Mosquito surveillance by sound – the impact of AI on species recognition

Mosquitoes transmit several pathogens of public health importance, including malaria, dengue, chikungunya, and Zika. These vector-borne diseases cause millions of infections and hundreds of thousands of deaths each year. The most effective way to address emerging or re-emerging vector-borne disease threats is through prevention through rigorous surveillance systems that help detect risks early and initiate mitigation efforts.for example mosquito control). In recent years, numerous techniques for monitoring and controlling vector-borne diseases have been developed, many of which rely on the use of deep learning, especially in species detection and classification. Acoustic data, in particular, can be used as passive acoustic monitoring to enable real-time monitoring of vector populations, potentially supporting timely public health decisions. Mosquitoes make sounds when they flap their wings during flight. The faster it flaps its wings, the higher the sound. Mosquito sounds vary based on several factors, including species. This is very useful as only a few species of interest need to be monitored. They often transmit diseases, but this can also apply to non-native species. AI-based algorithms already exist to identify mosquito species based on sound, and some work fairly well (up to 97%). However, there are several caveats: (1) accuracy tends to decrease when many species are included, (2) few species are available in the training dataset, and (3) the sounds of wild mosquito populations are much more variable than the sounds represented in the training data due to the influence of environmental (temperature, humidity, etc.) and biological factors (sex, age, size) on mosquito sounds. All these aspects reduce the applicability of AI-based species recognition based on mosquito sounds in the field. A study by researchers from the Hunlen Ecological Research Center, ELTE University of Budapest, and the University of Szeged investigates the last point, the influence of several environmental and biological factors on the variation in mosquito sounds between species and individuals.

The researchers captured and recorded hundreds of mosquitoes in Hungary and used recordings from the 10 most common mosquito species to assess how much mosquito sounds vary by species and individual. Additionally, we evaluated the influence of several factors, including temperature, humidity, time of day, sex, age, and size (represented by wing length), on sound variation. The sounds were fairly consistent between species and individuals. However, acoustic signals associated with specific species were more consistent when environmental and biological variables were controlled.

Both gender and temperature greatly affect mosquito sound. Women’s sounds were lower than men’s. This is not surprising since in most mosquito species females are usually larger compared to males. Temperature also affects mosquito sounds. Generally, the higher the temperature, the higher the sound. Higher temperatures increase the insect’s metabolism (to a certain extent). Therefore, higher temperatures cause the mosquito’s muscles to move faster, allowing it to flap its wings faster. However, the significance of this increase varies between species, meaning that different species respond differently to temperature. This may be explained by the species’ origin (temperate or subtropical) or by its preferred host and associated blood temperature (bird blood temperatures are usually lower compared to mammalian blood temperatures). This species-specific difference in temperature suggests that simple temperature correction rules cannot be applied to mosquito sounds, or at least that the same formula cannot be applied to all species.

“Our data show that intraspecific and intraindividual variation cannot be ignored in AI-based acoustic classification. One solution to better integrate natural variation is to better represent that environmental and biological variation in the training data. Unfortunately, such complete databases remain rare, and especially for invertebrates, building such extensive databases requires significant time and effort,” said Julie, first author of this publication. Augustine says. Alternatively, the classification system can control or include additional environmental information to improve classification accuracy. Although some studies have already implemented this, it requires a detailed understanding of the effects of environmental variables on all species included in the model, which is not yet known. In both cases, natural variation in the target population needs to be better understood and accounted for to improve the accuracy of classification models in real-world situations and increase the likelihood of their use for surveillance purposes.