Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Semi-supervised machine learning for livestock threat classification using gps data
Thesis (MEng)--Stellenbosch University, 2023.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2023
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613755086471168 |
|---|---|
| access_status_str | Open Access |
| author | De Swardt, Urs |
| author2 | Kamper, Herman |
| author_browse | De Swardt, Urs Kamper, Herman |
| author_facet | Kamper, Herman De Swardt, Urs |
| author_sort | De Swardt, Urs |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2023. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/128972 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA en_ZA |
| last_indexed | 2026-06-10T12:41:10.692Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2023 |
| publishDateRange | 2023 |
| publishDateSort | 2023 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/128972 Semi-supervised machine learning for livestock threat classification using gps data De Swardt, Urs Kamper, Herman Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Animal industry -- Technological innovations Animal radio tracking Tracking and trailing Rangelands Internet-of-things Alarm reaction Radio collars Thesis (MEng)--Stellenbosch University, 2023. ENGLISH ABSTRACT: South African livestock farmers face major challenges in the form of livestock theft and predation. In response to these concerns, farmers started using a collar that monitors the acceleration of an animal and, when specific parameters are met, triggers an alarm that transmits GPS data to the user’s mobile application. Typically, a collar is placed on one animal per flock of sheep or herd of cattle. In this work, our primary goal is to classify the GPS trajectories captured by these devices into four categories: theft, predation, own-handling and other. We lay particular emphasis on distinguishing theft alarms since these have direct implications for the safety and financial sustainability of farmers. Our secondary goal is distinguishing emergency events (theft and predation) from non-emergency events (own-handling and other). To date, just over one million of these alarms have been recorded. Unfortunately, these trajectories are not labelled with the four categories. Therefore, we start by collecting a small labelled data set that can be used for validation. As a first approach to distinguishing alarms from one another, we investigate pure unsupervised learning, answering the question, “what can the data tell us without any labels?” We find that a convolutional autoencoder can produce fixed-dimensional embeddings from the GPS trajectories that can be used to cluster the data. We achieve a cluster purity score of 60% when comparing the clustering results to the ground truth labels from the small labelled data set. We also report other quantitative metrics best suited for our context. We then turn to supervised and semi-supervised approaches for our actual goal of classifying the trajectories. Our semi-supervised approach shows the best results with performance comparable to human performance. The approach consists of three parts. First, building off of our initial unsupervised model, an autoencoder and classifier are jointly trained to produce fixed-dimensional embeddings from GPS trajectories. Second, these embeddings are clustered to produce cluster labels. And lastly, the cluster labels are added to human-engineered features and used to train a final classifier. For our primary goal, the semi-supervised approach achieves an overall classification accuracy of 69% (4% lower than human performance), with an F1 score of 56% for theft events (4% lower than human performance) and an F1 score of 95% for own-handling events (outperforming a human by 6%). For our secondary goal, both the supervised and semi-supervised approaches show comparable performance to a human expert with accuracies of 87% and 86%, respectively. The model that we introduce can be deployed to aid farmers in terms of safety and security by providing them with critical information in emergency situations. Ultimately, this thesis tackles a novel machine learning problem, highlights its challenges, and proposes a semi-supervised classification model that performs comparably to a human expert. AFRIKAANSE OPSPMMING: Suid-Afrikaanse veeboere staar groot uitdagings in die gesig in die vorm van vee-diefstal en predasie. As ’n reaksie op hierdie probleme, het boere begin om ’n halsband