FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching

AI summary

Overview

• Research area: Machine learning for weather forecasting — specifically probabilistic radar-based precipitation nowcasting using generative models (Conditional Flow Matching vs. diffusion).
• Technical level: Advanced. The paper assumes familiarity with diffusion models, continuous normalizing flows, latent generative modeling, VAEs, and standard nowcasting evaluation metrics (CSI, HSS, FSS, FAR, CRPS).
• Scope: The paper proposes FlowCast, a latent Conditional Flow Matching model for generating ensemble precipitation forecasts from past radar frames, and benchmarks it against deterministic and probabilistic baselines on two radar datasets.

What This Paper Is About

Precipitation nowcasting means predicting the next hour or so of rainfall from recent radar images, which matters for flood warnings and water management. Deep learning models produce better forecasts than traditional advection methods, but deterministic models trained with pixel-wise losses such as MSE generate blurry, over-smoothed forecasts, while diffusion models generate sharp ensembles at the cost of hundreds of iterative model evaluations per forecast. FlowCast asks whether Conditional Flow Matching — which learns a direct, near-straight noise-to-data mapping — can deliver diffusion-level (or better) forecast quality with far fewer sampling steps.

Key Contributions

1. FlowCast, described by the authors as the first fully probabilistic, end-to-end application of Conditional Flow Matching as a direct noise-to-data generative model for precipitation nowcasting, operating in a compressed latent space rather than as a refinement module on top of a deterministic forecast.
2. State-of-the-art claims on two radar datasets — the benchmark SEVIR dataset and the local ARSO dataset — reported as improvements in both probabilistic performance (CRPS) and predictive accuracy (CSI-M, FSS-P16-M, HSS-M).
3. A controlled ablation on the same architecture comparing the CFM objective against a diffusion (DDPM/DDIM) objective, which the authors report is both more accurate and substantially more efficient, holding quality at far fewer sampling steps.
4. A design argument for CFM's inductive bias in this domain: a straight-line ODE prior imposes the simplest possible transport between noise and the multi-modal but temporally consistent radar reflectivity distribution, whereas diffusion's stochastic denoising or curved probability-flow ODEs require more steps to resolve fine structure.

Main Findings

• Overall SEVIR performance: FlowCast records the lowest CRPS (0.0182), highest CSI-M (0.460), highest FSS-P16-M (0.767), and highest HSS-M (0.580) among the compared models. CasCast achieves a higher CSI-P16-M (0.520 vs. FlowCast's 0.506), but with a higher FAR-M (0.383 vs. FlowCast's 0.325), which the authors interpret as FlowCast striking a better balance between detection sensitivity and precision.
• SEVIR at long lead time (+65 min): FlowCast reaches CSI-M 0.324 and CSI-219 0.057, the highest of the compared models on those columns.
• Overall ARSO performance: FlowCast reports the lowest CRPS (0.0209), highest CSI-M (0.420), highest FSS-P16-M (0.738), and highest HSS-M (0.535). The authors state it outperforms all baselines in every metric except FAR-M.
• ARSO at long lead time (+65 min): FlowCast reaches CSI-M 0.315 and CSI-39 0.073.
• Extreme event detection on SEVIR (threshold 219): FlowCast achieves CSI 0.202, versus 0.137 for the best deterministic baseline (SimVP) — a reported improvement of over 47% — and 0.195 for the leading probabilistic baseline (CasCast). At threshold 181, FlowCast's CSI is 0.301 and HSS 0.443.
• Extreme event detection on ARSO (threshold 39 dBZ): FlowCast reports the highest CSI (0.183) and HSS (0.291), while maintaining a lower FAR than CasCast at all extreme thresholds. At threshold 36 dBZ, FlowCast's CSI is 0.254 and HSS 0.388.
• CFM vs. diffusion objective (ablation): With a single sampling step, FlowCast (CFM) reports CRPS 0.0207, CSI-M 0.454, and a runtime of 2.6 s per sequence — better on those metrics than a 100-step DDIM sampler (CRPS 0.0208, CSI-M 0.398, 239 s), which the authors describe as almost 100 times faster. At 10 steps, CFM reports CRPS 0.0168 and CSI-M 0.455 versus DDIM's CRPS 0.0262 and CSI-M 0.395 at the same 10 steps (both 24 s).
• Sampling efficiency: Each number of function evaluations (NFE) adds approximately 2.4 s per 8-member ensemble forecast. FlowCast approaches near-optimal CRPS and CSI-M in 3 to 10 steps, while the DDIM-based model requires 20 to 50 steps and degrades sharply below 10 NFE.
• ODE solver choice: An ablation over Euler (1 and 10 steps), Dormand-Prince 5, Midpoint (10 steps), Runge-Kutta 4 (10 steps), and Adaptive Heun showed no significant performance differences, so Euler with 10 steps was selected for computational efficiency.
• Qualitative behavior: The authors report that FlowCast produces sharp, perceptually realistic forecasts, avoiding the smoothness of deterministic models, with more realistic precipitation patterns than CasCast especially at longer lead times.
• Stability: The authors note that I-CFM training converged robustly with a simple regression loss and did not need the complex loss-weighting schedules that diffusion objectives often require.

