SynCast: Synergizing Contradictions in Precipitation Nowcasting via Diffusion Sequential Preference Optimization

TL;DR

This work tackles the conflicting evaluation of precipitation nowcasting by jointly optimizing Critical Success Index ($CSI$) and False Alarm Ratio ($FAR$) using a two-stage Diffusion Sequential Preference Optimization (Diffusion-SPO) framework. It introduces SynCast, a fully probabilistic, pixel-space diffusion model whose predictions are refined through frame-level preference pairs and a two-phase optimization: first aligning $FAR$ with Diffusion-DPO, then improving $CSI$ under a constraint to preserve the $FAR$ gain. The approach demonstrates superior or competitive performance across three radar datasets (SEVIR, MeteoNet, HKO-7), with evidence of generalization across score-based and flow-matching diffusion variants and clear qualitative improvements via visualizations. The results suggest that frame-level preference signals and staged metric alignment enable more reliable, detail-preserving nowcasts, offering a practical path toward robust probabilistic weather forecasting with multiple, potentially conflicting objectives.

Abstract

Precipitation nowcasting based on radar echoes plays a crucial role in monitoring extreme weather and supporting disaster prevention. Although deep learning approaches have achieved significant progress, they still face notable limitations. For example, deterministic models tend to produce over-smoothed predictions, which struggle to capture extreme events and fine-scale precipitation patterns. Probabilistic generative models, due to their inherent randomness, often show fluctuating performance across different metrics and rarely achieve consistently optimal results. Furthermore, precipitation nowcasting is typically evaluated using multiple metrics, some of which are inherently conflicting. For instance, there is often a trade-off between the Critical Success Index (CSI) and the False Alarm Ratio (FAR), making it challenging for existing models to deliver forecasts that perform well on both metrics simultaneously. To address these challenges, we introduce preference optimization into precipitation nowcasting for the first time, motivated by the success of reinforcement learning from human feedback in large language models. Specifically, we propose SynCast, a method that employs the two-stage post-training framework of Diffusion Sequential Preference Optimization (Diffusion-SPO), to progressively align conflicting metrics and consistently achieve superior performance. In the first stage, the framework focuses on reducing FAR, training the model to effectively suppress false alarms. Building on this foundation, the second stage further optimizes CSI with constraints that preserve FAR alignment, thereby achieving synergistic improvements across these conflicting metrics.

Figures (7)

• Figure 1: Different precipitation nowcasting methods. Compared with existing deterministic and probabilistic approaches for precipitation nowcasting, SynCast leverages a probabilistic generative model combined with reinforcement learning fine-tuning to preserve fine-grained details while approaching the upper performance bounds of the conflicting metrics FAR and CSI.
• Figure 2: Trade-off between CSI and FAR. When CSI is used as the sole preference target, the model consistently improves CSI but at the cost of increased FAR, and vice versa.
• Figure 3: The workflow of SynCast. The training of SynCast consists of three steps. First, a UNet-based diffusion model is trained in pixel space, with conditional information injected via a CNN-based feature extractor to facilitate denoising. Next, we construct preference pairs for conflicting metrics. Specifically, by varying the initial noise, we generate $n$ predictions for each input, along with their average, resulting in $n+1$ candidates. For each predicted frame, FAR and CSI are computed. The best frame at each lead time is treated as a win sample, while the worst is regarded as a lose sample. Finally, reinforcement fine-tuning is performed. We first use Diffusion-DPO to align the model with respect to FAR, and then apply Diffusion-SPO to optimize CSI under the constraint of preserving the FAR alignment achieved in the previous stage.
• Figure 4: Visual comparison of a set of SEVIR precipitation events. Columns from left to right represent lead times from 10 to 180 minutes, and rows from top to bottom show Target, SimVP, TAU, PreDiff, DiffCast, and SynCast predictions. The yellow-marked region highlights excessive false alarms produced by PreDiff, while the red-marked region indicates a medium-intensity precipitation patch missed by DiffCast.
• Figure 5: Visual comparison of precipitation events: (a) MeteoNet and (b) HKO-7.