Panagiotou et al. (2026) Investigating the mechanisms of flood susceptibility with the use of multi-basin machine learning models in data-scarce environments in Cyprus

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-06
• Authors: Constantinos F. Panagiotou, Giorgia Guerrisi, Davide De Santis, Fabio Del Frate, Marios Tzouvaras
• DOI: 10.1016/j.ejrh.2025.103075

Research Groups

• Eratosthenes Centre of Excellence, Limassol, Cyprus
• Department of Civil Engineering and Computer Science Engineering, Tor Vergata University of Rome, Rome, Italy

Short Summary

This study developed and compared multi-basin machine learning models (SVM, XGBoost, RF, MLP) for flood susceptibility assessment in data-scarce environments in Cyprus. It demonstrated that simplified Random Forest models, utilizing key topographical and land-use features, can provide rapid and accurate predictions for effective flood risk management.

Objective

• To develop multi-basin, within-region generalized machine learning models (MLMs) for flood susceptibility by merging data from eight watersheds and testing them on an "unseen" watershed in Cyprus.
• To comprehensively assess flood susceptibility by developing and comparing multiple MLMs and analyzing the influence of driving factors on flood emergence.
• To determine the minimum number of features required for developing simplified, cost-effective, and fast screening models suitable for data-scarce environments.

Study Configuration

• Spatial Scale: The island of Cyprus, focusing on nine small-scale watersheds (typically less than 100 square kilometers). Eight watersheds were used for training, and the Geroskipou watershed was used for testing.
• Temporal Scale: Flood susceptibility assessment based on physical characteristics. Input data included daily rainfall from 2011 to 2023/2024 and air temperature data from 2011 to 2024. Flood inventory data corresponded to a 500-year return period worst-case scenario.

Methodology and Data

• Models used: Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Multilayer Perceptron (MLP).
• Data sources:
  • Flood-related factors (7 features):
    • Rainfall Intensity (RI): Daily rainfall data from 90 ground-based stations (2011–2023), calculated as modified Fournier index (dMFI).
    • Euclidean distance from the drainage network (DD): Geoportal of Water Development Department.
    • Terrain Elevation (TE): Digital Elevation Model (DEM) with 25 meter resolution, Department of Land and Surveys (version 2023).
    • Terrain Slope (TS): Derived from DEM.
    • Soil (texture and depth): Geological Survey Department.
    • Land Use and Land Cover (LULC): CORINE Land Cover 2018 (version 2020_20u1).
    • Flow Accumulation (FA): Derived from DEM.
  • Flood Inventory Database: Geospatial information on high-risk flood areas (500-year return period) from the geoportal of the national water authorities (WDD, 2023), supplemented by an expected flood susceptibility map (Panagiotou, 2025a).
  • Training Dataset: 46,299 data points (31,311 non-flooded, 14,988 flooded) from eight watersheds.
  • Test Dataset: 898 data points (449 for each class) from the Geroskipou watershed.
  • Validation: Leave-One-Cluster-Out (LOCO) cross-validation strategy for training, and an independent test dataset for external validation.

Main Results

• All four MLMs achieved high performance, with Balanced Accuracy (BA) and F1 scores exceeding 0.95 for both training and test datasets.
• The SVM model exhibited the highest performance on the test dataset (BA = 0.995, F1 score = 0.996, Brier score = 0.003), closely followed by RF and MLP.
• Random Forest (RF) was selected for compiling multi-level susceptibility maps due to its robustness, ability to capture complex nonlinear relationships, and an optimal probability threshold (0.48) facilitating map interpretation.
• Feature importance analysis consistently identified Land Use and Land Cover (LULC), Terrain Slope (TS), Terrain Elevation (TE), and Flow Accumulation (FA) as the most significant flood-related factors across all MLMs. Rainfall Intensity (RI), Distance from Drainage Network (DD), and Soil properties showed minor influence.
• Approximately half (49.9%) of the Geroskipou watershed was classified as highly susceptible to flooding (36.9% "Very-High", 13% "High"), predominantly in urban and semi-urban coastal areas characterized by low terrain elevation and slope, and high LULC scores.
• Simplified RF models, built using only the four most important features (LULC, TS, TE, FA), achieved a Balanced Accuracy of 0.951 and a Brier score of 0.033 on the test dataset, demonstrating their effectiveness as computationally efficient flood screening tools.
• Partial dependence plots revealed critical threshold values: LULC scores greater than 0.8, terrain slopes less than 50%, flow accumulation values around 5.8, and terrain elevations below 350 meters above sea level were associated with increased flood susceptibility.

Contributions

• First study to apply multi-basin machine learning algorithms for flood susceptibility assessment in Cyprus, including explicit external-basin testing.
• Provides an operational and parsimonious methodology for flood susceptibility management in data-scarce environments.
• Implemented hyperparameter optimization for all MLMs, ensuring fair comparison, a step often omitted in similar studies.
• Assessed functional relationships between major input variables and flood susceptibility, identifying critical threshold values for key flood indicators.
• Demonstrated that shallow MLMs (e.g., RF, SVM) can serve as rapid, computationally efficient flood screening tools, offering a viable alternative to complex, data-intensive deep learning or physically-based models, especially in data-scarce regions.
• Offered a data-driven approach to feature importance analysis, reducing the influence of subjective expert opinions in decision-making.

Funding

• European Union’s Horizon Europe Framework Programme HORIZON-WIDERA-2021-ACCESS-03 (Grant Agreement No. 101079468) for the AI-OBSERVER project.
• European Union’s Horizon 2020 research and innovation programme (Grant Agreement No. 857510) for the EXCELSIOR H2020 Widespread Teaming project.
• Government of the Republic of Cyprus through the Directorate General for the European Programmes, Coordination and Development.
• Cyprus University of Technology.

Citation

@article{Panagiotou2026Investigating,
  author = {Panagiotou, Constantinos F. and Guerrisi, Giorgia and Santis, Davide De and Frate, Fabio Del and Tzouvaras, Marios},
  title = {Investigating the mechanisms of flood susceptibility with the use of multi-basin machine learning models in data-scarce environments in Cyprus},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2025.103075},
  url = {https://doi.org/10.1016/j.ejrh.2025.103075}
}