Extreme Weather Nowcasting via Local Precipitation Pattern Prediction

TL;DR

Extreme rainfall nowcasting is challenged by pronounced locality and the need for real-time inference. The authors introduce exPreCast, a deterministic transformer-based framework with a Cubic Dual Upsample (CDU) decoder and a Temporal Extractor (TE) to enable accurate, texture-preserving forecasts across short and extended horizons, complemented by a balanced KMA radar dataset. Across SEVIR, MeteoNet, and KMA, exPreCast achieves state-of-the-art or near-state-of-the-art performance while drastically reducing computational cost compared to diffusion-based ensembles. This combination of accuracy, efficiency, and robust generalization enables practical deployment for real-time warning systems and risk mitigation.

Abstract

Accurate forecasting of extreme weather events such as heavy rainfall or storms is critical for risk management and disaster mitigation. Although high-resolution radar observations have spurred extensive research on nowcasting models, precipitation nowcasting remains particularly challenging due to pronounced spatial locality, intricate fine-scale rainfall structures, and variability in forecasting horizons. While recent diffusion-based generative ensembles show promising results, they are computationally expensive and unsuitable for real-time applications. In contrast, deterministic models are computationally efficient but remain biased toward normal rainfall. Furthermore, the benchmark datasets commonly used in prior studies are themselves skewed--either dominated by ordinary rainfall events or restricted to extreme rainfall episodes--thereby hindering general applicability in real-world settings. In this paper, we propose exPreCast, an efficient deterministic framework for generating finely detailed radar forecasts, and introduce a newly constructed balanced radar dataset from the Korea Meteorological Administration (KMA), which encompasses both ordinary precipitation and extreme events. Our model integrates local spatiotemporal attention, a texture-preserving cubic dual upsampling decoder, and a temporal extractor to flexibly adjust forecasting horizons. Experiments on established benchmarks (SEVIR and MeteoNet) as well as on the balanced KMA dataset demonstrate that our approach achieves state-of-the-art performance, delivering accurate and reliable nowcasts across both normal and extreme rainfall regimes.

Figures (20)

• Figure 1: exPreCast Architecture: The encoder compresses input sequences, while the decoder uses the proposed Cubic Dual Upsample (CDU) blocks to reconstruct high-resolution precipitation patterns. The CDU block combines trilinear interpolation and 3D pixel shuffle to refine details and mitigate artifacts. Additionally, a dedicated Temporal Extractor (TE) module enables flexible forecasting across various lead times.
• Figure 2: Reconstruction results from different upsampling methods. CDU significantly mitigates checkerboard artifacts and preserves fine-grained radar textures, resulting in better reconstructions.
• Figure 3: Visualization of 1-hour predictions from different models on the KMA dataset.
• Figure 4: Visualization of 1-hour predictions from different models on the SEVIR dataset.
• Figure 5: Visualization of last-frame predictions from different models on the MeteoNet dataset.