Kaur et al. (2026) Leveraging deep learning and kernel density representations for transformation-based regional flood frequency analysis

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-17
• Authors: Sukhsehaj Kaur, Sagar Rohidas Chavan
• DOI: 10.1016/j.jhydrol.2026.136434

Research Groups

• Department of Civil Engineering, Indian Institute of Technology Ropar, India

Short Summary

The study introduces a deep learning-based approach using kernel density representations to automate the selection of regional frequency distributions for flood frequency analysis, significantly reducing computational time while maintaining or improving estimation accuracy.

Objective

• To develop a computationally efficient and robust regional goodness-of-fit (GOF) test by reformulating it as a classification problem to identify the most suitable regional frequency distribution for design flood estimation at ungauged or sparsely gauged sites.

Study Configuration

• Spatial Scale: Regional (applied to catchments in the United States and peninsular India).
• Temporal Scale: Not specified (focused on flood frequency return periods).

Methodology and Data

• Models used: Deep Neural Network (DNN) classifier, Kernel Density Estimator (KDE), and Transformation-based Regional Flood Frequency Analysis (T_RFFA).
• Data sources: Peak flow records from gauged sites, Monte Carlo simulation experiments, and real-world catchment data from the US and India.

Main Results

• The DNN classifier achieved high classification accuracies of 95.09% for the training dataset and 94.86% for the test dataset.
• The proposed DL-based GOF test is approximately 18 to 27 times faster than the conventional L-moment-based approach.
• Design flood estimates produced by the DL-based approach are comparable or superior to those obtained via the conventional T_RFFA approach.

Contributions

• Reformulates the regional GOF test as a classification problem using KDE plots as inputs to a DNN, eliminating the need for sample mean-based normalization which often violates RFFA assumptions.
• Provides a computationally efficient alternative to traditional L-moment-based methods for selecting regional frequency distributions (GEV, GP, GL, GN, and Pearson Type 3).

Funding

• Not specified in the provided text.

Citation

@article{Kaur2026Leveraging,
  author = {Kaur, Sukhsehaj and Chavan, Sagar Rohidas},
  title = {Leveraging deep learning and kernel density representations for transformation-based regional flood frequency analysis},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136434},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136434}
}