Changes in precipitation directly impact river runoff volume, subsequently influencing food production, and the security of downstream urban areas. In this study, we introduce a random velocity field (RVF) capable of performing multi-step predictions while providing interpretable insights into precipitation variations. The RVF leverages the gradient of a Gaussian random field to learn spatiotemporal velocity patterns and employs a predictive process to reduce dimensionality and enable multi-step forecasting. Bayesian parameter estimation is obtained using the Markov Chain Monte Carlo (MCMC) method. Our analysis reveals a noticeable shifting trend in annual precipitation based on diverse real datasets. This trend serves as a valuable foundation for further exploration of urban flood control and agricultural development strategies.
