Precipitation Nowcasting Exploring the Impact of Echo Top Heights in Generative Models

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Accurate short-term forecasting of rainfall, also known as precipitation nowcasting, is critical for a wide variety of sectors. From agriculture to early flood warning systems, reliable precipitation forecasts are essential for informed decision-making. Numerical Weather Prediction systems forecasting rainfall are not updated frequently enough and lack the spatial high-resolution required for early warning on short time scales. Deep learning nowcasting approaches can fill this gap and create fine-resolution forecasts. Generative Adversarial Networks (GAN) have lately shown promising results to improve forecasting of the challenging growth and dissipation processes of rainfall. However, forecasts of intense precipitation remain a challenge. This study investigates the prospect of improving deep learning nowcasting by the inclusion of Echo Top Height data. For this, the state-of-the-art Deep Generative Model of Radar was modified to include Echo Top Height data alongside precipitation input from the Netherlands. Both the original and the modified model were tested for accuracy on a range of continuous and categorical metrics as well as the Fraction Skill Score. It could be shown that the inclusion of ETH data improves the forecasting of low precipitation events on all metrics. The forecasting of high precipitation events was improved for large-scale applications, however, it did not improve small-scale evaluation as ETH inclusion tended to mislocate high precipitation events.
Faculteit der Sociale Wetenschappen