Abstract:
The formation and development of fog and low stratus clouds (FLS) depend on meteorological and land surface conditions and their interactions with each other. While analyses of temporal and spatial patterns of FLS in Europe exist, the interactions between FLS determinants underlying them have not been studied explicitly and quantitatively at a continental scale yet. In this study, a state-of-the-art machine learning technique is applied to model FLS occurrence over continental Europe, using meteorological and land surface parameters from geostationary satellite and reanalysis data. Spatially explicit model units are created to test for spatial and seasonal differences in model performance and FLS sensitivities to changes in predictors, and effects of different data preprocessing procedures are evaluated. The statistical models show good performance in predicting FLS occurrence during validation, with R2>0.9 especially in winter high pressure situations.The predictive skill of the models seems to be dependent on data availability, data preprocessing, time period, and geographic characteristics. It is shown that atmospheric proxies are more important determinants of FLS presence than surface characteristics, in particular mean sea level pressure, near-surface wind speed and evapotranspiration are crucial, together with FLS occurrence on the previous day. Higher wind speeds, higher land surface temperatures and higher evapotranspiration tend to be negatively related to FLS. Spatial patterns of feature importance show the dominant influence of mean sea level pressure on FLS occurrence throughout the central European domain. When only high pressure situations are considered, wind speed (in the western study region) and evapotranspiration (in the eastern study region) gain importance, highlighting the influence of moisture advection on FLS occurrence in the western parts of the central European domain. This study shows that FLS occurrence can be accurately modeled using machine learning techniques in large spatial domains based on meteorological and land surface predictors. The statistical models used in this study provide a novel analysis tool for investigating empirical relationships in the FLS – land surface system and possibly infer processes.
