Mosquitoes are carriers of dangerous communicable diseases such as dengue fever, malaria, and Zika virus, all of which pose significant public health concerns in many countries around the world. Therefore, monitoring and controlling mosquito populations is critical, especially in high-risk outbreak areas. Traditional mosquito surveillance typically involves trapping and species identification by experts, which is a labor-intensive and costly process.
However, in recent years, it has been discovered that the frequency of mosquito wingbeats can be used to distinguish between mosquito species and strains, as each species has its own unique wingbeat sound. This discovery inspired a research team from the Faculty of Information and Communication Technology, Mahidol University (ICT Mahidol), led by Asst. Prof. Dr. Akara Supratak, Instructor in the Computer Science Academic Group; Professor Dr. Peter Haddawy, Global Talent Professor; Dr. Myat Su Yin, ICT Senior Researcher; and a research team from the Faculty of Tropical Medicine, Mahidol University, to explore the use of wingbeat frequency in developing a model to identify the origin of mosquitoes and effectively control the spread of mosquito-borne diseases.
“MosquitoSong+: A noise-robust deep learning model for mosquito classification from wingbeat sounds” originated from the Faculty of Tropical Medicine, Mahidol University, which aimed to develop a strategy for controlling the spread of dengue fever transmitted by mosquitoes. The initiative began by focusing on hospitals with the highest number of dengue cases. However, a major challenge was the lack of information about mosquito breeding sites. Traditional surveys required significant manpower to inspect potential sources such as water jars, containers, and hoses. Therefore, the team sought a solution to identify the breeding sources, mosquito species, and even their sex—particularly to estimate areas with high populations of female mosquitoes, which are responsible for disease transmission, and target those areas for control. This project focused on analyzing mosquito wingbeat sounds and led to the development of an AI model named MosquitoSong+.
In previous studies, the research team collected mosquito samples by conducting fieldwork and setting various types of outdoor traps, such as net traps, to collect dead mosquito specimens for subsequent analysis of their sex and species. However, this method was time-consuming. Therefore, the research team changed their data collection method by installing small computers and high-quality microphones on the mosquito traps to record wingbeat frequency sounds. The computers would then count how many mosquitoes, including Aedes mosquitoes, entered the traps—without requiring experts to identify their sex or species. This approach helps reduce both the time and cost of surveillance. Moreover, if a large number of mosquitoes is detected, the team can immediately identify that the area should be monitored. Nevertheless, once sound was introduced into the study, the most challenging task for the research team became sound classification.
“The problem we encountered was that we couldn’t distinguish mosquito sounds because they are very faint, which initially resulted in limited data. The Faculty of Tropical Medicine, Mahidol University, addressed this issue by breeding mosquitoes in a laboratory and recording their sounds in a controlled environment. This allowed them to better manage various factors—such as knowing the mosquitoes’ sex, species, and age—thus generating more data for the study. Another issue we faced was that when placing the microphones alongside outdoor traps, the environment wasn’t quiet. There were interfering sounds from passing vehicles, birds, and wind, which disrupted the sound recordings.”
Since the recorded audio contained background noise, the research team had to process the recordings through a noise simulation procedure. This was done to enhance the model’s ability to distinguish mosquito wingbeat sounds even in noisy environments. Background noise is a major factor that reduces the model’s accuracy in real-world applications, so this process was used to minimize its impact and help the model learn how to better separate mosquito sounds from ambient noise, aiming for the lowest error rate possible. The study found that the MosquitoSong+ model could classify mosquito species with over 80% accuracy, even in noisy environments, and determine the sex of mosquitoes with 93.3% accuracy. The model performed well across various conditions and is suitable for field deployment.
Asst. Prof. Dr. Akara also discussed mosquito trends and future directions for developing this work further, stating:
“At this point, we have a model that we believe performs well even in noisy environments. However, there are still many factors we need to continue studying—such as humidity, seasonal temperatures, mosquito behavior, and mosquito development—all of which affect the wingbeat patterns. This project is being conducted in collaboration with researchers around the world, and everyone faces the same challenge: it is nearly impossible to record mosquito sounds from real outdoor environments. That’s why we need to use technology to help gather data while we continue our own data collection. Our expected outcome is to have traps that are durable for long-term outdoor use and can support increased surveillance. The data collected from each area could then be visualized on Google Maps. Moreover, with global warming causing higher temperatures in some countries, mosquito outbreaks are now occurring in regions that previously had no mosquitoes. This has led to growing global interest in mosquito-borne disease control.”
Did you know?
In the MosquitoSong+ research project, the team used mosquito wingbeat sounds to distinguish between male and female mosquitoes in order to identify potential breeding sites. Female mosquitoes typically have a lower wingbeat frequency than males. Researchers also tracked the flight behavior of female mosquitoes, as they seek blood from living hosts to develop their eggs and are often found near water sources where they lay those eggs. This insight allows the team to better pinpoint mosquito breeding grounds.
Follow the work of Asst. Prof. Dr. Akara Supratak at: https://akaraspt.github.io/
Download the published paper, MosquitoSong+: A noise-robust deep learning model for mosquito classification from wingbeat sounds, here: https://doi.org/10.1371/journal.pone.0310121