Applications of the Underwater Vision Profiler for Particle Annotation in the Oligotrophic North Pacific Subtropical Gyre
Date
2024
Authors
Contributor
Advisor
Department
Instructor
Depositor
Speaker
Researcher
Consultant
Interviewer
Narrator
Transcriber
Annotator
Journal Title
Journal ISSN
Volume Title
Publisher
Volume
Number/Issue
Starting Page
Ending Page
Alternative Title
Abstract
Machine learning algorithms (MLAs) are increasingly applied to optical imaging datasets of oceanic plankton and marine aggregates to obtain improved image annotation efficiency while preserving high annotation consistency. However, this process relies on an ever-decreasing number of expert morphological taxonomists to annotate training sets and validate MLA outputs. While recent attention has focused on training annotators on how to use machine learning algorithms, there has been limited effort to educate new annotators on how to annotate the datasets needed to train and verify such algorithms. By teaching new annotators how to create regionally-relevant and accurate training sets for MLAs, one could better utilize instruments, such as the Underwater Vision Profiler 5 HD (UVP), that are the basis for growing databases of images collected over an expanding set of temporal and regional studies. The UVP is a high resolution in situ camera-based instrument that samples particles above 0.064 mm up to ~54 mm, producing images for those particles that are > 0.5 mm. The UVP images a size fraction of fragile plankton and marine aggregates known to play an important role in the Biological Carbon Pump (BCP) and can quantify their vertical distribution, changing morphological characteristics, and visual interactions from the sea surface to ~6000 db. In this study, the UVP has been used to assess particle distributions on 14 Hawaiʻi Ocean Time-series (HOT) cruises between 2020 to 2023 at Station ALOHA (22.750N, 158.00 0W). Significant findings from this initial effort include - (1) seasonal changes in the slope of particle size distributions evidence summer (June – August) increases in the abundance of large particles, (2) subsurface peaks of large particles were frequently observed at the base of the euphotic zone between 100 to 150 db which we interpret to be the accumulation of sinking particles along isopycnals, and (3) in moving towards assessing organismal abundance, it became apparent that an annotation guide for the UVP was not available to the user community. To facilitate further research, we identified key classifiers for our region including 13 categories of organismal and detrital UVP images, including the genus Trichodesmium, Rhizaria, and marine aggregates. We then outlined a method and standards for development of a cooperatively annotated dataset with intra-annotator self-consistency. Individual annotations were made by two annotators and then compared to a cooperatively annotated dataset, displaying 87.6% and 88.1% agreement. Comparing the annotations made between annotator’s individual datasets, the agreement was 85.2%. In comparison, predictions by a machine learning algorithm tailored to the UVP, the EcoTaxa random forest, had only 31% precision. With this manually annotated dataset, the temporal and spatial patterns of aggregates and organisms were then assessed. One pronounced pattern observed was that marine aggregates were found in concentrations more than double that of organismal categories with peaks in concentration at the mixed layer boundary. Further research will investigate co-occurrence patterns and potential relationships with regional hydrodynamics and climate indices as the time-series of UVP data lengthens. Importantly, the standardized UVP annotation schema developed herein will allow increasingly large optical datasets collected by the HOT program to be annotated by multiple annotators. This will reduce the overall time spent manually annotating images and facilitate consistency across datasets. The classification guide and annotation pipeline described here maximizes the potential research questions that can be addressed with the large datasets generated by the UVP and outlines a path for new users in different regions to generate their own classification guides and annotation pipelines.
Description
Keywords
Biological oceanography, Machine Learning, Marine Aggregates, Optical Oceanography, Underwater Vision Profiler, Zooplankton
Citation
Extent
64 pages
Format
Geographic Location
Time Period
Related To
Related To (URI)
Table of Contents
Rights
All UHM dissertations and theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission from the copyright owner.
Rights Holder
Local Contexts
Collections
Email libraryada-l@lists.hawaii.edu if you need this content in ADA-compliant format.