FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching

TL;DR

FlowCast introduces a full probabilistic precipitation nowcasting framework using Conditional Flow Matching in a latent space, enabling rapid, high-fidelity forecasts. By combining a frame-wise VAE with Independent CFM and a Cuboid-Attention U-Net, FlowCast achieves state-of-the-art probabilistic performance while requiring far fewer sampling steps than diffusion-based methods. Extensive experiments on SEVIR and ARSO demonstrate superior metrics, especially for extreme events, and ablations show CFM’s accuracy and efficiency advantages. The approach promises practical real-time ensemble nowcasting, with potential enhancements through multi-modal data fusion and broader dataset evaluation.

Abstract

Radar-based precipitation nowcasting, the task of forecasting short-term precipitation fields from previous radar images, is a critical problem for flood risk management and decision-making. While deep learning has substantially advanced this field, two challenges remain fundamental: the uncertainty of atmospheric dynamics and the efficient modeling of high-dimensional data. Diffusion models have shown strong promise by producing sharp, reliable forecasts, but their iterative sampling process is computationally prohibitive for time-critical applications. We introduce FlowCast, the first end-to-end probabilistic model leveraging Conditional Flow Matching (CFM) as a direct noise-to-data generative framework for precipitation nowcasting. Unlike hybrid approaches, FlowCast learns a direct noise-to-data mapping in a compressed latent space, enabling rapid, high-fidelity sample generation. Our experiments demonstrate that FlowCast establishes a new state-of-the-art in probabilistic performance while also exceeding deterministic baselines in predictive accuracy. A direct comparison further reveals the CFM objective is both more accurate and significantly more efficient than a diffusion objective on the same architecture, maintaining high performance with significantly fewer sampling steps. This work positions CFM as a powerful and practical alternative for high-dimensional spatiotemporal forecasting.

Figures (11)

• Figure 1: The FlowCast architecture. A U-Net with Cuboid Attention blocks processes latent spatiotemporal data. Conditioning on the flow time $t$ enables the model to learn the time-dependent vector field for generating forecasts.
• Figure 2: CSI-M and CSI at the 219 threshold per lead time on the SEVIR dataset. FlowCast shows consistent improvement over baselines for CSI-M and avoids the oversmoothing of deterministic models at longer lead times (CSI-219).
• Figure 3: CSI-M and CSI at the 39 DBz threshold per lead time on the ARSO dataset. FlowCast shows significant improvements over probabilistic baselines for earlier lead times, and over deterministic baselines for later lead times.
• Figure 4: Qualitative comparison of FlowCast with other baselines on a SEVIR sequence. Columns show lead times from 10 to 60 minutes. Rows show the ground-truth, followed by the models.
• Figure 5: Performance vs. efficiency trade-off. Forecast quality (CRPS $\downarrow$, CSI-M $\uparrow$) as a function of NFE. FlowCast (CFM) achieves near-optimal performance with only 3 to 10 steps, while the DDIM-based model requires 20 steps to 50 steps, and degrades sharply at low NFE.