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

Flood susceptibility studies that utilize machine learning (ML) face challenges in interpretability and generalization due to the disconnection between feature selection and model training. This study proposed the C2MI-ML framework, which integrates the Chi-square test, mutual information, and ML to simultaneously optimize feature selection and model training. Using Shenzhen, China, as a case study, this study found that both natural and socioeconomic factors have significant impacts on waterlogging susceptibility. Elevation, vegetation coverage, precipitation, economic activity level, and building characteristics were identified as the main influencing factors, with elevation and gross domestic product (GDP) showing particularly strong effects. The C2MI-LightGBM model achieved the best performance, with an AUC of 0.965, while all ensemble models yielded AUC values exceeding 0.95, demonstrating high stability and strong applicability. Evaluation using an independent validation set yielded an AUC of 0.846 and a Kappa coefficient of 0.606, indicating a robust discriminative ability and consistent predictive performance. Spatial analysis of waterlogging susceptibility indicates that Shenzhen exhibits a pattern of higher susceptibility in the western regions, moderate susceptibility in the eastern regions, and lower susceptibility in the central urban areas. Approximately three-quarters of the city is classified as high-risk or very high-risk zones, with Baoan, Guangming, Dapeng, and Yantian districts exhibiting the highest levels of susceptibility. Kindergartens, primary schools, secondary schools, and nursing homes were identified as having high exposure risks, underscoring their vulnerability. These findings pinpoint major high-risk zones and vulnerable sites, enabling proactive early warning and targeted flood control actions.