Li et al. (2025) Introducing an explainable neural network framework for nonstationary flood frequency analysis

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-06
• Authors: Wenbin Li, Lihua Xiong, Yanlai Zhou, Mingze Li, Rongrong Li, Chong‐Yu Xu
• DOI: 10.1016/j.jhydrol.2025.134729

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, Wuhan University, China
• School of Hydraulic and Ocean Engineering, Changsha University of Science & Technology, China
• Department of Geosciences, University of Oslo, Norway

Short Summary

This study introduces an explainable neural network framework (XNN-NFFA) for nonstationary flood frequency analysis, integrating feedforward neural networks with SHAP to accurately estimate flood distributions and interpret the influence of environmental drivers. The framework demonstrates superior performance over traditional models and identifies key drivers of flood nonstationarity in the upper Yangtze River Basin.

Objective

• To propose an explainable neural network framework (XNN-NFFA) for nonstationary flood frequency analysis (NFFA).
• To accurately estimate nonstationary flood frequency distributions (NFFDs) under changing environmental conditions.
• To understand and interpret how changing environments and various covariates influence flood behavior and contribute to flood nonstationarity.

Study Configuration

• Spatial Scale: Four representative hydrological stations in the upper Yangtze River Basin.
• Temporal Scale: 1951–2019 (69 years of data).

Methodology and Data

• Models used:
  • eXplainable Neural Network framework for Nonstationary Flood Frequency Analysis (XNN-NFFA)
  • Feedforward Neural Networks (FNNs) optimized through Generalized Extreme Value (GEV) distribution likelihood
  • SHapley Additive exPlanations (SHAP) for interpretability
  • Generalized Additive Models for Location, Scale, and Shape (GAMLSS) (for comparison)
• Data sources:
  • Historical flood observations (1951–2019)
  • Covariates: Warm-season precipitation, temperature, reservoir operation data, and large-scale atmospheric circulation indices.

Main Results

• Warm-season precipitation, temperature, and reservoir operation are consistently identified as the dominant drivers of flood nonstationarity.
• Large-scale atmospheric circulation indices exhibit more complex and station-specific influences on flood behavior.
• The FNN-based nonstationary flood frequency distributions (NFFDs) demonstrate strong performance, reducing estimation deviance by over 10% in training and approximately 3% in validation compared to GAMLSS models.
• SHAP analysis quantifies the relative importance of each covariate and characterizes their effect forms, revealing three distinct patterns for large-scale circulation indices: consistent monotonic, non-monotonic, and pronounced only under extreme values.
• FNN-based estimates of nonstationary design floods are generally lower and more reasonable than those derived from stationary assumptions and GAMLSS models.

Contributions

• Introduces a novel, explainable neural network framework (XNN-NFFA) that integrates FNNs with SHAP for comprehensive and interpretable nonstationary flood frequency analysis.
• Provides a robust methodology for accurately estimating nonstationary flood distributions, outperforming existing GAMLSS models in both training and validation.
• Offers detailed insights into the complex contributions of various environmental drivers (e.g., precipitation, temperature, reservoir operation, large-scale atmospheric circulation indices) to flood nonstationarity.
• Characterizes the diverse forms of covariate effects on flood distributions, enhancing understanding beyond simple importance rankings.
• Delivers practical implications for flood risk management by providing more reasonable nonstationary design flood estimates.

Funding

Not specified in the provided text.

Citation

@article{Li2025Introducing,
  author = {Li, Wenbin and Xiong, Lihua and Zhou, Yanlai and Li, Mingze and Li, Rongrong and Xu, Chong‐Yu},
  title = {Introducing an explainable neural network framework for nonstationary flood frequency analysis},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134729},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134729}
}