A Hybrid MIM Model for Radar Echo Forecasting With Multi‐Scale Feature Extraction and Spatiotemporal Interaction

Radar echo maps are essential for precipitation forecasting, providing visual representations of rainfall patterns, including spatial distribution and intensity. To enhance radar echo prediction, this study introduces the MSIM–MIM model, which integrates the MFEF and SIM modules within the MIM framework. The MFEF module utilizes dilated convolutions to capture multi‐scale features while maintaining spatial details, improving contextual understanding, and boosting prediction accuracy, all without increasing computational cost. The SIM module employs a gating mechanism to selectively extract and process spatiotemporal context, thereby enhancing the model's ability to represent these patterns. This results in more refined state representations, allowing the MSIM–MIM model to retain and leverage context more effectively, thus reducing prediction errors. Experimental results demonstrate that MSIM–MIM outperforms other spatiotemporal models, achieving lower MSE and MAE in radar echo predictions across multiple datasets.