PlanktoShare: A large (50k+) and FAIR learning set for the Plankton Imager (Pi-10) for the Greater North Sea and NE Atlantic, based on a new flexible classification protocol

van Walraven, Lodewijk; Scott, James; Pitois, Sophie; Ribeiro, Joseph; Close, Hayden; Pettigrew, James; Liszka, Cecilia M.; Fileman, Elaine; Hoekendijk, Jeroen; Hovenkamp, Pieter; Jak, Robbert; van Dalen, Joost; van Oevelen, Dick

doi:10.5194/essd-2026-215

Preprints

https://doi.org/10.5194/essd-2026-215

Preprints

02 Apr 2026

| 02 Apr 2026

Status: this preprint is currently under review for the journal ESSD.

PlanktoShare: A large (50k+) and FAIR learning set for the Plankton Imager (Pi-10) for the Greater North Sea and NE Atlantic, based on a new flexible classification protocol

Lodewijk van Walraven, James Scott, Sophie Pitois, Joseph Ribeiro, Hayden Close, James Pettigrew, Cecilia M. Liszka, Elaine Fileman, Jeroen Hoekendijk, Pieter Hovenkamp, Robbert Jak, Joost van Dalen, and Dick van Oevelen

Abstract. The use of imaging techniques for the study of particles and plankton is a rapidly advancing field in marine sciences. The data the tools produce require automated classification solutions that are trained on learning sets of manually labelled images. In this study we present PlanktoShare, a comprehensive (50k+ images) database with manually labelled images captured by the vessel-mounted Plankton Imager (Pi-10), in the 200 – 2,000 μm size range and including phytoplankton, holoplankton, meroplankton and various gelatinous taxa. The Pi-10 images particles continuously in a flow-through mode and can operate alongside research operations and during transits making it a popular choice for plankton monitoring. PlanktoShare provides a robust resource for training classifiers as an open resource. A key challenge in developing classifiers such as these is that commonly arises when merging learning sets from different sources, because images are often organized in a folder-like structure with incompatible or inconsistent nomenclature. To address this, we propose a database approach which separates the taxonomic information from descriptive attributes. Each image is assigned to one of the classes ‘Organism’ (whole organism), ‘Taxo_particle‘ (particle with taxonomic information, such as exuvia) and ‘Non_taxo_particle’ (particle without taxonomic information, such as marine snow aggregates). Taxonomic information is standardised using the aphiaID system from the WOrld Register of Marine Species while additional descriptive information (e.g. Life_stage) is stored as attributes. This database approach ensures full interoperability across learning sets from research groups allowing rapid expansion of geographical coverage and improved classification performance. Finally, we provide open-source code to apply pre-trained classifiers for users and outline future directions for collaborative plankton imaging.

How to cite. van Walraven, L., Scott, J., Pitois, S., Ribeiro, J., Close, H., Pettigrew, J., Liszka, C. M., Fileman, E., Hoekendijk, J., Hovenkamp, P., Jak, R., van Dalen, J., and van Oevelen, D.: PlanktoShare: A large (50k+) and FAIR learning set for the Plankton Imager (Pi-10) for the Greater North Sea and NE Atlantic, based on a new flexible classification protocol, Earth Syst. Sci. Data Discuss. [preprint], https://doi.org/10.5194/essd-2026-215, in review, 2026.

Received: 20 Mar 2026 – Discussion started: 02 Apr 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Status: final response (author comments only)

RC1:
'Comment on essd-2026-215', Anonymous Referee #1, 04 Jul 2026
This manuscript presents PlanktoShare, an open plankton-image learning set for the Plankton Imager, comprising more than 50,000 manually labelled images from the Greater North Sea and the NE Atlantic. It also proposes a flexible classification and database framework. Overall, the manuscript has clear value as a data paper. The dataset is timely and useful for promoting Pi-10 data sharing, classifier training, and regional plankton monitoring. I support publication after appropriate revision.
Major comments
The manual annotation and quality-control procedures should be described in greater detail.The manuscript repeatedly emphasizes that the learning set is manually labelled and of high quality, but the current description of the annotation workflow, expert involvement, cross-checking, and consistency control remains limited.

The model evaluation relies mainly on random splitting, which may overestimate cross-cruise or cross-region generalization.Images collected from the same cruise, time period, or nearby locations may be highly similar. A random train–validation split may therefore lead to overly optimistic performance estimates when the classifier is applied to independent cruises, different regions, different vessels, or different seasons. I recommend adding a lightweight but more informative validation, such as a blocked split by cruise, vessel, month, or region, or a leave-one-cruise-out test. If this is not feasible because of sample-size limitations or class imbalance, the authors should at least state clearly in the discussion that the random-split results primarily reflect within-dataset performance and should not be interpreted as general performance across all Pi-10 observational settings.

Class imbalance and the performance of rare taxa should be reported more transparently.Appendix A shows large differences in sample size among classes: some dominant classes contain thousands to tens of thousands of images, whereas some rare taxa are represented by very few images. Under such imbalance, the macro F1-score is useful but insufficient to support the conclusion that all classes perform consistently well. I recommend providing class-wise precision, recall, F1-score, and support, preferably in a table or supplementary material. For classes with very limited support, model performance should be described cautiously to avoid implying that all classification units have equal reliability.

The representativeness and application boundaries of the dataset should be more clearly defined.The spatial coverage of PlanktoShare is mainly concentrated along the Dutch coast, the Greater North Sea, the Celtic Sea, and selected NE Atlantic transects. This coverage is highly valuable for European marine monitoring, but the resulting classifiers may not directly generalize to other ocean regions, turbidity regimes, biogeographic provinces, or plankton community structures. I suggest adding a concise table summarizing the number of images and major class composition by cruise, vessel, year, season, and region.

