A Two-Stage, Semi-Supervised Deep Learning Framework for the Detection and Classification of Ambient Pollen using...

Airborne pollen, an aeroallergen, impacts human health, particularly as climate change shifts pollen dynamics, requiring accurate monitoring to support public health interventions. The present study utilizes evanescent wave scattering microscopy (EWSM) for label-free identification of airborne pollen. EWSM provides unique scattering signatures for various particles based on morphological characteristics. Bright-field (BF) microscopy offers limited pollen morphological details (spines, muri, and perforations); EWSM effectively visualizes these details and is used for the first time to identify airborne pollen. Two-stage semi-supervised framework is implemented for the identification and classification of pollen particles. Three models (YOLOv8n, YOLOv8l, and RT-DETR) were selected for the purpose of pollen localization; RT-DETR was selected as the stage 1 detector due to its 82.3% recall rate. Classification was performed using an EfficientNetB0 architecture. "Human-in-the-loop" label refinement was implemented in two phases. Phase 1 improved the F1 score from 0.277 to 0.537 for 41 species (80% recall). Phase 2 consolidated these species into 19 morphological groups with a precision of 0.648, a recall of 0.744, and an F1-score of 0.693, optimizing detection thresholds to 0.80. These refinements reduced false-positive detections by 63%. This framework demonstrates a scalable approach for the autonomous analysis of ambient pollen in heterogeneous urban environments.