Detecting Aedes aegypti mosquitoes through audio classification with convolutional neural networks.

Area of Science:

• Medical Entomology
• Machine Learning
• Public Health

Background:

• Mosquito-borne diseases pose a significant threat, particularly in resource-limited regions.
• Effective mosquito control is hampered by a lack of resources for monitoring and intervention.
• Community-based strategies, like crowdsourced mapping, can enhance awareness and control efforts.

Purpose of the Study:

• To investigate the feasibility of identifying Aedes aegypti mosquitoes using audio analysis from smartphones.
• To develop and evaluate machine learning models for mosquito detection based on wingbeat sounds.
• To assess the potential of crowdsourced data for building live maps of mosquito incidences.

Main Methods:

• Downsampling of Aedes aegypti wingbeat recordings.
• Training a convolutional neural network (CNN) using supervised learning.
• Utilizing spectrograms of recordings as visual features for wingbeat frequency analysis.
• Comparing the performance of binary, multiclass, and ensemble classifiers.

Main Results:

• The binary classifier achieved the highest accuracy at 97.65% (±0.55).
• The ensemble classifier demonstrated the best sensitivity (96.82% ± 1.62) for detecting Aedes aegypti.
• Binary and multiclass classifiers showed a strong balance between precision and recall, with F1-measures near 90%.

Conclusions:

• Machine learning analysis of smartphone-captured wingbeat audio is a viable method for identifying Aedes aegypti.
• The developed models show potential for real-time mosquito monitoring and public health applications.
• This approach could facilitate community-driven initiatives for mosquito control and disease prevention.