Two-Stage Prototypical Networks Reveal Mosquito Flight Patterns

ICLR 2026 Conference Submission17481 Authors

Keywords: Mosquito Behavior, Movement Density Map, YOLO, Few-Shot Learning, Prototypical Networks

Abstract: Understanding behavioral movements in mosquitoes is fundamental for monitoring arbovirus transmission. Most existing Artificial Intelligence (AI) methods recognize tiny insects as background and fail to extract correct features from video frames. To address this issue, we propose a two-stage few-shot classification by Movement Density Map (MDM) prototyping. A novel approach that integrates object detection with two-stage prototype training to analyze and identify mosquito behavior from videos. In the first stage, mosquitoes are detected using a fine-tuned YOLO, achieving a maximum mean Average Precision (mAP50) of 97.8% after 100 training sessions. The detected areas with eliminated backgrounds are then aggregated into MDMs. This mechanism enables encoding hundreds of frames into a single spatiotemporal representation that reveals biologically meaningful flight patterns over time. The MDMs are then mapped into a Vision Transformer (ViT) embedding environment, where class-level prototypes are generated for few-shot classification under 1 and 5 exposures using prototypical networks. Results on datasets of dengue and Zika-carrier mosquitoes, as well as non-carrier ones, collected over 13 days and nights show that our approach significantly extracts more accurate features than a common single-stage prototypical network, leading to an overall performance accuracy of 85.86% . These findings reveal that two-stage prototyping is a reliable and scalable solution for analyzing tiny-object biological videos and holds promise for other spatiotemporal recognition tasks where motion aggregation is critical.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 17481