Extreme Weather Nowcasting via Local Precipitation Pattern Prediction

AI summary

Overview

Research area: Machine learning for precipitation nowcasting — radar-based spatiotemporal forecasting, with emphasis on extreme rainfall events.

Technical level: Advanced. The paper assumes familiarity with transformer architectures (Video Swin Transformer), upsampling techniques (pixel shuffle, trilinear interpolation), attention mechanisms, and meteorological verification metrics (CSI, HSS).

One-sentence scope: The authors propose exPreCast, a deterministic radar nowcasting architecture built on a 3D Swin Transformer backbone with a Cubic Dual Upsample (CDU) decoder and a Temporal Extractor (TE), evaluated on SEVIR, MeteoNet, and a newly constructed balanced KMA dataset spanning 2014–2023.

Availability: Code is publicly available at https://github.com/tony890048/exPreCast. The paper is licensed CC BY 4.0.

What This Paper Is About

Precipitation nowcasting is hard because rainfall is spatially local, has fine-scale structure, and must be predicted over varying time horizons — and extreme events are exactly the cases that matter most but appear least often in data. Existing diffusion-based generative ensembles perform well but are too computationally expensive for real time, while deterministic models are cheap but biased toward ordinary rainfall; additionally, the standard benchmarks are skewed (SEVIR toward storms, MeteoNet toward normal rain), so generalization across the full intensity spectrum is hard to measure. The authors build a deterministic model that aims to preserve sharp, small-scale high-intensity rainfall patterns at low computational cost, and introduce a balanced KMA radar dataset covering both ordinary and extreme precipitation to test that generalization.

Key Contributions

1. exPreCast architecture: A deterministic radar-based forecasting model that reports state-of-the-art accuracy across KMA, SEVIR, and MeteoNet while maintaining substantially lower computational cost than the diffusion-based baselines it is compared against.

2. The Cubic Dual Upsample (CDU) block: A decoder upsampling module that fuses trilinear interpolation with 3D pixel shuffle to mitigate checkerboard artifacts and retain small-scale high-intensity rainfall patterns. The authors report that CDU consistently improves CSI scores, especially with pooling.

3. A new balanced benchmark dataset: A large-scale KMA radar dataset spanning 2014 to 2023 at 10-minute intervals, containing both ordinary and extreme precipitation, offering more balanced meteorological coverage than SEVIR or MeteoNet.

4. The Temporal Extractor (TE) block: A module appended after the decoder that adjusts the temporal dimension to the desired forecast lead time, allowing the same encoder to be shared between short-term and long-term prediction via transfer learning.

Main Findings

• KMA short-term (1-hour) results: exPreCast reaches CSI-M of 0.2794 / 0.3721 / 0.4841 at pooling sizes 1 / 4 / 16, CSI-20 of 0.1786 / 0.3013 / 0.4490, CSI-80 of 0.0348 / 0.0835 / 0.1488, and the highest HSS of 0.4042 on this balanced dataset. CasCast obtains CSI-M of 0.2573 / 0.3742 / 0.4837, CSI-20 of 0.1702 / 0.3066 / 0.4444, CSI-80 of 0.0346 / 0.0926 / 0.1695, and HSS 0.3806.

• Efficiency gap on KMA: exPreCast uses 32.0M parameters and 55 GFLOPs, versus 391.0M parameters and 1,729 GFLOPs for CasCast — described as over 30 times lower inference GFLOPs and 10 times fewer parameters. CasCast is slightly better than exPreCast on CSI-80 with pooling.

• SEVIR results: exPreCast records CSI-M of 0.4179 / 0.4527 / 0.5427 (pools 1 / 4 / 16), CSI-181 of 0.2568 / 0.2989 / 0.4104, CSI-219 of 0.1468 / 0.1859 / 0.2910, and HSS 0.5430, with 32.0M parameters and 208 GFLOPs. The paper states that for the extreme thresholds, exPreCast's scores with pooling 4 and 16 are only exceeded by CasCast (392.9M parameters, 4,567 GFLOPs). EarthFormer reaches HSS 0.5411.

• MeteoNet results: exPreCast reaches CSI-M of 0.2861 / 0.3452 / 0.4446, CSI-40 of 0.1500 / 0.2145 / 0.3584, CSI-47 of 0.0856 / 0.1393 / 0.2525, and HSS 0.4116, with 32.0M parameters and 199 GFLOPs. CasCast was excluded from this experiment because the authors report its results were highly unstable — consistently noisy and inaccurate even after converging under a fair training setup.

