Mosquito monitoring through sound - implications for AI species recognition

Mosquitoes transmit several pathogens of public health importance, including malaria, dengue, chikungunya and ZIKA. These vector-borne diseases are responsible for millions of cases every year, and hundreds of thousands of deaths. The most effective way to cope with the threat of emerging or re-emerging vector borne diseases is the prevention by rigorous surveillance system, which can help early detection of risk and the initiation of mitigation efforts (e.g. mosquito control). In recent years, numerous technologies have been developed to monitor and control vectors and vector-borne diseases, many of which rely on the use of deep-learning, specifically for the detection and classification of species. Acoustic data in particular, used as passive acoustic monitoring, could allow the surveillance of vector populations in real-time, and assist timely public health decisions. Mosquitoes emit sound when they flap their wings during flight; the faster they beat their wings, the higher the sound. Mosquito sound varies based on several factors, including species, which is extremely useful because we only need to monitor a few species of interest; often those that transmit the diseases, but it could also apply to invasive species. AI-based algorithms already exist for the identification of mosquito species based on sound, and some perform rather well (up to 97%). However, there are a few caveats: (1) accuracies tend to decrease when many species are included, (2) few species are available in the training datasets, (3) the sounds of wild mosquito populations are likely much more variable than those represented in the training data, due to the impact of environmental (e.g. temperature, humidity) and biological factors (sex, age, size) on mosquito sound. All these aspects reduce the field applicability of AI-based species recognition based on mosquito sound. A study of researchers from HUN-REN Centre for Ecological Research, ELTE University, Budapest and University of Szeged investigates the last point, i.e. the impact of several environmental and biological factors on the variability of mosquito sound between species and between individuals.    

The researchers captured and recorded hundreds of mosquitoes in Hungary, and used the recordings from the 10 most abundant species to evaluate how much mosquito sound varies among species and individuals. In addition, they assessed the impact of several factors on sound variability: temperature, humidity, time of day, sex, age and size (represented by wing length). Sound was fairly consistent among species and among individuals. However, the acoustic signal related to a given species was even more consistent when environmental and biological variables were controlled for.

Sex and temperature both affected significantly mosquito sound. Females had a lower sound compared to males; this is not surprising, as females are usually bigger compared to the males in most mosquito species. Temperature also affected mosquito sounds; usually a higher temperature resulted in a higher sound. Higher temperature increases insect metabolism (up to a point); thus mosquito muscles move faster when temperature is higher, and they can beat their wings faster. However, the importance of this increase varied between species, meaning that different species responded differently to temperature. This could be explained by the origin of the species (temperate vs subtropical), or their preferred host and the associated blood temperature (bird blood’s temperature is usually lower compared to mammal’s blood). This species-specific difference in response to temperature suggests that we cannot apply a simple temperature correction rule for mosquito sounds, or at least that we cannot apply the same mathematical formula to all species. 

"Our data demonstrates that we cannot ignore intra-specific and intra-individual variability for AI based acoustic classification. One solution for better integration of natural variance would be to adequately represent that environmental and biological variability in the training data. Unfortunately, such complete databases remain rare, especially for invertebrates, and building these extensive databases require a lot of time and effort" - said Julie Augustin, the first author of the publication.  Alternatively, classification systems could control for or include additional environmental information to improve classification accuracy. Some studies already implement this, but it requires a detailed understanding of the impact of environmental variables on all species included in the model, which we do not have yet. In all cases, in order to improve classification models’ accuracy in real life conditions, and the chance that we can use them for monitoring purposes, we need to better understand and account for natural variability in the target populations.

END