EN FR
EN FR


Section: New Results

Estimating the volume of a copepod from a single image with Deep Learning

Participants : Cédric Dubois, Eric Debreuve.

This work was made in collaboration with Jean-Olivier Irisson (Laboratoire d'Océanographie de Villefranche).

Ecologists and biogeochemists are interested in estimating the volume of copepods (to then convert it into a biomass), a subclass of zooplankton, in order to estimate how much carbon it can store and how much it will store in the future. Those studies are made thanks to the online database EcoTaxa, which gives access to a large number of plankton images. The standard method used in ecology produces partially incorrect results due to geometric approximations and projection issues (from 3D to the 2D image plane). We first proposed a study of the error made by this method on the volume estimation of copepods. Then we proposed a new method based on the deep learning framework. Its performances have been analyzed on simulated data (Fig. 14) and preliminary tests have been made on a subset of the data of the UVP5hd GreenEdge 2016 acquisition campaign available on EcoTaxa. Our work pointed out the limitations of both methods, indicating that a broader study is needed to improve the computation of copepod volumes.

This work formed the basis for Cédric Dubois's PhD which began on October 1st 2019 with a Ministère de la Recherche funding.

Figure 14. Left: synthetic 3D model of a typical copepod. Middle: a simulated 2D observation of the model. Right: a real observation.
IMG/cedric.png