Betula pollen observation: integration of automated device records into long-term datasets

Technological progress and the timely availability of Earth Observation (EO) data have rapidly changed pollen research. Innovative solutions enable the development of instruments for airborne pollen identification, and integrating an increasing amount of remotely acquired EO data improves pollen forecasts. Nowadays, data about pollen in the air are available from different types of devices. Pollen data gathered using Hirst-type volumetric spore traps is especially valuable as they are long-term and can be used to evaluate climate peculiarities. The air samples were collected over 20 years. Meanwhile, data on pollen spread has recently been monitored using more sophisticated devices. The new generation devices collect data about airborne pollen automatically in near real-time. Multiple statistical methods are essential in data homogenisation, especially when integrating heterogeneous data to handle long-term observation challenges. This study aims to demonstrate the feasibility of statistical methods used to integrate records about airborne pollen from automated devices into long-term data collected with Hirst-type volumetric spore traps.

The research is based on 20 years of Betula pollen data (2005-2024) collected with a Hirst-type trap (so-named manual data) and the short-term pollen data from SwisensPoleno Mars records (so-named automatic data) covered by several years. Both devices are operational and located in Vilnius, Lithuania. Overlapping datasets from 2022 to 2024 were used in this research. We chose the data modelling pathway to assess the integration of automated device records with long-term data. Several statistical modelling approaches were tested: simple linear regression, polynomial multiple regression, generalised additive model, Prophet model, random forest model and their combinations.

Multivariate polynomial regression enables the estimation of non-linear relationships and local data heterogeneity. Heterogeneity in local pollen data records can be caused by peculiarities of flowering time and/or local weather patterns, which require models to evaluate differences. Generalized Additive Model (GAM’s) handless non-linear and seasonal patterns of airborne pollen. The Prophet model concept was also applied to estimate long-term trends and seasonality of datasets. Statistical data analysis of long-term Betula pollen data was used to make corrections to the data collected by the automatic devices and to compare the corrected bias with the observational data to assess the performance and applicability of the methods. Considering variations in mean bias error (MBE), mean absolute error (MAE) and root mean square error (RMSE), the tested models were demonstrated to highlight different causal pathways for the inconsistency between the long-term manual data and the short-term automatic data. The knowledge gained is valuable for integrating observational data into current forecasting tools, such as PASYFO, which forecasts allergy symptoms, as well as homogenising heterogeneous airborne pollen data.

This research is supported by the projects EO4EU and SYLVA, funded by the Horizon Europe RIA Programme under Grant Agreements No. 101060784 and No. 101086109.