Comparison of gridded ambient temperature exposure estimates derived from various temperature models and weather station networks for epidemiological research

We evaluate systematic differences between weather station-based temperature exposures and exposures derived from a number of different spatially resolved temperature databases in the context of short and long-term epidemiological research. We compared daily ambient temperature data across multiple European cities from the following four sources: i) weather station networks (Ta_WS); ii) land surface temperature (LST); iii) ERA5-land (Ta_ERA5); and iv) statistical models (Ta_EXP). We calculated the spatial and temporal variability for each of the four temperature datasets and their pairwise agreement using correlation coefficients, mean bias error (MBE) and root mean squared error (RMSE). We found very high temporal agreement between all pairs of temperature datasets. In contrast, spatial correlations were only high for LST and Ta_EXP (r: 0.89, other pairs r < 0.4). LST and Ta_EXP showed higher spatial variability linked to urban topography when compared to Ta_ERA5 and Ta_WS. During extreme heat days, Ta_EXP and LST showed average spatial temperature variability above 2C° and 4C°. However, LST temperature variability and pairwise agreement against ambient temperature datasets showed seasonal differences with LST overestimating temperatures and thermal contrasts in summer and underestimating Ta during winter. For citywide time-series studies product choice has a limited effect on epidemiological research as all tested products showed similar daily trends. For studies focusing on individual or small-area levels, higher resolution products are required to capture spatial temperature contrasts. Statistical models show a good balance between using LST as predictor to tap its abundant spatial information and limiting LST season-specific over- and underestimation of temperature and temperature contrasts by calibrating predictors with weather stations data.