DI-UMONS : Dépôt institutionnel de l’université de Mons

Recherche transversale
(titres de publication, de périodique et noms de colloque inclus)
2020-02-01 - Colloque/Article dans les actes avec comité de lecture - Anglais - page(s)

Wacquet Guillaume, Lefebvre Alain, Blondel Camille, Louchart Arnaud, Grosjean Philippe , Neaud-Masson Nadine, Belin Catherine, Artigas Luis Felipe, "Combination of “machine learning” methodologies and imaging- in-flow system to detect Harmful Algae semi-automatically" in 18th International Conference on Harmful Algae, 1, 46-50, Nantes, France, 2018

  • Codes CREF : Environnement et pollution (DI3840), Océanographie biologique (DI3191), Statistique appliquée (DI1133)
  • Unités de recherche UMONS : Ecologie numérique des milieux aquatiques (S807)
  • Instituts UMONS : Institut de Recherche sur les Systèmes Complexes (Complexys)
Texte intégral :

Abstract(s) :

(Anglais) In recent years, improvements in data acquisition techniques have been carried out to sample, characterize and quantify phytoplankton communities at high temporal and geographical resolution, with a special focus on potential harmful algae, during oceanographic campaigns or in the frame of monitoring networks (to support knowledge but also for EU Directives and Regional Sea Convention needs). These acquisition and digitization techniques, including ”imaging-in-flow” systems, allow to process a high number of samples and, consequently, generate an important quantity of data in which the presence of target events might not be detected. Indeed, as for traditional samples analysis with inverted microscope, a full manual quantification of the particles based on a simple visual inspection can be time-consuming, tedious and consequently lead to erroneous or wrong identifications. For this purpose, the ZooImage R-package was and is still being developed to allow greater automation in data classification and analysis while also permitting some user-interaction during the process. The proposed methodology consists in combining few expert knowledge and machine learning algorithms at different levels: (i) to classify particles into different groups based on the definition and the adaptation of a specific training set through the use of ”contextual data”; (ii) to detect and partially validate the ”most suspect” predictions, based on a probability of misclassification; (iii) to estimate the number of cells for each colonial form thanks to the building of specific predictive models. These different semi-automated tools were applied to the in vivo image dataset acquired with the FlowCam instrument during the September-October CAMANOC 2014 (Ifremer) cruise in the English Channel, in order to evaluate their operational ability to monitor the diversity of samples for the microphytoplankton, and especially to detect, track and count the most frequent potentially harmful algae found in this area at that period, like species belonging to Pseudo-nitzschia, Dinophysis, Prorocentrum and Phaeocystis genera. A distribution of these target groups was computed which highlights different sub-regions in the English Channel during the late summer-fall transition.

Mots-clés :
  • (Anglais) Machine learning
  • (Anglais) User-interaction
  • (Anglais) English Channel
  • (Anglais) Semi-automated classification
  • (Anglais) HABs