Stephanie Fiedler

Mineral dust is one of the most abundant atmospheric aerosol species and has various far‐reaching effects on the climate system and adverse impacts on air quality. Satellite observations can provide spatio‐temporal information on dust emission and transport pathways. However, satellite observations of dust plumes are frequently obscured by clouds. We use a method based on established, machine‐learning‐based image in‐painting techniques to restore the spatial extent of dust plumes for the first time. We train an artificial neural net (ANN) on modern reanalysis data paired with satellite‐derived cloud masks. The trained ANN is applied to cloud‐masked, gray‐scaled images, which were derived from false color images indicating elevated dust plumes in bright magenta. The images were obtained from the Spinning Enhanced Visible and Infrared Imager instrument onboard the Meteosat Second Generation satellite. We find up to 15% of summertime observations in West Africa and 10% of summertime observations in Nubia by satellite images miss dust plumes due to cloud cover. We use the new dust‐plume data to demonstrate a novel approach for validating spatial patterns of the operational forecasts provided by the World Meteorological Organization Dust Regional Center in Barcelona. The comparison elucidates often similar dust plume patterns in the forecasts and the satellite‐based reconstruction, but once trained, the reconstruction is computationally inexpensive. Our proposed reconstruction provides a new opportunity for validating dust aerosol transport in numerical weather models and Earth system models. It can be adapted to other aerosol species and trace gases.