192 - Artificial intelligence solutions for next-generation mosquito surveillance

Abstract: The mosquito is the deadliest animal on the planet, infecting over 700 million and killing nearly 1 million people every year. However, only 3% of the 3,500 species of mosquitoes are known to transmit diseases to humans, making identification the critical bottleneck for surveillance and control. Currently, mosquito trapping and identification is a purely manual process that is inefficient, time-consuming, and cognitively demanding, and requires entomological expertise that is becoming scarcer. Our solution is a hardware and software system that enables accurate, automated, and explainable species identification—bringing the power of AI to mosquito control districts for surveillance that is economical, efficient, and scalable. This smart trap is a low-cost device (< $100) built with off-the-shelf components (e.g., Raspberry Pi) and 3D-printed materials, including an adapter for the CDC miniature light trap. Methodological innovations are discussed, including explainable AI (XAI) approaches such as grad-CAM/CAM heat maps for visualization and validation, leveraging focal loss to handle class imbalances, and masking techniques useful for anatomically informed deep learning. Together, this system provides real-time, cloud-based analyses via AI algorithms that identify mosquito species and gonotrophic stage, and we report initial results from lab and field trials. For web-based training of both these algorithms and personnel, we have also developed an anatomical annotation dashboard. We also present our updated Global Mosquito Observations Dashboard (mosquitodashboard.org), an interactive mapping tool that integrates mosquito data from citizen science apps, as well as our traps. Finally, based on such citizen science observations, we share our AI case study involving the potential first detection of invasive malaria vector Anopheles stephensi in Madagascar [publication: doi.org/10.3390/insects16111098].