MosquitoSong+: A noise-robust deep learning model for mosquito classification from wingbeat sounds.

Area of Science:

• Entomology
• Machine Learning
• Bioacoustics

Background:

• Accurate mosquito population data is crucial for assessing disease risk and guiding vector control.
• Traditional methods using traps and manual identification are labor-intensive and costly.
• Existing machine learning models for mosquito identification struggle with environmental noise.

Purpose of the Study:

• To develop a robust deep learning model for identifying mosquito species and sex from wingbeat sounds.
• To address the challenge of background noise and volume variations in field recordings.
• To create a practical tool for automated mosquito surveillance.

Main Methods:

• Proposed MosquitoSong+, a novel deep learning model based on 1D-CNN architecture.
• Implemented noise augmentation and wingbeat volume variation data augmentation techniques.
• Trained and evaluated the model on diverse wingbeat sound datasets with varying noise levels.

Main Results:

• MosquitoSong+ achieved over 80% accuracy for species classification across datasets with background noise.
• The model demonstrated 93.3% accuracy for combined species and sex classification with various noises.
• The model shows excellent generalizability and robustness to environmental factors.

Conclusions:

• MosquitoSong+ offers a practical and accurate method for automated mosquito species and sex classification.
• The approach has the potential to significantly improve vector surveillance and disease control strategies.
• Further development could lead to widespread field application for public health.