🤖 machine learning

Extreme Weather Nowcasting via Local Precipitation Pattern Prediction

This paper introduces exPreCast, an efficient deterministic framework enhanced with local spatiotemporal attention and a texture-preserving decoder, alongside a newly constructed balanced radar dataset from the Korea Meteorological Administration, to achieve state-of-the-art nowcasting performance across both ordinary and extreme rainfall events.

Original authors: Changhoon Song, Teng Yuan Chang, Youngjoon Hong

Published 2026-02-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Changhoon Song, Teng Yuan Chang, Youngjoon Hong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict exactly where a sudden, heavy downpour will hit your neighborhood in the next hour. This is the challenge of extreme weather nowcasting.

The paper introduces a new tool called exPreCast (Extreme Precipitation Nowcasting) designed to solve this problem. Here is how it works, explained through simple analogies:

The Problem: The "Blurry Camera" and the "Expensive Crystal Ball"

Currently, scientists use two main types of tools to predict rain, and both have flaws:

The "Blurry Camera" (Deterministic Models): These are fast and cheap, like a standard security camera. But they tend to smooth out the picture. If a tiny, intense storm cell is forming, these models often blur it out, making it look like a light drizzle instead of a dangerous flood. They are biased toward "normal" rain and miss the scary, extreme stuff.
The "Expensive Crystal Ball" (Diffusion Models): These are incredibly accurate and can see the tiny details, but they are like a crystal ball that requires a supercomputer to look into. They take too long to calculate, making them useless for real-time warnings when you need an answer now.

Also, most of the "training data" (the past weather records used to teach these AI models) is unbalanced. Some datasets are full of only storms, while others are full of only light rain. It's like trying to learn how to drive a car by only practicing on empty highways or only in a demolition derby; you won't be ready for the real mix of traffic.

The Solution: exPreCast

The authors built a new system that acts like a smart, high-speed photographer that is both fast and sharp.

1. The "Local Detective" (Video Swin Transformer)
Instead of trying to look at the entire sky at once (which is too much information), the model acts like a detective focusing on small neighborhoods. It uses a "shifted window" approach, zooming in on local patterns. Rain is a local phenomenon; a storm in one street doesn't necessarily mean it's raining on the next. By focusing locally, the model learns the specific "personality" of the rain in that specific area.

2. The "Texture Preserver" (The CDU Block)
This is the paper's secret sauce. When AI tries to make a small, blurry image bigger (upsampling), it usually does one of two things:

The Smoothie Maker (Interpolation): It blends the pixels together. Great for smooth skies, but it destroys the sharp edges of a storm.
The Pixel Shuffler: It rearranges pixels to keep edges, but often creates a weird "checkerboard" pattern that looks fake.

The authors created a Cubic Dual Upsample (CDU) block. Think of this as a two-person construction crew:

Worker A (Trilinear Interpolation): Builds the smooth, solid foundation of the building (the general shape of the rain).
Worker B (Pixel Shuffle): Adds the detailed brickwork and windows (the intense, high-frequency details of the storm).
By having them work together, the model gets a building that is both structurally sound and full of sharp, realistic details, without the weird checkerboard glitches.

3. The "Time Traveler" (Temporal Extractor)
Sometimes you need to know what will happen in 1 hour; other times, you need to know about 6 hours from now. The model has a special module called the Temporal Extractor (TE). Imagine this as a flexible lens on a camera. You can twist it to focus on the immediate future (short-term) or zoom out to see the longer trend (long-term) without needing to rebuild the whole camera. This makes the system flexible for different needs.

The New "Training Ground" (The KMA Dataset)

To teach this new model, the authors didn't just use the old, unbalanced datasets. They built a new one using data from the Korea Meteorological Administration (KMA).

Why it's special: South Korea has a unique climate with light rain in some seasons and massive typhoons in others. This dataset is like a balanced gym where the model trains on light jogging, heavy lifting, and everything in between. It ensures the AI learns to handle both a gentle breeze and a hurricane, making it a true generalist.

The Results: Fast, Sharp, and Reliable

When tested against other top models on three different datasets (including the new Korean one, plus US and French data), exPreCast won:

Accuracy: It predicted the location and intensity of extreme rain better than almost everyone else, especially when it came to those tiny, dangerous storm cells that other models blurred out.
Speed: It was 30 times faster (in terms of computer power needed) than the most accurate "expensive crystal ball" models.
Efficiency: It achieved these results with a fraction of the computing cost, making it practical for real-world use.

In a Nutshell

The paper presents exPreCast, a weather prediction tool that combines the speed of a standard camera with the sharp detail of a high-end microscope. By using a clever "two-worker" system to preserve rain textures and training on a perfectly balanced dataset, it can accurately predict dangerous, extreme rainfall events in real-time, offering a practical solution for saving lives and property.

Technical Summary: exPreCast for Extreme Weather Nowcasting

Problem Statement

Accurate forecasting of extreme weather events, such as heavy rainfall and storms, is critical for risk management and disaster mitigation. While high-resolution radar observations have spurred research into precipitation nowcasting, the task remains challenging due to pronounced spatial locality, intricate fine-scale rainfall structures, and variability in forecasting horizons.

Current approaches face significant trade-offs:

Deterministic models are computationally efficient but often exhibit bias toward normal rainfall, failing to capture extreme events.
Diffusion-based generative ensembles show promising results in capturing fine details but are computationally expensive, rendering them unsuitable for real-time applications.
Benchmark Datasets are often skewed. For instance, the SEVIR dataset is dominated by storm events, while MeteoNet consists mostly of normal rainfall. This lack of balance hinders the evaluation of a model's generalization capabilities across the full spectrum of precipitation intensities.

