![[2511.17628] Rectifying Distribution Shift in Cascaded Precipitation Nowcasting](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2511.17628] Rectifying Distribution Shift in Cascaded Precipitation Nowcasting
Summary
This article presents RectiCast, a novel framework for improving precipitation nowcasting by addressing distribution shifts in deep learning models, enhancing forecast accuracy.
Why It Matters
Accurate precipitation forecasting is crucial for various sectors, including agriculture and disaster management. This research addresses a significant limitation in current models, potentially leading to better decision-making and resource allocation during weather events.
Key Takeaways
- RectiCast decouples mean-field shift rectification from local stochasticity generation.
- The framework improves accuracy in precipitation forecasts over longer lead times.
- Experiments show significant performance enhancements compared to existing methods.
Computer Science > Machine Learning arXiv:2511.17628 (cs) [Submitted on 19 Nov 2025 (v1), last revised 22 Feb 2026 (this version, v3)] Title:Rectifying Distribution Shift in Cascaded Precipitation Nowcasting Authors:Fanbo Ju, Haiyuan Shi, Qingjian Ni View a PDF of the paper titled Rectifying Distribution Shift in Cascaded Precipitation Nowcasting, by Fanbo Ju and 2 other authors View PDF HTML (experimental) 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 deter...
