Improvement of plankton biovolume estimates derived from image-based automatic sampling devices: application to FlowCAM

Abstract

The most commonly used biomass estimate for microalgae is obtained from cell biovolume, usually calculated from microscopically measured linear dimensions. Although reliable, this is a highly time-consuming and specialized technique. Automated sampling devices that acquire images of cells and use pattern recognition techniques to identify the images have been developed as an alternative to microscopy-based methods. There are some aspects of automatic sampling and classification methods, however, which can be improved for the analysis of field samples including living and non-living particles. In this work, we demonstrate how the accuracy of a state-of the-art technique for plankton classification (Support Vector Machine) can be improved up to 86% if a previous automated step designed to remove non-living images is included. There is a tendency with the currently applied automatic methods to misestimate cell biovolume due to the two-dimensionality of the images. Here, we present a data set of more than 500 samples to show that the greatest effect is caused by the incorrect estimation of biovolume of the chain-forming diatoms. This results in an overestimate of biomass of between 20 and 100% where chain-forming diatoms represent more than the 20% of the biomass of the sample. We show how the classification method can be adapted to provide not only taxonomic but also the morphological classification of cells in order to obtain a more reliable estimate of biovolume according to the predicted cell shape, in a way comparable with microscopy-based estimates.