Municipality-Scale Flood Risk Mapping in 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

·

---

Overview

A reproducible, open-access framework that delivers municipality-level flood risk statistics across multiple departments of Colombia. Each department study processes Sentinel-1 C-band SAR scenes (2015–2025) within Google Earth Engine using adaptive Otsu thresholding, then integrates predictor variables into a weighted ensemble of Random Forest, XGBoost, and LightGBM.

Departments

Department	Municipalities	Study Area	Ensemble AUC-ROC	Subdirectory
Antioquia	125	63,612 km²	0.94 ± 0.02	departments/antioquia
Bolívar	46	25,978 km²	—	departments/bolivar
Cauca	42	29,308 km²	—	departments/cauca
Chocó	30	46,530 km²	—	departments/choco
Guajira	15	20,848 km²	—	departments/guajira
Magdalena	30	23,188 km²	—	departments/magdalena
Nariño	64	33,268 km²	—	departments/narino
Total	352	242,732 km²

Guajira includes a specialized Sand Exclusion Layer for arid/semi-arid terrain adaptation.

Repository Structure

colombia-flood-risk/
├── README.md
├── departments/
│   ├── antioquia/          # Full pipeline + manuscript
│   ├── bolivar/            # Full pipeline + manuscript
│   ├── cauca/              # Full pipeline + manuscript
│   ├── choco/              # Full pipeline + manuscript
│   ├── guajira/            # Full pipeline + manuscript (arid adaptation)
│   ├── magdalena/          # Full pipeline + manuscript
│   └── narino/             # Full pipeline + manuscript

Each department subdirectory contains:

• scripts/ — Processing and analysis pipeline (SAR water detection, ML susceptibility, population exposure, climate analysis)
• overleaf/ — Manuscript in LaTeX (preprint format)
• gee_config.py — Google Earth Engine configuration
• requirements.txt — Python dependencies
• README.md — Department-specific results and metrics

Methodology

1. SAR Water Detection — Sentinel-1 scenes processed with adaptive Otsu thresholding → monthly water extent composites at 10 m resolution
2. Feature Engineering — 18 predictor variables (HAND, elevation, slope, SAR flood frequency, land cover, population density, etc.)
3. Ensemble ML — Weighted ensemble of Random Forest, XGBoost, and LightGBM with spatial five-fold cross-validation
4. Population Exposure — Susceptibility surface overlaid with 100 m population data
5. Climate Analysis — ENSO influence on flood extent (La Nina vs El Nino)

Citation

If you use this work, please cite the corresponding department preprint. See each subdirectory's README for specific citation details.