An Individual Motion-Driven Artificial Intelligence Method for Precipitation Forecasting Using Radar Image Sequencing

Precipitation forecasting, encompassing high-resolution regional and accurate intensity forecasting, has long been a central focus in the artificial intelligence (AI) field. Generally, AI models forecast future radar precipitation sequences from historical radar precipitation sequences. However, current precipitation forecasting suffers from three issues: 1) mismatched precipitation patterns, for example, motion speed; 2) blurred precipitation field generation; and 3) inaccurate intensity forecasts. It is owing that AI models: 1) tend to simulate the average motion states and overlook individual motion (defined as the description of the motion speed, trajectory, and direction for a single precipitation sample), leading to either overestimation or underestimation of the motion speed for a specific precipitation field and 2) are inclined to filter out high-frequency components to reduce noise in the information transmission, resulting in low resolution and blurred precipitation fields. The modeling challenges impose constraints on achieving high-resolution regional and accurate intensity forecasting. Here, to the former, we propose an individual motion-driven AI (IMD-AI) method, incorporating motion matching, alignment, and refinement through constructed pattern groups. It effectively solves the mismatch in motion estimation for an individual precipitation field, ensuring global and intact regional precipitation forecasting. To the latter, we put forward a schedule sampling, patch embedding, and adversarial strategies (SPA) training mechanism, which eliminates sharp and local information loss issues during the filtering of high-frequency components. Extensive experimental results demonstrate that IMD-AI achieves accurate motion estimation of individual precipitation fields and generates high-resolution regional and accurate intensity forecasting.