Machine-learning based observational cloud products for process-oriented climate model evaluation

The importance of clouds in regulating the Earth’s energy balance as well as moisture and heat distributions cannot be overstated. Consequently, clouds have a considerable influence on the trajectory of anthropogenic climate change, of which possible scenarios are being studied with global climate models (GCMs). Uncertainties from the representation of clouds in GCMs have been identified as a leading cause of inter-model spread in climate projections.
Our current understanding of clouds and the processes relevant to their formation and effect on climate is informed partly by observations from remote sensing instruments aboard orbital satellites. This thesis introduces new methods of characterizing clouds from space with the help of machine learning and neural networks. The purpose of these methods is to improve the understanding of and reduce the uncertainties in climate projections by providing satellite products that are objectively interpretable and consistently comparable to GCM output.
The methods explored in this thesis highlight that machine learning and especially neural networks have the potential to improve multiple aspects of climate science. The presented results show that cloud classes can be reliably obtained from low-resolution data to improve their interpretability. They further show that comparison between climate models and observations can potentially be simplified with machine learning.