DeepPollenCount: a swin-transformer-YOLOv5-based deep learning method for pollen counting in various plant species

Accurate identification and quantification of pollens (e.g., pollen of a flower, airborne pollens) is essential to understand plant pollination and reproductive biology, pollen aerobiology, and plant–insect interactions. Currently, a couple of methods are available for pollen counting, such as manual counting, flow cytometry-based and image software-based counting. However, due to inconsistent results and experimental repeatability, a more accurate, consistent, and high-throughput quantification approach is required. This study evaluated and compared the performance between a proposed Swin-transformer-YOLOv5 (S-T-YOLOv5) and common YOLO models in pollen detection and quantification. The present study demonstrated that the S-T-YOLOv5 outperformed other YOLO models, including YOLOv3, YOLOv4, YOLOR, and YOLOv5 for alfalfa (Medicago sativa L.) pollen detection and quantification, with excellent precision (99.6%), recall (99.4%), F₁-score (0.995), mAP50 (99.4%), and mAP50-95 (76.2%) values. The mAP50-95 (mAP at an IoU of 0.5–0.95) of S-T-YOLOv5 was 9.9, 58.7, 25.3 and 8.2% higher than those of YOLOv3, YOLOv4, YOLOR, and YOLOv5, respectively. Additionally, the S-T-YOLOv5 showed a good transferability in quantifying pollen with varied sizes and shapes in different plant species, including annual fleabane, camelina, Canadian goldenrod, Indian lettuce, mustard, and oilseed rape. In summary, our results showed that the S-T-YOLOv5 is an accurate, robust, and widely adaptable pollen quantification approach, with minimizing errors and labor expense. We would like to highlight the potential application of S-T-YOLOv5 in quantifying samples of airborne pollens from a known pollen source or insect-dispersed pollens (e.g., alfalfa) in supporting the environmental risk assessment of genetically engineered (GE) plants.