Mapping mosquito flight dynamics and directional responses: A scalable deep learning model for behavioural research

Abstract

Mosquitoes are important vectors of pathogens that affect millions of people worldwide. Understanding their flight patterns and behaviours is crucial for developing novel control and surveillance tools. Mosquito flight track analysis using traditional manual methods can be time-consuming and laborious. In this study, we employed advanced image processing and deep learning techniques, specifically using a GRU model, to analyse 2D video recordings of three medically important mosquito vector species. Videos were recorded in controlled laboratory settings using simple, low-tech equipment. Our model integrates background subtraction, YOLOv5 detection, and the DeepSORT matching strategy to detect and track mosquitoes with high accuracy, achieving detection rates between 99.7% and 99.9% depending on the mosquito species, effectively mitigating challenges posed by background noise, occlusions, and tracking labels inconsistencies. In addition, a Gated Recurrent Unit (GRU) model was employed to classify mosquito movement directions with ∼97.6% accuracy. The system also generates visual outputs, including heatmaps and videos that illustrate mosquito flight trajectories, facilitating the interpretation of mosquito behavioural responses under experimental conditions. These findings demonstrate that integrating computer vision with deep learning techniques provides an effective method for tracking mosquito flight paths, classifying movement patterns, and assessing behavioural responses. This approach offers a promising avenue for automating video analysis in mosquito research and may be adapted for studying other small flying insect species.