Long-term Radar Echo Extrapolation Method Based on Optical Flow and Deep Learning Hybrid Model

Radar echo extrapolation plays a crucial role in nowcasting. With the advancement of deep learning techniques in recent years, numerous advanced models for radar echo extrapolation have been introduced, resulting in great improvement in results and performance metrics. However, the extrapolation results of many existing LSTM-based networks tend to become increasingly blurred over time, losing precipitation details and failing to meet the requirements for refined forecasting. The complexity of network models also results in high memory consumption, which limits the size of input data and sequence length. Therefore, this paper proposes a radar echo extrapolation model that integrates the optical flow method with deep learning to address these issues. Within the 0-1 hour range, radar echoes, as spatiotemporal sequences, allow neural networks to extract more detailed features, thus enhancing performance. For the 1-2 hour range, the optical flow method, which consumes less memory, is used to avoid excessively blurred extrapolation results. Experimental results show that this hybrid model outperforms single models in the long-term radar echo extrapolation task within 0-2 hours.