Rectifying Distribution Shift in Cascaded Precipitation Nowcasting

TL;DR

RectiCast tackles distribution shift in cascaded precipitation nowcasting by decoupling mean-shift rectification from local stochastic generation. It introduces a two-stage, Flow Matching–based framework with a Deterministic Predictive Backbone, a Rectifier to rectify mean shifts, and a Generator to model local stochasticity, all trained in a staged manner. On SEVIR and MeteoNet, RectiCast yields substantial gains over strong cascaded baselines, especially at longer lead times, while maintaining perceptual and structural fidelity. This approach demonstrates the value of explicitly separating mean-shift correction from stochastic detail synthesis in multi-stage weather forecasting and points to broader applicability in multi-stage generative forecasting.

Abstract

Precipitation nowcasting, which aims to provide high spatio-temporal resolution precipitation forecasts by leveraging current radar observations, is a core task in regional weather forecasting. Recently, the cascaded architecture has emerged as the mainstream paradigm for deep learning-based precipitation nowcasting. This paradigm involves a deterministic model to predict posterior mean, followed by a probabilistic model to generate local stochasticity. However, existing methods commonly overlook the conflation of the systematic distribution shift in deterministic predictions and the local stochasticity. As a result, the distribution shift of the deterministic component contaminates the predictions of the probabilistic component, leading to inaccuracies in precipitation patterns and intensity, particularly over longer lead times. To address this issue, we introduce RectiCast, a two-stage framework that explicitly decouples the rectification of mean-field shift from the generation of local stochasticity via a dual Flow Matching model. In the first stage, a deterministic model generates the posterior mean. In the second stage, we introduce a Rectifier to explicitly learn the distribution shift and produce a rectified mean. Subsequently, a Generator focuses on modeling the local stochasticity conditioned on the rectified mean. Experiments on two radar datasets demonstrate that RectiCast achieves significant performance improvements over existing state-of-the-art methods.

Figures (4)

• Figure 1: Illustration of the distribution shift in deterministic predictions and its impact on cascaded models. The deterministic forecast (Posterior $\mu$) suffers from temporal distribution shift, leading to progressively blurry prediction. Consequently, the probabilistic component conditioned on raw $\mu$ ($\mu$-based Prediction) fails to accurately capture local precipitation morphology and intensity, especially in the red boxes. In contrast, our Rectified Posterior $\mu$ corrects this shift, enabling our final forecast to deliver substantially more accurate intensity and patterns in the highlighted areas.
• Figure 2: The overall architecture of our proposed framework.
• Figure 3: CSI curves against lead time for cascaded models on the SEVIR dataset.
• Figure 4: A case study comparing forecasts from different cascaded models on the SEVIR dataset.