Plankton support the marine ecosystem & are extremely sensitive to environmental change. The Marine Biological Association of the UK have been exploring new, autonomous imaging technology for rapid estimation of zooplankton abundance in order to improve monitoring & reporting speed to meet legislative monitoring requirements. In collaboration with EPCC, development of robust, automated plankton classification systems are being explored. The Continuous Plankton Recorder Survey has been used to generate an image dataset of zooplankton species generated via a digital imaging system; this project outlines the development, training & validation of Convolutional Neural Networks (CNN) on highly imbalanced datasets for rapid classification. Using a ready-labelled training dataset of 20 zooplankton species, a number of architectures & ensemble methods have been explored to obtain high accuracy classification; robust strategies for handling extreme class imbalance have been developed such that species that occur very infrequently & in low numbers can be reliably classified in near real-time. By optimising the CNNs for use on GPUs on EPCC’s CIRRUS HPC system, we will show how we have improved on existing work in this rapidly evolving field.
