A spatially interpretable machine learning framework for urban waterlogging risk mapping in Beijing.

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning

Background:

• Urban waterlogging is a growing problem due to urbanization and climate change.
• Accurate spatial prediction is difficult due to complex factors and varying conditions.
• Existing methods struggle with nonlinear drivers and spatial heterogeneity.

Purpose of the Study:

• To develop a spatially interpretable machine learning framework for urban waterlogging prediction.
• To integrate remote sensing and geospatial data for improved risk assessment.
• To enhance urban flood resilience through better planning and governance.

Main Methods:

• A hybrid machine learning framework combining XGBoost and MGWR was developed.
• Remote sensing and geospatial data (topographic, hydrologic, land cover, proximity) were used.
• Four algorithms (RF, SVM, KNN, XGBoost) were evaluated, with XGBoost showing initial promise.

Main Results:

• XGBoost achieved high classification performance (AUC = 0.913).
• The MGWR-XGBoost hybrid model demonstrated superior probabilistic accuracy (Brier = 0.289) and PR-AUC (0.576).
• This hybrid approach provided a spatially stable risk map with high specificity (0.734).

Conclusions:

• The proposed MGWR-XGBoost framework offers high-resolution, spatially explicit waterlogging risk mapping.
• This approach provides practical support for urban drainage planning and adaptive flood governance.
• The framework shows potential for improving resilience in data-scarce urban areas globally.