Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
An analysis of hybrid hidden Markov models for cetacean detection
Edwardsr, D. J. E. 2025. An analysis of hybrid hidden Markov models for cetacean detection. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/6d13104b-e940-4f3d-aa3d-7ec19d1ee487
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613912588877824 |
|---|---|
| access_status_str | Open Access |
| author | Edwards, David James Erich |
| author2 | Versfeld, Daniel J. |
| author_browse | Edwards, David James Erich Versfeld, Daniel J. |
| author_facet | Versfeld, Daniel J. Edwards, David James Erich |
| author_sort | Edwards, David James Erich |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Edwardsr, D. J. E. 2025. An analysis of hybrid hidden Markov models for cetacean detection. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/6d13104b-e940-4f3d-aa3d-7ec19d1ee487 |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/132171 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:43:40.919Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/132171 An analysis of hybrid hidden Markov models for cetacean detection Edwards, David James Erich Versfeld, Daniel J. Du Preez, Johan Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Hidden Markov models Pattern recognition systems Cetacea -- Identification Acoustic localization UCTD Edwardsr, D. J. E. 2025. An analysis of hybrid hidden Markov models for cetacean detection. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/6d13104b-e940-4f3d-aa3d-7ec19d1ee487 Thesis (MEng)--Stellenbosch University, 2025. ENGLISH ABSTRACT: Anthropogenic activities in the ocean have increasingly been found to be detrimental to sea life. Cetaceans are negatively affected by these activities which include environmental pollution, noise pollution and whaling. Research is underway to determine the impact of anthropogenic activities on cetaceans by monitoring behavior and tracking population changes. A popular method for collecting information is known as passive acoustic monitoring. Although useful, this method is limited since it involves manual processing of large amounts of data. The development of algorithms is therefore essential to process data efficiently for detection, classification and tracking of cetaceans. The hidden Markov model (HMM) is a useful machine learning model in this context. Developed in the parallel field of speech recognition, the HMM is well suited to processing sound. However, the HMM is still being developed in the field of cetacean detection. This field is challenging as there is a need to detect different cetacean species accurately in various ocean habitats, with multiple sources of noise interference and therefore varying quality of recorded data. It is possible that extensions developed for the HMM can be adapted and utilized in this context. The unique advantages of these extensions have the potential to improve the performance of the HMM in the ocean environment. In this thesis, two extensions to the HMM were explored. These are the factorial hidden Markov model (FHMM) and the Bayesian hidden Markov model (Bayesian HMM). In literature, the FHMM has proved capable of modelling several sources at once, allowing the model to both pre-process and handle noise. The Bayesian HMM has useful characteristics as well, with literature showing that in some cases a Bayesian approach allows the model to fit data more accurately. For the development of the FHMM extension, signals were combined with noise to determine the efficacy with which the FHMM model could detect a specific signal. Results of the FHMM were compared to more classic approaches such as the Gaussian mixture model HMM and Gaussian HMM. Two speech databases and one whale database were used for testing. Results were collected using an isolated speech approach and the Viterbi approach. It was found that the FHMM consistently produced higher performance metrics than the other models. The isolated speech approach produced the best results, which was to be expected since it is a popular approach for speech recognition systems. The whale database showed the strongest results, suggesting that this technique will have promising applications for cetacean classification and research. The Bayesian HMM expansion was explored using both synthetic and real data. It was shown in synthetic tests that the Bayesian approach has several advantages such as the ability to automatically select the number of components. Additionally, the Bayesian approach was able to fit models more accurately on synthetic data. However, the success of this model was limited to real data. When multi-class classifiers were created on the whale database, no clear increase in performance metrics were found. Additionally, the classic expectation maximization HMM approach took less time to train with a similar level of accuracy. The advantage of the Bayesian approach was therefore limited to being able to automatically select the number of components during training. A system named the adaptive noise-robust HMM (ANR-HMM) system was built using both the FHMM and the Bayesian HMM. This system combined the ability of the FHMM to model multiple sources to handle noise and the Bayesian HMM’s ability to train and select the number of components automatically. Threshold based re-training was implemented which allowed the system to update and retain accuracy in varying noise conditions. Final testing showed the ANR-HMM was useful in high to mid-range signal-to-noise ratio contexts. The extensions and functions were compiled into a library and packaged in a practical way for future detection and research. Further applications of these libraries would involve testing larger databases to determine how accurately the extensions and ANR-HMM perform in a variety of ocean habitats. Functionality should be tested in conjunction with other extensions to determine the most effective and accurate approach for the detection of cetaceans. This will more effectively support Ocean conservation. AFRIKAANSE