A U-Net-Based Minute-Level Precipitation Nowcasting Model with Multimodal Meteorological Variables over North China

In order to strengthen the application of deep learning in precipitation nowcasting over North China, a three-year dataset of 10-minute quantitative precipitation estimation (QPE) observations was employed to develop a minute-level nowcasting model based on the U-Net architecture. This model enables rolling precipitation forecasts with a 10-minute update interval for the next 0 to 2 hours. To evaluate its performance, we conducted verification using long-term series from June to September in 2020 and 2025, and further analyzed four heavy precipitation events that occurred on August 12, 2020, July 1, 2021, July 30, 2024, and August 27, 2025. A set of evaluation metrics—including threat score (TS), bias score (BIAS), probability of detection (POD), success ratio (SR), and false alarm rate (FAR)—was adopted for comprehensive assessment.

The results demonstrate that the U-Net model produces predictions close to observations, albeit with some degree of false alarms. Its overall forecasting performance is markedly superior to that of the optical flow method, persistent forecast, and the CMA-MESO numerical model. Specifically, when the minute-level precipitation intensity does not exceed 10 mm per 10 minutes, the U-Net model outperforms both the optical flow method and the persistent forecast. Similarly, for hourly precipitation not exceeding 25 mm·h⁻¹, the U-Net model shows better performance than the CMA-MESO model and the optical flow method. It should be noted, however, that heavy precipitation events with intensities exceeding these thresholds are relatively scarce in the training dataset, resulting in insufficient samples for the model to adequately capture such extreme patterns; thus, special attention is required when applying the model to heavy rainfall scenarios.

Despite these advantages, the U-Net model exhibits certain limitations, including forecast blurring and an excessively broad precipitation area. To mitigate these issues, we further introduced multimodal meteorological variables by incorporating 10-minute 850 hPa temperature fields and u/v wind components into the model input. Experimental results indicate that this multimodal approach effectively reduces the extent of false alarms, enhances positional accuracy, and yields more realistic precipitation intensity, ultimately leading to a further improvement in the TS score.