Methodology in Plain English

The task is framed as video prediction: given 13 past radar frames (65 minutes), predict the next 12 frames (1 hour).

FloodCast works in two stages:

1. Compress the frames. A Variational Autoencoder is trained separately for each dataset to squeeze each radar frame into a small latent representation. The VAE uses an L1 reconstruction loss, a KL-divergence regularizer with weight 1e-4, and a PatchGAN adversarial loss with hinge loss (adversarial weight 0.5) to keep images perceptually sharp. Its weights are then frozen. SEVIR latents are 48×48×4 spatial-channel maps; ARSO latents are 38×52×4 (ARSO's native 301×401 grid is replication-padded to 304×416 to be divisible by the VAE's downsampling factor of 8, then cropped back before evaluation).

2. Generate in latent space. A U-Net built from Earthformer's Cuboid Attention blocks is trained to predict a vector field. During training, a noisy latent is formed by linearly interpolating between random Gaussian noise and the true future latent, with a small added Gaussian perturbation (σ = 0.01, the "independent CFM" setting that thickens the training trajectory for regularization). The network is conditioned on the interpolation time t, injected as an embedding at every U-Net level, and must regress the difference between the true future latent and the noise. At inference, the learned ODE is solved with the Euler method over 10 steps starting from fresh Gaussian noise, repeated eight times to build an ensemble, and decoded back to radar-frame space.

Training specifics: 200 epochs, AdamW with learning rate 5e-4 and cosine scheduling, exponential moving average of weights with decay 0.999, checkpoint selection by validation CSI-M, global batch size 12, on 4 NVIDIA H100 GPUs for 7 days. A separate DDPM with 1000 timesteps paired with a DDIM sampler serves as the same-architecture diffusion comparison. All ablations were run on the first 10% of the SEVIR test set on a single NVIDIA A100.

Evaluation follows the SEVIR Nowcasting Challenge protocol: threshold-based categorical metrics (FAR, CSI, HSS, plus max-pooled CSI and FSS over 16×16 km neighborhoods), a mean-across-thresholds summary (denoted "-M"), and CRPS for probabilistic skill. Ensemble size is N = 8, and categorical scores are computed on the ensemble mean.

Why This Matters

Operational nowcasting is time-critical: a forecast that takes minutes of compute to produce may already be too late for an evacuation or a flash-flood warning. Diffusion models are the current standard for sharp probabilistic nowcasting, but their iterative sampling is expensive. FlowCast's claim is that a nearly straight-line noise-to-data mapping retains the sharpness and ensemble calibration of diffusion while collapsing sampling to a handful of steps — with the reported 1-step CFM result being both more accurate and far faster than a 100-step DDIM sampler.

Real-world applications:

• Flood warning and risk management, where issuing timely alerts depends on rapid ensemble generation — the motivation the paper opens with.
• Water resource management, including reservoir and drainage operations that need short-term precipitation outlooks.
• Operational meteorological deployment, since the ARSO experiments target a national weather service setting (Slovenia) rather than only a large US benchmark.
• Ensemble forecasting under compute constraints, where the reported 3-to-10-step regime and roughly 2.4 s per NFE per 8-member ensemble make probabilistic forecasting feasible with limited hardware.

Industry relevance: the result is relevant to weather-service vendors, hydrology and flood-risk platforms, and any group running diffusion-based spatiotemporal forecasting who wants a cheaper sampler without retraining the backbone — the ablation isolates the training objective on an identical architecture.

Future Directions

• Multi-modal data fusion. The authors explicitly identify the reliance on radar data alone as a limitation and propose incorporating satellite observations and numerical weather prediction (NWP) inputs.
• Broader generalization testing. Evaluation was limited to SEVIR and ARSO because of computational cost; the authors call for a wider study across more meteorological regimes to confirm generalizability.
• Scaling and refinement of the sampling regime. The reported behavior — near-optimal CRPS and CSI-M at 3 to 10 steps, with quality still improving between 1 and 10 steps — leaves open how low the step budget can go before degradation, and whether the 10-step Euler setting is optimal across datasets and lead times.
• Public release of the ARSO dataset. The authors state they are actively collaborating with the Slovenian Environment Agency to facilitate public release, which would let others test the benchmark.

Target Audience

Machine learning researchers working on generative models for spatiotemporal and scientific data; meteorological and hydrology researchers evaluating probabilistic nowcasting systems; and practitioners building operational forecasting pipelines who care about the performance-versus-compute trade-off. Readers without a background in flow matching or diffusion models will need to consult the cited references (Lipman et al., Tong et al., Ho et al.) first, since the method section relies on their notation.

Authors’ 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.