Skilful precipitation nowcasting using deep generative models of radar

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence and Machine Learning
• Geospatial Data Analysis

Background:

• Precipitation nowcasting (forecasting up to two hours) is vital for weather-dependent sectors.
• Current radar-based methods struggle with non-linear events like convective initiation.
• Deep learning models lack physical constraints, leading to blurry forecasts and poor performance on heavy rain events.

Purpose of the Study:

• To develop a deep generative model for probabilistic precipitation nowcasting from radar data.
• To address limitations of existing methods, particularly for rare, medium-to-heavy rain events and longer lead times.
• To improve forecast quality, consistency, and value in operational weather prediction.

Main Methods:

• Development of a deep generative model for probabilistic precipitation nowcasting.
• Utilizing radar data for high-resolution, spatiotemporally consistent predictions.
• Evaluation using statistical, economic, cognitive measures, and expert meteorologist assessments.

Main Results:

• The generative model produces realistic and consistent predictions over large regions (up to 1,536 km × 1,280 km) and lead times (5-90 min).
• Expert meteorologists ranked the model first for accuracy and usefulness in 89% of cases against two competitive methods.
• Quantitative verification shows skillful nowcasts without blurring, improving forecast value and operational utility.

Conclusions:

• Deep generative models offer a promising approach for accurate and valuable precipitation nowcasting.
• The proposed model overcomes limitations of existing methods, especially for extreme weather events.
• Generative nowcasting enhances operational decision-making by providing reliable, high-resolution, probabilistic forecasts.