DiffREE: feature-conditioned diffusion model for radar echo extrapolation

In recent years, deep learning has become integral to short-term precipitation forecasting through radar echo extrapolation. However, as the extrapolation time increases, radar echo intensity diminishes, leading to a significant decline in forecast accuracy, particularly for strong echoes. To address these challenges, we propose a novel diffusion radar echo extrapolation algorithm, DiffREE. This algorithm integrates spatio-temporal information from radar echo frames using a conditional encoding module and employs a Transformer encoder to automatically extract these features. The extracted features then guide a conditional diffusion model to reconstruct the current radar echo frame. Validation experiments demonstrate that DiffREE produces high-precision and high-quality radar echo forecasts. When compared to four other models using public datasets, DiffREE significantly improves the critical success index, equitable threat score, Heidke skill score, and probability of detection by 21.5%, 27.6%, 25.8%, and 21.8%, respectively, underscoring its superiority.