• 6-hour long-horizon forecasting on KMA: exPreCast (trained from scratch, 33.6M parameters, 229 GFLOPs) reaches CSI-M of 0.0879 / 0.1271 / 0.2262, CSI-20 of 0.0323 / 0.0646 / 0.1542, CSI-80 of 0.0042 / 0.0130 / 0.0415, and HSS 0.1486. The fine-tuned variant exPreCast† (25.5M parameters) improves on this with CSI-M of 0.1214 / 0.1572 / 0.2289, CSI-20 of 0.0510 / 0.0931 / 0.1708, CSI-80 of 0.0116 / 0.0282 / 0.0642, and HSS 0.1987. The authors note AlphaPre* was trained on a half-scaled dataset due to computational limits.

• CDU ablation: Compared against pixel shuffle (PS), trilinear interpolation (TI), bicubic (BIC), and area (AREA) interpolation, CDU gives the best overall CSI. CDU achieves CSI-M of 0.2794 / 0.3721 / 0.4841 and CSI-M at the final 1-hour frame of 0.1963 / 0.2772 / 0.4053; the next closest at pool 16 is BIC with 0.4814 for CSI-M. The authors report that other upsampling methods are unable to deliver reliable performance at the final frame.

• TE ablation: Fine-tuned variants labeled TE-6f through TE-36f maintain accuracy as the horizon extends. TE-36f records CSI-M (pool 16) of 0.3923 at 1h, 0.2510 at 2h, 0.1773 at 3h, 0.1347 at 4h, 0.1075 at 5h, and 0.0923 at 6h, showing the TE block extends lead times without sacrificing short-term accuracy.

• Qualitative comparison: The authors report that SimVP, EarthFormer, and AlphaPre produce blurry predictions that fail to identify heavy rainfall, and that CasCast preserves more detail but still misses some heavy rainfall regions, whereas exPreCast produces the sharpest predictions and identifies the majority of rainfall events.

Methodology in Plain English

The model treats radar rainfall as a video: a sequence of past frames goes in, future frames come out. The backbone is a Video Swin Transformer, adapted to 3D radar volumes, which uses shifted-window attention so that the network learns local spatial correlations rather than global ones — motivated by the fact that rainfall is governed by local meteorological phenomena. The network is an encoder-decoder with skip connections: the encoder partitions the input into non-overlapping 3D patches and downsamples through several stages with 3D Swin Transformer blocks and Patch Merging; a bottleneck of two 3D Swin Transformer blocks sits in the middle; the decoder mirrors the encoder using 3D Swin Transformer blocks and CDU blocks.

The CDU block addresses a specific failure mode. Simple interpolation smooths out small, high-intensity rainfall regions — which a generic video model might treat as noise but which matter for extreme events — while pixel shuffling preserves high-frequency detail but introduces unrealistic checkerboard artifacts. CDU runs two parallel branches on the same input: a trilinear interpolation branch that preserves channels, and a 3D pixel-shuffle branch that expands channels to reconstruct high-frequency detail, each using channel-mixing 3D convolutions and a PReLU activation. The two branch outputs are concatenated and fused by a final 3D convolution.

The Temporal Extractor sits after the decoder and reshapes the temporal dimension to whatever forecast horizon is needed using a spatiotemporal 3D convolution that slides over the height, width, and channel dimensions. It is designed to pair with the CDU's temporal scaling: small temporal scale factors for short-term prediction, larger ones for long-term prediction. Because both short-term and long-term tasks use the same past observations, the encoder is shared — the authors train on a short-term task, freeze the encoder, and fine-tune only the decoder and TE block for long-term forecasting.

Training uses the AdamW optimizer with a warm-up scheduler and the Fourier Amplitude Correlation Loss. Evaluation uses CSI (including CSI-p at intensity thresholds and averaged CSI-M), reported with and without spatial pooling, plus HSS, parameter counts, and GFLOPs.

