Kareem et al. (2025) Improving River Flood Mapping with Adaptive Sampling and Artificial Intelligence Techniques for Enhanced Flood Risk Assessment

Identification

• Journal: Water Resources Management
• Year: 2025
• Date: 2025-12-29
• Authors: Kola Yusuff Kareem, Innkyo Choo, Seungoh Lee, Younghun Jung
• DOI: 10.1007/s11269-025-04366-5

Research Groups

• Department of Advanced Science and Technology Convergence, Major in Civil and Environmental Engineering, Kyungpook National University, Korea
• Department of Civil Engineering, Hongik University, Seoul, Korea

Short Summary

This study develops and evaluates AI-assisted adaptive sampling techniques, driven by precipitation and topographic factors, to enhance river flood mapping and risk assessment. It demonstrates that elevation-adapted sampling significantly improves the accuracy of AI models, particularly Random Forest, in delineating flood extents and identifying vulnerable assets for 200-year and 500-year flood events.

Objective

• To evaluate the applicability of coupling adaptive sampling techniques, specifically those driven by precipitation and topographic factors, with AI-based data-driven approaches to enhance and complement existing flood design maps for improved flood risk assessment.

Study Configuration

• Spatial Scale: Miho River basin, South Korea, with a flow length of 89.2 kilometers and a basin area of 1,854 square kilometers. The study focuses on the main river channel.
• Temporal Scale: Ground truth data from 1% annual exceedance probability (100-year) flood extent; predictions for 200-year and 500-year flood recurrence extents; precipitation data derived from frequency analysis (Generalized Extreme Value distribution) of historical records; land cover data from 2024 Sentinel-2 timeseries.

Methodology and Data

• Models used:
  • Machine Learning (ML) classification models: Random Forest (RF), eXtreme Gradient Boosting (XGB), Logistic Regression (LogR).
  • Deep Learning (DL) model: UNet (for object-based modeling).
  • Statistical distribution: Generalized Extreme Value (GEV) distribution for precipitation frequency analysis.
• Data sources:
  • Ground truth: Korea Ministry of Environment’s Flood Risk Map Information System (KME-FRMIS) flood recurrence maps (1% AEP/100-year extent).
  • Topographic features: 30 meter Shuttle Radar Topography Mission Elevation (SRTM-DEM), from which aspect, slope, topographic position index (TPI), topographic wetness index (TWI), terrain roughness index (TRI), curvature, flow accumulation (FlowAcc), flow direction (FlowDir), and stream power index (SPI) were derived.
  • Satellite-derived features: Landsat 8 surface reflectance for Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), and Normalized Difference Vegetation Index (NDVI).
  • Precipitation data: 105 Automated Synoptic Observation System (ASOS) stations.
  • Land Use/Land Cover (LULC): 2024 Sentinel-2 timeseries dataset.
  • Asset values: Korea Development Institute.
  • Sampling techniques for pseudo-absences: Random sampling, Elevation sampling, Precipitation sampling, Inverse Occurrence sampling.

Main Results

• Model Performance: Elevation-adapted models consistently outperformed other sampling methods. The Random Forest with Elevation sampling (RFElev) achieved the best performance for point-based modeling with an Area Under Curve (AUC) of 0.98, Recall of 0.99, 48.52% true flood positives, 50.93% true non-flood negatives, and a total error of 0.567% on test data. The object-based UNetElev also performed best among UNet variants.
• Flood Recurrence Forecasts: The RFElev model was selected for predicting 200-year and 500-year flood extents due to its robustness in modeling non-linear relationships and generalizing to unseen events, outperforming the UNetElev.
• Feature Importance: Changes in precipitation were identified as the strongest influence on RFElev's predictions, followed by TPI, CN, TRI, LULC, Flow_Dir, and TWI.
• Land Use Vulnerability: Built areas exhibited the highest vulnerability, accounting for 57.2% to 64.2% of affected land cover in low to high flood risk zones for the 200-year event, and 59.8% to 62.3% for the 500-year event. Agricultural, industrial, and residential assets were most affected in downstream areas.
• Flood Depth: Maximum modeled flood depths reached 3.94 meters for the 200-year event and 4.82 meters for the 500-year event, with higher depths concentrated near the Miho River mouth due to low-lying topography.
• Damage Cost: Estimated total economic losses were approximately 262,000 USD for the 200-year event and 330,000 USD for the 500-year event. Residential buildings accounted for 56–58% of the total damages. Hotspot analysis identified concentrated high-damage building clusters in the deepest downstream zones.

Contributions

• Introduces novel AI-assisted adaptive sampling techniques that dynamically link sampling with physical, spatially-varied flood drivers (elevation and precipitation), explicitly incorporating physics-guided hydrological reasoning into pseudo-absence selection.
• Demonstrates that elevation-adapted pseudo-absence samples create a stronger flood-no-flood feature pipeline, significantly improving model performance and the delineation of transitional flood zones critical for infrastructure protection and emergency planning.
• Provides a more comprehensive assessment of unpriced asset vulnerability outside existing flood insurance zones, offering a basis for updating flood tax systems and improving disaster preparedness.
• Offers a parsimonious and efficient data-driven approach for enhancing existing design flood maps, particularly beneficial in regions where computationally expensive hydrodynamic modeling is impractical.
• Integrates geomorphological and meteorological information to improve predictive accuracy and provide interpretability for flood probability mapping.

Funding

• Korea Environment Industry & Technology Institute (KEITI) through R&D Program for Innovative Flood Protection Technologies against Climate Crisis, funded by the Korea Ministry of Environment (MOE) (2022003470001).

Citation

@article{Kareem2025Improving,
  author = {Kareem, Kola Yusuff and Choo, Innkyo and Lee, Seungoh and Jung, Younghun},
  title = {Improving River Flood Mapping with Adaptive Sampling and Artificial Intelligence Techniques for Enhanced Flood Risk Assessment},
  journal = {Water Resources Management},
  year = {2025},
  doi = {10.1007/s11269-025-04366-5},
  url = {https://doi.org/10.1007/s11269-025-04366-5}
}