[en] We present here an original method to calculate optimal set of taxonomical groups with automatic plankton classifiers in order to achieve the best trade-off between higher recognition rate and higher taxonomical separation. Groupings can be modulated to keep pools of related groups together or to isolate target species (e.g., toxic algae) as much as possible. At the end of the process, an interactive dendrogram graphically presents how the initial taxonomical groups are progressively pooled together, and it is possible to select the final level of groupings by clicking on the plot. As a practical application, an initial classifier of North Sea phytoplankton, able to discriminate 29 taxa with an accuracy of 78%, was optimized with the proposed method to separate 23 groups with an accuracy of 84%. We have also shown that, for several classification algorithms, post-processing of the original classification with our method yields higher recognition rate than direct classification into the final groups. That method will be available in Zoo/PhytoImage (http://www.sciviews.org/zooimage/).
