Abstract
Succession of necrophagous Diptera is the predictable sequence of species arrival and colonisation on carrion. This sequence helps estimate the period of insect activity or the minimum postmortem interval (mPMI). However, successional patterns can be variable due to the influence of environmental factors, season, location, perimortem treatments, and the dynamics of species interactions. Traditional methods of building successional databases, such as physically capturing insects, are labour-intensive, invasive, and time-consuming.This study explores a multidisciplinary approach combining bioacoustics and machine learning as a non-invasive alternative for studying local succession. By analysing species-specific wing beat frequencies (WBF) produced during flight, species can be identified with minimal disruption to decomposition processes. To achieve this, we aimed to develop a tool that combines audio recordings of flies with a machine learning model trained to recognise species-specific wingbeats in field recordings collected near carrion. While not suitable for crime scene application, such a tool can be valuable for creating local species databases. These databases can support forensic entomologists by enabling comparisons between casework specimens and local species communities, while also enhancing our understanding of ecological interactions during decomposition.
In the laboratory, five species of forensically important flies were bred and recorded under a variety of controlled conditions. The species included Calliphora augur, Calliphora stygia, Lucilia cuprina, Sarcophaga impatiens, and Musca domestica. The results showed that WBF varied depending on environmental and physiological conditions. Temperature was the most influential, with higher temperatures leading to higher WBFs across all species. However, the rate of WBF change with temperature differed between species. Other factors such as sex, size, and age, had species-specific effect. While interspecies differences in WBF were observed, there was also considerable overlap. In many cases, individual variability within a species exceeded the differences between species. Thus, WBF alone was insufficient for species discrimination.
To improve discrimination between species, additional audio characteristics were extracted from the audio recordings. These included spectral centroid, rugosity, and Mel Frequency Cepstral Coefficients (MFCC) in combination with the WBF. Machine learning models, including k-nearest neighbour, Support Vector Machine, Random Forest and Extreme Gradient Boosting (XGBoost), were trained on this data, achieving classification accuracies between 72% and 76%. XGBoost, in particular, showed strong performance due to its robustness and efficiency under varying conditions.
Applying the XGBoost model in field recordings presented new challenges. Background noise from sources like wind, vehicles, and other anthropogenic activities often interfered with the acoustic detection and resulted in a high number of misclassifications. Variability in acoustic equipment used to collect the laboratory database and field data further complicated the model’s ability to accurately classify the species. These results highlight the need for collecting field audio data in noise-free environments or alternatively, training the model with augmented (noisy) datasets to improve its ability to match patterns across the two datasets. Additionally, a larger dataset is necessary to better understand species overlap and to estimate the extent of misclassification caused by similar wingbeats. Finally, exploring advanced deep learning algorithms is crucial, as recent research has shown promising results in addressing these complexities in field data.
Despite these challenges, the findings highlight the potential of acoustic tools for forensic applications. While identifying all visiting flies may not be crucial for successional studies, since only a subset of species breed on carrion, there are other valuable uses for this tool in entomology. Monitoring fly activity using audio recordings offers several advantages, such as tracking the timing of insect arrivals, activity patterns, and responses to environmental factors such as temperature, rainfall, and daylight. This information can help refine mPMI estimations, such as by exploring scenarios that delay insect arrival, which are often influenced by environmental or perimortem factors.
While this study focuses on forensic entomology, the methodologies developed here have much broader applications. Acoustic monitoring combined with machine learning could be used for biodiversity studies, agricultural pest management, monitoring pollinator populations, and assessing the ecological impacts of environmental disasters. Forensic entomology serves as a valuable model system, demonstrating how bioacoustics and machine learning can be applied to diverse challenges in both forensic and ecological research. Although further refinement is needed for practical field use, this work establishes a foundation for developing non-invasive, scalable alternatives to traditional monitoring techniques in forensic entomology.
| Date of Award | 2025 |
|---|---|
| Original language | English |
| Supervisor | Jurian HOOGEWERFF (Supervisor) & James ROBERTSON (Supervisor) |