Minor comments
In Section 3, the phrase “a specific? bespoke? learning set instance” appears to contain residual editing marks. Please remove the question marks and revise the wording.

In the conclusion, the manuscript refers to “Fig. 7” when discussing the spatial coverage of the three vessels, but this should be Fig. 9. Please correct the figure number.

The labels in the confusion matrices in Figs. 7 and 8 are very dense and difficult to read at the current page size. Please improve the resolution, enlarge the font, or move the full matrices to the supplementary material while retaining a simplified summary figure in the main text.

4 shows the spatial distribution of image locations, but it would be helpful to distinguish different vessels, years, or data sources using different colours, so that readers can better understand the composition of the dataset.

Section 5.3 states that the OSPAR classifier uses 53,538 images, whereas the full learning set is earlier reported to contain 52,882 images. Please explain why the classifier training set exceeds the total reported size of the learning set, and clarify whether this reflects duplicate images, augmented samples, excluded/added categories, or version differences.
Citation: https://doi.org/10.5194/essd-2026-215-RC1
RC2:
'Comment on essd-2026-215', Jonas Mortelmans, 24 Jul 2026
First of all, i have to excuse myself for the brief comments, often prone to spelling errors. My initial revision was accidentally deleted so i had to redo it all (and to save time i have quite some spelling errors).
The manuscript then! It provides a good contribution useful to many (including myself). On one end, the provided training set is open-access and allows users to start building their own classifiers quickly. On the other end, it provides a good scheme to work on image datasets.
The contribution is very much needed, especially when working on high-throughput imaging sensors. Despite the nice contribution, the textual and graphical aspect remains quite poor.
all figures need (major) updates

the need for interoperability - perhaps mentions how the scheme can work for other sensors? We should indeed go beyound the mere 'taxonomic' scope.

there is a strong need for additional tables and documentation.

there is a strong mismatch between a training set on on end (relatively well tackled); but on modeling, CNN, classificiatin tools on the other end which is not tackled at all and just comes in like that.

i have strong issues with the current architecture of the paper - it reads quite difficult and lacks a more conventient style (m&m, results and discussion, conclusion).

the numbers, figures, and actual images in the training set are not matching at all - or it is poorly described; or there are several versions of PlanktoShare which are not published as yet? This needs to be fixed. Similarly - a canonical record manifest, exact file formats, folder structure, identifiers, metadata files, checksums, licences and a version-specific DOI is entirely missing.

i have strong issues for the lack of scientific backgorund. To elevate the paper to a worthy release, much more scientific background is needed (read and catch up on recent literature).

there are many typographic errors and remnant of draft versions throughout the manuscript (i stopped commenting them after a while i have to admit). I hope authors were equally involved in proof-reading as obvious typos are still there.

consistent terminology. The manuscript alternates between database, learning set, training set, library and folder. These terms should be clearly defined and used consistently.

the annotation quality controll eneds to be addressed in detail. Who, how, more then one expert? CNN only? I strongly believe the scheme proposed needs to be revised. On one end, to include such information (essential as in time we evolve to a non-human validated dataset); but also to avoid the ZERO/NA thing which is poorly explained?

The confusion matrices are too small and insufficient by themselves. Per-class precision, recall, F1 and support, as well as macro and weighted F1, balanced accuracy, confidence intervals and major confusion patterns...

audit the bibliography - quite some missing.

explain way more thoroughly the applicability for users; and similarly (as said above) make sure to explain what is in the scheme and in what format (table!).

you speak of a database, but i see nothing hierarchical? how are thing coupled? how are images saved - are these in a noSQL server database? I have the feeling I "miss" something?

the model software, Plankton Imager Classifier by geojoost; is not documented at all here.

Many of these, are elaborated in comment, throughout the manuscript.
I advice a major revision.
Citation: https://doi.org/10.5194/essd-2026-215-RC2
- RC3: 'Reply on RC2', Jonas Mortelmans, 24 Jul 2026
  
  So just to be sure: please, download the attachment to see all comments.
  
  Citation: https://doi.org/10.5194/essd-2026-215-RC3

Data sets

PlanktoShare: A large (50k+) and FAIR learning set for the Plankton Imager (Pi-10) for the Greater North Sea and NE Atlantic, based on a new flexible classification protocol Lodewijk van Walraven et al. https://doi.org/10.5281/zenodo.19119233

Model code and software

Plankton Imager Classifier Joost van Dalen https://github.com/geoJoost/planktoshare

Viewed

Total article views: 661 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
499	127	35	661	34	29

HTML: 499
PDF: 127
XML: 35
Total: 661
BibTeX: 34
EndNote: 29

Views and downloads (calculated since 02 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	279	65	21	365
May 2026	118	40	3	161
Jun 2026	17	2	2	21
Jul 2026	57	14	4	75
Aug 2026	28	6	5	39

Cumulative views and downloads (calculated since 02 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	279	65	21	365
May 2026	118	40	3	161
Jun 2026	17	2	2	21
Jul 2026	57	14	4	75
Aug 2026	28	6	5	39

Viewed (geographical distribution)

Total article views: 637 (including HTML, PDF, and XML) Thereof 637 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 12 Aug 2026

Short summary

PlanktoShare offers a large collection of carefully labelled plankton images collected using the Pi-10 Plankton Imager in the North-East Atlantic. It was created to overcome inconsistent naming across datasets and to support reliable classification. By standardizing taxonomic details and storing extra traits separately, the database enables sharing and combining learning sets. This helps expand global monitoring efforts and strengthens future plankton imaging research.


Total:	0
HTML:	0
PDF:	0
XML:	0