Nowcasting of dust and convective storms via diffusion-model predictions of SEVIRI RGB imagery

Convective storms in the tropics are inherently unpredictable on the scales typical of global Numerical Weather Prediction (NWP) models. Rapid-refresh short term predictions, so called nowcasts, provide added value and can improve forecasts on short timescales. Nowcasts provide the most up to date predictions, making them particularly relevant for rapidly developing high impact weather that is not reproduced in global NWP models. While NWP nowcasts exist for USA, UK, and continental Europe, this is not the case for Africa where nowcasts are primarily observation-based. Here we focus on West Africa, an area where convective storms are frequent weather events. Besides direct impacts from the convective storms, outflow from these storms frequently causes large-scale dust storms. Dust storms are high impact weather and directly impact human life. Early warning is crucial for mitigating their adverse impacts. While dust storms are poorly forecast in currently operational weather prediction models, they are well observed from space, posing great potential for nowcasting. The desert dust red-green-blue (RGB) composite highlights dust and convective systems in bright colours, making it a useful product for a trained observer for identifying and tracking dust storms. In this study, we introduce DustCast, a diffusion model for image-based nowcasting of dust storms by predicting the SEVIRI desert dust RGB composite up to 6 h ahead. Our nowcasts can reproduce convectively generated dust storms that currently operational NWP do not reliably reproduce. We also predict convective storms that are contained in the same imagery and provide useful context information for a forecaster assessing the weather situation. Our model shows limited capability of reproducing entirely new features that are not contained in the input data. This primarily poses limitations for cases with convection initiation. On average, our model achieves useful skill (Fractions Skill Score > 0.5) for predicting dust storms up to 5 h lead time, and for convective systems for up to 4 h. DustCast is the first model of its kind for nowcasting dust, and extends skill for nowcasting convective storms by more than 2 h compared to conventional methods based on optical flow. Deployment during a nowcasting testbed has shown that our nowcast provides an easy-to-use product for operational forecasters. Our method could also be adapted to predict other RGB composites such as those specifically for convection, ash or fog, and indeed other products using observation data from geostationary satellites, opening potential for a large variety of applications.