Municipality-Scale Flood Risk Mapping in Magdalena, Colombia

Using Sentinel-1 SAR and Ensemble Machine Learning (2015--2025)

Cristian Espinal Maya · Santiago Jimenez Londono

School of Applied Sciences and Engineering, Universidad EAFIT, Medellin, Colombia

· ·

---

Abstract

We present a reproducible, open-access framework that delivers municipality-level flood risk statistics for all 30 municipalities of the Department of Magdalena, Colombia (23,188 km²; ~1.3 million inhabitants). The department encompasses diverse terrain ranging from the Sierra Nevada de Santa Marta (5,775 m a.s.l.) to the coastal lowlands of the Cienaga Grande de Santa Marta and the Rio Magdalena floodplain. We processed Sentinel-1 C-band SAR scenes (2015–2025) within Google Earth Engine using adaptive Otsu thresholding to produce monthly water extent maps at 10 m resolution. Eighteen predictor variables were integrated into a weighted ensemble of Random Forest, XGBoost, and LightGBM, evaluated under spatial five-fold cross-validation. HAND, SAR flood frequency, and elevation were identified as dominant predictors via SHAP analysis. Overlaying the susceptibility surface with 100 m population data quantifies the population residing in high or very high susceptibility zones. La Nina years amplify mean flood extent relative to El Nino years.

Study Area

The Department of Magdalena is located in northern Colombia along the Caribbean coast. It features:

• Sierra Nevada de Santa Marta: The world's highest coastal mountain range (5,775 m)
• Cienaga Grande de Santa Marta: Colombia's largest coastal lagoon system (Ramsar wetland)
• Rio Magdalena floodplain: Extensive fluvial lowlands and the Depresion Momposina
• Caribbean coast: Coastal and estuarine dynamics
• Bimodal precipitation: Two wet seasons (MAM, SON) with peak flooding in October-November

5 Subregions (30 Municipalities)

Subregion	Municipalities	Key Features
Santa Marta	1	Capital city, Sierra Nevada foothills
Norte	7	Zona Bananera, Cienaga Grande
Rio	9	Rio Magdalena floodplain, wetlands
Centro	6	Transitional terrain
Sur	7	Depresion Momposina, El Banco

Repository Structure

.
├── overleaf/                  # Manuscript (preprint format)
│   ├── main.tex               # Main LaTeX source
│   ├── arxiv.sty              # Preprint style file
│   ├── references.bib         # Bibliography
│   └── figures/               # All figure PDFs
├── scripts/                   # Processing and analysis pipeline
│   ├── 01_sar_water_detection.py
│   ├── 02_jrc_water_analysis.py
│   ├── 03_flood_susceptibility_features.py
│   ├── 04_ml_flood_susceptibility.py
│   ├── 05_population_exposure.py
│   ├── 06_climate_analysis.py
│   ├── 07_visualization.py
│   ├── 08_generate_tables.py
│   ├── 09_quality_control.py
│   └── utils.py
├── bib/                       # Additional bibliography files
│   └── references.bib
├── gee_config.py              # Central GEE configuration
├── utils.py                   # Root-level utility functions
├── requirements.txt           # Python dependencies
└── README.md

Data Sources

All data are open-access and processed via Google Earth Engine:

• Sentinel-1 GRD (ESA/Copernicus) — 10 m SAR flood detection
• JRC Global Surface Water v1.4 — 38-year water dynamics
• SRTM DEM v3 — Topographic features (30 m)
• MERIT Hydro — HAND computation (90 m)
• CHIRPS / ERA5-Land — Precipitation and soil moisture
• ESA WorldCover / Sentinel-2 — Land cover and NDVI
• WorldPop — Population density (100 m)
• FAO GAUL — Administrative boundaries

Reproducing the Analysis

Requirements

• Python 3.10+
• Google Earth Engine account (sign up)
• Libraries: earthengine-api, scikit-learn, xgboost, lightgbm, shap, matplotlib, geopandas

Pipeline

# 1. SAR water detection (runs on GEE)
python scripts/01_sar_water_detection.py

# 2. JRC validation analysis
python scripts/02_jrc_water_analysis.py

# 3. Feature engineering
python scripts/03_flood_susceptibility_features.py

# 4. ML model training and evaluation
python scripts/04_ml_flood_susceptibility.py

# 5. Population exposure analysis
python scripts/05_population_exposure.py

# 6. ENSO and seasonal analysis
python scripts/06_climate_analysis.py

# 7. Generate all figures
python scripts/07_visualization.py

# 8. Generate all tables
python scripts/08_generate_tables.py

# 9. Quality control
python scripts/09_quality_control.py

Magdalena-Specific Considerations

1. Dual SAR orbits: Both ASCENDING and DESCENDING passes are recommended due to radar shadow effects from the Sierra Nevada de Santa Marta.
2. Coastal dynamics: The Cienaga Grande and Caribbean coast introduce tidal and estuarine water fluctuations that must be distinguished from fluvial flooding.
3. Flat terrain: The extensive lowlands (HAND < 10 m) require adjusted thresholds for flood susceptibility mapping compared to mountainous regions.
4. Seasonal flooding: The Rio Magdalena and its tributaries produce extensive seasonal inundation, particularly during the October-November wet season.

Citation

If you use this work, please cite:

@article{EspinalMaya2026Magdalena,
  author  = {Espinal Maya, Cristian and Jim\'enez Londo\~no, Santiago},
  title   = {Municipality-Scale Flood Risk Mapping in {Magdalena}, {Colombia},
             Using {Sentinel-1} {SAR} and Ensemble Machine Learning (2015--2025)},
  year    = {2026},
  note    = {Available at SSRN}
}

License

Source code: MIT License. Manuscript and figures: CC BY 4.0.