Predicting the Olea pollen concentration with a machine learning algorithm ensemble.

Area of Science:

• Environmental science
• Biometeorology
• Computational biology

Background:

• Air pollution causes millions of deaths annually, and pollen exposure is linked to allergic diseases.
• Predicting airborne pollen concentrations is challenging due to complex environmental and biological interactions.
• Accurate pollen forecasting is vital for assessing aeroallergen risk.

Purpose of the Study:

• To forecast daily Olea pollen concentrations in Madrid, Spain, using machine learning.
• To predict the timing of the peak pollen season.
• To assess the risk levels associated with olive pollen exposure.

Main Methods:

• Employed supervised machine learning algorithms, including Light Gradient Boosting Machine (LightGBM) and Artificial Neural Network (ANN).
• Developed individual models to predict the day of the year (DOY) for peak pollen season.
• Utilized an ensemble of two-step Generalized Additive Models (GAM) followed by LightGBM and ANN for daily pollen concentration prediction.

Main Results:

• Individual models accurately estimated the average peak pollen season date (around 149 DOY).
• Ensemble models achieved a coefficient of determination (r²) above 0.75 for daily pollen concentration prediction.
• Key predictors included meteorological variables, phenological metrics, site characteristics, and prior pollen counts.

Conclusions:

• State-of-the-art machine learning models can effectively predict Olea pollen concentrations.
• These predictive models are valuable tools for understanding and forecasting pollen risk levels.
• The findings contribute to better management of allergic diseases exacerbated by airborne pollen.