Exploring effective ways to increase reliable positive samples for machine learning-based urban waterlogging...

Area of Science:

• Environmental Science
• Geographic Information Science
• Urban Planning

Background:

• Machine learning models for urban waterlogging susceptibility often face class imbalance, with limited positive samples.
• Improving training data quality is crucial for enhancing model performance and accuracy in waterlogging prediction.
• Existing oversampling techniques may not adequately represent the physical mechanisms of waterlogging.

Purpose of the Study:

• To investigate effective methods for increasing reliable positive samples in urban waterlogging susceptibility studies.
• To compare the performance of oversampling (SMOTE) and physical simulation (OSSA) approaches for sample generation.
• To evaluate the impact of different sample generation techniques on classifier performance and waterlogging susceptibility map (WSM) accuracy.

Main Methods:

• Employed Synthetic Minority Over-Sampling Technique (SMOTE) and Optimized Seed Spread Algorithm (OSSA) to generate synthetic positive samples.
• Conducted a case study of waterlogging in Shenzhen using eight spatial variables.
• Compared classifier performance and WSM accuracy using original samples, SMOTE-generated samples, and OSSA-generated samples.

Main Results:

• Classifiers trained with SMOTE-generated samples performed worse than those trained with original samples.
• OSSA significantly improved the performance of trained classifiers compared to original samples.
• SMOTE did not improve WSM accuracy, while OSSA markedly increased it.

Conclusions:

• SMOTE is unreliable for generating valid positive samples in Shenzhen's waterlogging analysis due to its feature-space-based approach.
• OSSA, by simulating water flow mechanisms, effectively generates reliable positive samples, enhancing waterlogging susceptibility models.
• Physical simulation methods like OSSA are superior to data-driven oversampling for improving the accuracy of waterlogging susceptibility assessments.