<p>In recent years, the advancement in short-term precipitation forecasting has been significantly driven by the implementation of deep neural networks in processing radar echo data for predictive modeling. However, the current radar echo extrapolation algorithms generally suffer from the problems of fuzzy extrapolated echo, intensity recession, and inaccurate prediction of evolutionary trend. In order to alleviate the above problems, a radar echo extrapolation model (M2Net) based on multi-scale context fusion and multi-branch memory recall is proposed in this article. The network integrates a module for extracting multi-scale features, which adequately captures contextual features of different scales, enhances the echo detail texture and improves the echo visual quality. In addition, a multi-branch memory recall structure is designed, and a memory module for the long-term echo evolution law is introduced in three different branches to fully capture the spatiotemporal characteristics of precipitation and simulate the real echo evolution process. Comprehensive experimental results demonstrate that the developed model surpasses current advanced deep learning-based prediction models with respect to prediction precision and visual fidelity.</p>