Skip to main navigation menu Skip to main content Skip to site footer

Investigation of Unsupervised Auto-segmentation for Weak Phytoplankton Annotations


When certain phytoplankton and algae bloom, they can be a threat to both freshwater and marine ecosystems. To overcome these challenges many industry players are looking for improved harmful algal bloom (HAB) monitoring. One such promising approach is the use of supervised deep learning to automatically identify and count plankton species in microscope images. However, the main drawback of supervised learning is that it requires a large labeled dataset, which requires significant time from subject matter experts (SMEs) to meticulously annotate images. In this paper we propose an unsupervised auto-segmentation approach to automatically generate noisy segmentations of four common marine phytoplankton species: Prorocentrum lima, Alexandrium catenella, Heterosigma akashiwo, and Dolichospermum sp. We demonstrate the efficacy of our method by measuring the Intersection over Union (IoU) metric between the unsupervised auto-segmentation and the ground truth annotations generated by a team of professional taxonomists. Specifically, the average IoU for the four species mentioned were 0.7619, 0.7191, 0.6978, and 0.6495, respectively. Therefore, the feasibility of unsupervised auto-segmentation is viable to create weak labels, which contributes towards building a more sustainable phytoplankton monitoring system.