Allergenic pollen forecasting with ensemble machine learning: Toward spatially transferable models in sparse monitoring networks

• Weighted ensembles were developed for multi-day allergenic pollen forecasting. • The approach showed a robust performance and spatial transferability to new locations. • Models showed promise for regions with sparse monitoring using non‐local pollen data. • The impact of weather and phenological indicators on pollen levels was assessed. • Methodology provides a solution for allergy management in data-scarce environments. Airborne allergenic birch (Betula) and grass (Poaceae) pollen poses significant public health challenges, as high concentrations can trigger allergic rhinitis and exacerbate asthma for millions of people globally. Accurate forecasts of pollen concentrations enable vulnerable individuals to take preventive measures and support healthcare planning. In this study, we developed a spatially transferable forecasting methodology based on weighted ensemble machine learning models to predict daily birch and grass pollen concentrations up to five days ahead. Historical daily pollen measurements (2006–2022) from a Hirst-type trap in Wrocław, Poland, were combined with Weather Research & Forecasting (WRF) meteorological outputs (e.g., temperature, humidity, wind speed, precipitation), phenological indices (e.g., growing degree days), calendar features (e.g., day of year, week), and lagged pollen predictors. Four decision tree-based base learners – Random Forest, ExtraTrees, XGBoost, and LightGBM – were tuned within nested leave‐year‐out cross-validation and combined via a weighted‐average ensemble. Separate ensembles were trained for lead times from +1 to +5 days for both taxa. Variable importance analysis revealed that lagged pollen predictors dominated short‐term forecasts, while phenological and meteorological variables gained prominence at longer lead times. Independent testing on 2023–2024 data covered Wrocław and three other Polish cities, with two scenarios: (A) using local lagged pollen inputs and (B) replacing local inputs with non-local data from Wrocław (applied only to external sites). Successful spatial transferability was demonstrated in both scenarios. Ensemble models consistently outperformed individual algorithms, achieving test coefficients of determination (R²) of ∼0.77 (birch) and ∼0.72 (grass) at +1 day, declining to ∼0.55 and ∼0.66 at +5 days (scenario A).The findings illustrate that a weighted ensemble can produce reliable short‐ and medium‐term pollen forecasts in data‐scarce regions even if only a single well‐monitored site is available. The approach is readily adaptable for operational implementation and may improve allergy prevention and public‐health interventions in regions with limited pollen monitoring.