Bioacoustic approaches for ecosystem monitoring and conservation
Background
Effective conservation depends on being able to monitor ecosystems continuously and at scale.
In many remote areas, however, conventional monitoring is difficult because of limited infrastructure, power, connectivity, and personnel.
On the other hand, manual collection of audio and subsequent processing inhibits real time monitoring of the ecosystem.
Bird sounds provide a practical way to monitor these environments.
Changes in bird activity and species presence can provide useful indicators of ecosystem health, making passive acoustic monitoring a promising approach for long-term conservation.
This project investigates how artificial intelligence can make such monitoring more practical in remote, resource-constrained settings,
and how such monitoring systems can offer real time updates from remote locations.
We use knowledge distillation to transfer the capabilities of a large bird-classification model, Perch, into smaller models that can run locally on low-cost edge devices.
The resulting pipeline covers field audio processing, model training, evaluation, and deployment. For real time updates, we integrate LoRaWAN transmissions for live telemetry and develop a dashboard that shows both live updates and detections history.
Accomplishments
- Deployed a prototype bioacoustic monitoring system at the DeKUT Conservancy, using Raspberry Pi-based edge hardware and the PiWild platform to support autonomous field monitoring.
- Integrated LoRaWAN telemetry to enable low-power communication of monitoring data from remote deployments.
- Developed a live monitoring dashboard, providing a central view of deployed devices and their activity.
- Built an end-to-end model distillation pipeline, covering audio preprocessing, knowledge distillation, experiment tracking, and MLflow-based experiment management.
- Trained and evaluated several lightweight student models, including EfficientNet-B0, ShuffleNet, SqueezeNet, and MobileNet, to explore different trade-offs between model size, efficiency, and performance.
- Demonstrated the feasibility of moving bird-call analysis closer to the point of recording, reducing the need to continuously transfer raw audio from remote sites.
Next Steps
- Conduct systematic ablation studies to understand which components of the distillation process contribute most to model performance and efficiency.
- Deploy the student models on edge devices and evaluate their performance under real-world field conditions.
- Fine-tune models on Kenyan bird recordings to improve their suitability for local species and acoustic environments.