Drivers of fog and low stratus - a satellite-based evaluation with machine learning

In this study the distribution of fog and low stratus (FLS) relative to land cover and meteorological conditions in continental Europe is investigated using a state-of-the-art machine learning technique  and geostationary satellite data. While analyses of the spatial and temporal patterns of FLS exist, the relationships to land cover and meteorological conditions have not been studied explicitly, quantitatively and on a continental scale.
The machine learning model is built using daily means of a FLS dataset from Egli et al. (2017) as the predictand, and different land surface and meteorological parameters as predictors over continental Europe from 2006-2015. The application of a machine learning approach provides the ability to explicitly, synchronously and quantitatively link suspected determinants of FLS to its occurrence.
The model shows good performance with R² values ranging between 0.5 and 0.9, depending on model grid size, season and further settings such as the exclusion of low pressure values and subtraction of seasonality. It is thus able to adequately represent the dynamics that drive FLS development. Based on a systematic analysis of this model, the most important features for FLS prediction are mean surface pressure, wind speed, and FLS on the previous day. High mean surface pressure, high FLS cover on the previous day, low evapotranspiration, wind speed and land surface temperature lead to higher predicted FLS values.
Generally the results show that it is possible to predict FLS occurrence over continental Europe using meteorological as well as land surface parameters with good performance indicating the benefits of using machine learning in the analysis of non-linear, multivariate systems such as the land-atmosphere system. Further studies will integrate the machine learning model into a land surface based model grid and implement fog and low cloud properties as predictands.