(context of the research) Pollen modeling is a crucial field for understanding the impact of pollen on human health and the environment. Deterministic models are the most common type of pollen model. They are generally more accurate than probabilistic models, but they require a good understanding of the physical laws that govern pollen production, release mechanism, and dispersion; further, a precise distribution of plants is needed, and it is not always possible with invasive species or seasonal plants that are continuously spreading.

This study aims to develop a neural network model that can accurately predict pollen concentration and the start of the pollen seasons in France.

We tested different neural network models combining phenological observations, pollen data from a Hirst network, and meteorological data, selecting two locations (Avignon and Brest). Over the 20 years of data, several random sets were selected, excluding different weeks, to test the ability to predict pollen concentration.

The “best” model was able to predict with accuracy above 80% pollen concentrations and the arrival of the grass season with an accuracy of 6 days. This is significantly better than the accuracy of either deterministic or probabilistic models alone. Notably, our findings indicate that pollen concentrations from Hirst samplers facilitate reliable predictions at one week.

The model is a promising tool for predicting the arrival of the pollen season and improving forecasting pollen concentration. It is more accurate than traditional models and can provide early warnings to people with allergies.