Monday, 29 January 2024: 2:00 PM
338 (The Baltimore Convention Center)
Handout (3.1 MB)
This presentation will discuss efforts to use a combination of deep learning models for nowcasting of model-like cloud layers. Both networks use data fusion methods to combine satellite data with numerical weather prediction data. The first model, a generative adversarial network, fuses model data and cloud observations to nowcast cloud fields. The second model then uses the cloud layers created by the generative model and determines various cloud properties in the vertical, including low/middle/cloud pressures and heights as well as cloud base and cloud top. The first model is trained on geostationary observations while the second model is trained on co-located space-borne LIDAR observations with geostationary cloud observations. We discuss in depth the architectures used for each component of the combined model. Future directions of merging these fields with NWP output and creating line-of-sight algorithms derived from the model output will be also discussed.

