Robust methods of unsupervised clustering to discover new planktonic species in-situ

Plankton species are of vital importance to the marine food chain. They are susceptible to minor changes in their environment, which can lead to rapid and devastating changes in the global ecosystem. Thus, monitoring plankton species and their population dispersion is crucial to understanding the dynamics of their community abundance as well as their consumers in higher trophic levels. Technological advancements of systems providing high-resolution imaging augmented by powerful computing devices made it possible to infer the distribution from sampling millions of planktonic images at low cost. Yet, this requires an extensive and time consuming manual labeling effort. The process of training to distinguish different species on manually labeled data is called supervised learning. The objective of this paper is to find new algorithms capable of minimizing the training supervision and assisting in discovering unseen classes. We explore the use of unsupervised classes of models for in-situ classification and identification of planktonic images. The aim is to embed those models into existing robotic imaging platforms to enhance the classification ability and to allow the discovery of new classes without any prior knowledge or exhaustive labeling effort. This work compares different models and shows their abilities to learn essential data structures over the National Science Bowl planktonic dataset.