Tongtiegang et al. (2026) Relating precipitation extremes to streamflow peaks using percentile-based correspondence and interpretable machine learning

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-10-19
• Authors: Zhao Tongtiegang, Xiaodong Li
• DOI: 10.1016/j.jhydrol.2026.136647

Research Groups

• Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), School of Civil Engineering, Sun Yat-Sen University, Guangzhou, China
• State Key Laboratory of Water Resources Engineering and Management, Wuhan University, Wuhan, China

Short Summary

This paper quantifies the percentile-based correspondence between upper 90th percentile precipitation extremes and streamflow peaks, and diagnoses the influence of catchment attributes using interpretable machine learning. It reveals that median proportions of paired streamflow peaks exceeding the 80th, 90th, and 95th percentiles are approximately 70%, 57%, and 45%, respectively, with snow fraction being a dominant attribute, especially when interacting with mean elevation.

Objective

• To quantify the percentile-based correspondence between upper 90th percentile precipitation extremes and streamflow peaks above different percentile levels.
• To diagnose the effects of catchment attributes, including feature importance and interaction effects, on this correspondence using interpretable machine learning.

Study Configuration

• Spatial Scale: Continental United States, utilizing the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset.
• Temporal Scale: Large-sample study across multiple catchments, implicitly covering historical periods captured by the CAMELS dataset.

Methodology and Data

• Models used: Random forest models coupled with SHapley Additive exPlanations (SHAP) for feature importance and interaction effect analysis.
• Data sources: Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset, comprising hydro-climatic and catchment attribute data.

Main Results

• The median proportions of paired streamflow peaks exceeding the 80th, 90th, and 95th percentiles are approximately 70%, 57%, and 45%, respectively.
• For upper 90th percentile streamflow peaks, snow fraction is identified as the dominant catchment attribute, with its negative impacts significantly shaped by its interaction with mean elevation.
• In catchments with the highest 20% of snow fraction, median hit rates for streamflow peaks remain 0.53 below 500 m and 0.57 between 500 m and 1000 m, but sharply decrease to 0.14 above 1000 m.
• After accounting for snow fraction-elevation interactions, the five leading catchment attributes influencing the correspondence shift to mean high-flow duration, longitude, mean precipitation, streamflow elasticity, and silt fraction.

Contributions

• Provides a novel percentile-based correspondence metric (hit rate) to quantify the relationship between precipitation extremes and streamflow peaks.
• Utilizes interpretable machine learning (Random Forest with SHAP) to comprehensively diagnose the effects of catchment attributes, including complex interaction effects, on this correspondence.
• Identifies the critical role of snow fraction and its interaction with mean elevation in modulating the translation of precipitation extremes to streamflow peaks, offering specific quantitative thresholds.
• Offers insights valuable for improving streamflow forecasting and flood warning systems by informing the interpretation of catchment-specific precipitation extremes.

Funding

Not specified in the provided text.

Citation

@article{Tongtiegang2026Relating,
  author = {Tongtiegang, Zhao and Li, Xiaodong},
  title = {Relating precipitation extremes to streamflow peaks using percentile-based correspondence and interpretable machine learning},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136647},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136647}
}