Zhong et al. (2026) FADiff: A Frequency-Aware Diffusion Model Based on Hybrid CNN–Transformer Network for Radar-Based Precipitation Nowcasting

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

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-04-02
• Authors: Jiandan Zhong, Wei Deng, Guanru Lyu, Jingbo Zhai, Yong Li, Yajuan Xue, Ziheng Yang
• DOI: 10.3390/rs18071061

Research Groups

[Not explicitly mentioned in the provided text.]

Short Summary

This paper proposes FADiff, a novel frequency-aware diffusion model based on a hybrid CNN–Transformer network, to address challenges in deep learning-based precipitation nowcasting such as blurry predictions and signal–noise confusion. FADiff significantly outperforms state-of-the-art methods, particularly in generating high-fidelity meteorological structures under high-intensity precipitation thresholds.

Objective

• To develop a deep learning model for radar-based precipitation nowcasting that overcomes limitations of existing models, specifically by mitigating blurry predictions, effectively modeling both local and long-range spatial dependencies, and resolving signal–noise confusion to improve structural fidelity.

Study Configuration

• Spatial Scale: Regional/local, based on radar data for precipitation nowcasting.
• Temporal Scale: Short-term (nowcasting), typically within a few hours, but specific lead times are not provided.

Methodology and Data

• Models used: FADiff (Frequency-Aware Diffusion model) incorporating a hybrid CNN–Transformer backbone, a Frequency-Aware Module (FAM) utilizing Discrete Cosine Transform (DCT) with a learnable gating mechanism, and a latent diffusion model.
• Data sources: CIKM and SEVIR datasets (radar-based precipitation data).

Main Results

• The proposed FADiff model outperforms state-of-the-art methods across a comprehensive suite of evaluation metrics for precipitation nowcasting.
• FADiff demonstrates remarkable robustness and stability, particularly under high-intensity precipitation thresholds.
• The model exhibits superior capability in generating meteorologically critical structures with high fidelity, addressing the issue of blurry predictions.

Contributions

• Introduction of FADiff, a novel frequency-aware diffusion model specifically designed for radar-based precipitation nowcasting.
• Development of a hybrid CNN–Transformer backbone that effectively integrates CNNs and Transformers for joint local and global feature extraction.
• Proposal of a novel Frequency-Aware Module (FAM) that leverages the Discrete Cosine Transform (DCT) and a learnable gating mechanism to mitigate signal–noise confusion by adaptively filtering noise-dominant frequency components while preserving high-frequency meteorological signals.
• Achievement of state-of-the-art performance, particularly in generating high-fidelity meteorological structures during high-intensity precipitation events.

Funding

[Not explicitly mentioned in the provided text.]

Citation

@article{Zhong2026FADiff,
  author = {Zhong, Jiandan and Deng, Wei and Lyu, Guanru and Zhai, Jingbo and Li, Yong and Xue, Yajuan and Yang, Ziheng},
  title = {FADiff: A Frequency-Aware Diffusion Model Based on Hybrid CNN–Transformer Network for Radar-Based Precipitation Nowcasting},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18071061},
  url = {https://doi.org/10.3390/rs18071061}
}