Speaker
Description
Clouds represent a major source of uncertainty in Earth system models. In particular, the altitude and vertical distribution of clouds strongly influence their radiative properties and impact Earth’s climate. As such, constructing real-time 3D global cloud maps with a high temporal frequency could result in a significant improvement in these models. Satellites such as CloudSat and, more recently, EarthCARE help us in gaining information about the vertical distribution of cloud properties. However, with narrow track width and long revisit times (about 25 days), we only obtain a snapshot of our atmosphere and cannot directly study how clouds evolve on a timescale of minutes to hours. In comparison, imagery from passive sensors onboard geostationary satellites is available with high temporal cadence (15 mins) but only provide the “top down” view without directly probing atmospheric profiles. In this study, we extend the sparse vertical profiles in both space and time and create 3D cloud maps from 2D satellite imagery using machine learning. More specifically, we use imagery in 11 spectral channels from the MSG/SEVIRI instrument, collocated with vertical profiles measured by CloudSat’s Cloud Profiling Radar (CPR) for 2010. To take into account the vast amount of data from MSG/SEVIRI, we use unsupervised pre-training by using a masked autoencoder architecture to improve upon existing methods.
| If your abstract is not accepted for an oral presentation, would you be interested in presenting it as a poster instead? | Yes |
|---|---|
| If accepted by the Scientific Board, I agree to have my presentation/poster and abstract published on the Belgian Climate Centre websites and social media. | Yes |
