Precipitation nowcasting over Western Africa using transfer learning with WeatherGenerator

Accurate nowcasting of high-intensity precipitation is essential for flood modeling, disaster management, and decision making. Due to the nature of precipitation, the intensity and timing can strongly vary spatially. While some areas of the world have dense networks of openly available automated weather stations or weather radars, these are not available everywhere. In sub-Sahara Western Africa, high-intensity precipitation has a high risk of causing hazardous flash floods, and with very little radar data available in the region, nowcasting is mostly restricted to available satellite products. 

Using WeatherGenerator atmospheric foundation model, we explore the viability of training a machine learning model to accurately nowcast heavy precipitation in Western Africa. We investigate fine-tuning the pre-trained WeatherGenerator to SEVIRI output, training a tail network that predicts rainfall retrieval from the MSG-CPP product. We also explore transfer learning with WeatherGenerator, using a decoder trained to EURADCLIM over the European continent with SEVIRI input and assessing its accuracy over the target region. 

This effort adds to our understanding of the flexibility and added value of WeatherGenerator as a foundation model for weather and climate. It also serves as a pilot for upcoming service projects that the WeatherGenerator consortium will offer to the earth-scientific community, focusing on a broad range of applications and stakeholders.