Methodology: exPreCast

The authors propose exPreCast, an efficient deterministic framework designed to generate finely detailed radar forecasts. The architecture is built upon a Video Swin Transformer backbone, extended with domain-specific modules to address the unique properties of weather forecasting.

Core Architectural Components

Encoder-Decoder 3D Swin Transformer:
- The model employs an encoder-decoder structure with skip connections to facilitate multi-scale feature flow.
- Unlike global attention mechanisms, the model utilizes shifted window attention to focus on local meteorological phenomena, restricting attention blocks to learn local patterns rather than global relations.
- The encoder partitions input radar volumes into non-overlapping 3D patches, processing them through multiple stages of 3D Swin Transformer blocks with patch merging for downsampling.
Cubic Dual Upsample (CDU) Block:
- Traditional upsampling methods (e.g., linear interpolation or standard pixel shuffle) often fail in this context: interpolation smooths out high-intensity small-scale patterns, while pixel shuffle can introduce checkerboard artifacts.
- The CDU block fuses trilinear interpolation and 3D pixel-shuffle in parallel branches.
  - The interpolation branch preserves channel information to capture smooth, low-frequency content.
  - The pixel-shuffle branch expands channels to reconstruct high-frequency details.
- These branches are concatenated and fused via a final 3D convolution. This design mitigates checkerboard artifacts while preserving fine-grained, high-intensity rainfall patterns essential for extreme event prediction.
Temporal Extractor (TE) Block:
- To address the need for flexible forecasting horizons (from short-term warnings to long-term preparedness), a TE block is appended after the decoder.
- Using a spatiotemporal 3D convolution, the TE block transforms the decoder's output temporal dimension to match the desired forecast lead time ( $T^*$ ).
- This allows the model to adapt to varying time horizons with minimal computational cost. The authors also demonstrate a transfer learning approach where the encoder is shared and frozen, while only the decoder and TE block are fine-tuned for long-term predictions.

Dataset Construction

The authors introduce a new, large-scale dataset from the Korea Meteorological Administration (KMA) covering 2014–2023. Unlike SEVIR (storm-biased) or MeteoNet (moderate-rain-biased), the KMA dataset offers a balanced distribution of rainfall intensities, ranging from ordinary precipitation to extreme events, facilitated by South Korea's distinct seasonal monsoons and typhoons.

Key Contributions

exPreCast Architecture: A deterministic, radar-based forecasting model that achieves state-of-the-art (SOTA) accuracy across KMA, SEVIR, and MeteoNet datasets while maintaining substantially lower computational costs compared to diffusion-based models.
Cubic Dual Upsample (CDU) Block: A novel upsampling mechanism that fuses trilinear interpolation with 3D pixel shuffle. It effectively mitigates checkerboard artifacts and retains small-scale, high-intensity rainfall patterns, consistently improving Critical Success Index (CSI) scores, particularly with pooling.
Balanced KMA Dataset: The construction of a new dataset encompassing both ordinary and extreme precipitation events, providing a more balanced meteorological coverage than existing benchmarks and enabling rigorous evaluation of model generalization.

Experimental Results

The model was evaluated on three datasets (KMA, SEVIR, MeteoNet) using metrics including Critical Success Index (CSI), Heidke Skill Score (HSS), and computational cost (GFLOPs/Parameters).

Performance on KMA: exPreCast achieved the highest HSS score and competitive CSI scores compared to baselines like UNet, ConvLSTM, and EarthFormer. While the computationally intensive CasCast (a diffusion-based model) slightly outperformed exPreCast in specific CSI-80 metrics with pooling, exPreCast required over 30 times lower inference GFLOPs and had 10 times fewer parameters.
Performance on SEVIR & MeteoNet: exPreCast demonstrated superior performance in preserving precipitation structures and achieving high CSI scores with pooling, indicating better local pattern fidelity. It outperformed most baselines and remained competitive with CasCast, which was noted to be computationally prohibitive (requiring >4,500 GFLOPs on SEVIR).
Long-term Prediction: On 6-hour prediction tasks, exPreCast (using the TE block and transfer learning) significantly outperformed baselines like ConvLSTM and SimVP, maintaining the ability to capture heavy rainfall events where other models failed.
Ablation Studies: Experiments confirmed that the CDU block outperforms standard upsampling methods (Pixel Shuffle, Trilinear Interpolation, Bicubic, Area) in both short-term and long-term scenarios. The TE block was shown to successfully adapt the model to various lead times without sacrificing short-term accuracy.

Significance and Claims

The paper claims that exPreCast offers a compelling solution to the accuracy-efficiency trade-off in precipitation nowcasting. By combining a deterministic framework with specialized architectural components (CDU and TE), the model delivers accurate and reliable nowcasts across both normal and extreme rainfall regimes.

The authors emphasize that while diffusion models offer high accuracy, their computational cost limits real-time applicability. In contrast, exPreCast provides a practical alternative that is significantly more efficient while maintaining SOTA performance. The inclusion of the balanced KMA dataset is highlighted as a crucial step toward evaluating generalization capabilities that existing benchmarks fail to provide.

The work concludes that exPreCast is well-suited for real-world forecasting systems, particularly for early warning systems where timely and accurate detection of localized extreme rainfall is vital for public safety. The authors acknowledge limitations regarding the quantification of predictive uncertainty and the challenges of long-term extreme event forecasting due to data scarcity, suggesting these as directions for future work.