Use this URL to cite or link to this record in EThOS:
Title: Machine learning for acoustic mosquito detection
Author: Kiskin, Ivan
ISNI:       0000 0005 0292 0684
Awarding Body: University of Oxford
Current Institution: University of Oxford
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Mosquitoes are responsible for over one billion cases of disease and over one million deaths each year. The data produced during mosquito surveillance are needed to identify emerging insecticide resistance, facilitate effective and evidence-led insecticide intervention programmes as well as model current and future vector-borne disease transmission. Traditional mosquito survey methods are time-consuming, expensive, and spatially limited. Consequently, many mosquito distribution models that map the range of these insects rely on small quantities of poorly spatially distributed occurrence data. There is therefore an urgent need to develop new mosquito survey methods that can provide real-time species-specific occurrence and abundant data without human risk. Here we consider an acoustic detection paradigm, in which the distinctive buzz of mosquito flight is used as a characteristic signature for detection and subsequent species identification. We show it is possible to achieve high classification accuracy even in data-scarce scenarios, using a combination of deep learning and wavelet encoding. Additionally, we develop and deploy a smartphone app that allows mosquito detection at scale and can discriminate between species with high accuracy. We garner low-resolution labelling for parts of our data via crowdsourcing, supplementing the high-resolution labels obtained from experts and publicly release a baseline model and dataset. The technical materials that underpin this thesis detail development of machine learning approaches for detecting and identifying events in data, with the primary focus of finding mosquito flight tones in acoustic time series. Solutions specific to such audio detection are developed, tested and applied to field-gathered data. Although the research is specific to one focal application domain, the approaches developed are generic and were motivated by canonical problems of low signal-to-noise ratio, sparse data, multiple resolution labelling, class imbalance, and decision bias. They thus apply to a far wider set of detection problems, in audio time series and beyond.
Supervisor: Roberts, Steve Sponsor: Engineering and Physical Sciences Research Council
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: Engineering ; Deep learning ; Acoustic models ; Machine learning