Geng et al. (2026) Three-Dimensional Radar Echo Extrapolation Using a Physics-Constrained Deep Learning Model

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

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-01-08
• Authors: Liangchao Geng, Jinzhong Min, Huantong Geng, Xiaoran Zhuang
• DOI: 10.3390/rs18020206

Research Groups

Not explicitly stated in the provided text.

Short Summary

This paper proposes DIFF-3DRformer, a novel deep learning framework for 3D radar echo extrapolation, which unifies mesoscale and convective-scale networks with physical constraints. It significantly improves the nowcasting accuracy of severe convective storms by utilizing 19 vertical levels of radar data, outperforming conventional models.

Objective

• To develop a novel deep learning framework (DIFF-3DRformer) for 3D radar echo extrapolation to improve the nowcasting accuracy of severe convective storms, addressing limitations of existing models that use single-level radar data and lack physical constraints.

Study Configuration

• Spatial Scale: Mesoscale to convective-scale, focusing on severe storm events over Jiangsu, China.
• Temporal Scale: Short-term nowcasting of convective system evolution.

Methodology and Data

• Models used: DIFF-3DRformer, an end-to-end deep learning framework comprising:
  • A mesoscale evolution network embedded with 3D advection equation neural operators.
  • A 3D continuity equation-informed loss function.
  • A convective-scale denoising generative network based on a diffusion model.
• Data sources: 19 vertical levels of radar data and composite reflectivity.

Main Results

• DIFF-3DRformer demonstrates robust predictive skill across various convective scales for severe storm events.
• It outperforms NowcastNet, improving the comprehensive score by 44.8% for reflectivity thresholds greater than or equal to 35 dBZ.
• Utilizing 19 vertical levels of radar data as input enhances morphology and intensity prediction of convective echoes, boosting performance by 4.63% compared to using only composite reflectivity.
• The incorporation of physical constraints further refines the forecasted echo structure and spatial placement, yielding additional improvements.

Contributions

• Proposes DIFF-3DRformer, a novel deep learning framework that unifies mesoscale and convective-scale networks with embedded 3D advection equation neural operators and a 3D continuity equation-informed loss function.
• Introduces a diffusion model-based generative network for convective-scale denoising within an end-to-end architecture.
• Demonstrates significant improvement in severe convective storm nowcasting by directly characterizing the 3D structure of convective storms using multi-level radar data and physical constraints.
• Provides a promising solution for developing nowcasting methods that leverage the full 3D radar information.

Funding

Not explicitly stated in the provided text.

Citation

@article{Geng2026ThreeDimensional,
  author = {Geng, Liangchao and Min, Jinzhong and Geng, Huantong and Zhuang, Xiaoran},
  title = {Three-Dimensional Radar Echo Extrapolation Using a Physics-Constrained Deep Learning Model},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18020206},
  url = {https://doi.org/10.3390/rs18020206}
}