Automated identification of larval mosquito species with computer vision

Abstract. Mosquitoes kill nearly half-a-million people each year. Due to lack of effective vaccines for most mosquito-borne diseases, prevention of mosquito bites remains the primary focus of disease mitigation. Larval surveillance - monitoring potential breeding sites to understand larval species composition, abundance, and spatial distribution - is key to enabling precision vector control. If medically relevant larvae are detected, targeted larvicide treatments can eliminate the population prior to emergence as adult mosquitoes, when disease transmission occurs. Unfortunately, due to its resource-intensive nature, few mosquito control organizations (MCOs) have the capacity to conduct the full-process of larval surveillance with identification to species. Technicians need to travel to the potential breeding site, collect specimens, return to the lab for visual morphological identification under a microscope, and if medically relevant species are found, return to the field for larvicide treatments. Delayed, inaccurate, or missing larval species data can miss an opportunity for intervention prior to emergence of adults, or misinform costly unnecessary treatment. While significant efforts have been made to explore crowdsourcing larval data and unmanned aircraft systems (UAS) for rapid assessments of potential breeding sites, implementation and technical challenges have limited practical use of these approaches. No automated solutions exist to build larval species identification capacities. We propose to develop the first commercially available image recognition platform for operational larval surveillance of the 20 most relevant mosquito species in the US. This proposal will first validate the optical requirements for species identification of larvae. This will enable the development of a standardized imaging configuration that will be used to build a high-quality image dataset of mosquito larvae of different species capturing diagnostic morphological features. The image dataset will be used to train a computer vision system to identify species of mosquito larvae. Ultimately the approaches developed here will enable new larval surveillance products, from an identification field-tool to inform technician larviciding at the point of specimen collection, to remote sensing systems that will continuously monitor historical breeding sites and alert mosquito control organizations (MCOs) when intervention is needed.