Fan et al. (2025) Applications of Attention‐Enhanced CNN Models to Regional Precipitation Downscaling

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

Identification

• Journal: Earth and Space Science
• Year: 2025
• Date: 2025-11-25
• Authors: Lei Fan, Xiaoning Xie, Cailing Wang, Jianing Guo, Heng Liu, Xiyue Mao, Zhengguo Shi
• DOI: 10.1029/2025ea004465

Research Groups

[Information not available in the provided abstract.]

Short Summary

This study evaluates three attention-enhanced Convolutional Neural Networks (CNNs) for regional precipitation downscaling in the Middle Reaches of the Yellow River, China. The findings demonstrate that these models significantly improve spatio-temporal precipitation simulations and better capture extreme precipitation events compared to conventional CNNs, with the AttLap model showing the most notable improvements.

Objective

• To evaluate and compare the performance of three attention-enhanced Convolutional Neural Networks (AttLap, ACMix, and MAN) against a conventional CNN for regional precipitation downscaling, focusing on spatio-temporal simulation accuracy and extreme precipitation representation.

Study Configuration

• Spatial Scale: Regional (Middle Reaches of the Yellow River in China)
• Temporal Scale: Daily, monthly, and annual timescales

Methodology and Data

• Models used: Attention-based Laplacian Pyramid Network (AttLap), Attention and Convolutional Mix Network (ACMix), Multi-scale Attention Network (MAN), and a conventional Convolutional Neural Network (CNN).
• Data sources: ERA5 atmospheric variables, Global Precipitation Measurement (GPM) data.

Main Results

• All attention-enhanced CNNs improved spatio-temporal precipitation simulations across daily, monthly, and annual timescales compared to the conventional CNN model.
• The AttLap model showed the greatest improvements, reducing the root-mean-square error of daily precipitation by 10.1% and increasing the correlation coefficient by 16.7% for the regional mean.
• The attention mechanism enhanced the models' ability to simulate extreme precipitation, with the 95th and 99th percentiles of predicted precipitation showing closer agreement with GPM data.
• The probability density function for daily precipitation in attention-enhanced CNNs exhibited better agreement with GPM data, particularly for heavy precipitation.

Contributions

• Demonstrates the significant potential of attention-enhanced Convolutional Neural Networks for improving regional precipitation downscaling accuracy.
• Highlights the specific advantage of attention mechanisms in capturing spatio-temporal precipitation features and accurately simulating extreme precipitation events.
• Provides valuable tools for climate projection and water resource management, particularly in complex terrain regions.

Funding

[Information not available in the provided abstract.]

Citation

@article{Fan2025Applications,
  author = {Fan, Lei and Xie, Xiaoning and Wang, Cailing and Guo, Jianing and Liu, Heng and Mao, Xiyue and Shi, Zhengguo},
  title = {Applications of Attention‐Enhanced CNN Models to Regional Precipitation Downscaling},
  journal = {Earth and Space Science},
  year = {2025},
  doi = {10.1029/2025ea004465},
  url = {https://doi.org/10.1029/2025ea004465}
}