Chers partenaires et clients, venez découvrir notre nouveau site institutionnel

Guide méthodologique. Version actualisée de ZooPhytoImage avec refonte de l’interface graphique. Action 9. FlowCam ZooPhytoImage. Livrable n°1. Rapport final

Zoo/PhytoImage 4 is an « open source » software based on R and ImageJ. It processes numerical images of plankton particles digitized using a FlowCAM, a flat-bed scanner, microor macrophotos, etc. The general concept consists in the individual outlining of particles on the pictures, followed by their measurements (so-called « attributes ») such the size, the shape, transparency, textures, etc. These attributes are then used by a classification tool to automatically predict the taxonomic group the particles belong to. The classifier is obtained after a learning stage using a machine learning algorithm and a training set of pre-identified particles. The algorithm learns to recognize the taxonomic group from the set of attributes measured on the picture. The training set is manually build from a few hundreds to a few thousands of representative items by an operator that visually identifies the candidate particles on the images. Before use, the performances of a classifier are checked on another, independent, set of known particles called the test set, or by using an unbiased technique such cross-validation. Results are collected in a two-way contingency table that compares automatic and manual classification of the particles of the test set. This is the confusion matrix. Zoo/PhytoImage 4 proposes a series of numerical and graphical diagnostic tools to assess the performances of the classifiers, including new visualizations like the « F-score » by group plot. If the performances of the classifier are too low, Zoo/PhytoImage allows to test how the reworking of the groups can possibly increase these. The automatic classification of plankton is a difficult exercise. Currently, the residual error rate still remains too high for a practical use of the fully automatic approach (typically around 20-30 % of global error, but error by group vary much and can reach 90-100 % for rare or difficult groups). Hence, a manual (visual) validation of the classification by a trained operator is required to ultimately lower the error, global and by group. This work is long and fastidious. It greatly limits the interest of the automation provided by this approach. A new technique to detect suspect particles (those with a greater probability to be wrongly classified), combined with a modeling, and then, a statistical correction of the residual error, is implemented in Zoo/PhytoImage 4. The fraction of vignettes that need to be manually validated is lower (for instance, between 20 and 40% only), while keeping residual error per group around 10 % or less. A reworked graphical user interface is added to this version 4 to facilitate partial validation of the vignettes with the higher efficiency and ease of use. In addition to this process, Zoo/PhytoImage 4 also manages metadata. It optimizes storage on the disk and the speed to gather any information back. Finally, this software computes and extracts derived variables of interest, like abundances, biomasses, biovolumes or sizespectra per total or per taxonomic group. It operates in batch and can deal with a whole series of samples in one pass. The deliverable is the user manual of Zoo/PhytoImage 4 that details all these features from the user point of view, showing how to apply these techniques for plankton analysis, including with little examples and advices where there are commonly identified pitfalls. Corresponding binaries and source code of Zoo/PhytoImage 4 has been delivered to IFREMER too. The FlowCAM / ZooPhytoImage is becoming an operational tool in 2014. Completely adapted to the phytoplankton observations performed in the context of the French monitoring network REPHY, it will allow answering more accurately to the questions of WFD and MSFD concerning the evaluation of marine water quality. For instance, the first benefit will be for the acquisition of data necessary to the calculation of abundance index, part of the phytoplankton index for WFD in Channel and Atlantic water bodies : as a matter of fact, this index is based on the proportion of micro-phytoplankton taxa which are very abundant in a water sample.

Métadonnées du document