Machine-Learned Cloud Classes From Satellite Data for Process-Oriented Climate Model Evaluation

Abstract

Although clouds are crucial in controlling climate change, they are challenging to replicate in Earth system models (ESMs). Enhancing cloud representation is a crucial step towards more reliable climate change forecasts. In order to enhance comprehension of cloud representation and associated processes in climate models, this work presents a novel machine-learning framework based on satellite data. Coarse data may be assigned distributions of known cloud kinds using the suggested technique. It enhances the consistency of cloud process analysis and makes it easier to evaluate clouds in ESMs more objectively. Using cloud type labels from Cloud Sat as ground truth, the technique is based on deep neural networks labeling satellite data from the MODIS instrument with cloud categories established by the World Meteorological Organization (WMO). The technique works with datasets that provide physical cloud variable information at a temporal resolution that is high enough to be equivalent to MODIS satellite data. We use the technique using alternative satellite data, coarse-grained to usual resolutions of climate models, from the Cloud_cci project (ESA Climate Change Initiative). Our technique works with the common horizontal resolutions of ESMs, and the resultant cloud type distributions are physically consistent. We suggest that important variables needed by our approach be produced for next analysis of ESM data. This will make it possible to assess clouds in climate models more methodically by using tagged satellite data.