Ecoacoustics for Environmental Monitoring
Ecoacoustics uses sound recordings from the environment to monitor biodiversity, ecosystem health, and ecological change over time. In Australia’s diverse landscapes—from rainforests and eucalypt woodlands to arid zones and coastal systems—passive acoustic sensors provide a non-invasive, scalable way to track species and ecosystem dynamics.
This project focuses on:
- Deploying autonomous recorders across key Australian habitats.
- Building annotated datasets of iconic and ecologically important species.
- Developing and evaluating machine learning methods for spectrogram-based classification of vocalisations.
- Translating acoustic insights into practical tools for conservation and land management.
By combining field ecology, signal processing, and AI, the project aims to improve how we detect, quantify, and respond to environmental change across the continent.
Target Species
The following species are focal points for acoustic monitoring and algorithm development. Their vocalisations are distinctive, ecologically informative, and representative of different habitats and taxonomic groups.
Laughing kookaburra
Known for its iconic “laughing” territorial call, often heard at dawn and dusk in open woodlands and urban edges across eastern Australia.
Australian magpie
Produces complex, melodious carolling and a range of contact calls. Highly adaptable, found in open habitats from rural to suburban landscapes.
Glossy black cockatoo
Emits soft, rolling contact calls and higher-pitched alarm calls. Strongly associated with specific food trees (e.g. Allocasuarina) in eastern Australia.
Koala
Males produce deep, resonant bellows during the breeding season; females emit higher-pitched calls. Vocalisations are important for monitoring populations in eucalypt forests.
Common ringtail possum
Produces soft chirps, squeaks, and occasional harsh calls, mainly at night in forests, woodlands, and urban gardens along the east coast.
Methods and Algorithms
Spectrogram classification
A spectrogram is a visual representation of sound showing how frequency content changes over time, with amplitude encoded as colour or intensity. In ecoacoustics, spectrograms are widely used as inputs to machine learning models for automated species detection and soundscape analysis.
A typical spectrogram-based classification pipeline includes:
- Preprocessing: trimming silence, normalising volume, and segmenting recordings into fixed-length clips (e.g. 1–5 seconds).
- Feature extraction: converting audio clips to spectrograms or mel-spectrograms, optionally applying denoising or data augmentation.
- Modeling: training classifiers (e.g. convolutional neural networks) to map spectrograms to species or sound-type labels.
- Evaluation: measuring performance using metrics such as precision, recall, F1-score, and area under the ROC curve on held-out test data.
For Australian ecoacoustics, models must handle diverse backgrounds (wind, rain, insects, anthropogenic noise) and variation in call structure across regions and individuals. Careful dataset curation and domain-adaptation strategies are critical for robust performance in the field.