Integration of reference data from different Rapid-E devices supports automatic pollen detection in more locations

Area of Science:

• Environmental science
• Aerobiology
• Computational biology

Background:

• Seasonal allergies affect over 33% of Europeans, with pollen as a primary trigger.
• Real-time atmospheric pollen information is vital for public health and economic reasons.
• Automatic particle analyzers are increasingly used, but model transferability between devices is challenging due to device-specific noise.

Purpose of the Study:

• To investigate the performance of pollen classification models across different devices and locations.
• To evaluate methods for improving model transferability and accuracy in pollen identification.
• To address the challenge of device-specific noise in laser-induced bioaerosol data.

Main Methods:

• Collected pollen data using two Rapid-E bioaerosol identifiers in Serbia and Italy.
• Implemented a multi-modal convolutional neural network for pollen classification.
• Tested model performance with data from single and multiple devices, including attempts at data augmentation and domain adaptation.

Main Results:

• Models trained on data from one device performed poorly when tested on data from another, even compared to manual methods.
• Including missing pollen classes from other locations did not significantly improve performance.
• Combining all reference data improved classification for more pollen types, and domain adaptation offered partial improvements.

Conclusions:

• Pollen data from different devices exhibit significant variations, impacting model generalizability.
• A unified approach combining data and domain adaptation shows potential for cross-device pollen classification.
• Further research is needed to optimize domain adaptation for robust, real-time pollen monitoring systems.