Airborne pollen grain detection from partially labelled data utilising semi-supervised learning.

Area of Science:

• Aerobiology
• Computer Vision
• Machine Learning

Background:

• Airborne pollen monitoring is crucial for climate studies, forensics, and allergy warnings.
• Current pollen detection relies on manual methods, which are accurate but time-consuming.
• Automated pollen classification exists, but automated detection remains a challenge.

Purpose of the Study:

• To develop and evaluate deep learning models for automated pollen detection using a new sampler (BAA500).
• To compare the performance of deep learning algorithms against the commercial BAA500 algorithm.
• To assess the effectiveness of a semi-supervised training scheme for improving detection accuracy with limited labeled data.

Main Methods:

• Utilized a new-generation automated pollen sampler (BAA500) with raw and synthesized microscope images.
• Employed two-stage deep neural network object detectors for pollen detection.
• Implemented a semi-supervised learning approach with a teacher-student model to address partial labeling.

Main Results:

• Deep learning models (supervised and semi-supervised) outperformed the commercial algorithm on a manually curated test set, achieving an F1 score of up to 76.9% (vs. 61.3%).
• Achieved a maximum mean Average Precision (mAP) of 92.7% on an automatically created, partially labeled dataset.
• Experiments with raw microscope images showed comparable performance, suggesting potential simplification of image processing.

Conclusions:

• Developed deep learning models that significantly advance automated pollen detection performance.
• Closed the performance gap between manual and automated pollen detection methods.
• Results support the potential for more accurate and efficient near-real-time airborne pollen monitoring.