Exploiting algal strains for robust cross‐domain phytoplankton classification via deep learning
Limnology and Oceanography: Methods · 2025-11 · 23 (11) · 815-833
DOI: 10.1002/lom3.10723 ↗Kurz gefasst
Fachartikel in der Zeitschrift Limnology and Oceanography: Methods, online erschienen am 21.08.2025. Verfasst von Ladislav Hodač und 10 weiteren Autor:innen. Die Arbeit wurde bisher 3-mal zitiert (Crossref) und nennt 87 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Bundesministerium für Umwelt, Naturschutz, nukleare Sicherheit und Verbraucherschutz, Deutsche Forschungsgemeinschaft.
Abstract (Originalsprache)
Phytoplankton species are essential bioindicators for evaluating the status of freshwater ecosystems in accordance with the EU Water Framework Directive. However, manual identification of phytoplankton is time‐consuming and requires taxonomic expertise. Deep learning (DL) offers promising tools for automating the identification, but challenges remain due to imaging biases, morphological diversity, and the lack of validated benchmark datasets. In this study, we trained a DL model on microphotographs of controlled laboratory strains from 20 phytoplankton species and tested its performance on independent environmental image datasets. We assessed which species are suitable for cross‐dataset classification and explored whether computer vision–based image representations (DL features) reflect species similarity across datasets. Additionally, we combined shape analysis with DL features to determine whether feature‐based species distances correspond to morphological similarity. The model trained on strain images achieved reliable cross‐dataset classification for over half of the species. Classification performance declined with increasing feature/domain shifts between training and test images but improved when environmental images enriched the training set. Morphologically distinctive species, such as star‐like forms and those with lobes or bristles, exhibited higher classification rates, whereas rectangular or roundish forms posed greater challenges. DL features consistently clustered species across datasets, and the distances in DL feature space aligned with those in simplified shape space. Our findings demonstrate that using strains as references in DL models enables effective cross‐dataset classification while capturing morphological patterns. Integrating taxonomic expertise with computer vision is crucial for developing robust, interpretable phytoplankton bioindicator systems for ecological monitoring and biodiversity research.
Förderung
- Bundesministerium für Umwelt, Naturschutz, nukleare Sicherheit und Verbraucherschutz
- Deutsche Forschungsgemeinschaft
Zitieren
APA
Hodač, L., Dunker, S., Schmal, M., Carreño, E., Mäder, P., Lorenz, M., Jamroszczyk, M., Šubrt, D., Meier, S., Dürselen, C., & Wäldchen, J. (2025). Exploiting algal strains for robust cross‐domain phytoplankton classification via deep learning. Limnology and Oceanography: Methods, 23(11), 815-833. https://doi.org/10.1002/lom3.10723
Vancouver
Hodač L, Dunker S, Schmal M, Carreño E, Mäder P, Lorenz M, et al. Exploiting algal strains for robust cross‐domain phytoplankton classification via deep learning. Limnology and Oceanography: Methods. 2025;23(11):815-833. doi:10.1002/lom3.10723
BibTeX
@article{Hoda2025,
author = {Hodač, Ladislav and Dunker, Susanne and Schmal, Matthias and Carreño, Edwin and Mäder, Patrick and Lorenz, Maike and Jamroszczyk, Marie and Šubrt, David and Meier, Sandra and Dürselen, Claus‐Dieter and Wäldchen, Jana},
title = {Exploiting algal strains for robust cross‐domain phytoplankton classification via deep learning},
journal = {Limnology and Oceanography: Methods},
year = {2025},
volume = {23},
number = {11},
pages = {815--833},
doi = {10.1002/lom3.10723}
}
EUVIMED