OPSOMMING: Menslike aktiwiteite in die oseaan is toenemend gevind om nadelige uitwerking op seelewe to veroorsaak. Walvisse word negatief be¨ınvloed deur hierdie aktiwiteite wat mgewingsbesoedeling, geraasbesoedeling en walvisjag insluit. Navorsing word gedoen om die impak van menslike aktiwiteite op walvisse te bepaal deur monitering van gedrag, sowel as bevolkingsveranderinge. Passiewe akoestiese monitoring is ’n gewilde metode vir die insameling van inligting. Alhoewel nuttig, is hierdie metode beperk aangesien dit handmatige verwerking van groot hoeveelhede data behels. Die ontwikkeling van algoritmes is dus noodsaaklik om data doeltreffend te verwerk vir die opsporing, klassifikasie en waarneming van walvisse. Die hidden Markov model (HMM) is ’n nuttige masjienleermodel in hierdie konteks. HMM is ontwikkel in die parallelle veld van spraakherkenning en is dus geskik vir klank verwerking. Op die gebied van walvisnavorsing bly HMM egter steeds onder ontwikkeling. Hierdie veld is uitdagend aangesien daar ’n behoefte is om alle walvissoorte in hul verskeie oseaanhabitats akkuraat op te spoor, en moet dit geskied tussenin verskeie bronne van geraasbesoedeling en dus wisselende data kwaliteit. Dit word voorgestel dat vordering en navorsing in die gebruik van HMM uitbreidings vir hierdie doeleinde aangepas kan word. Die unieke voordele van hierdie uitbreidings het die potensiaal om die prestasie van HMM in die oseaanekosisteem drasties te verbeter. In hierdie proefskrif is twee uitbreidings van die HMM ondersoek, die factorial hidden Markov model (FHMM) en die Bayesian hidden Markov model (Bayesian HMM). In die literatuur het die FHMM bewys dat dit verskeie bronne tegelyk kan modelleer, wat die model toelaat om beide voorverwerking en verwerking van geraas te hanteer. Die Bayesian HMM het ook nuttige eienskappe, met literatuur wat toon dat ’n Bayesian benadering in sommige gevalle die model toelaat om data meer akkuraat te pas. Vir die ontwikkeling van die FHMM-uitbreiding is seine met geraas gekombineer. Hierdie metode kon die doeltreffendheid waarmee die FHMM-model ’n spesifieke sein opspoor bepaal. Resultate van die FHMM is vergelyk met meer klassieke benaderings soos die Gaussian mixture model HMM en Gaussian HMM. Twee spraakdatabasisse en een walvisdatabasis is vir toets doeleindes gebruik. Resultate is ingesamel deur ’n ge¨ısoleerde spraakbenadering en die Viterbi-benadering. Daar is gevind dat die FHMM konsekwent ho¨er prestasiemetrieke as die ander modelle gelewer het. Die ge¨ısoleerde spraakbenadering het die beste resultate opgelewer, wat verwag is aangesien dit ’n gewilde benadering vir sprekerherkenningstelsels is. Die walvisdatabasis het die sterkste resultate getoon, wat daarop dui dat hierdie tegniek belowende toepassings vir walvisklassifikasie en navorsing sal bied. Die Bayesian HMM-uitbreiding is ondersoek deur beide sintetiese en werklike data te gebruik. Dit is in sintetiese toetse getoon dat die Bayesian benadering verskeie voordele het, soos die vermo¨e om outomaties die aantal komponente te kies. Daarbenewens kon die Bayesian benadering modelle meer akkuraat op sintetiese data pas. Die sukses van hierdie model was egter beperk tot werklike data. Toe multi-klas klassifiseerders op die walvisdatabasis geskep is, is geen duidelike toename in prestasiemetrieke gevind nie. Daarbenewens het die klassieke verwagtingsmaksimerings-HMM-benadering minder tyd geneem om op te lei met ’n soortgelyke vlak van akkuraatheid. Die voordeel van die Bayesian benadering was dus beperk tot die vermo¨e om outomaties die aantal komponente tydens opleiding te kies. ’n Stelsel genaamd die adaptive noise-robust HMM (ANR-HMM) stelsel is gebou met behulp van beide die FHMM en die Bayesian HMM. Hierdie stelsel het die vermo¨e van die FHMM om verskeie bronne te modelleer om geraas te hanteer, gekombineer met die Bayesian HMM se vermo¨e om outomaties die aantal komponente tydens opleiding te kies. Drempelgebaseerde heropleiding is ge¨ımplementeer wat die stelsel toegelaat het om akkuraatheid in verskillende geraasomstandighede op te dateer en te behou. Finale toetsing het getoon dat die ANR-HMM nuttig was in ho¨e tot middel-reeks sein-tot-geraasverhouding kontekste. Die uitbreidings en funksies is in ’n biblioteek saamgestel en op ’n praktiese manier verpak vir toekomstige beskikbaarheid en navorsing. Verdere toepassings van hierdie biblioteke kan die gebruik en toets van ’n groter databasis behels. Die proses sal onder andere hulp bepaal hoe akkuraat die uitbreidings en ANR-HMM in ’n verskeidenheid oseaanhabitats presteer. Funksionaliteit moet in samewerking met ander uitbreidings getoets word om die mees effektiewe en akkurate benadering vir die opsporing van walvisse te bepaal. Dit sal oseaanbewaring meer effektief ondersteun. Masters 2025-05-28T12:43:03Z 2025-05-28T12:43:03Z 2025-03 Thesis https://scholar.sun.ac.za/handle/10019.1/132171 en Stellenbosch University xvii, 122 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Hidden Markov models Pattern recognition systems Cetacea -- Identification Acoustic localization UCTD Edwards, David James Erich An analysis of hybrid hidden Markov models for cetacean detection |
| title | An analysis of hybrid hidden Markov models for cetacean detection |
| title_full | An analysis of hybrid hidden Markov models for cetacean detection |
| title_fullStr | An analysis of hybrid hidden Markov models for cetacean detection |
| title_full_unstemmed | An analysis of hybrid hidden Markov models for cetacean detection |
| title_short | An analysis of hybrid hidden Markov models for cetacean detection |
| title_sort | analysis of hybrid hidden markov models for cetacean detection |
| topic | Hidden Markov models Pattern recognition systems Cetacea -- Identification Acoustic localization UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/132171 |
| work_keys_str_mv | AT edwardsdavidjameserich ananalysisofhybridhiddenmarkovmodelsforcetaceandetection AT edwardsdavidjameserich analysisofhybridhiddenmarkovmodelsforcetaceandetection |