| 28 Jul 2026
Technical note: From coccolith counts to quantitative carbonate fluxes: three decades of AI-assisted automated coccolith analysis
Abstract. Automated analysis of calcareous nannofossils has progressively transformed coccolith-based paleoceanography by enabling the quantitative investigation of extremely abundant microfossil assemblages. Over the past three decades, the SYRACO (SYstème de Reconnaissance Automatique de COccolithes) framework evolved from one of the earliest neural-network–based recognition systems for micropaleontology into an integrated platform combining automated acquisition, object detection, morphometry and coccolith-mass estimation. Here, we retrace this methodological evolution from the first artificial neural networks developed in the 1990s to the current YOLOv8-based deep-learning framework.
We describe the successive developments in imaging, segmentation and object detection that improved over time coccolith recovery, quantitative robustness and operational scalability. Particular attention is given to the transition from segmentation-dependent workflows to integrated object detection, which fundamentally changed the balance between false positives and false negatives in quantitative coccolith analysis. To address the specific constraints of paleoceanographic applications, we use class-specific confidence thresholds to control inference reliability.
Beyond taxonomic recognition, the integration of bidirectional circular polarization imaging and automated morphometry progressively transformed SYRACO into a quantitative optical framework capable of estimating coccolith size, thickness and calcite mass on large datasets. As an illustration of the scientific applications enabled by these developments, we present a reanalysis of the EUMELI sediment-trap series from the Mauritanian upwelling system. This reanalysis reveals two short-lived coccolithophore export events during which Emiliania huxleyi and Gephyrocapsa oceanica dominated carbonate production and export, producing several grams of CaCO₃ m⁻² at 2500 m depth within about a month.
Together, these developments illustrate how successive generations of artificial intelligence and automated microscopy transformed coccolith analysis from a labor-intensive counting procedure into a scalable quantitative approach suitable for ecological, biogeochemical and paleoceanographic investigations.
I very much enjoyed reading the historical overview of the technical development of the SYRACO system. As noted in the text, this has been the work of decades, with parallel developments in sample processing, microscopy, image capture and segmentation and classification algorithms. The text is well illustrated with high quality figures that clearly demonstrate the improvements in imaging, but also the more subtle changes in data quality over time – a consistency in taxonomic compositional analysis but with a gradual increase in the accurate identification of total number of coccolith particles (reduction in false negatives). The return to the original test cases of early SYRACO systems clearly illustrates these changes in image and data quality through time.
My one concern is the about the nature of the paper. It is titled a ‘Technical note’ and, through a historical presentation of the evolution of the SYRACO system, does seem to be serving the purpose of documenting the SYRCAO5 system and methodology. I would suspect that it will be used as a, if not the, key reference paper for this methodology in future analyses using SYRACO5. In this context, however, although there is clear methodological content in the description of training datasets, acquisition approaches and validation metrics of SYRACO5, there is not a fully explained methodology of how this data is acquired. Aspects of this may be so familiar to the author team, that they seem obvious, but it for the external community the more clearly these components are documented, the greater the value and legacy of this major contribution to nannofossil research.
I think there are three areas to this:
Overall I would strongly support the publication of this manuscript, both as a concise, clear and engaging introduction to coccolith automated microscopy and analysis and as the methodological reference point for publications using the current SYRACO5 ‘system’ but would strongly encourage the inclusion of some more detailed methods for each component of the workflow, and an explanation of the linked image processing MANTA code. The presentation as a historical narrative makes methodological detail seem out of place, but the paper needs a bit more detail to underpin the results that it does present using the SYRACO5 system.
Minor issues:
It would be useful to have an estimate of throughput numbers in terms of slide scanning and image processing rates.
Fig. 6 caption says the black curves are SYRACOv2 counts; the text says the comparison is v3 vs v5 – which is right? (line 357)
Bibliography: Beaufort & Dollfus 2004a and 2004b are the same paper