Machine Learning for the Study of Plankton and Marine Snow from Images

Abstract

Quantitative imaging instruments produce a large number of images of plankton and marine snow, acquired in a controlled manner, from which the visual characteristics of individual objects and their in situ concentrations can be computed. To exploit this wealth of information, machine learning is necessary to automate tasks such as taxonomic classification. Through a review of the literature, we highlight the progress of those machine classifiers and what they can and still cannot be trusted for. Several examples showcase how the combination of quantitative imaging with machine learning has brought insights on pelagic ecology. They also highlight what is still missing and how images could be exploited further through trait-based approaches. In the future, we suggest deeper interactions with the computer sciences community, the adoption of data standards, and the more systematic sharing of databases to build a global community of pelagic image providers and users. Expected final online publication date for the Annual Review of Marine Science, Volume 14 is January 2022. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

A flow-through imaging system for automated measurement of ichthyoplankton

“…Much effort has gone into high-volume imaging and automated classification of non-ichthyous planktons, including the use of convolutional neural networks – see for instance [9] for a review including imaging methods and [10] for more recent work in classification. However, automated imaging and analysis of fish eggs and larvae has until now received relatively little attention, with no dedicated published data sets.…”

Basin-Scale Underway Quantitative Survey of Surface Microplankton Using Affordable Collection and Imaging Tools Deployed From Tara

“…Initial tests generated data of a quality comparable to that produced by the FlowCam, an automated commercial microscope taking digital image of microscopic particles flowing through a capillary imaging chamber (Sieracki et al, 1998). The reliability of medium/high throughput imaging instruments for quantitative analysis of marine plankton is evidenced by a growing number of studies in the scientific community using these methods (Irisson et al, 2022). Notably FlowCam data have been compared and validated against microscopy analyses as regard to organismal size (Sieracki et al, 1998;Buskey and Hyatt, 2006;Ide et al, 2007;Álvarez et al, 2014;Le Bourg et al, 2015) and biovolume (Hrycik et al, 2019).…”

Machine learning techniques to characterize functional traits of plankton from image data

“…Human labeling has remained the major bottleneck limiting scientists' ability to analyze image data or use it for monitoring purposes (MacLeod et al 2010). Scientists have started to look toward supervised machine learning (ML) methods-algorithms that learn to classify new data from a set of human-generated training examples-to expedite classification efforts (Irisson et al 2022). Until the 2010s, such methods typically involved extracting hand-engineered numerical features from a curated set of training images followed by tuning an ensemble or margin-based classifier (Simpson et al 1991;Blaschko et al 2005;Sosik and Olson 2007;Gorsky et al 2010).…”