Sun et al. (2025) Multi-source precipitation product fusion strategy based on a novel ensemble validation framework

Identification

• Journal: Atmospheric Research
• Year: 2025
• Date: 2025-10-15
• Authors: Jianping Sun, Xungui Li, Qiyong Yang
• DOI: 10.1016/j.atmosres.2025.108563

Research Groups

• State Key Laboratory of Featured Metal Materials and Life-cycle Safety for Composite Structures, College of Civil Engineering and Architecture, Guangxi University, Nanning, China
• Guangxi Provincial Engineering Research Center of Water Security and Intelligent Control for Karst Region, College of Civil Engineering and Architecture, Guangxi University, Nanning, China
• Key Laboratory of Disaster Prevention and Structural Safety of Ministry of Education, College of Civil Engineering and Architecture, Guangxi University, Nanning, China

Short Summary

This study develops a novel Ensemble Validation Precipitation Framework (EVPF) using a CNN-LSTM deep learning architecture to address significant validation randomness in multi-source precipitation data fusion. The EVPF robustly fuses six precipitation products, eliminating the "validation set gambling" phenomenon and achieving high accuracy for precipitation estimation in the Yujiang River Basin, China.

Objective

• To develop a novel Ensemble Validation Precipitation Framework (EVPF) that systematically addresses the substantial impact of validation set selection on model performance assessment in multi-source precipitation data fusion.

Study Configuration

• Spatial Scale: Yujiang River Basin, China, utilizing data from 15 meteorological stations.
• Temporal Scale: Daily (implied by RMSE and MAE units of mm d⁻¹).

Methodology and Data

• Models used:
  • Ensemble Validation Precipitation Framework (EVPF)
  • CNN-LSTM deep learning architecture (baseline)
  • Complex attention mechanisms and Transformer architectures (for comparison)
• Data sources:
  • Six satellite precipitation products: CMORPH, ERA5, PERSIANN, GsMAP, IMERG, CHIRPS.
  • Ground-based observation data from 15 meteorological stations.

Main Results

• Traditional validation approaches exhibit enormous performance variations, with a maximum RMSE difference of 14.4 % between optimal and worst validation configurations.
• The EVPF, based on a CNN-LSTM architecture and employing cyclical training with different stations as validation sets, effectively eliminates validation set gambling and achieves robust performance with:
  • RMSE = 8.14 mm d⁻¹
  • MAE = 3.07 mm d⁻¹
  • CC = 0.78
  • R² = 0.60
  • POD = 0.88
  • FAR = 0.39
  • CSI = 0.57
  • ETS = 0.41
  • HSS = 0.58
• Significant geographical differentiation patterns exist: coastal stations show larger absolute errors but lower coefficients of variation, while inland mountainous stations exhibit smaller average errors but higher variability.
• Complex attention mechanisms and Transformer architectures provided limited marginal benefits over the baseline CNN-LSTM framework.
• Ablation experiments confirmed that all six precipitation products contribute meaningful information to the fusion process.

Contributions

• Introduces a novel Ensemble Validation Precipitation Framework (EVPF) that provides a robust methodological foundation for multi-source remote sensing data fusion.
• Systematically addresses and effectively eliminates the "validation set gambling" phenomenon, a significant issue in traditional precipitation data fusion validation.
• Offers a highly accurate and reliable precipitation estimation method with significant application value for regional hydrological forecasting, agricultural decision-making, and disaster risk assessment.

Funding

• Not specified in the provided text.

Citation

@article{Sun2025Multisource,
  author = {Sun, Jianping and Li, Xungui and Yang, Qiyong},
  title = {Multi-source precipitation product fusion strategy based on a novel ensemble validation framework},
  journal = {Atmospheric Research},
  year = {2025},
  doi = {10.1016/j.atmosres.2025.108563},
  url = {https://doi.org/10.1016/j.atmosres.2025.108563}
}