Modelling daily air temperature at a fine spatial resolution dealing with challenging meteorological phenomena and topography in Switzerland

Abstract

Epidemiological studies investigating the relationship between air temperature or heat and health, still, by and large, rely on either information from the nearest weather station or on coarse gridded temperature predictions, thereby ignoring small-scale intra-urban variations. Recent methodological advances show promise in achieving high spatiotemporal temperature predictions, thus improving the characterization of spatial variations in temperature and decreasing bias in health studies. Here, we applied a two-stage approach using random forest to (a) impute missing moderate resolution imaging spectroradiometer (MODIS) land surface temperature at a 1 × 1 km resolution and (b) to use the gap-filled MODIS data to explain spatiotemporal variation in the measured ground-based air temperature data at a 100 × 100 m resolution across Switzerland using a range of predictor variables, including meteorological parameters, normalized difference vegetation index, impervious surface and altitude. Models presented here managed to capture temporal and spatial variations in air temperature in Switzerland from 2003 to 2018 at a fine spatial resolution of 100 × 100 m. Stage 1 models achieved an overall R2 of 0.98 and a root mean squared error (RMSE) of 1.49°C (independent validation), and the stage 2 model performed well for all years with R2 and RMSE ranging from 0.94 to 0.99 and 1.05 to 1.86°C, respectively. We were also able to capture the urban heat island effect and some typical weather phenomena caused by Switzerland's complex topography, like the foehn effect and inversion conditions. The resulting daily temperature surfaces for 2003–2018 will facilitate ongoing epidemiological research investigating the health effects of heat.