Community-based mosquito surveillance: an automatic mosquito-on-human-skin recognition system with a deep learning...

Area of Science:

• Entomology
• Computer Science
• Public Health

Background:

• Community engagement is vital for effective mosquito surveillance programs.
• Public participation can be enhanced by enabling individuals to identify disease-carrying mosquitoes.
• Developing automated systems supports community-level mosquito identification.

Purpose of the Study:

• To develop an automatic mosquito recognition system for public community use.
• To identify three mosquito species, including those in damaged conditions.
• To compare deep learning model performance with expert identification.

Main Methods:

• A custom image dataset of three mosquito species was created, including damaged specimens.
• Two deep learning architectures were explored: a partially trainable convolutional base and a fully trainable model.
• Weighted feature maps were used to visualize classification regions, comparing them to expert morphological keys.

Main Results:

• The model using the second architecture with the Adam optimizer achieved over 98% accuracy in identifying mosquitoes and their condition.
• The Xception model demonstrated the best generalization on an independent dataset with 0.7775 accuracy and 0.795 precision.
• Visualizations showed that the model's identified regions largely corresponded with expert morphological keys.

Conclusions:

• A customized deep learning model for pest mosquito taxonomy identification was developed.
• Visualization techniques revealed that computer-identified regions could potentially be integrated into systematic identification processes.
• The study provides a foundation for enhanced community-level mosquito surveillance through automated identification.