@article{fdi:010095038,
  title = {{M}achine learning approach to study microboring assemblage dynamics in two living massive coral genera},
  author = {{A}laguarda, {D}iego and {B}rajard, {J}. and {L}guensat, {R}edouane and {T}ribollet, {A}line},
  editor = {},
  language = {{ENG}},
  abstract = {{T}he coral host comprises various microorganisms including those living in its skeleton. {I}n coral skeletons, bioeroding microflora (cyanobacteria, algae, and fungi), which play an important role in reefs and coral resilience, produce specific traces (microborings) by actively dissolving the carbonate. {T}o date, only a few highly time-consuming methods that rely on the observer allow microborings' study, limiting the number of samples that can be analyzed. {R}ecently, a machine-learning approach based on the analysis of scanning electronic microscope images via a modified convolutional neural network ({CNN}) was developed to evaluate accurately microborings abundance along a core of a massive {D}iploastrea sp. from {M}ayotte. {T}he aim here was to test this {CNN} on another massive coral species, {P}orites sp., to verify that it can be applied to other massive corals. {W}e found that the classification accuracy decreased by 8% while the other metrics dropped significantly down to 26% on average. {T}o improve our {CNN} (training step especially), we tested diverse loss functions. {W}e also developed a specific {CNN} for {P}orites sp. and obtained a similar accuracy (94%) to that for the {CNN} for {D}iploastrea sp. (93%). {D}espite this result, we developed a {CNN}-{M}ixed model combining images collected from both coral genera to propose a unique and accurate model. {A}s we obtained an accuracy above 90% when applying the mixed model to either massive coral genus, we strongly suggest that it can be used to better understand the role of microborers in living massive corals and reefs over long term.},
  keywords = {{MAYOTTE} ; {REUNION}},
  booktitle = {},
  journal = {{L}imnology and {O}ceanography : {M}ethods},
  volume = {[{E}arly access]},
  numero = {},
  pages = {[20 p.]},
  ISSN = {1541-5856},
  year = {2025},
  DOI = {10.1002/lom3.10714},
  URL = {https://www.documentation.ird.fr/hor/fdi:010095038},
}