Mabrouk et al. (2025) Artificial intelligence evaluation of nature based flood resilience in hilly terrain

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-10-10
• Authors: Abdelkader Mabrouk, Inamullah Inam, Muhammad Zeeshan Qureshi, Tariq Ali, Nadir Murtaza, Mohamed Ouda, Ahmed A. Alawi Al-Naghi, Dany Marcelo Tasán Cruz
• DOI: 10.1038/s41598-025-19629-9

Research Groups

• Department of Civil Engineering, College of Engineering, Northern Border University, Arar, Saudi Arabia
• Department of Civil Engineering, Engineering Faculty, Laghman University, Mehtarlam, Afghanistan
• Department of Civil Engineering, University of Engineering and Technology, Taxila, Pakistan
• Department of Civil Engineering, Swedish College of Engineering and Technology, Wah, Pakistan
• Architecture Department, College of Architecture & Planning, King Khalid University, Abha, Saudi Arabia
• Civil Engineering Department, University of Ha’il, Ha’il, Saudi Arabia
• Universidad Nacional de Chimborazo, Riobamba, Ecuador
• Department of Civil Engineering, CECOS University of IT and Emerging Sciences, Peshawar, Pakistan

Short Summary

This study evaluates the effectiveness of nature-based solutions (NBS), specifically flexible and rigid vegetation, in mitigating flash floods in hilly terrain by using artificial intelligence (AI) models to predict peak discharge. It found that flexible vegetation reduced peak discharge by 8% more than rigid vegetation, with the Random Forest model demonstrating superior predictive accuracy (R² of 0.9809 for flexible and 0.9906 for rigid vegetation).

Objective

• To investigate flood mitigation on hilly terrain under varying rainfall intensities and slopes, utilizing flexible versus rigid vegetation.
• To develop high-accuracy AI models (Random Forest and Support Vector Regression) for the precise prediction of peak discharge.
• To assess the influence of various input parameters on peak discharge prediction through interpretable AI techniques, including SHAP analysis, partial dependence plots, Monte Carlo simulation, cumulative accuracy curves, and error distribution.

Study Configuration

• Spatial Scale: Laboratory-scale experiments simulating hilly terrain with slopes ranging from 0% to 2%.
• Temporal Scale: Data series incorporating time ratio (total time / time of concentration) from simulated rainfall-runoff events.

Methodology and Data

• Models used: Random Forest (RF), Support Vector Regression (SVR). Interpretable AI techniques included SHAP (SHapley Additive exPlanations) analysis, Monte Carlo simulation, partial dependence plots (PDP), cumulative accuracy profile (CAP) curves, error distribution plots, and permutation feature importance.
• Data sources: A total of 344 laboratory-scale experimental data series derived from Rehman et al. (2023), including terrain slope (0% to 2%), rainfall intensity (0.05 mm/s to 0.083 mm/s), and time ratio (T/Tc), for both flexible and rigid vegetation conditions.

Main Results

• Flexible vegetation (FV) achieved an 8% greater reduction in peak discharge compared to rigid vegetation (RV), attributed to its superior surface resistance and infiltration capability.
• The Random Forest (RF) model demonstrated superior predictive performance with higher coefficients of determination (R²) of 0.9809 for FV and 0.9906 for RV, outperforming the Support Vector Regression (SVR) model (R² of 0.6806 for FV and 0.7922 for RV).
• 10-fold cross-validation confirmed the RF model's robustness (NSE values ranging from 0.898 to 0.968 for RV, and 0.650 to 0.962 for FV), while the SVR model showed significant inconsistency (NSE values ranging from 0.011 to 0.313 for RV, and 0.044 to 0.350 for FV).
• SHAP analysis indicated that the time ratio (T/Tc) was the most influential factor on peak discharge for both vegetation types (SHAP range: ±0.000417 m³/s for FV, ±0.000500 m³/s for RV), followed by rainfall intensity (moderate impact: ±0.000083 m³/s for FV, ±0.000117 m³/s for RV), with terrain slope having the minimum influence.
• Monte Carlo simulation revealed that flexible vegetation resulted in a more predictable, unimodal peak discharge distribution, whereas rigid vegetation exhibited higher variability and a wider, right-skewed distribution of peak discharge.
• Permutation feature importance for the RF model highlighted the time ratio as the most critical parameter for rigid vegetation, while terrain slope was most critical for flexible vegetation.

Contributions

• This study is the first to integrate AI methods with laboratory-scale experimental datasets to systematically assess the impact of different vegetation patterns (flexible and rigid) on peak discharge.
• It provides a novel comparative analysis of flexible and rigid vegetation's effectiveness in restricting runoff using advanced AI techniques, addressing a gap in existing literature.
• The research enhances model interpretability by extensively utilizing explainable AI (XAI) tools (SHAP analysis, PDP, permutation feature importance, error distribution), moving beyond mere predictive capability.
• It incorporates Monte Carlo simulation for uncertainty quantification and probabilistic distribution analysis, improving the reliability of nature-based solutions assessment under diverse terrain scenarios.

Funding

• Deanship of Research and Graduate Studies at King Khalid University (Large Research Project, grant number RGP2/121/46)
• Deanship of Scientific Research at Northern Border University, Arar, KSA (Project number NBU-FFR-2025-2507-10)

Citation

@article{Mabrouk2025Artificial,
  author = {Mabrouk, Abdelkader and Inam, Inamullah and Qureshi, Muhammad Zeeshan and Ali, Tariq and Murtaza, Nadir and Ouda, Mohamed and Al-Naghi, Ahmed A. Alawi and Cruz, Dany Marcelo Tasán},
  title = {Artificial intelligence evaluation of nature based flood resilience in hilly terrain},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-19629-9},
  url = {https://doi.org/10.1038/s41598-025-19629-9}
}