DPI and one for the CPICS, and both achieved an F1-score of 80.1 %, showing that these classifiers can be reliably used on newly collected ISIIS-DPI and CPICS data in this region. All annotated images, as well as location and environmental data, are shared publicly (10.5281/zenodo.21737141, Hovenkamp et al., 2026c) following a flexible taxonomic classification scheme to ensure flexible use and inter-operability with similar (future) sets. The annotated data and classifiers that we present should greatly facilitate zooplankton studies employing in situ imaging in the Greater North Sea. This improves our ability to assess how zooplankton communities are likely to respond to the ongoing anthropogenic pressure on the North Sea and addresses a critical knowledge need.
Dear Editor and Authors,
Thank you for giving me the opportunity to review the preprint entitled “A FAIR annotated set of 60k+ in situ zooplankton images from the Greater North Sea and Greenland”. Please find my detailed comments below.
Sincerely,
--- General comments ---
Overall, this preprint provides clear, well-structured, and efficient documentation accompanying a new dataset of in situ zooplankton images. The dataset, alongside the associated Python scripts, is clean, well organised, and readily accessible via the provided Zenodo repository. By encompassing a wide range of regions, the dataset enables convolutional neural networks (CNNs) to learn robust, generalisable features of North Sea plankton. Consequently, the dataset and CNN classifiers presented here have strong potential to serve as a regional benchmark to support future plankton classification efforts across the Greater North Sea.
However, the dataset's utility for the broader marine ecological community would be substantially enhanced if the metadata required to reconstruct plankton abundances/biovolumes were provided (i.e. total sampled volume, or at least the number of frames – including empty frames; see (I) below). Without these metadata, the dataset's reusability remains largely restricted to computer vision and machine learning applications.
An overarching issue with the manuscript is that it oscillates between technical validation and ecological interpretation. The latter is often overextended, as raw counts of annotated images cannot be directly equated to true in situ abundances, and key sampling and seasonal biases are not explicitly addressed (see (II) below). I elaborate on these aspects below.
--- Specific comments ---
(I) Complete the dataset with number of frames or volumetric metadata to make it possible to retrieve plankton abundances
At the end of the abstract, the authors state: "The annotated data and classifiers that we present should greatly facilitate zooplankton studies employing in situ imaging in the Greater North Sea. This improves our ability to assess how zooplankton communities are likely to respond to the ongoing anthropogenic pressure on the North Sea and addresses a critical knowledge need.” (Line 16-18)
While I fully agree with the first sentence (which succinctly summarises the primary merit of this data paper), achieving the ambition set out in the second sentence strictly requires deriving in situ abundances. With the metadata currently available, this is not possible (unless I missed the information).
Furthermore, the authors note: “Finally, useful exchange of annotated data between research groups, if available, is commonly inhibited due to a lack of provided metadata (e.g., sampling location)” (Line 59-61)
Sampled volume is fundamental metadata for any quantitative imaging dataset. Here, as information about the volume per frame is available, providing the total frame count (or total volume sampled per deployment directly) would allow end-users to account for frames where target organisms were absent, thereby preventing severe overestimation of plankton abundances.
Finally, regarding the statement: "However, this can be corrected for by dividing the resulting taxon densities by the class-specific recall (Hu and Davis, 2006).” (Line 236)
To enable users to perform this correction accurately, class-specific recall values (and associated classification performance metrics) should be explicitly tabulated in the published dataset metadata alongside the volumetric information.
(II) Technical validation vs ecological interpretation
In section 3.1.2, I am concerned by several ecological oversimplifications. For a number of taxa, image counts predominantly originate from a single region. Regarding species distributions, because sampling effort is not accounted for in Figure 5, the ecological assertions appear overstated. For instance, for Corymorpha nutans, the vast majority of images originate from the Northern North Sea, which was also the most heavily sampled region, during cruises conducted in May and June. Seasonality likely plays a major role here, particularly given that this species exhibits strong phenology, typically characterised by a late-May bloom followed by rapid senescence. A comparison with OBIS data confirms that its distribution is not restricted to the Northern North Sea. Regarding Physonect zooids, both phenology and sampling depth likely influence their observed presence. Transitioning from image interpretation to ecological inference must therefore be handled with caution. The importance of the seasonality should be highlighted, which would actually complete the author’s argument about “the importance of basin-wide sampling for constructing a taxonomically representative annotated set” (Line 210-211).
