Waterlogging susceptibility assessment in developed urban area using explainable machine learning methods with different negative sampling strategies

Exacerbated by global climate change, rainstorm-induced waterlogging has become a frequent hazard threatening sustainable development, causing disruptions from traffic paralysis to casualties. Therefore, it is essential to assess urban waterlogging susceptibility and identify high-risk areas. Although explainable machine learning (ML) shows promise in such assessments, its effectiveness is often limited by data limitation. For instance, rainfall and drainage network data are difficult to obtain, while the commonly used method of subjectively selecting random negative samples (non-waterlogged points) may compromise model classification performance and generalization ability. To address these challenges, we conducted a catchment unit-based framework based on morphological characteristics and proposed three data-driven negative sample sampling strategies (one kernel density estimation (KDE)-based and two K-means-based). Multiple explainable ML models (random forest/extreme gradient boosting/categorical boosting) were trained to assess waterlogging susceptibility in Changsha City. The results indicate that catchment units achieve high evaluation accuracy. The proposed negative sampling methods (mean test AUC ≈0.85) significantly outperform conventional random sampling (mean test AUC <0.75). Notably, the K-means-I clustering strategy achieved superior performance with a mean test AUC exceeding 0.90 across all three machine learning models. The Shapley additive explanation (SHAP) values appeals that the scale and topographic features of catchment units are the primary drivers of waterlogging in Changsha City. This study demonstrates strong applicability in identifying waterlogging risks for urban underpass roads, providing scientific support for flood prevention design in critical urban infrastructures. • Catchment morphometrics perform well in waterlogging susceptibility assessment. • Data-driven negative sampling surpasses traditional random methods. • Changsha's waterlogging susceptibility and flood-prone traffic nodes identified. • Explainable machine learning reveals key drivers of urban waterlogging.