DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting

Penghui Wen, Mengwei He, Patrick Filippi, Na Zhao, Feng Zhang, Thomas Francis Bishop, Zhiyong Wang, Kun Hu · 2024

Read DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting by Penghui Wen, Mengwei He, Patrick Filippi, Na Zhao, Feng Zhang, Thomas Francis Bishop, Zhiyong Wang, Kun Hu (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-01091.txt.

Introduction

Penghui Wen and seven co-authors built a new tool for very short-term rain forecasts in 2024. Rain forecasts for the next few hours help farms, travel, and flood response. Past deep learning tools often kept the broad storm shape but lost fine detail. Their system, DuoCast, splits the task into two parts.

One part models slow, large-scale rain trends with a network tuned to weather fronts. The other part sharpens fine detail with a self-attention design. The authors prove on paper that this split cuts forecast error. Tests on four radar data sets beat strong prior tools. DuoCast held both the broad shape and the fine detail of storms as they moved and changed over time.

Catalog record

CREATORPenghui Wen, Mengwei He, Patrick Filippi, Na Zhao, Feng Zhang, Thomas Francis Bishop, Zhiyong Wang, Kun Hu
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-01091

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting Penghui Wen, Mengwei He, Patrick Filippi, Na Zhao, Feng Zhang, Thomas Francis Bishop, Zhiyong Wang, Kun Hu Abstract Accurate short-term precipitation forecasting is critical for weather-sensitive decision-making in agriculture, transportation, and disaster response. Existing deep learning approaches often struggle to balance global structural consistency with local detail preservation, especially under complex meteorological conditions. We propose DuoCast , a dual-diffusion framework that decomposes precipitation forecasting into low- and high-frequency components modeled in orthogonal latent subspaces. We theoretically prove that this frequency decomposition reduces prediction error compared to conventional single branch U-Net diffusion models. In DuoCast , the low-frequency model captures large-scale trends via convolutional encoders conditioned on weather front dynamics, while the high-frequency model refines fine-scale variability using a self-attention-based architecture. Experiments on four benchmark radar datasets show that DuoCast outperforms state-of-the-art baselines, achieving superior accuracy in both spatial detail and temporal evolution. 1 School of Computer Science, The University of Sydney, Sydney, Australia 2 School of Life and Environmental Sciences, The University of Sydney, Sydney, Australia 3 School of Mathematics, Shanghai University of Finance and Economics, Shanghai, China 4 Department of Atmospheric and Oceanic Sciences, Institutes of Atmospheric Sciences, Fudan University, Shanghai, China 5 School of Science, Edith Cowan University, Perth, Australia {penghui.wen, mengwei.he, patrick.filippi}@sydney.edu.au, [email protected], [email protected], {thomas.bishop, zhiyong.wang}@sydney.edu.au, [email protected] Code — https://github.com/ph-w2000/DuoCast Appendix — https://arxiv.org/abs/2412.01091 Introduction Precipitation plays a vital role in weather systems, influencing temperature, humidity, and overall atmospheric conditions. Precipitation nowcasting, which predicts short-term rainfall using recent radar observations, is increasingly important for applications such as agriculture, transportation, and disaster response ( Zhang et al. 2023 ) . Its importance was underscored during the July 2025 Texas flash floods: National Weather Service (NWS) observed localized totals over 5 inches and rainfall rates of 2–3 inches per hour at many locations, yet forecasts slightly mispredicted the heaviest cells, resulting in flash flooding in central Texas and trapping hundreds of people ( Moore et al. 2025 ) . Conventional approaches to nowcasting rely on numerical weather prediction (NWP), which simulates atmospheric dynamics by solving partial differential equations (PDEs) derived from physical laws ( Bauer et al. 2015 ; Lorenc 1986 ; Alley et al. 2019 ) . However, the complexity of NWP renders them computationally expensive and inefficient for real-time forecasting ( Bi et al. 2023 ) . Figure 1: Challenges in precipitation nowcasting: a) precipitation with weather (warm) front patterns, and b) high-frequency features (micro-scale variability) within edge regions shown in pink parallelograms. Recent advances in data-driven deep learning techniques have shown greater potential than traditional NWP methods in precipitation prediction by leveraging large datasets and avoiding the need to solve complex physical equations. Deterministic models ( Shi et al. 2017 ; Gao et al. 2022 ; Yan et al. 2024 ; Lin et al. 2025 ) aim to capture the overall motion of precipitation systems by generating mean-value forecasts. However, they often produce blurry predictions and lack fine-grained spatial detail ( Ravuri et al. 2021 ) . In contrast, probabilistic