Lu et al. (2026) Application and Comparison of Two Transformer-Based Deep Learning Models in Short-Term Precipitation Nowcasting

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water
• Year: 2026
• Date: 2026-03-23
• Authors: Chuhan Lu, Qilong Pan
• DOI: 10.3390/w18060757

Research Groups

[Information not available in the provided text.]

Short Summary

This study systematically compares Earthformer and LLMDiff, two Transformer-based deep learning models, for short-term extreme precipitation nowcasting using the SEVIR dataset, finding Earthformer excels for rapid early warning of light precipitation at shorter lead times (0-30 minutes) while LLMDiff is better for high-accuracy nowcasting of heavy precipitation at longer lead times (up to 60 minutes).

Objective

• To systematically compare the performance of Earthformer and LLMDiff, two Transformer-based deep learning models, for short-term extreme precipitation nowcasting.

Study Configuration

• Spatial Scale: High spatiotemporal resolution (specific spatial extent not detailed in text), utilizing the SEVIR dataset.
• Temporal Scale: Short-term nowcasting, with lead times evaluated from 0 minutes to 60 minutes.

Methodology and Data

• Models used: Earthformer (Transformer-based with Cuboid Attention mechanism), LLMDiff (Transformer-based with diffusion-based probabilistic modeling and a frozen large language model module).
• Data sources: SEVIR dataset.

Main Results

• For 0–30 minute lead times, Earthformer more efficiently captures both local and long-range spatiotemporal dependencies and shows a slight advantage for low-intensity precipitation.
• As the lead time extends to 60 minutes, LLMDiff demonstrates stronger longer-horizon skill due to its diffusion-based probabilistic modeling and frozen large language model module, enhancing the representation of uncertainty and longer-term evolution of precipitation systems.
• LLMDiff tends to produce a higher false-alarm rate.
• Earthformer is better suited for rapid early warning of light precipitation.
• LLMDiff is more appropriate for high-accuracy nowcasting of heavy precipitation.
• Model performance was evaluated using the Critical Success Index (CSI), Probability of Detection (POD), and Success Ratio (SUCR).

Contributions

• Provides a systematic comparison of two advanced Transformer-based deep learning models (Earthformer and LLMDiff) for extreme precipitation nowcasting.
• Offers practical insights for selecting appropriate models based on lead time and precipitation intensity for intelligent extreme weather forecasting.

Funding

[Information not available in the provided text.]

Citation

@article{Lu2026Application,
  author = {Lu, Chuhan and Pan, Qilong},
  title = {Application and Comparison of Two Transformer-Based Deep Learning Models in Short-Term Precipitation Nowcasting},
  journal = {Water},
  year = {2026},
  doi = {10.3390/w18060757},
  url = {https://doi.org/10.3390/w18060757}
}