Open Collections

Blair, Jarrett

University of British Columbia

The world is experiencing a biodiversity crisis. Ecological data is urgently required to grasp the full extent and severity of this crisis, while also guiding conservation management efforts to alleviate its impacts. Categorizing specimens plays a pivotal role in ecological studies, biodiversity monitoring, and conservation efforts. However, the traditional method of manually identifying specimens can be exceedingly labour-intensive and costly, thus constraining the volume and velocity of data that ecological studies can produce. Computer vision, a branch of machine learning, presents a solution to these challenges by allowing efficient, accurate, and automated classification of specimen images. Despite its potential benefits, computer vision has its own set of challenges when applied to biodiversity data, such as class imbalance, model generalizability, and obfuscation. In this thesis I address the challenges and explore potential synergies between computer vision and biodiversity monitoring through a series of case studies. Most of the thesis works with the National Ecological Observatory Network’s (NEON) ground beetle (Family: Carabidae) and terrestrial invertebrate datasets. The NEON dataset spans the continental United States, Alaska, and Puerto Rico, and includes other potentially relevant metadata, thus making it an excellent, practical study system for computer vision applications in large-scale biomonitoring projects. Using these datasets, I explore the advantages of hierarchical classification, benefits of pairing images with contextual metadata, and the efficacy of various machine learning algorithms. In collaboration with other researchers working with NEON, I also used the invertebrate dataset to prototype a novel approach to combining computer vision and DNA metabarcoding data to improve the accuracy and taxonomic granularity of specimen classifications. Additionally, I also worked with the Canadian Museum of Nature’s carnivore skull collection to explore synergies between natural history collections and computer vision, as well as address challenges of domain generalisation between synthetic images and photographs. Through the work presented in this thesis I aim to progress computer vision as a tool for specimen processing in large-scale ecological monitoring efforts, with the goal of expediting ecological and conservation research to safeguard the world’s biodiversity.

Text

2024-07-04

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/88570

Zoology

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/