Cao et al. (2025) Enhancing Machine Learning Models for Nowcasting and Short‐Term Forecasting of Precipitation With a Novel Probability‐Matching Loss Function

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

Identification

• Journal: Geophysical Research Letters
• Year: 2025
• Date: 2025-12-16
• Authors: Yuan Cao, Shuaiyi Li, Jie Feng, Lei Chen, Hao Li, Yijun Zhang
• DOI: 10.1029/2025gl119442

Research Groups

Not specified in the abstract.

Short Summary

This paper proposes a novel probability-matching (PM) based loss function for machine learning precipitation forecasting to overcome the limitations of mean squared error (MSE) loss, demonstrating improved skill, reduced bias across rainfall intensities, and better preservation of small-scale variability.

Objective

• To propose and evaluate a novel probability-matching (PM) based loss function for machine learning precipitation nowcasting and short-term forecasting to address the limitations of traditional MSE-based losses, particularly regarding overly smoothed predictions and systematic underestimation of heavy rainfall frequency.

Study Configuration

• Spatial Scale: Regional to local, focusing on small-scale precipitation variability.
• Temporal Scale: Nowcasting and short-term forecasting (e.g., minutes to hours).

Methodology and Data

• Models used: Machine learning (ML) precipitation forecasting models, optimized using a novel probability-matching (PM) based loss function, compared against classical loss functions (e.g., Mean Squared Error).
• Data sources: Not specified, but implies observed precipitation data for comparison and evaluation.

Main Results

• The PM-based loss offers more balanced and consistent performance across various metrics, exhibiting lower forecast bias from light to heavy rainfall intensities.
• Spectral power analysis indicates that the PM-based loss better preserves small-scale precipitation variability throughout the forecast period.
• The PM-based loss results in a forecast frequency distribution of precipitation that more closely aligns with the observed distribution.
• These findings demonstrate consistent improvements in the predictive skill and reliability of ML precipitation forecasts when trained with the PM-based loss.

Contributions

• Introduction of a novel probability-matching (PM) based loss function specifically designed to mitigate the "double penalty" effect in machine learning precipitation forecasting.
• Demonstration of improved predictive skill, reduced bias across rainfall intensities, and better preservation of small-scale precipitation variability compared to classical loss functions like MSE.
• Enhancement of the reliability of ML precipitation forecasts by producing frequency distributions that more closely match observations.

Funding

Not specified in the abstract.

Citation

@article{Cao2025Enhancing,
  author = {Cao, Yuan and Li, Shuaiyi and Feng, Jie and Chen, Lei and Li, Hao and Zhang, Yijun},
  title = {Enhancing Machine Learning Models for Nowcasting and Short‐Term Forecasting of Precipitation With a Novel Probability‐Matching Loss Function},
  journal = {Geophysical Research Letters},
  year = {2025},
  doi = {10.1029/2025gl119442},
  url = {https://doi.org/10.1029/2025gl119442}
}