Urban waterlogging susceptibility assessment based on feature selection optimization and interpretable machine...

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning Applications

Background:

• Machine learning (ML) in flood susceptibility studies often suffers from poor interpretability and generalization.
• A disconnect between feature selection and model training hinders the effectiveness of existing ML models.

Purpose of the Study:

• To develop and validate the C2MI-ML framework for simultaneous optimization of feature selection and model training in flood susceptibility assessment.
• To identify key natural and socioeconomic factors influencing waterlogging susceptibility in Shenzhen, China.

Main Methods:

• Integration of Chi-square test, mutual information, and ML algorithms within the C2MI-ML framework.
• Case study application in Shenzhen, China, utilizing diverse datasets for training and validation.
• Performance evaluation using Area Under the Curve (AUC) and Kappa coefficient.

Main Results:

• The C2MI-LightGBM model achieved a high AUC of 0.965; all ensemble models surpassed an AUC of 0.95.
• Elevation and Gross Domestic Product (GDP) were identified as the most significant factors influencing waterlogging.
• Spatial analysis revealed high susceptibility in western Shenzhen, with specific districts and vulnerable sites like schools identified.

Conclusions:

• The C2MI-ML framework offers improved interpretability and generalization for flood susceptibility modeling.
• Findings highlight the critical role of both natural and socioeconomic factors in urban waterlogging.
• The study provides actionable insights for targeted flood risk management and early warning systems in Shenzhen.