Reproduce “A review of radar-based nowcasting of precipitation and applicable machine learning techniques” on OpenResearch

About this paper

Rachel Prudden, Samantha Adams, Dmitry Kangin, Niall Robinson, Suman Ravuri, Shakir Mohamed, Alberto Arribas

This review paper from the Met Office Informatics Lab, University of Exeter, and DeepMind synthesizes traditional radar-based precipitation nowcasting methods with applicable machine learning techniques. It identifies the limitations of current extrapolation models, particularly their inability to capture the evolving life cycle of precipitation, and highlights how deep learning, especially generative models, can address these challenges to produce more accurate and probabilistic short-term forecasts.

Key insights

• Traditional radar-based nowcasting, predominantly using optical flow and advection, is effective for very short timescales but struggles with predicting the temporal evolution, growth, or decay of precipitation, which is a major source of error.
• Machine learning, especially deep learning, offers a pathway to overcome these limitations by capturing complex non-linear spatio-temporal patterns and integrating diverse data sources without rigid physical assumptions.
• Generative models within machine learning are crucial for producing sharp, realistic, and probabilistic precipitation forecasts, addressing the tendency of deterministic models to yield blurry predictions when multiple outcomes are possible.

Summary by alphaXiv.