RainODE: Continuous-Time Precipitation Forecasting with Latent Neural ODEs

2026-06-29

AI Summary

Traditional precipitation forecasting struggles to simultaneously achieve high temporal resolution, continuity, and sharpness due to sparse observations and the high computational cost of discrete modeling. This work proposes a continuous-time precipitation prediction framework that models precipitation evolution as a Neural Ordinary Differential Equation (Neural ODE) in a latent space to capture large-scale advection dynamics, while incorporating a Brownian bridge stochastic source module to represent non-advective intensity variations. By uniquely integrating latent-space Neural ODEs with Brownian bridge stochastic processes, the method enables high-fidelity inference at arbitrary time points. Evaluated on the SEVIR benchmark and a newly introduced RAPID dataset, the approach significantly outperforms existing methods, producing forecasts that exhibit both strong temporal consistency and fine spatial detail.

Abstract

In precipitation forecasting, not only accuracy but also temporal resolution is critical. However, increasing temporal resolution is constrained by observational limitations and the computational cost of dense discrete modeling. To overcome this limitation, we reformulate precipitation forecasting as a continuous-time dynamical system and propose RainODE, a framework that models precipitation evolution in latent space using a Neural ODE. This formulation enables derivative-consistent temporal dynamics and captures the dominant large-scale advective motion of precipitation systems. Nevertheless, a purely deterministic ODE struggles to represent non-advective intensity changes such as localized growth, decay, and sub-grid variability, often leading to over-smoothed predictions. To address this issue, we introduce a stochastic source modeling module based on a Brownian Bridge formulation, which refines residual intensity variations and restores fine-grained structures while preserving advective consistency. By combining deterministic continuous dynamics with stochastic refinement, RainODE enables arbitrary-time inference while maintaining sharp predictions. Experiments on SEVIR and the newly introduced Radar-based Precipitation Integrated Dataset (RAPID) demonstrate consistent improvements across multiple temporal intervals and precipitation regimes. The code is available at https://github.com/SeongYE/RainODE.

Problem

• precipitation forecasting
• temporal resolution
• non-advective intensity changes
• over-smoothed predictions
• continuous-time modeling

Innovation

• Neural ODE
• continuous-time forecasting
• stochastic refinement
• Brownian Bridge
• latent dynamics