A comparative study on urban waterlogging susceptibility assessment based on multiple data-driven models

Area of Science:

• Environmental Science
• Urban Planning
• Geographic Information Systems

Background:

• Urban waterlogging is a growing concern, necessitating accurate prediction and susceptibility assessment.
• Data-driven models offer an alternative to complex mechanistic models, incorporating socio-economic factors.
• Existing research often lacks comprehensive model comparisons and interpretability analyses.

Purpose of the Study:

• To compare the performance of four data-driven models for urban waterlogging susceptibility mapping.
• To analyze the interpretability of these models and identify key influencing factors.
• To develop an integrated approach for reducing prediction uncertainty.

Main Methods:

• Four models were constructed: Random Forest (RF), Support Vector Machine with Radial Basis Function (SVM-RBF), Particle Swarm Optimization-Weakly Labeled Support Vector Machine (PSO-WELLSVM), and Maximum Entropy (MaxEnt).
• Twelve explanatory variables were used to predict waterlogging susceptibility in Beijing's central area.
• The Distance between Indices of Simulation and Observation (DISO) was employed for comprehensive model performance evaluation, and a geographical detector was used for interpretability analysis.

Main Results:

• PSO-WELLSVM demonstrated the highest performance (DISOtest = 0.63), outperforming MaxEnt (DISOtest = 0.78).
• MaxEnt excelled at identifying highly susceptible areas, while RF and SVM-RBF showed suboptimal performance and overfitting.
• Building Density (BD) was the most influential factor, followed by Distance to Road and Frequency of Heavy Rainstorms (FHR). Interactions between factors, like BD and FHR, non-linearly increased susceptibility prediction power.

Conclusions:

• The study highlights the effectiveness of PSO-WELLSVM for waterlogging susceptibility mapping and the importance of considering multiple factors.
• Integrating multiple models significantly reduces prediction uncertainty compared to single models.
• Understanding the interplay of factors like Building Density and heavy rainfall is crucial for effective urban waterlogging risk management.