Computational study demonstrates deep learning improves short-term precipitation nowcasting, highlighting the potential of neural networks for extreme weather prediction.
Key Points
To evaluate whether deep learning architectures can treat radar-based precipitation nowcasting as an image-to-image translation problem to generate accurate, high-resolution short-term rain forecasts.
Trained a UNET convolutional neural network using radar imagery to generate 1-hour precipitation forecasts at a 1 km by 1 km spatial resolution.
Benchmarked model performance against persistence, optical flow, and NOAA's numerical High-Resolution Rapid Refresh (HRRR) nowcasting prediction.
The UNET architecture outperformed both persistence and traditional optical flow methods for 1-hour localized precipitation forecasts.
The deep learning approach showed favorable predictive skill when compared directly to NOAA's operational numerical HRRR nowcast.