Power-efficient AI
for Bird Call Classification
New Zealand's unique native bird species face increasing threats, making conservation efforts critical. In remote regions like Fiordland, AI-powered bird monitoring plays a vital role in tracking bird populations with precision. To be effective in such challenging environments, bird call classification algorithms must deliver high accuracy while remaining energy-efficient.
Harsh and Remote Operating Conditions: In locations where battery replacement and solar power are impractical, devices must run on ultra-low-power consumption.
Challenges Faced
Developed and validated high-accuracy classification models including Supported Vector Machine, Random Forest, k-Nearest Neighbours, and Convolutional Neural Networks for call classification of native bird species.
Performed in-depth analysis of key acoustic features and temporal signal statistics for optimised feature extraction. Features include MFCC, spectrograms, Local Binary Patterns and Local Phase Quantization.
Conducted systematic benchmarking using performance and computational metrics to evaluate classification model performance under variable noise conditions and limited training data. Performance metrics include precision, recall, and F1-score.
