A machine-learning Strategy for Generating Pollen Estimates and Forecasts in North America

Abstract

BACKGROUND AND AIM: Pollen information is important for risk communication, diagnosis of allergic disease, medical management, and surveillance. Pollen observations are sparse, however, and existing models have relatively low skill in estimating pollen concentrations for certain taxa and locations. The aim of this work is to leverage recently developed supervised Random Forest machine learning models to estimate daily pollen concentrations for multiple allergenic pollen taxa in the continental US. METHODS: We built upon a previously developed statistical supervised Random Forest machine learning model with demonstrated skill in estimating and forecasting four pollen types (Quercus, Cupressaceae, Ambrosia, and Poaceae) to make retrospective forecasts of seven other allergenic pollen taxa for the contiguous US. Meteorological, vegetation, and web search information were input to the models at city and regional scales and geographic coverage was expanded using data augmentation techniques. The models were further developed to estimate pollen concentrations in locations where there are no observations. Daily pollen concentration estimates were made for each of the 11 taxa. RESULTS:Daily estimates of pollen concentrations for 11 allergenic taxa for the period 2000 to 2020 were developed. Model forecast skill for the seven newly modeled taxa was assessed and compared with skill for the four previously modeled pollen types and available diagnostics for other pollen forecasting models. CONCLUSIONS:Weather, vegetation, and web search data can be used to estimate airborne pollen concentrations for prevalent allergenic taxa in the contiguous US. Reanalysis data products may be of use in diagnostic and management activities for people with allergic disease and for epidemiological analyses. Forecast models can be used in risk communication to facilitate exposure avoidance. Model forecasting skill is limited by availability of pollen observations and could be augmented by development of additional observation sites. KEYWORDS: pollen, random forest machine learning, allergic rhinitis, allergic asthma, forecasting, climate