Bird Species Identification Using Spectrogram Based on Multi-Channel Fusion of DCNNs

Area of Science:

• Bioacoustics
• Machine Learning
• Ornithology

Background:

• Deep convolutional neural networks (DCNNs) show promise in bird species identification from vocalizations.
• Imbalanced datasets pose a challenge for accurate bird sound classification.

Purpose of the Study:

• To develop an improved bird species identification model addressing dataset imbalance.
• To enhance identification accuracy using multi-channel fusion techniques.

Main Methods:

• Proposed a single feature identification model (SFIM) with residual blocks and a modified, weighted, cross-entropy function.
• Developed two multi-channel fusion methods (feature and result fusion) using three SFIMs trained on different spectrogram types (STFT, MFCC, Chirplet Transform).
• Employed transfer learning to manage model parameters and evaluated spectrogram durations (100 ms, 300 ms, 500 ms).

Main Results:

• The result fusion mode model achieved the highest mean average precision (MAP) of 0.914 on the custom dataset.
• A spectrogram duration of 300 ms was found to be optimal for the custom dataset.
• The model demonstrated generalization ability with a classification mean average precision (cmAP) of 0.135 on the BirdCLEF2019 dataset.

Conclusions:

• The proposed multi-channel fusion model, particularly the result fusion mode, effectively improves bird species identification accuracy.
• Optimal spectrogram duration is dataset-dependent, suggesting analysis of syllable duration distribution.
• The model exhibits promising generalization capabilities for real-world applications.