Modelling Europe-wide fine resolution daily ambient temperature for 2003-2020 using machine learning

Area of Science:

• Environmental epidemiology
• Climate science
• Geospatial analysis

Background:

• Epidemiological studies investigating health impacts of temperature extremes require accurate, high-resolution ambient temperature (Ta) data.
• Existing data often lacks the necessary spatiotemporal resolution and harmonization for large-scale European multi-cohort studies.

Purpose of the Study:

• To develop daily mean, minimum, and maximum ambient temperature (Ta) surfaces at 1x1 km resolution for Europe.
• To cover the period from 2003 to 2020, facilitating harmonized exposure assessment for European multi-cohort studies.

Main Methods:

• A two-stage random forest modeling approach was employed.
• Stage 1: Imputed missing satellite-derived Land Surface Temperature (LST) using vegetation and weather data.
• Stage 2: Modeled Ta variation using gap-filled LST, land use, and meteorological data, validated with random and block methods.

Main Results:

• Models explained over 81% (LST) and 93% (Ta) of variability in block validation sets.
• Average Root Mean Square Error (RMSE) was 1.3°C (mean Ta), 1.9°C (min Ta), and 1.7°C (max Ta).
• Model performance was generally stable but showed lower accuracy in areas with sparse weather station data.

Conclusions:

• The developed high-resolution temperature surfaces provide a valuable resource for European epidemiological research.
• These datasets enable improved exposure assessment in studies examining the health effects of ambient temperature.
• The methodology offers a robust approach for generating harmonized environmental exposure data at a continental scale.