A deep learning-based pipeline for mosquito detection and classification from wingbeat sounds

Mosquito vector-borne diseases such as malaria and dengue constitute some of the most serious public health burdens in tropical and sub-tropical countries. Effective targeting of disease control efforts requires accurate estimates of mosquito vector population density. The traditional, and still most common, approach to this involves the use of traps along with manual counting and classification of mosquito species. This process is costly and labor-intensive, which hinders its widespread use. In this paper we present a software pipeline for detection and classification of mosquito wingbeat sounds. Since our target platform is low-cost IoT devices, we explore the tradeoff between accuracy and efficiency. When a fast binary mosquito detector precedes the classifier, we can reduce the computational demand compared with use of the classifier alone by a factor of 10. While the accuracy of traditional machine learning model drops from 90% to 64% when reducing the sample rate from 96 kHz to 8 kHz, our deep-learning models maintain an accuracy of almost 83%, even when additionally reducing the bit depth from 24 to 16 bits. We conclude that the combination of an efficient mosquito detector with a convolutional neural network provides for an excellent trade-off between accuracy and efficiency to detect, classify and count mosquitoes.

Automated summary

Mosquito vector-borne diseases pose serious public health burdens in tropical and sub-tropical countries. Accurate estimates of mosquito vector population density are crucial for effective disease control efforts. This paper presents a software pipeline that uses a combination of efficient mosquito detector and convolutional neural network for detecting and classifying mosquito wingbeat sounds. The results show a good trade-off between accuracy and efficiency.