Applied Plankton Image Classification for Imaging FlowCytobot Data

Abstract

As the ability to gather vast quantities of data from oceanographic bioimaging sensors increases, so too does the need to process, analyze, and store that data in a consistent, standard way that enables replicability and accessibility for future studies. The Imaging FlowCytobot (IFCB), an automated submersible flow cytometer, produces high resolution images of plankton at rates up to 10 Hz for months or years, resulting in billions of images. This project compares various methods to categorize incoming images of plankton gathered by the IFCB - Convolutional Neural Nets (CNNs), Vision Transformers (ViT), and self-supervised learning (MAE). The benefits and downsides of each model are analyzed and discussed for future IFCB operators to process their data using the methods that best align with their research questions, along with step-by-step explanations about the pros and cons of each method depending on the use case.