A0705
Title: Nowcasting of high precipitation events with deep learning
Authors: Yuliya Shapovalova - Radboud University (Netherlands) [presenting]
Abstract: Precipitation nowcasting is critical for weather-dependent decisions but remains challenging despite active research. Combining radar data and deep learning has created new research opportunities. Radar data, with high space-time resolution, are ideal for nowcasting, while deep learning exploits possible nonlinearities in the precipitation process. Deep learning approaches have matched or outperformed optical flow methods for low-intensity precipitation, but high-intensity events nowcasting remains difficult. Two different deep learning architectures, deep generative model and diffusion model, are built upon and are extended to enhance the nowcasting of heavy precipitation. Specifically, different loss functions and the effect of adding temperature data as an additional feature are explored. Using KNMI radar data and 590-min lead times, the model with these enhancements outperforms state-of-the-art models, effectively nowcasting high rainfall intensities up to 60-min lead times.