Data specifics:

• KMA: Hybrid Surface Rainfall composite radar images, natively generated at 5-minute intervals covering 1152 km × 1440 km at 500 m resolution; the authors collected all images from 2014 to 2023 at 10-minute intervals (excluding timestamps with no data) and downscaled to 4 km by uniform subsampling, then cropped a central 1024 × 1024 km square yielding 256 × 256 inputs. Split: 2014–2021 train, 2022 validation, 2023 test. Input is 7 frames (one hour of past observations); targets are 6 future frames (1 hour) or 36 frames (6 hours). Reflectivity is converted to mm/h using the Marshall-Palmer Z-R relation Z = 200R^1.6, with thresholds [1, 4, 8, 10, 20, 40, 80].
• SEVIR: 384 × 384 km region, 5-minute intervals, up to 4 hours, used at 1 km resolution VIL; input 13 frames, predict 12 frames; thresholds [16, 74, 133, 160, 181, 219], with reported columns at 181 and 219; values in [0, 256).
• MeteoNet: Radar data from Météo-France spanning January 2016 to October 2018 at 5-minute resolution; the authors use the south-eastern region and a 416 × 416 km top-left corner; input 12 frames, predict 12 frames; thresholds 19, 28, 35, 40, 47 dBZ, with 40 and 47 as the two most extreme.

Why This Matters

Impact on research: The paper argues that benchmark skew is a real methodological problem — SEVIR is dominated by storm events while MeteoNet consists mostly of normal rains — and offers a balanced alternative spanning ten years. It also positions deterministic architectures as a credible alternative to expensive diffusion ensembles: the reported comparison on KMA shows comparable CSI with over 30 times lower inference GFLOPs and 10 times fewer parameters than CasCast, and evaluation is performed at the original resolution of SEVIR and MeteoNet rather than downscaled.

Real-world applications:

• Flash flood and landslide early warning: The paper frames accurate extreme rainfall nowcasting as critical for risk management and disaster mitigation, citing floods and landslides as consequences of intense precipitation.
• Operational real-time nowcasting: The low inference cost (55 GFLOPs on KMA for the 1-hour model) makes real-time deployment plausible where diffusion ensembles are impractical.
• Multi-horizon forecast products: Because the TE block adapts to lead times from 1 hour to 6 hours, a single encoder can serve both immediate-warning and longer-lead preparedness use cases.
• Regional monitoring in monsoon and typhoon zones: The KMA dataset's balanced intensity distribution reflects South Korea's seasonal monsoons and typhoons, providing a testbed for regions with mixed ordinary and extreme rainfall regimes.

Industry relevance: The efficiency-versus-accuracy tradeoff is the central practical claim — meteorological agencies, disaster-response organizations, and forecasting service providers that need high-frequency radar updates are constrained by compute budgets, and a model that preserves small-scale extremes at a fraction of the cost is directly relevant to operational deployment. The authors also note that a deterministic model cannot quantify predictive uncertainty, which matters for early warning systems where confidence estimates inform decisions.

Future Directions

1. Long-term extreme rainfall forecasting: The authors state that strong performance on long-term extreme rainfall events remains challenging, which they attribute to both the scarcity of extreme events and the fundamentally ill-posed nature of forecasting from local radar observations alone, and suggest reliable extrapolation or interpolation to compensate for data scarcity as a future direction.

2. Uncertainty quantification: As a deterministic method, exPreCast cannot quantify predictive uncertainty; the authors propose hybrid approaches that integrate probabilistic methods to improve applicability in early warning systems.

3. Hybrid deterministic-generative systems: The paper describes exPreCast as a potential practical complement to real-world forecasting systems rather than a full replacement, suggesting integration paths with ensemble or generative approaches.

4. Broader generalization testing: The KMA dataset spans over 10 years and is presented as a highly generalized benchmark; extending evaluation to other regions with different precipitation regimes would test whether the balanced-coverage argument holds beyond South Korea.

Target Audience

This paper is most useful to machine learning researchers working on spatiotemporal forecasting and video prediction architectures; meteorological and atmospheric science researchers interested in radar-based nowcasting; and practitioners building operational forecasting systems who need to weigh accuracy against inference cost. Readers should be comfortable with transformer attention mechanisms, upsampling operations, and verification metrics like CSI and HSS; the paper's core architectural ideas (dual-branch upsampling, temporal extraction from a shared encoder) are accessible at an intermediate level, but the evaluation setup and baseline comparisons assume advanced background.

Authors’ 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.