Li et al. (2025) A super-resolution network based on dual aggregate transformer for climate downscaling

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-09-29
• Authors: Meng Li, Yijing Chen, Zhihui Song
• DOI: 10.1038/s41598-025-17234-4

Research Groups

• College of Statistics and Mathematics, Hebei University of Economics and Business, Shijiazhuang, China

Short Summary

This paper proposes a novel Climate Downscaling Dual Aggregation Transformer (CDDAT) model that integrates a lightweight CNN and a dual aggregation transformer with multimodal fusion to enhance high-resolution climate downscaling. The CDDAT achieves state-of-the-art performance in rainfall image restoration and dew point reconstruction by effectively capturing complex details and dynamically reassigning the importance of different rainfall variables.

Objective

• To develop a novel deep learning model, the Climate Downscaling Dual Aggregation Transformer (CDDAT), capable of generating high-resolution climate data, particularly rainfall images, by effectively capturing complex details and dynamically reassigning the importance of different rainfall variables through multimodal fusion.

Study Configuration

• Spatial Scale: Super-resolution downscaling by factors of 2 and 4 for radar rainfall images. Additional evaluation on urban climate data downscaling from 2.5 kilometers to 250 meters.
• Temporal Scale: Data collected over a period of 2014–2019 for 268 precipitation events; downscaling applied to instantaneous climate level images.

Methodology and Data

• Models used:
  • Proposed: Climate Downscaling Dual Aggregation Transformer (CDDAT), a hybrid model comprising a Lightweight CNN Backbone (LCB) with High Preservation Blocks (HPBs) and a Dual Aggregation Transformer Backbone (DATB) equipped with adaptive self-attention (Adaptive Spatial Self-Attention and Adaptive Channel Self-Attention), and a CNN-based multimodal fusion operation.
  • Compared with: Bicubic interpolation, SRCNN, SRGAN, SCNet, OSEDiff, and DAT (Dual Aggregation Transformer).
• Data sources:
  • NJU-CPOL dataset: C-band dual polarization weather radar data from Nanjing University (2014–2019), including horizontal reflectance factor (ZH), differential reflectance factor (ZDR), and differential phase factor (KDP).
  • Additional dataset: Urban climate information (dew point) at 2.5 kilometers (low-resolution) and 250 meters (high-resolution) resolutions, obtained by physical modeling paradigms.

Main Results

• CDDAT achieved state-of-the-art performance in climate downscaling tasks, demonstrating superior high texture restoration and detail preservation.
• For 2x super-resolution on the NJU-CPOL dataset, CDDAT achieved the highest Peak Signal-to-Noise Ratio (PSNR) of 38.02 dB and Structural Similarity Index (SSIM) of 0.9607.
• For 4x super-resolution on the NJU-CPOL dataset, CDDAT achieved the highest PSNR of 32.92 dB and SSIM of 0.8728.
• Visually, CDDAT effectively suppressed texture noise, preserved small rainfall information, and generated more detailed high-frequency information in heavy rainfall areas compared to other models, which often produced blurry images, artifacts, or lost small-scale information.
• Multimodal fusion of ZH, ZDR, and KDP indicators proved effective, with KDP specifically enhancing structural details and edge information of rainfall images.
• On an additional dataset for dew point reconstruction, CDDAT outperformed other methods with the lowest Mean Absolute Error (MAE) of 0.9659 ± 0.0187 and Root Mean Squared Error (RMSE) of 1.4102 ± 0.0263, generating more refined small-scale features.
• CDDAT demonstrated a good trade-off between performance and model complexity, with 33.9 million parameters and 100.76 Giga-FLOPs, showing improved performance over the baseline DAT model with only a slight increase in complexity.

Contributions

• Proposed CDDAT, a novel dual aggregation Transformer model, for high-resolution climate prediction, marking a first attempt to adopt the Transformer model in the super-resolution domain for climate downscaling.
• Introduced a CNN-based multimodal fusion method to integrate multiple radar metrics (ZH, ZDR, KDP), enabling the model to mine dynamic structural information of convective precipitation more effectively than single-variable approaches.
• Conducted extensive comparative experiments on the NJU-CPOL dataset and an additional urban climate dataset, verifying the effectiveness and superiority of CDDAT in accurately recovering echo edges and details of severe convective weather and dew point data.

Funding

• Science Research Project of Hebei Education Department (Grant ZD2021319)
• Hebei University of Economics and Business (Grant 2024ZD10)

Citation

@article{Li2025superresolution,
  author = {Li, Meng and Chen, Yijing and Song, Zhihui},
  title = {A super-resolution network based on dual aggregate transformer for climate downscaling},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-17234-4},
  url = {https://doi.org/10.1038/s41598-025-17234-4}
}