Development of a machine learning model to automatically identify key zooplankton species and stages from in-situ imaging, and its application to study zooplankton responses to phytoplankton dynamics in the Canadian Arctic


Supervisor: Prof. Louis Fortier, Co-Supervisor: Prof Marcel Babin, and committee members: Prof. Jean-Éric Tremblay and Prof. Frédéric Maps


Arctic mesozooplankton organisms (0.2 – 2 cm), largely dominated by copepods, transfer the bulk of energy and carbon from primary producers to the vertebrate fauna of the Arctic Ocean (fish, marine mammals, seabirds). Traditional net sampling, which integrates or roughly stratifies the water column, does not provide the necessary resolution for studying the fine-scale vertical distribution of key zooplankton components.

The lack of resolution from traditional zooplankton samplers can be overcome with recent in-situ optical imaging systems such as the Lightframe On-sight Keyspecies Investigation System (LOKI) for zooplankton identification. LOKI allows determining the exact vertical distribution of organisms in the water column, as well as corresponding environmental data (e.g. T, S, Chl a).

For the first part of my PhD I developed a detailed automatic zooplankton species identification model based on LOKI imagery and the Random Forests algorithm (Schmid et al. 2016) as well as ZOOMIE v1.0 (Zooplankton Multiple Image Exclusion), in order to study the coupling between phyto- and zooplankton.

The coupling is being studied in two main areas. First and foremost on a transect spanning from Northern Baffin Bay to the Lincoln Sea in the Canadian Arctic, an area which is sampled yearly by the ArcticNet mission onboard the CCGS Amundsen (for my PhD I am working on data from 2013 and 2014). Especially in 2014, sampling followed a very strong southward pattern in the development of a phytoplankton bloom (the development was for example visible in chlorophyll a and dimethyl sulfide (DMS) readings). This gradient developed from the North (pre bloom conditions) to the South (post bloom conditions). My main goal is now to investigate the fine scale vertical distribution (~30-40 cm vertical resolution) of single zooplankton species and even developmental stages under these conditions and how zooplankton dynamics are coupled with phytoplankton dynamics (e.g. species composition of phyto- vs zooplankton, cell size, vertical distribution patterns).

Secondly, I deployed LOKI during the SUBICE2014 (Chief scientist: Prof. Kevin Arrigo, Stanford University) mission onboard the USCGC Healy very early in the year (May and June 2014 in the Chukchi Sea). Based on LOKI and Imaging FlowCytobot (IFCB, which gives similar results to LOKI but for phytoplankton) data collected during this mission, we are investigating the coupling of phyto- and zooplankton under the ice as well as in the open water.

With this new level of ecological data we hope to contribute substantially to the understanding of zooplankton and their coupling with other trophic levels in the Arctic.



Vertical distribution for C. glacialis stage 3 and 5 as automatically identified by my RandomForests model, based on in-situ LOKI imagery. Data in the figure comes from a LOKI deployment during the ArcticNet 2013 mission in the North of Baffin Bay. The fluorescence data comes from the LOKI chl a sensor.