Integrated nowcasting of convective precipitation with Transformer-based models using multi-source data

Abstract

Precipitation nowcasting is crucial for mitigating the impacts of severe weather events and supporting daily activities. Conventional models predominantly relying on radar data have limited performance in predicting cases with complex temporal features such as convection initiation, highlighting the need to integrate data from other sources for more comprehensive nowcasting. Unlike physics-based models, machine learning (ML)-based models offer promising solutions for efficiently integrating large volumes of diverse data. We present EF4INCA, a spatiotemporal Transformer model for precipitation nowcasting that integrates satellite- and ground-based observations with numerical weather prediction outputs. EF4INCA provides high-resolution forecasts over Austria, accurately predicting the location and shape of precipitation fields with a spatial resolution of 1 kilometre and a temporal resolution of 5 minutes, up to 90 minutes ahead. Our evaluation shows that EF4INCA outperforms conventional nowcasting models, including the operational model of Austria, particularly in scenarios with complex temporal features such as convective initiation and rapid weather changes. EF4INCA maintains higher accuracy in location forecasting but generates smoother fields at later prediction times compared to traditional models. Interpretation of our model showed that precipitation products and SEVIRI infrared channels CH7 and CH9 are the most important data streams. These results underscore the importance of combining data from different domains, including physics-based model products, with ML approaches. Our study highlights the robustness of EF4INCA and its potential for improved precipitation nowcasting. We provide access to our code repository, model weights, and the dataset curated for benchmarking, facilitating further development and application.