Short-term rainfall prediction based on radar echo using an efficient spatio-temporal recurrent unit

Accurate short-term precipitation prediction is critical for agricultural production, transportation safety, and water resource management. In this paper, we propose an Efficient Spatio-Temporal Recurrent Unit (ESTRU) for short-term precipitation prediction based on radar echoes. The ability of the model to process spatio-temporal information is enhanced by fusing two ConvGRU units while controlling the complexity. The trajectory tracking structure (TTS) facilitates the capture of rotational and scaling motions and improves the model's adaptability in complex meteorological conditions. The combined effect of the Self-Attention (SA) mechanism and convolution allows the model to focus on both global and local dependencies in spatial information, improving the clarity of the generated images. ESTRU demonstrated the best performance on the radar echo dataset compared to the other nine classical models. Quantitative and qualitative results show that ESTRU can efficiently model complex spatio-temporal relationships in radar echoes.