Quantifying Land-Surface Effects on Cloud Occurrence Using Neural Networks

The aim of this study is to investigate the effect of land surface conditions on cloud occurrence by quantifying how they modulate the influence of large-scale meteorological conditions.
The land surface can modulate clouds through its influence on surface heat fluxes, local moisture availability, and surface roughness. However, quantifying these effects from observations remains challenging, as the temporal and spatial variability of cloud occurrence is large and influencing factors covary.
Here, we employ a convolutional neural network (CNN) to predict satellite-observed cloud fraction over Europe for the period 1983–2020. Cloud fraction is taken from the CM SAF Cloud Fractional Cover dataset based on Meteosat First and Second Generation observations (COMET). Predictors are derived from the ERA5 reanalysis, including ERA5-Land as well as ERA5 fields on single and pressure levels. To delineate the land surface impact on cloud occurrence predictability, we develop two model configurations: one driven solely by large-scale meteorological conditions, and a second one that additionally incorporates land surface variables. Both models achieve high predictive skill (R² > 0.8), with a slight increase in performance when land surface conditions are included. Sensitivity analyses using permutation feature importance and partial dependency indicates that cloud occurrence is primarily controlled by large-scale meteorological drivers, while soil moisture and surface sensible heat flux emerge as the most influential land surface variables.
Future work will use this framework to quantify the impact of land cover change on cloud occurrence and extend the framework beyond Europe.