Bioacoustic Pest Detection System
Published:
Challenge
Pests feeding inside plants produce signals too faint and noisy for conventional monitoring. The system needed to capture those vibrations reliably and distinguish meaningful activity from background noise.
System
As project lead, I took the Insect Eavesdropper from concept through system validation. The architecture connects contact-microphone acquisition, signal preprocessing, self-supervised feature extraction, and classification in an end-to-end detection pipeline.
Engineering contributions
- Implemented DINO, autoencoder, and transformer-based representation learning for large, sparsely labeled acoustic datasets.
- Integrated low-cost LiDAR acquisition with NeRF-based 3D reconstruction, improving species-identification accuracy by 30% under challenging capture conditions.
- Defined repeatable acquisition, preprocessing, feature-extraction, classification, and evaluation stages for low-amplitude vibrational signals.
Result
The classification pipeline reached 96% precision on in-plant pest detection. The validated prototype and performance data supported $350,000 in project funding, a first-place pitch award, and selection for the World Agri-Tech Innovation Summit.