Machine learning methods for low-cost pollen monitoring - Model optimisation and interpretability.

Area of Science:

• Environmental Science
• Data Science
• Allergy Research

Background:

• Pollen allergies affect up to 40% of the global population, necessitating improved monitoring.
• Current pollen monitoring methods are often slow, laborious, or costly.
• There is a need for timely, localized airborne pollen concentration data.

Purpose of the Study:

• To enhance machine learning models for pollen concentration estimation using Optical Particle Counter (OPC) data.
• To investigate the impact of methodical hyperparameter tuning on model performance.
• To utilize explainable Artificial Intelligence (XAI) for interpreting model predictions.

Main Methods:

• Utilized low-cost Optical Particle Counter (OPC) sensors to collect particle data.
• Applied machine learning algorithms with methodical hyperparameter tuning to predict pollen concentrations.
• Employed SHAP (SHapley Additive exPlanations) for model interpretability and feature analysis.

Main Results:

• Hyperparameter tuning significantly improved model performance, with average R2 scores for total pollen models at least doubling.
• Models successfully predicted concentrations for Poaceae, Quercus, Betula, Pinus, and total pollen.
• SHAP analysis revealed specific particle size correlations, e.g., Quercus pollen with 1.7-2.3 μm particles.

Conclusions:

• Optimized machine learning models show significant potential for accurate, low-cost pollen monitoring.
• Explainable AI provides valuable insights into the relationship between particle size and pollen type.
• Further research is needed to assess model generalizability across different environments.