Convening Expert Taxonomists to Build Image Libraries for Training Automated Classifiers

Abstract

Digital imaging technologies are increasingly used to study life in the ocean. To deal with the large volume of image data collected over space and time, scientists employ various machine learning and deep learning algorithms to perform automated image classification. Training of classifiers requires a large number of expertly curated sets of images, a time‐consuming process that requires taxonomic knowledge and understanding of the local ecosystem. The creation of these labeled training sets is the critical bottleneck for building skillful automated classifiers. Here, we discuss how we overcame this barrier by leveraging taxonomic knowledge from a group of specialists in a workshop setting and suggest best practices for effectively organizing image annotation efforts. In our experience, this 2 day workshop proved very insightful and facilitated classification of over 4 years of plankton images obtained at Scripps Pier (La Jolla, CA), focusing on diatoms and dinoflagellates. We highlight the importance of facilitating a dialog between taxonomists and engineers to better integrate ecological goals with computational constraints, and encourage continuous involvement of taxonomic experts for successful implementation of automated classifiers.

Comparison of advanced methodologies for diatom identification within dynamic coastal communities

“…While FlowCam and traditional light microscopy provide more easily interpretable quantitative units, they can be more labor intensive and rely on the professional expertise in morphological identification from the user (Kenitz et al 2023). FlowCam results require some processing time to fully derive community composition, the imager does provide real-time images which may be a large advantage in the field for rapid bloom or harmful algae detection for example.…”

“…In addition, the FlowCam is well suited to analyzing live samples which may help to resolve several biases introduced by cell shrinkage and distortion associated with preservation for light microscopy (Zarauz and Irigoien 2008). Like the traditional light microscopy approach, the FlowCam relies primarily on cell features for phytoplankton identification, and thus is best suited to large, morphologically distinct phytoplankton groups (Lombard et al 2019;Kenitz et al 2023).…”