Loukili et al. (2025) Enhancing Flood Mapping Accuracy in North and West Africa Using Multi-Sensor Satellite Data and Machine Learning in Google Earth Engine

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-09-09
• Authors: Yassine Loukili, Younes Lakhrissi, Imad Bourian, Lahcen Hassine, Zakaria Kourab, Tarig Ali, Ahmed F. Elaksher, Rabin Chakrabortty
• DOI: 10.1007/s41748-025-00796-8

Research Groups

• LaRSI Laboratory, Sidi Mohamed Ben Abdellah University, Fez, Morocco
• Engineering Science Laboratory, Ibn Toufail University, Kénitra, Morocco
• LaGeS-SIRC Laboratory, Hassania School of Public Works, Casablanca, Morocco
• CCPS Laboratory, Hassan II University, Casablanca, Morocco
• Department of Civil Engineering, American University of Sharjah, Sharjah, United Arab Emirates
• Department of Engineering Technology and Survey Engineering, New Mexico State University, Las Cruces, NM, United States

Short Summary

This study evaluates Sentinel-1, Sentinel-2, and Landsat satellite data for flood detection using Random Forest (RF) and Minimum Distance (MD) classifiers within Google Earth Engine across urban (Tetouan, Morocco) and rural (Matam, Senegal) environments. It found that Sentinel-2 excelled in rural areas, while Sentinel-1 performed better in urban settings, with the RF-MD combination enhancing overall accuracy validated against UNOSAT benchmark maps.

Objective

• To evaluate the performance of Sentinel-1, Sentinel-2, and Landsat datasets and Random Forest (RF) and Minimum Distance (MD) machine learning algorithms for flood mapping in Google Earth Engine across urban and rural environments.
• To identify which combination of satellite data and machine learning algorithm provides the highest accuracy for flood mapping in diverse geographical contexts.

Study Configuration

• Spatial Scale:
  • Tetouan City, Morocco (urban environment): approximately 50 square kilometers.
  • Matam Region, Senegal (rural floodplain environment): approximately 2,700 square kilometers.
• Temporal Scale:
  • Tetouan flood event: January 1–18, 2021.
  • Matam flood event: October 10–30, 2024.
  • Data collection periods: Sentinel-1 (2014–2024), Sentinel-2 (2017–2024), Landsat (2013–2024).

Methodology and Data

• Models used: Random Forest (RF) and Minimum Distance (MD) machine learning classifiers.
• Data sources:
  • Sentinel-1 (Synthetic Aperture Radar, 10 meters spatial resolution, VH and VV polarization).
  • Sentinel-2 (Optical, 10 meters spatial resolution, multispectral bands including Green and Near-Infrared).
  • Landsat (Optical, 30 meters spatial resolution, multispectral bands including Green and Near-Infrared).
  • UNOSAT benchmark flood maps (Pleiades for Tetouan, Sentinel-2 for Matam) for external validation.
  • Google Earth Engine (GEE) cloud computing platform for data processing and analysis.

Main Results

• In rural Matam, Sentinel-2 demonstrated superior performance, achieving an AUC-ROC of approximately 99%, Intersection over Union (IoU) of approximately 85%, and minimal commission (approximately 1.4%) and omission errors (approximately 14%). Sentinel-1 significantly underestimated flood extents (approximately 54–60% error) in this environment.
• In urban Tetouan, Sentinel-1 outperformed optical datasets, providing inundated area estimates closest to the UNOSAT benchmark despite moderate omission errors (approximately 40–44%). Sentinel-2 and Landsat showed considerable underestimation in urban areas (approximately 43–46% and 57–60% respectively).
• The integration of Random Forest (RF) and Minimum Distance (MD) classifiers significantly improved flood detection accuracy, leveraging complementary strengths. RF consistently outperformed MD across all evaluation metrics.
• Iterative training sample optimization enhanced model stability, with the best results observed at 300–400 training samples.

Contributions

• Demonstrated that optimal satellite dataset and classifier selection for flood mapping is strongly environment-dependent (urban vs. rural).
• Introduced and validated a hybrid Random Forest (RF) and Minimum Distance (MD) classification strategy within Google Earth Engine, showing improved accuracy and stability compared to individual classifiers.
• Provided external validation of the flood mapping workflow using authoritative UNOSAT benchmark flood maps, enhancing confidence in its operational readiness.
• Developed a reproducible and scalable workflow using open-access multi-sensor data and cloud-based processing, making it transferable to other flood-prone regions without specialized computational infrastructure.

Funding

The authors declare that no funds or grants were received during the preparation of this manuscript.

Citation

@article{Loukili2025Enhancing,
  author = {Loukili, Yassine and Lakhrissi, Younes and Bourian, Imad and Hassine, Lahcen and Kourab, Zakaria and Ali, Tarig and Elaksher, Ahmed F. and Chakrabortty, Rabin},
  title = {Enhancing Flood Mapping Accuracy in North and West Africa Using Multi-Sensor Satellite Data and Machine Learning in Google Earth Engine},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00796-8},
  url = {https://doi.org/10.1007/s41748-025-00796-8}
}