A061-0001
Determinants of fog and low stratus clouds in continental central Europe quantified using machine learning and satellite data
Abstract:
FLS can alter the radiation balance in the climate system and provide water and nutrients to ecosystems, but the relationship to land cover and meteorological conditions have not been studied explicitly, quantitatively and on a continental scale yet. In a previously developed Gradient Boosting Regression Trees (GBRTs) model, FLS distribution based on geostationary satellite data is predicted using various land surface and meteorological parameters. Spatially explicit model units, seasonal and full-year models are created to test for spatial and seasonal differences in model performance and sensitivities. Permutation importance, partial dependencies and SHAP values are investigated to identify important features for the model and to distinguish and individually analyze different FLS regimes.
Performance of the model is good (R² ~ 0.6-0.9), which means that it can represent the fog and low stratus conditions adequately. To some extent, performance depends on model data availability, season and pressure filtering, with high performance in winter high-pressure situations. The sensitivity analysis shows that mean surface pressure (MSP), wind speed (WS) and FLS cover on the previous day are most important for the model. Evapotranspiration (ET) is especially important in high pressure situations and in spring, whereas land surface temperature (LST) gains importance in summer. Furthermore high FLS on the previous day increase the prediction especially in high pressure situations. Increasing values of ET decrease the prediction especially in high pressure situations.
In future studies, the developed model set up will be applied to explain FLS properties (cloud top height, liquid water path) to gain further insights into the relationship of FLS and the land surface.