Doctoral thesis (Dissertations and theses)
Improvement and use of supervised classification applied to plankton digital images in Zoo/PhytoImage: (real-time) study of plankton communities in the North Sea
Denis, Kevin
2014
 

Files


Full Text
These-K.Denis.pdf
Publisher postprint (5.97 MB)
Request a copy

All documents in ORBi UMONS are protected by a user license.

Send to



Details



Abstract :
[en] Marine plankton constitutes the basis of pelagic food webs, it contains toxic species that could negatively impact environment and is a good bio-indicator of environmental changes. In addition, the composition of plankton communities is an indicator of water quality: the presence or absence of some plankton taxa could indicate perturbations of the marine ecosystem (pollution, eutrophication, etc.). Despite its ecological importance, plankton remains difficult to study because of its high diversity and spatio-temporal variability. The analysis of plankton samples under a microscope is time-consuming and requires several trained taxonomists. Currently, several approaches have been investigated to partly automate this manual analysis. Among them, one is used to improve analysis of plankton samples collected at sea by combining specific digitization devices for plankton images analysis (e.g. FlowCAM) with statistical techniques of supervised classification (e.g. 'Random Forest'). This approach largely decreases duration of plankton analysis but resulting enumerations still contain too much errors and must be manually checked before use by an ecologist. In this context, the present thesis proposes several ways to improve the treatment of plankton digital images and to allow the real-time analysis of plankton communities. The first chapter presents a completely new version of Zoo/PhytoImage software (version 3) and defines the basis of all developments proposed in this work. This new version implements the different steps needed to analyze plankton samples by supervised classification of digital images and to provide results that are usable by ecologists. The second chapter is focused on the association between FlowCAM and Zoo/PhytoImage in order to ensure that results obtained by different teams/devices are inter-comparable. In this context, a standardized protocol has been developed and tested during an intercalibration experiment of three FlowCAMs to determine the possibilities and limits of results exchanges and comparisons between different laboratories. The third chapter is focused on a new method to select plankton classes to keep in the classification, based on dendrogram representations. Through this approach, final classes to use for automatic classification can be automatically defined by the recombination of initial classes in order to achieve a target maximum error rate. This method can be additionally modulated by different options to match specific goals, such as target groups separation and/or to promote merging according to taxonomic or ecological similarities. The fourth chapter presents an iterative method of error correction based on the manual inspection of particles suspected to be wrongly classified by Zoo/PhytoImage. The manual identification of suspect particles allows to reduce error of classification until a level convenient for ecologists without requiring the inspection of all particles that compose the studied samples. In addition, this partial validation is used to further reduce the error by using a method of statistical correction allowing to better approximate real sample composition. Lastly, the fifth and last chapter is focused on the presentation and deployment of a real-time monitoring system dedicated to plankton communities of the Belgian Coastal Zone. This system has been tested aboard the research vessel 'Belgica' and allowed to estimate phytoplankton distribution along transects realized in the North Sea. This estimation was validated using parallel measurements of the Chl a concentration that are totally independent from images analysis approach. The present work contributes to improve several critical steps of the automatic classification of plankton images by opening the way for new approaches that could ease this methodology application for large-scale monitoring programs.
Disciplines :
Phytobiology (plant sciences, forestry, mycology...)
Mathematics
Author, co-author :
Language :
English
Title :
Improvement and use of supervised classification applied to plankton digital images in Zoo/PhytoImage: (real-time) study of plankton communities in the North Sea
Defense date :
02 October 2014
Number of pages :
364
Institution :
Université de Mons
Promotor :
Grosjean, Philippe  ;  Université de Mons > Faculté des Sciences > Service d'Ecologie numérique
President :
Eeckhaut, Igor  ;  Université de Mons > Faculté des Sciences > Biologie des Organismes Marins et Biomimétisme
Jury member :
Artigas, Luis Felipe
Lancelot, Christiane
Flammang, Patrick  ;  Université de Mons > Faculté des Sciences > Service de Biologie des Organismes Marins et Biomimétisme
Wattiez, Ruddy  ;  Université de Mons > Faculté des Sciences > Service de Protéomie et Microbiologie
Research unit :
S807 - Ecologie numérique
Research institute :
R150 - Institut de Recherche sur les Systèmes Complexes
R100 - Institut des Biosciences
Available on ORBi UMONS :
since 26 January 2015

Statistics


Number of views
16 (0 by UMONS)
Number of downloads
0 (0 by UMONS)

Bibliography


Similar publications



Contact ORBi UMONS