Machine learning for classification of marine zooplankton

Zooplankton are mostly tiny animals that live in the water column, feeding on algae and other organisms and eaten by many fish species. Hence, zooplankton are a key component of many marine ecosystems all over the world. However, zooplankton is particularly sensitive to climate change. This makes it vital to closely monitor zooplankton communities, and to gain a better understanding of their role in these changing ecosystems. Traditionally, zooplankton is sampled using nets, but a strong limitation is that analysis of these samples is time consuming. To avoid this problem, we work with underwater microscopes that take photos of zooplankton.

Over the past years, we developed machine learning techniques that can identify and count plankton species from such images automatically. For this we make use of Convolutional Neural Networks, and we are now able to classifiy observed zooplankton groups with an accuracy of around 90%. The combination of these imaging instruments and machine learning enables us more and more to collect zooplankton data at the pace and scale that we need.

However, a major problem in further application of these techniques is in the variety of instruments that are being used: some are used to pull through the water from a ship (in situ), some are used on deck, and different instruments target different size ranges of organisms. Each type of instrument produces different types of images and this means that currently, every instrument requires the training of a separate machine learning classifier. This is a major bottleneck, because each new classifier requires large amounts of training data.

In this project, we aim to develop a single machine learning classifier that is applicable for zooplankton images from different types of instruments. This could be approached by exploring and modifying a variety of modern state-of-the-art machine learning techniques. In addition to exploring and developing techniques, an important component is benchmarking of new results against existing classifiers using various (provided) labelled datasets. Datasets will be made available, so data collection or manual labelling is not part of this project.

Requirements

For a research project of 6+ months (40-45 EC), we seek a highly motivated Masters student in Computer Science or a related field with a strong interest in, and experience with, modern AI methods, who is keen to work on these techniques to enable relevant scientific applications. Good Python skills are required, as well as strong analytical skills and being able to work independently. Affinity with ocean research would be a plus.

Starting date of the project is September - November 2024. The project takes place at NIOZ in Yerseke, Netherlands, where housing can be arranged. Remote work is possible as well.

Contact

For more information contact Pieter Hovenkamp (pieter.hovenkamp@nioz.nl).