Regarding Figure 6, using the bounding box diagonals of the images to compute size spectra provides a useful technical sanity check of the instruments and processing pipeline. However, asserting that this distribution “adequately represents the expected size distribution of plankton in the natural environment” (Line 215-216) is scientifically inaccurate. A true Sheldon size spectrum requires volume-normalised biovolume distributions.
In section 3.1.3, the observed environmental variability is heavily dependent on sampling effort and timing. This should be explicitly clarified, in a similar manner to the caveat regarding sampling depth (Line 225). Figure 7 displays the environmental envelope of the sampling transects and campaigns, and must not be presented as the baseline climatology of the regions themselves.
Please consider adding a figure showing the spatio-temporal distribution of the sampling effort (e.g. a month-by-region heatmap) to the supplementary materials, to clarify the seasonality aspect.
(III) Taxonomic classification confidence
To maximise the reusability of this dataset as a regional foundation model, it is critical that the authors provide a clearly defined, homogeneous taxonomic list accompanying the dataset. This list should explicitly state the lowest taxonomic rank to which each group can be identified with high confidence using these specific imaging instruments. Otherwise, future users will struggle to seamlessly align the classifier's outputs with external ecological datasets or alternative imaging platforms.
--- Technical comments ---
Line 28-29: “zooplankton was used in several quality indicators in the latest OSPAR’s quality status report (Holland et al., 2023)” -> To my knowledge, there are 3 OSPAR indicators regarding Pelagic Habitats. Replace “several” by “three”. Why quoting only the work about one indicator here? Add “e.g.” or cite also the other two:
Line 30: “novel monitoring of zooplankton dynamics is critically needed” -> see also Ratnarajah et al., (2023), studies about augmented observatories and even the Plankton Manifesto for arguments
Line 41: “Continuous Plankton Recorder” is written twice
There are quite a lot of auto-citations, especially for reference Hovenkamp et al., (2026a). While it is highly relevant for the technical part, I think it can be avoided in the state-of-the-art, where foundational work should be cited instead:
Line 54: “a manually annotated set of images is required for all taxonomic groups of interest that can be expected (e.g. Orenstein et al., 2022)” -> citing a trait-based study doesn’t really serve the argument here
Line 56: “can reduce the number of living images to as little 1% of all data” -> “can reduce the proportion of images containing living organisms to as little as 1% of all data."
Line 66: “the framework presented by (Van Walraven et al., 2026, under review)” -> “the framework presented by Van Walraven et al. (2026, under review)”
Line 92: "The ISIIS-DPI (Bellamare, Cowen and Guigand, 2008; Greer et al., 2025)” -> does not coincide with references. Seems to be Cowen and Guigand, 2008 instead of Bellamare, Cowen and Guigand, 2008, according to the “References” section.
Table 1: “Number of deployements” -> unjustify the text or add line break, so that the “of” does not stand alone
Line 117-118: “This frame was towed from the aft of a vessel with a towing speed of 2 to 5 knots while undulating between the surface and sea floor (i.e. ‘tow-yo’-ing), it or was undulated while the vessel was stationary (i.e. ‘yo-yo’-ing).” -> remove “it”, or invert “it” and “or”.
Line 200: “A total of 15,659 CPICS images was annotated among 42 classes (Figure 3, and examples in Figure 4b).” -> inverting as follows: (Figure 4b, and examples in Figure 3) feels more logical regarding the figures.
Line 210: “Siphonophore zoids” -> typo: zooids
Line 223: “due to meltwater influence in Greenland (Hovenkamp et al., unpublished)” -> replace with (or at least add) a published study to back-up the argument, for instance: Dukhovskoy et al., 2019; https://doi.org/10.1029/2018JC014686.
Line 234-235: “a trade-off between increasing precision and decreasing precision” -> typo: decreasing recall
Line 238: “the ISIIS-DPI set is comparable to the classifier from (e.g.) Hovenkamp et al. (2026a)” -> remove (e.g.)