Self-supervised deep-learning of cloud spatio-temporal features to improve understanding of processes and evolutions of cloud organizations.

In 2050, in Europe, damages due to climate change and flooding are expected to reach € 45 billion annually. Severe storms produce heavy rain, often causing landslides, and orography is crucial in triggering such events, especially in the foothills of the Alps. Due to the difficulties in conducting ground-based remote sensing observations over complex terrains, satellite observations represent a valuable alternative for monitoring high-resolution natural hazards over the Alpine region.

The most recent approaches to detecting severe storms rely on integrating multiple datasets and machine learning to improve weather prediction and develop accurate nowcasting. Most of the time, the temporal dimension is used to make short-term predictions; this is done, for instance, by applying atmospheric motion vectors to radar images or by exploiting machine learning approaches. However, in the cloud's spatiotemporal evolution patterns, there is still some information on cloud systems that remains largely unexplored.

In the past, deep learning self-supervised techniques have shown exciting developments in identifying spatial cloud patterns from satellite images and characterizing the conditions under which such organization occurs. In this contribution, we utilize recent deep-learning self-supervised algorithms developed to identify motions in space-time to study the evolution of cloud patterns. We aim to characterize and distinguish cloud systems based on their spatiotemporal evolution, exploiting the features that these algorithms can extract. We will characterize the rates of change in the environment and cloud parameters of the patterns identified by the deep learning architecture. We are currently testing the algorithm on Meteosat Second Generation (MSG)/OPERA radar data available on European Weather Cloud (EWC). However, we plan to exploit Meteosat Third Generation (MTG) and Deutsche Wetterndiest (DWD) radar data on a smaller domain to test the algorithm's performance of the new highly resolved MTG satellite data.