The Gnu Frontier: deploying machine learning and open-source electronics for the study of ungulate movement in the Anthropocene

Kavwele, Cyrus Mutunga (2024) The Gnu Frontier: deploying machine learning and open-source electronics for the study of ungulate movement in the Anthropocene. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2024KavwelePhD.pdf] PDF
Download (40MB)


The Anthropocene epoch has ushered in unprecedented and irreversible changes in many biomes, resulting in the disruption of ecological functions and processes. These changes are largely driven by the increased human footprint on a planetary scale and global warming. Consequently, various impacts have been documented, including the extinction of flora and fauna, modification of ecosystems into more homogeneous covers (e.g., farmlands), increase in human-dominated landscapes, disruptions of animal migrations, species range shifts, invasions leading to the extermination of native species, and encroachment of protected areas. These widespread ecosystem changes have become a primary concern for researchers and policymakers who must maintain a delicate balance between the persistence of species and their habitats and the promotion of sustainable development. Furthermore, the rate at which these changes are occurring outpaces the evolutionary response of many species. Consequently, gaining insight into how species respond to various ecological disruptions, both within and outside protected areas, is imperative. However, a thorough understanding of animal behaviour and their responses to rapid ecosystem changes remains challenging due to the lack of robust tools for collecting fine-grained data.

To address this methodological gap, I first use camera trap data to demonstrate how migratory species in the Serengeti ecosystem are spatially distributed in relation to human activities occurring in the immediate landscapes adjoining protected areas. The results reveal that the species tend to avoid areas transitioning into human-dominated landscapes as opposed to those bordering buffer zones. The results hold significant conservation value and illuminate population-level responses to anthropogenic disturbances. However, camera trap data does not provide individual-level behavioural insights. Consequently, it remains unclear which additional factors and social cues animals may be observing when traversing across habitats with varying threat levels and how these factors influence their behaviour. Camera traps and telemetry tools such as GPS alone cannot provide data required to answer such questions; therefore, a different set of tools is necessary. Leveraging the capabilities of open-source electronics, I present a low-cost system for automated and repeated observation of collared animals. This system consists of a GPS collar, a long range network (LoRa) radio transmitter, and a commercially available low-flying unmanned aerial vehicle (UAV), taking advantage of its built-in capacity to track a stream of GPS points. The system was tested on a small group of ponies and demonstrated its efficacy and performance by collecting data on focal individual as well as information about its nearest neighbours.

Furthermore, automated tracking system collects data in bursts of approximately 20 minutes, aligning with the flight time capacity of a fully charged battery. As such, obtaining behavioural data for longer periods is difficult which necessitates a different approach. Given the rapid ecological changes, it is crucial to understand animal behaviour and its perception of the immediate surroundings. For instance, where do animals spend more time being vigilant as opposed to engaging in other restorative activities like resting? Such areas could be regarded as risky from an animal’s perspective. In this study, I developed a near real-time animal behaviour classifier using low-cost open-source electronics, a low-power long-range wide-area network (LoRaWAN) for connectivity, and edge machine learning. The custom-designed animal tracking system records behavioural data, preprocesses it, and classifies it into four classes: grazing, lying, standing, and walking. The predicted behaviour classes are transmitted to the end-user via servers in near real-time. The tracking tool was tested on Serengeti wildebeest and demonstrated its performance by sending both behavioural classes alongside positional data of the collared animal.

In this study, I have demonstrated the utility of existing remote tracking tools as well as their limitations in addressing evolving ecological questions in relation to animal behaviour and response to ecosystem perturbations. The methodological approaches presented here have the potential to greatly enhance our understanding of animal ecology. Importantly, the application of novel technologies will empower scientists to enhance existing tools, generate complementary data streams, improve data resolution and quantity, and enrich their overall capabilities to study complex questions. For instance, it improves our ability to collect behavioural and positional data, monitor focal individuals, and track nearest neighbours, and potentially opens up other avenues for scientific applications. The application of open-source electronics creates an opportunity for other researchers to customise the tools as an alternative to commercial devices to address specific questions and potentially result in other valuable innovations.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Additional Information: Supported by funding from the University of Glasgow’s Lord Kelvin/Adam Smith PhD scholarship, and the African Research Fellowship from the American Society of Mammalogists.
Subjects: G Geography. Anthropology. Recreation > GE Environmental Sciences
Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Q Science > QH Natural history > QH301 Biology
Q Science > QL Zoology
T Technology > T Technology (General)
Colleges/Schools: College of Medical Veterinary and Life Sciences > Institute of Biodiversity Animal Health and Comparative Medicine
Supervisor's Name: Hopcraft, Professor Grant and Torney, Professor Colin
Date of Award: 2024
Depositing User: Theses Team
Unique ID: glathesis:2024-84206
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 08 Apr 2024 12:55
Last Modified: 11 Apr 2024 10:15
Thesis DOI: 10.5525/gla.thesis.84206
Related URLs:

Actions (login required)

View Item View Item


Downloads per month over past year