Yousefi et al. (2025) A reinforcement learning approach with explainable AI for spatial flood susceptibility analysis

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-12-11
• Authors: Saleh Yousefi, Sara Mardanian, Abolfazl Jaafari, Zahra Tavangar
• DOI: 10.1016/j.ejrh.2025.103035

Research Groups

• Soil Conservation and Watershed Management Research Department, Chaharmahal and Bakhtiari Agricultural and Natural Resources Research and Education Center, AREEO, Shahrekord, Iran
• Research Institute of Forests and Rangelands, Agricultural Research, Education and Extension Organization (AREEO), Tehran, Iran
• Shahid Beheshti University, Faculty of Architecture and Urban Planning, Landscape Architecture Department, Tehran, Iran

Short Summary

This study develops and compares reinforcement learning (RL) models, including a novel RL-Stack ensemble, for spatial flood susceptibility mapping in a semi-arid mountainous region, finding that the Proximal Updating (PU) model achieved the highest accuracy and stability, with snow depth identified as the primary hydrological control.

Objective

• To develop and evaluate a spatially explicit framework integrating reinforcement learning (RL) algorithms (Q-Learning, Deep Q-Learning, Proximal Updating, and a novel RL-Stack ensemble) with Geographic Information Systems (GIS) for flood susceptibility mapping.
• To enhance interpretability of the RL models and quantify the contribution of explanatory variables using SHapley Additive exPlanations (SHAP).

Study Configuration

• Spatial Scale: Chaharmahal and Bakhtiari Province, western Iran, covering approximately 16553 square kilometers, with elevations ranging from 778 meters to 4203 meters above sea level. All raster layers were resampled to a standardized resolution of 30 meters.
• Temporal Scale: Flood inventory data and explanatory variables were collected and selected to correspond to the period from 1983 to 2023.

Methodology and Data

• Models used: Q-Learning (QL), Proximal Policy Optimization (PPO)-based Proximal Updating (PU), Deep Q-Learning (DQL), and RL-Stack (a novel ensemble approach combining QL, PU, and DQL with a logistic regression meta-learner). SHapley Additive exPlanations (SHAP) was used for model interpretability.
• Data sources:
  • Flood inventory data: 545 location points (346 flood, 199 non-flood) compiled from historical records of the regional water administrative, field surveys, and local disaster management authorities (1983–2023).
  • Explanatory variables: 22 distinct factors (selected from an initial 28 after correlation analysis) encompassing environmental, topographic, climatic, and anthropogenic characteristics. Sources include a 30-meter Digital Elevation Model (DEM), Iran Meteorological Organization (IRIMO), Soil Grid, Landsat 8 and 9 operational land imager (OLI) - USGS, and regional datasets.

Main Results

• The Proximal Updating (PU) model consistently demonstrated the best performance, achieving a validation Area Under the Curve (AUC) of 0.9642 and a Kappa coefficient of 0.7523, indicating superior accuracy, sensitivity, and generalization capacity.
• The novel RL-Stack ensemble model ranked second, with a validation AUC of 0.9523 and high sensitivity (0.9468), effectively leveraging the complementary strengths of its base RL learners.
• Deep Q-Learning (DQL) showed strong performance (validation AUC = 0.9331) but was characterized by the highest modeling error (RMSE) and overestimation bias in local predictions.
• Q-Learning (QL) exhibited the weakest performance (validation AUC = 0.5871), demonstrating poor generalization and a strong bias towards positive classifications.
• SHAP analysis identified snow depth (SD) as the most influential variable (mean SHAP = 2.108), playing a dual role in intensifying or attenuating floods. Other key contributors included flow accumulation (FA), maximum 24-hour rainfall (R24HR), and distance from residential areas (DRA).
• Strong interactions were observed between short-duration intense rainfall (R24HR) and snow depth (SD) and flow accumulation (FA), highlighting the significance of rain-on-snow processes in generating large floods.

Contributions

• Expansion of Reinforcement Learning (RL) applications from flood management to the domain of spatial flood susceptibility mapping.
• Introduction of RL-Stack, a novel ensemble hybrid model specifically designed to capture the complex flood dynamics in semi-arid, mountainous environments by combining multiple RL algorithms.
• Integration of Explainable AI (SHAP) to provide transparent, instance-level, and dataset-wide insights into RL model predictions, enhancing trust and supporting informed decision-making.
• Generation of highly accurate and actionable flood susceptibility maps, offering critical guidance for policymakers, urban planners, and disaster managers in reducing flood risks and enhancing community resilience.

Funding

• The Research Institute of Forests and Rangelands (RIFR) under the National Research Project No. 0–09–09–002–000095.

Citation

@article{Yousefi2025reinforcement,
  author = {Yousefi, Saleh and Mardanian, Sara and Jaafari, Abolfazl and Tavangar, Zahra},
  title = {A reinforcement learning approach with explainable AI for spatial flood susceptibility analysis},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.103035},
  url = {https://doi.org/10.1016/j.ejrh.2025.103035}
}