Underwater dual-magnification imaging for automated lake plankton monitoring

Deep Learning Classification of Lake Zooplankton

“…The diversity and abundance of plankton is generally measured using labor intensive sampling and microscopy, which suffer from a number of limitations, such as high costs, specialized personnel, low throughput, high sample processing time, subjectivity of classification and low traceability and reproducibility of data. These limitations have stimulated the development of a multitude of alternative and automated plankton monitoring tools (Lombard et al, 2019 ), some of which were recently applied in freshwater systems (Spanbauer et al, 2020 ; Merz et al, 2021 ; Tapics et al, 2021 ). Recently developed methods like eDNA hold a lot of promise in particular to monitor biodiversity at large spatial and temporal scales, to identify cryptic species (not detectable morphologically), and to account for genetic/functional diversity (Deiner et al, 2017 ) but are not yet implemented for high frequency on-site monitoring.…”

“…Among automated plankton monitoring approaches, imaging techniques have the highest potential to yield standardized and reproducible quantification of abundance, biomass, diversity and morphology of plankton across scales (Lombard et al, 2019 ; Merz et al, 2021 ). Currently, several in-situ digital imaging devices exists such as, Imaging FlowCytobot (Olson and Sosik, 2007 ), Scripps Plankton Camera (SPC) (Orenstein et al, 2020 ), Video Plankton Recorder (Davis, 1992 ), SIPPER and a dual-magnification modified SPC ( www.aquascope.ch ) (Merz et al, 2021 ).…”

“…These digital imaging systems can produce very large volumes of plankton images, especially if deployed in-situ for automated continuous monitoring (Orenstein et al, 2020 ; Merz et al, 2021 ). While the extraction of image features that describe important plankton traits like size and shape are well-established (Orenstein et al, 2020 ; Merz et al, 2021 ), classifying large volumes of objects into different plankton taxonomic categories is still an ongoing challenge, and represents the most important component for plankton monitoring (MacLeod et al, 2010 ). Automated classification of imaged plankton objects may help taxonomists annotating images and allow sampling and counting taxa at high temporal and spatial resolution.…”

“…In this paper, we study the classification of plankton organisms from lake ecosystems, on a novel dataset of lake plankton images that we make freely accessible, together with a code that allows to easily train and deploy our deep neural networks. We analyze plankton images from the Dual-magnification Scripps Plankton Camera (DSPC), which is a dark field imaging microscope, currently deployed in Lake Greifensee (Switzerland) (Merz et al, 2021 ), and specifically the images from the 0.5x magnification, which targets zooplankton and large colony-forming phytoplankton taxa in the ranges of 100 μm to 1 cm. We manually annotated 17943 images consisting of n c = 35 unevenly distributed categories (classes), which were collected in-situ using the DSPC deployed at 3 m depth in Lake Greifensee.…”

“…We evaluate the performances of our models on publicly available datasets, obtaining a slight but systematic increase in performance with respect to the previous literature. The simplest of the presented models were used to analyze part of the data in Merz et al ( 2021 ).…”