RainOps: Multi-Source Availability-Aware Nowcasting with Residual Probabilistic Refinement

Precipitation nowcasting requires accurate, high-resolution forecasts while accounting for the uncertainty of rapidly evolving weather systems and heterogeneous observations. We introduce RainOps, a probabilistic nowcasting model that forecasts from multiple observation sources whose availability may vary over time. RainOps combines an availability-conditioned forecasting backbone with a lightweight stochastic residual refiner and post-hoc intensity calibration, enabling efficient ensemble generation from a shared forecast. On the SEVIR benchmark, RainOps achieves state-of-the-art performance across a broad range of precipitation nowcasting metrics, with particularly large improvements for localized and high-intensity precipitation. A single model remains effective when observations are missing, offering a practical approach to reliable, low-cost probabilistic nowcasting for operational use.