Elephant Sound Classification Using Deep Learning Optimization.

Area of Science:

• Wildlife conservation
• Bioacoustics
• Machine learning for ecology

Background:

• Elephant vocalizations are vital for understanding behavior and conservation efforts.
• Accurate identification of elephant sounds is challenging, especially for resource-constrained devices.
• Current methods often rely on spectrograms or binary classification.

Purpose of the Study:

• To develop and evaluate lightweight models for elephant sound classification directly from raw audio.
• To introduce and test a novel model, ElephantCallerNet, for this task.
• To compare raw audio processing against spectrogram-based methods for elephant sound identification.

Main Methods:

• Exploration of lightweight models (MobileNet, YAMNET, RawNet) and a novel model, ElephantCallerNet.
• Direct classification of raw audio data without spectrogram conversion.
• Optimization of model parameters using Bayesian optimization techniques.
• Comparative analysis with spectrogram-based training approaches.

Main Results:

• ElephantCallerNet achieved 89% accuracy in classifying raw elephant sounds.
• Raw audio processing demonstrated superior performance compared to spectrogram-based methods.
• The model successfully classified three distinct elephant vocalization types: roar, rumble, and trumpet.

Conclusions:

• Direct raw audio processing with ElephantCallerNet offers a highly accurate and efficient approach for elephant sound classification.
• This method is suitable for deployment on edge devices, aiding real-time conservation monitoring.
• The ability to differentiate between roar, rumble, and trumpet calls provides deeper insights into elephant communication and social structures.