Integrating geospatial intelligence and machine learning for flood susceptibility mapping.

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning

Background:

• Flood susceptibility mapping is crucial for disaster risk reduction.
• Remote sensing and machine learning offer powerful tools for this purpose.

Purpose of the Study:

• To evaluate five machine learning algorithms (XGBoost, DT, RF, LightGBM, GLM) for flood susceptibility mapping.
• To assess the performance of an ensemble voting model integrating these algorithms.

Main Methods:

• Utilized flood extent data (2000-2018) from the Global Flood Database (GFD).
• Incorporated diverse ancillary spatial data (climate, topography, hydrology, land cover).
• Compared individual model performance (XGBoost, RF, LightGBM, DT, GLM) and an ensemble voting model using AUC values.

Main Results:

• XGBoost (AUC=0.985), RF (AUC=0.984), and LightGBM (AUC=0.982) showed strong predictive performance.
• The ensemble voting model achieved the highest accuracy (AUC=0.994), outperforming all individual models.
• DT (AUC=0.972) showed moderate accuracy, while GLM (AUC=0.879) had the lowest.

Conclusions:

• Machine learning, especially ensemble frameworks, significantly enhances flood susceptibility mapping accuracy and reliability.
• These advanced techniques are valuable tools for effective flood risk management and spatial analysis.