te gebruik wat die versnelling van ’n dier monitor en wanneer spesifieke parameters ontmoet word, ’n alarm maak wat GPS-data na die gebruiker se selfoon toepassing stuur. Tipies word ’n halsband op een dier in ’n trop skape of beeste geplaas. In hierdie navorsing is ons primˆere doel om die GPS-trajekte wat deur hierdie toestelle opgeneem word in vier kategorie¨e te klassifiseer: diefstal, predasie, eie-hantering en ander. Ons lˆe spesiale klem op die onderskeiding van diefstal-alarms, omdat dit direkte implikasies het vir die veiligheid en finansi¨ele volhoubaarheid van boere. Ons sekondˆere doel is om noodgevalle (diefstal en predasie) te onderskei van nie-noodgevalle (eie-hantering en ander). Tot dusver is net oor ’n miljoen van hierdie alarms opgeneem. Ongelukkig is hierdie trajekte nie met die vier kategorie¨e geannoteer nie. Daarom begin ons deur ’n klein geannoteerde datastel te versamel wat vir validering gebruik kan word. As ’n eerste benadering tot die onderskeiding van alarms, ondersoek ons suiwer onbegeleide masjienleer, waar ons die vraag beantwoord, “wat kan die data vir ons sˆe sonder enige annotasies?” Ons vind dat ’n konvolusionele outo-enkodeerder model vastedimensionele inbeddings van die GPS-trajectorie¨e kan produseer wat gebruik kan word om die data te groepeer in sinvolle klusters. Ons bereik ’n kluster suiwerheid van 60% wanneer ons die kluster resultate vergelyk met die ware annotasies van die klein geannoteerde datastel. Ons rapporteer ook ander kwantitatiewe metings wat die beste geskik is vir ons konteks. Ons draai dan na begeleide en semi-begeleide benaderings vir ons werklike doel van die klassifisering van die trajekte. Ons semi-begeleide benadering toon die beste resultate met ’n prestasie wat vergelykbaar is met menslike prestasie. Die benadering bestaan uit drie dele. Eerstens, gebaseer op ons aanvanklike onbegeleide model, word ’n outo-enkodeerder en klassifiseerder saam afgerig om vaste-dimensionele inbeddings van GPS-trajekte te produseer. Tweedens word hierdie inbeddings gekluster om kluster annotasies te produseer. En laastens word die kluster annotasies by mens-ontwerpte kenmerke gevoeg en gebruik om ’n finale klassifiseerder af te rig. Vir ons primˆere doel bereik die semi-begeleide benadering ’n algehele klassifikasie akkuraatheid van 69% (4% laer as menslike prestasie), met ’n F1-telling van 56% vir diefstal gevalle (4% laer as menslike prestasie) en ’n F1-telling van 95% vir eie-hantering gevalle (wat die prestasie van ’n mens met 6% oortref). Vir ons sekondˆere doel wys beide die begeleide en semi-begeleide benaderings vergelykbare prestasie met ’n mens, met akkuraathede van 87% en 86%, onderskeidelik. Die model wat ons voorstel, kan ingespan word om boere te ondersteun in terme van veiligheid en sekuriteit deur kritiese inligting in nood situasies te voorsien. Op die ou end pak hierdie tesis ’n nuwe masjienleer-probleem aan, beklemtoon die uitdagings daarvan en stel ’n semi-begeleide klassifikasiemodel voor wat vergelykbaar presteer in vergelyking met ’n menslike kenner. Masters 2023-11-21T09:49:58Z 2024-01-08T17:44:48Z 2023-11-21T09:49:58Z 2024-01-08T17:44:48Z 2023-12 Thesis https://scholar.sun.ac.za/handle/10019.1/128972 en_ZA en_ZA Stellenbosch University 88 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Animal industry -- Technological innovations Animal radio tracking Tracking and trailing Rangelands Internet-of-things Alarm reaction Radio collars De Swardt, Urs Semi-supervised machine learning for livestock threat classification using gps data |
| title | Semi-supervised machine learning for livestock threat classification using gps data |
| title_full | Semi-supervised machine learning for livestock threat classification using gps data |
| title_fullStr | Semi-supervised machine learning for livestock threat classification using gps data |
| title_full_unstemmed | Semi-supervised machine learning for livestock threat classification using gps data |
| title_short | Semi-supervised machine learning for livestock threat classification using gps data |
| title_sort | semi supervised machine learning for livestock threat classification using gps data |
| topic | Animal industry -- Technological innovations Animal radio tracking Tracking and trailing Rangelands Internet-of-things Alarm reaction Radio collars |
| url | https://scholar.sun.ac.za/handle/10019.1/128972 |
| work_keys_str_mv | AT deswardturs semisupervisedmachinelearningforlivestockthreatclassificationusinggpsdata |