Machine learning surrogates for flash flood hazard mapping: CSI-optimized feature selection and spatial transferability
Abstract
Flood Hazard Mapping is constrained by the computational cost of high-resolution hydrodynamic simulation. This study evaluates three Machine Learning (ML) models -Random Forest (RF), eXtreme Gradient Boosting (XGB), and a Feed-Forward Neural Network -as surrogates, addressing two gaps: retrieval of deterministic flood hazard extents from probabilistic maps, and transferability to unseen catchments. Using a 5metre FLOODOS 2D simulation of the 1,000-year flood in southeastern France, models were trained on the Argens basin and tested on the Gapeau basin. A performancedriven feature selection algorithm -a greedy forward search maximizing the Critical Success Index (CSI) reduced 21 to four factors: discharge, Height Above Nearest Drainage, Distance to Stream, and river slope. Within the training domain, the RF achieved CSI = 0.76 (AUROC = 0.99). During extrapolation, all models maintained AUROC = 0.96, but the CSI declined to ~0.61, with XGB most robust. Notably, high threshold-independent discrimination (AUROC) did not guarantee accurate retrieval of the deterministic flood extent (CSI). Sub-catchment analysis revealed a global-local paradox: aggregate CSI masks systematic failures in upstream tributaries. Covariate shift was the principal barrier to spatial generalization; although the transferability assessment remains preliminary and basin-specific, results point to hybrid physics-informed ML as promising for operational flash-flood hazard assessment.
Related
- A machine learning model for early season wildfire risk in Ghana’s Guinea savannah assessing discrimination, operational prevalence, spatial transferability, and CMIP6 climate scenarios
- Machine learning-based temperature mapping in Sydney: Accuracy, transferability, and scalability
- Optimized hybrid machine learning for interpretable prediction of crack-healing percentage in self-healing concrete
- A novel integrated workflow for deep geothermal reservoir characterization using core data, optimized coupled petrophysical–seismic modeling, and machine learning: A case study from the SW Paris Basin, France
- Zero‐Shot Transferability of Deep Learning Models for Precipitation Super‐Resolution Across Diverse Climate Regimes
BibTeX
@article{arora2026machine,
title = {{Machine learning surrogates for flash flood hazard mapping: CSI-optimized feature selection and spatial transferability}},
author = {Aman Arora and Pierre Nicolle and Olivier Payrastre},
journal = {Geomatics, Natural Hazards and Risk},
year = {2026},
month = sep,
doi = {10.1080/19475705.2026.2707816},
url = {https://hal.science/hal-05778088v1},
note = {CC BY-NC 4.0},
}
RIS
TY - JOUR
TI - Machine learning surrogates for flash flood hazard mapping: CSI-optimized feature selection and spatial transferability
AU - Aman Arora
AU - Pierre Nicolle
AU - Olivier Payrastre
PY - 2026
DA - 2026/09/30
JO - Geomatics, Natural Hazards and Risk
DO - 10.1080/19475705.2026.2707816
UR - https://hal.science/hal-05778088v1
AB - Flood Hazard Mapping is constrained by the computational cost of high-resolution hydrodynamic simulation. This study evaluates three Machine Learning (ML) models -Random Forest (RF), eXtreme Gradient Boosting (XGB), and a Feed-Forward Neural Network -as surrogates, addressing two gaps: retrieval of deterministic flood hazard extents from probabilistic maps, and transferability to unseen catchments. Using a 5metre FLOODOS 2D simulation of the 1,000-year flood in southeastern France, models were trained on the Argens basin and tested on the Gapeau basin. A performancedriven feature selection algorithm -a greedy forward search maximizing the Critical Success Index (CSI) reduced 21 to four factors: discharge, Height Above Nearest Drainage, Distance to Stream, and river slope. Within the training domain, the RF achieved CSI = 0.76 (AUROC = 0.99). During extrapolation, all models maintained AUROC = 0.96, but the CSI declined to ~0.61, with XGB most robust. Notably, high threshold-independent discrimination (AUROC) did not guarantee accurate retrieval of the deterministic flood extent (CSI). Sub-catchment analysis revealed a global-local paradox: aggregate CSI masks systematic failures in upstream tributaries. Covariate shift was the principal barrier to spatial generalization; although the transferability assessment remains preliminary and basin-specific, results point to hybrid physics-informed ML as promising for operational flash-flood hazard assessment.
SN - 1947-5705
SN - 1947-5713
ER -