models ( Ravuri et al. 2021 ; Zhang et al. 2023 ; Gao et al. 2024 ) introduce stochastic latent variables to better capture local-scale variability. While effective at modeling uncertainty, treating the entire system as fully stochastic can introduce excessive randomness and degrade forecasting precision ( Yu et al. 2024 ) . To address this, hybrid approaches such as DiffCast ( Yu et al. 2024 ) combine global deterministic forecasting with local probabilistic refinement. Nevertheless, its CNN-based diffusion model struggles to preserve high-frequency information due to limited capacity ( Yang et al. 2023 ) , resulting in less spatially accurate precipitation intensity, as illustrated in Fig. 1 (b). Therefore, we introduce DuoCast , a dual-diffusion framework for precipitation nowcasting. DuoCast decomposes the forecasting task into two frequency-specific subspaces, where separate stochastic diffusion models capture low- and high-frequency components. We theoretically show that frequency decomposition reduces prediction error compared to standard U-Net-based diffusion with formal proof. Specifically, the low-frequency diffusion model captures large-scale precipitation trends via convolutions, conditioned on weather front representations extracted from historical radar data. These features integrate spatial modeling of air masses and temporal modeling of front evolution, with intensity-aware enhancement for structural regions. The high-frequency model operates in latent space using a self-attention backbone to model fine-scale variability, conditioned on both historical inputs and the low-frequency forecast. Evaluated on four benchmark radar datasets, DuoCast achieves state-of-the-art performance and yields substantial improvements in precipitation nowcasting accuracy. Our work’s key contributions are summarized as follows: • We propose DuoCast , a dual-diffusion framework that decomposes precipitation forecasting into low- and high-frequency subspaces for frequency-aware modeling. • We provide a theoretical analysis showing that frequency decomposition reduces prediction error compared to standard single-branch diffusion models. • Extensive experiments on four benchmark radar datasets demonstrate that DuoCast achieves state-of-the-art performance in short-term precipitation nowcasting. Related Work Precipitation Nowcasting Precipitation nowcasting has traditionally relied on numerical weather prediction (NWP), which solves atmospheric PDEs ( Skamarock et al. 2008 ) . Although effective, NWP is computationally intensive, often requiring hours on supercomputers ( Bi et al. 2023 ) . Recent deep learning models ( Andrychowicz et al. 2023 ) offer competitive performance with significantly lower latency. They are generally categorized as either deterministic or probabilistic. Deterministic Precipitation Models aim to predict the mean-based evolution of precipitation by modeling spatial and temporal dynamics. ConvLSTM ( Shi et al. 2015 ) captures these dynamics via convolutional and recurrent structures, while PhyDNet ( Guen and Thome 2020 ) introduces a physics-guided decomposition into stochastic and deterministic motion. FACL ( Yan et al. 2024 ) replaces standard L2 loss with a loss tailored for signal-based spatiotemporal modeling. AlphaPre ( Lin et al. 2025 ) disentangles phase and amplitude to separately model motion and intensity. Probabilistic Precipitation Models use latent variables to capture the inherent uncertainty of future weather, enabling finer modeling of micro-scale phenomena. Representative methods include DGMR ( Ravuri et al. 2021 ) , which employs a generative adversarial network (GAN) with spatial and temporal discriminators, and NowcastNet ( Zhang et al. 2023 ) , which integrates physical priors into GAN-based prediction. A diffusion model generates images by progressively denoising random noise into coherent visuals ( Wen et al. 2024 ; Lu et al. 2024 ) . Building on this, Prediff ( Gao et al. 2024 ) incorporates knowledge control to guide the sampling process. Although such approaches enhance forecast diversity, their fully stochastic nature can also introduce excessive variability, undermining predictive reliability. Hybrid Precipitation Models combine deterministic backbones for global precipitation trends with probabilistic components to capture local variability, aiming to balance accuracy and diversity. For example, CasCast ( Gong et al. 2024 ) and DiffCast ( Yu et al. 2024 ) integrate trajectory forecasting with stochastic refinement. However, their limited treatment of frequency-specific structures and micro- and macro-patterns may hinder long-range performance. Frequency Analysis and Modeling Frequency-domain analysis has shown strong effectiveness in image analysis and generation ( Tan et al. 2024 ; Woo and others 2022 ; Wen et al. 2023 ; Chen et al. 2025 ) , serving as a valuable complement to spatial-based approaches, which often struggle with artifacts. For instance, ( Doloriel and Cheung 2024 ) apply frequency-based masking to

Read the complete text