Extreme Weather Nowcasting via Local Precipitation Pattern Prediction

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.

💡 Research Summary

The paper addresses two major shortcomings in radar‑based precipitation nowcasting: the bias of existing benchmark datasets toward either normal or extreme rainfall, and the prohibitive computational cost of recent diffusion‑based generative ensembles. To remedy the first issue, the authors construct a large‑scale balanced dataset from the Korea Meteorological Administration (KMA), covering a ten‑year period (2014‑2023) with 10‑minute resolution and a uniform distribution of rainfall intensities ranging from light drizzle to heavy storm events.

The proposed model, exPreCast, is a deterministic architecture built on a 3‑D Video Swin Transformer backbone. It employs local window attention with shifted windows to capture spatial‑temporal correlations that are inherently local in precipitation processes. The decoder introduces a novel Cubic Dual Upsample (CDU) block that fuses trilinear interpolation and 3‑D pixel‑shuffle upsampling, followed by a convolutional fusion layer. This design preserves high‑frequency texture and eliminates the smoothing or checkerboard artifacts typical of standard upsampling methods, which is crucial for accurately reproducing small, high‑intensity rain cells.

A Temporal Extractor (TE) module sits after the decoder and reshapes the temporal dimension via a 3‑D convolution, enabling flexible forecasting horizons from a few minutes up to several hours without retraining the entire network. Training uses AdamW with a warm‑up scheduler and the Fourier Amplitude Correlation loss, encouraging spectral consistency between predictions and ground‑truth radar fields.

Extensive experiments on SEVIR, MeteoNet, and the newly released KMA dataset show that exPreCast achieves state‑of‑the‑art performance while consuming roughly one‑third of the FLOPs of comparable diffusion models. Critical Success Index (CSI) scores across multiple thresholds (CSI‑p) are consistently higher, especially in the extreme‑rainfall regime (>30 mm h⁻¹) where the model improves CSI‑p by 6‑8 % over prior art. The model also demonstrates minimal performance degradation when the same parameters are used for short‑term (10‑30 min) and long‑term (up to 3 h) forecasts, confirming the effectiveness of the TE block.

In summary, the work contributes (1) a balanced, publicly available KMA radar dataset that enables rigorous evaluation across the full spectrum of precipitation intensities, and (2) the exPreCast framework, which combines local spatio‑temporal attention, texture‑preserving dual upsampling, and flexible temporal extraction to deliver accurate, real‑time nowcasts for both ordinary and extreme weather events. The approach promises practical deployment in operational warning systems and offers a solid foundation for future research on climate‑driven extreme precipitation forecasting.