Kim et al. (2025) Ensemble artificial neural network and generalized additive model for data-scarce regional frequency analysis in design flood estimation

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-31
• Authors: Ji-Eun Kim, Seoyoung Shin, Daeryong Park, Kichul Jung
• DOI: 10.1016/j.jhydrol.2025.134886

Research Groups

• Department of Civil, Environmental and Plant Engineering, Konkuk University, Seoul 05029, South Korea
• Water Management Division, Korea Environment Institute, Sejong 30147, South Korea

Short Summary

This study applied ensemble artificial neural networks (EANN) and generalized additive models (GAM) with canonical correlation analysis (CCA) for regional frequency analysis (RFA) to estimate design floods in data-scarce small streams in South Korea, finding that CCA-GAM outperformed EANN and river basin area was the most influential variable.

Objective

• To develop and evaluate regional frequency analysis (RFA) models using ensemble artificial neural networks (EANN) and generalized additive models (GAM) with canonical correlation analysis (CCA) for design flood estimation in data-scarce small streams in South Korea.
• To identify influential input variables for RFA models using Morris sensitivity analysis, thereby improving model efficiency and addressing the scarcity of such analyses.

Study Configuration

• Spatial Scale: 67 watersheds in South Korea, focusing on small streams.
• Temporal Scale: Design flood estimation for various return periods (implied by "design frequencies"), based on available hydrological data.

Methodology and Data

• Models used: Ensemble Artificial Neural Network (EANN), Generalized Additive Model (GAM), Canonical Correlation Analysis (CCA), Morris sensitivity analysis.
• Data sources: Observed regional hydrological data from 67 watersheds in South Korea, characterized by data-scarce conditions.

Main Results

• Both EANN and GAM models, enhanced with CCA, demonstrated satisfactory performance for design flood estimation, achieving high Nash–Sutcliffe efficiency (NSE) values.
• The CCA-GAM model consistently outperformed the EANN model, exhibiting lower error and bias in its predictions.
• Morris sensitivity analysis revealed that the river basin area was the most influential variable in estimating design floods.
• The mean river basin slope was identified as the least influential variable among the input factors.
• The study concluded that reliable flood discharge estimations for ungauged basins can be achieved using limited observational data through model training with observed regional data.

Contributions

• Development and comparative evaluation of EANN and GAM, integrated with CCA, for regional frequency analysis in data-scarce small stream environments, specifically in South Korea.
• Introduction and application of Morris sensitivity analysis to RFA, providing insights into the relative importance of input variables for design flood estimation and addressing a gap in existing literature.
• Demonstration of a robust methodology for reliable flood discharge estimation in ungauged basins using limited observational data, enhancing hydrological prediction capabilities in data-poor regions.

Funding

• The provided paper text does not contain information regarding specific funding projects, programs, or reference codes.

Citation

@article{Kim2025Ensemble,
  author = {Kim, Ji-Eun and Shin, Seoyoung and Park, Daeryong and Jung, Kichul},
  title = {Ensemble artificial neural network and generalized additive model for data-scarce regional frequency analysis in design flood estimation},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134886},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134886}
}