Forecasting Grass Pollen with Satellite Sensor Time-series, Meteorology data, and Machine Learning Tools

Grass pollens are a major cause of seasonal allergic asthma and rhinitis (hay fever) globally. Current pollen forecasts are based on pollen concentrations at sparse monitoring stations and rely on empirical models based on weather data. It is crucial to provide more accurate and timely warnings, with greater lead times, for allergy sufferers and public health providers. In this study, we first extracted grass phenology metrics from time-series satellite greenness measurements to track grass pollen sources in advance of their pollen emission periods. Next, the spatially detailed grass phenological information that are related to the flowering phenophase was analysed together with the pollen concentration collected by pollen traps deployed in New South Wales region of Australia where pollen emissions are year-round, released by both warm-season and cool-season grasses. The state-of-art deep learning tools like long short-term memory (LSTM), and popular machine learning tools like artificial neural network (ANN) were used to analyse the relationships between pollen source, weather conditions, and pollen concentrations at the monitoring stations. Pollen concentrations collected at one monitoring site coupled with the meteorology and grass phenology data were used to train the model, which was then used to predict pollen concentrations at various locations. The prediction was done at different lead times (one day, one week, and three weeks), and the forecasted pollen concentration was validated using pollen data collected at a different site that was not involved in the training dataset. The accuracy of the model was evaluated using several statistical metrics. Our model shows that meteorological factors outweigh phenological factors in predicting the short-term pollen concentrations, while longer lead time predictions were more related to grass source phenology. The results demonstrate that our proposed pollen forecast model using remote sensing-based grass phenology information, meteorology data, and pollen concentration collected by pollen traps significantly improves grass pollen forecasting.