UC San Diego

To understand ocean health, it is crucial to carefully monitor and analyze marine plankton – the microorganisms that form the base of the marine food web and are responsible for the uptake of atmospheric carbon. With the recent development of in situ microscopes, that collect images of these organisms in vast quantities, the use of deep learning methods to taxonomically identify them has come to the forefront. Given this data, two questions arise: 1) How well do deep learning methods such as Convolutional Neural Networks (CNNs) identify these marine organisms using data from in situ microscopes? 2) How well do CNN derived estimates of abundance agree with established net- and bottle-based sampling?

Using images collected by the in situ Scripps Plankton Camera (SPC) system, we trained a CNN to recognize 10 species of phytoplankton that are identified as associated with Harmful Algal Blooms (HABs). The success of the CNNs is characterized using standard evaluation metrics. To compare abundance estimates, we fit a linear model between the number of organisms of each species counted in a known volume in the lab with the number of organisms collected by the in situ microscope sampling at the same time.

The CNNs evaluated on 26 independent natural samples collected at Scripps Pier achieved an averaged accuracy of 92%, with 7 of 10 target categories above 85%. The linear fit between lab and in situ counts of several key HAB species suggests in the case of these dinoflagellates there is good correspondence between the two methods. The linear relationship derived for other organisms that were not as abundant is not as conclusive as well as the failure of the SPC systems to successfully detect the diatom chain, Pseudo-nitzschia.

Given the excellent correlation between lab counts and in situ microscope counts on key species, the methodology proposed here provides a way to estimate an equivalent volume in which the employed microscope can identify in focus organisms and obtain reasonably confident estimates of abundance. Given the ease of collection and success of the method, we are hopeful that future systems can automatically monitor HAB formers and other plankton.