Deep transfer learning-based bird species classification using mel spectrogram images.

Area of Science:

• Ornithology
• Machine Learning
• Bioacoustics

Background:

• Traditional bird classification methods are time-consuming and require expert knowledge.
• Audio-based classification offers an automated and efficient alternative for species identification.
• Environmental monitoring relies heavily on accurate bird species classification.

Purpose of the Study:

• To develop an automated audio-based bird species classification system for 264 Eastern African species.
• To leverage deep transfer learning, specifically EfficientNet, for enhanced classification accuracy.
• To integrate Recurrent Neural Networks (RNNs) for capturing temporal audio patterns.

Main Methods:

• Utilized a modified deep transfer learning approach with the pre-trained EfficientNet model.
• Adapted EfficientNet to learn patterns from mel spectrogram images of bird sounds.
• Combined fine-tuned EfficientNet with Gated Recurrent Unit (GRU) and Long short-term memory (LSTM) RNNs.
• Trained the models on a dataset of approximately 17,000 bird sound recordings.

Main Results:

• The EfficientNet-B7 model combined with GRU achieved the highest performance.
• This model reached an accuracy of 84.03%.
• A macro-average precision score of 0.8342 was obtained.

Conclusions:

• The proposed audio-based system effectively classifies Eastern African bird species.
• Deep transfer learning, particularly EfficientNet-B7 with GRU, shows significant promise for automated bioacoustic monitoring.
• This approach enhances the efficiency and accuracy of ornithological surveys.