Vojtek et al. (2026) Transferability of machine/deep learning-based prediction of fluvial flood extent to distinct river sections in Slovakia based on benchmark flood maps and high-resolution spatial data

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-13
• Authors: Matej Vojtek, Dávid Držík, Jozef Kapusta, Jana Vojteková
• DOI: 10.1016/j.ejrh.2026.103339

Research Groups

• Department of Geography, Geoinformatics and Regional Development, Faculty of Natural Sciences and Informatics, Constantine the Philosopher University in Nitra, Slovakia
• Institute of Geography, Slovak Academy of Sciences, Slovakia
• Department of Informatics, Faculty of Natural Sciences and Informatics, Constantine the Philosopher University in Nitra, Slovakia
• Institute of Security and Computer Science, University of the National Education Commission, Krakow, Poland

Short Summary

This study investigates the transferability of machine learning (ML) and deep learning (DL) models for predicting fluvial flood extent across distinct river sections in Slovakia under three flood scenarios. It finds that transferability is most effective between similarly sized river sections, with HAND, distance from river, and slope being the most influential predictors, offering high potential for near real-time flood mapping.

Objective

• To assess the ability of ML/DL models (Random Forest, XGBoost, Neural Networks, U-Net) to predict and transfer fluvial flood extent between distinct river sections in Slovakia.
• To identify which river sections are most suitable for training ML/DL models for transferability.
• To determine the predictors with the highest influence on the prediction and transferability of fluvial flood extent.
• To compare the differences in predictive performance and training time between ML and DL approaches for estimating and transferring fluvial flood extent.

Study Configuration

• Spatial Scale: Four distinct river sections in Slovakia: Kysuca (23.6 km length, 35.9 km² domain), Torysa (32.2 km length, 49.7 km² domain), Topľa (20.9 km length, 37.0 km² domain), and Gidra (3.1 km length, 1.5 km² domain). Predictors were processed at 1 meter spatial resolution.
• Temporal Scale: Flood scenarios for Q10, Q100, and Q1000 (10, 100, and 1000-year return periods). Data sources span various periods: precipitation (1981–2010), discharge (1961–2023), LULC (2023), and orthophotos (2021–2023).

Methodology and Data

• Models used:
  • Machine Learning (ML): Random Forest (RF), Extreme Gradient Boosting (XGBoost), Neural Networks (NN).
  • Deep Learning (DL): U-Net.
• Data sources:
  • Official benchmark flood maps for Q10, Q100, and Q1000 scenarios, created by the Slovak Water Management Enterprise using a 2D hydraulic approach (MIKE+ model).
  • Airborne laser scanned (LiDAR) Digital Elevation Model (DEM) with 1 meter spatial resolution.
  • Land Use/Land Cover (LULC) vector layer from the Basic Data Base for the Geographic Information System (ZBGIS) for 2023.
  • Orthophotos (20 centimeter resolution for Kysuca, Topľa, Torysa; 15 centimeter resolution for Gidra).
  • Seven high-resolution physical-geographic and land cover predictors derived from DEM and LULC: slope, stream power index (SPI), topographic wetness index (TWI), height above the nearest drainage (HAND), distance from river, roughness (Manning's n values), and normalized difference vegetation index (NDVI).
  • Multicollinearity among predictors was assessed using Pearson correlation (threshold ≤ 0.7) and Variance Inflation Factor (VIF) (threshold ≤ 5).
  • Training/testing strategy involved using three river sections for training and the remaining one for testing, with data balancing applied using 100 rectangle bins and weighted non-flood pixel selection.
  • Model performance was evaluated using Recall, Precision, and F1-score metrics, along with training time.

Main Results

• All predictors showed acceptable independence with Pearson correlation values ≤ 0.7 and VIF values ≤ 5.
• Transferability of fluvial flood extent models performed best when trained and tested on similarly long and large river sections, resulting in lower and more balanced numbers of false positive (FP) and false negative (FN) pixels (e.g., Torysa/Kysuca, Topľa/Torysa, Kysuca/Torysa combinations).
• Training on larger rivers and testing on smaller rivers led to higher FP rates, while the opposite scenario (training on smaller, testing on larger) resulted in higher FN rates.
• The F1-score ranged from 0.30 to 0.89, with the highest values achieved when models were trained on the Topľa or Torysa river sections. The lowest F1-scores were observed when training on the shortest Gidra river section.
• For RF and XGBoost models, the most influential predictors were HAND (36.3–71.7% importance), distance from river (8.3–31.2% importance), and slope (2.7–14.4% importance). TWI was consistently the least important predictor.
• XGBoost exhibited the shortest training times, followed by RF, U-Net, and NN. Optimized U-Net models achieved reasonable training times, often shorter than NN, despite higher computational intensity. U-Net showed better F1-score performance when trained on the Gidra river section, while performance was similar across all models for other river sections.

Contributions

• Presents a novel methodological approach for assessing the transferability of ML/DL-based fluvial flood extent prediction across distinct river sections, which has not been widely explored in existing literature.
• Demonstrates the effectiveness of ML/DL models for rapid flood extent estimation, particularly when applied to river sections with similar physical-geographical characteristics.
• Identifies the most influential physical-geographic predictors (HAND, distance from river, slope) for fluvial flood extent modeling, enhancing interpretability.
• Provides a comparative analysis of ML and DL models regarding predictive performance, computational efficiency, and spatial representation, offering guidance for their application in near real-time flood mapping and early warning systems.

Funding

• EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project No. 09I03-03-V03-00085.

Citation

@article{Vojtek2026Transferability,
  author = {Vojtek, Matej and Držík, Dávid and Kapusta, Jozef and Vojteková, Jana},
  title = {Transferability of machine/deep learning-based prediction of fluvial flood extent to distinct river sections in Slovakia based on benchmark flood maps and high-resolution spatial data},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103339},
  url = {https://doi.org/10.1016/j.ejrh.2026.103339}
}