Yun et al. (2025) Addressing class imbalance extends the performance frontier of classification–regression satellite-gauge precipitation fusion

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-31
• Authors: Zhaode Yun, Yintang Wang, Pan Liu, Lingjie Li, Yong Liu, Yang Liu
• DOI: 10.1016/j.jhydrol.2025.134898

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, Wuhan University, Wuhan, China
• Hubei Provincial Key Lab of Water System Science for Sponge City Construction, Wuhan University, Wuhan, China
• The National Key Laboratory of Water Disaster Prevention, Nanjing Hydraulic Research Institute, Nanjing, China
• Yangtze Institute for Conservation and Development, Nanjing, China

Short Summary

This study introduces ImbCRPF, a novel classification-regression framework for satellite-gauge precipitation fusion that explicitly addresses class imbalance, significantly improving the accuracy of precipitation estimates, particularly for heavy rainfall events.

Objective

• To develop a classification-regression satellite-gauge precipitation fusion framework that explicitly addresses sample imbalance to enhance the accuracy of precipitation estimates, especially for relatively heavy rainfall.

Study Configuration

• Spatial Scale: Chaohu and Taihu Lake basins, with output datasets at 0.01 degrees and 0.1 degrees spatial resolution.
• Temporal Scale: Focus on heavy rainfall events.

Methodology and Data

• Models used:
  • Classification stage: Random Forest (RF), K-Nearest Neighbors (KNN), combined with oversampling, undersampling, and hybrid sampling strategies to balance class distribution (rain/no rain, relatively heavy/normal precipitation).
  • Regression stage: Random Forest (RF) incorporating virtual precipitation samples generated through sampling, spatial autocorrelation variables, and auxiliary variables.
  • The final fused dataset, ImbCRPF, combines optimal classification and regression results.
• Data sources: Satellite-based precipitation products and rain gauge observations.

Main Results

• The proposed ImbCRPF framework significantly outperforms the mainstream RF-RF model in heavy rainfall events.
• For the 0.01° dataset:
  • F1 scores for no rain and relatively heavy rainfall improved by 0.28 and 0.13, respectively.
  • Kling-Gupta Efficiency (KGE) increased by 0.26.
• For the 0.1° dataset:
  • F1 scores for no rain and relatively heavy rainfall improved by 0.18 and 0.06, respectively.
  • Kling-Gupta Efficiency (KGE) increased by 0.14.

Contributions

• This framework is the first to explicitly address sample imbalance within a classification-regression precipitation fusion approach.
• It effectively extends the upper limits of estimation accuracy for satellite-gauge precipitation fusion, particularly for extreme precipitation events.

Funding

• Not specified in the provided text.

Citation

@article{Yun2025Addressing,
  author = {Yun, Zhaode and Wang, Yintang and Liu, Pan and Li, Lingjie and Liu, Yong and Liu, Yang},
  title = {Addressing class imbalance extends the performance frontier of classification–regression satellite-gauge precipitation fusion},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134898},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134898}
}