In the Sahara and Sahel, where airborne dust poses serious hazards to public health, transportation, and energy infrastructure, this thesis explores how satellite data and machine learning might enhance short-term forecasting of mineral dust storms. It creates DustCast, a generative machine-learning nowcasting model that outperforms traditional satellite-image extrapolation by producing deterministic and probabilistic forecasts of dust storms up to five hours ahead and convective storms up to four hours ahead. The study also assesses the main atmospheric reanalysis datasets and concludes that MERRA-2 better captures diurnal cycles, while ERA5 typically offers the best accurate picture of wind speeds and variability. Lastly, it presents a process-focused approach to assessing dust simulations, pointing out shortcomings in the way dust sources, surface conditions, and wind dynamics are represented. The results promote cooperation with African meteorological agencies, show the potential of satellite-based machine learning for operational early warning, and offer methodological underpinnings for enhancing weather and climate models.
Dust storms in the Sahara and Sahel: From Nowcasting to Climate – White Rose eTheses Online