Birdsong classification based on ensemble multi-scale convolutional neural network.

Area of Science:

• Ornithology
• Ecosystem health assessment
• Bioacoustics

Background:

• Birds are key indicators of ecosystem health, making ornithological research crucial.
• Deep learning shows promise for birdsong recognition, but traditional CNNs face limitations with increasing network depth.
• Loss of semantic and detailed information, along with potential loss of global context in CNN operations, hinders classification performance.

Purpose of the Study:

• To address limitations in traditional CNNs for birdsong recognition.
• To explore and propose advanced deep learning frameworks for improved birdsong species identification.
• To enhance the accuracy, stability, and efficiency of birdsong classification for ecological monitoring.

Main Methods:

• Feature spectrograms were generated using wavelet transform for birdsong analysis.
• A multi-scale convolution neural network (MSCNN) was explored.
• An ensemble multi-scale convolution neural network (EMSCNN) classification framework was proposed and evaluated.
• Performance was compared against established CNN models like LeNet, VGG16, ResNet101, MobileNetV2, EfficientNetB7, Darknet53, and SPP-net.

Main Results:

• The MSCNN model achieved an accuracy of 89.61%.
• The proposed EMSCNN model demonstrated superior performance with an accuracy of 91.49%.
• Both models showed high stability and efficiency in recognizing 30 species of birds, indicating strong generalization ability.

Conclusions:

• The developed MSCNN and EMSCNN models significantly improve birdsong species recognition compared to traditional CNNs.
• The proposed frameworks offer a robust and efficient methodological scheme for bird classification research.
• These advancements contribute to better ecological quality assessment and bird protection strategies.