Distribution selection in at-site flood frequency analysis with machine-learning support: family identifiability and design-quantile consequences
Flood frequency analysis (FFA) is sensitive to the assumed probability distribution, yet short records and overlap among plausible families often make at-site distribution selection uncertain. This study evaluates whether machine learning can support this task under a controlled Monte Carlo framework in which the true generating family is known. Six candidate distributions were assessed using supervised multiclass classification based on sample-derived descriptors combining L-moments, L-moment-ratio-diagram acceptance features, goodness-of-fit and information-criterion summaries, and Bayes-factor-calibrated transformations of goodness-of-fit p-values. Results show that distribution-family identifiability is strongly heterogeneous across families, sample sizes, and parameter regimes. A parsimonious feature set based on classical L-moment descriptors captured most of the recoverable discriminatory signal, whereas richer scenarios produced only modest average gains. Quantile-consequence analysis showed that the impact of imperfect family selection increases with return period and depends on the confusion channel. These findings support ML as a conditional decision-support tool for at-site FFA.
Authors
- Wilson dos Santos Fernandes (ORCID: https://orcid.org/0000-0002-9731-2320)
Institutions
- Universidade Federal de Minas Gerais (BR)
Publication Details
- Journal
- Hydrological Sciences Journal
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1080/02626667.2026.2767652
- Primary Topic
- Flood Risk Assessment and Management
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Fundação de Amparo à Pesquisa do Estado de Minas Gerais
- Humboldt-Universität zu Berlin