Exploring the Performance and Interpretability of an Enhanced Data-Driven Model to Assess Surface Flooding Susceptibility

The effects of climate change and increasing urbanization mean that urban areas are facing a greater risk of serious flooding. The paper aimed to adopt a data-driven approach to capture surface flood-prone features, providing a basis for surface flood susceptibility. This research developed an enhanced framework En-XGBoost, which consists of three modules: the core module, preprocessing module, and postprocessing module. Data augmentation, random extraction strategies, and local enhancement were introduced to improve the model’s performance. En-XGBoost was tested in Fuzhou, China. The main findings were as follows: (1) Neighborhood information extraction strategy outperformed information extraction strategy in extracting detailed flood-prone features, producing clearer boundaries between different flood susceptibility levels, and refining the flood risk areas. (2) Crucial explanatory variables were identified as major drivers of flood risk, with location-specific factors influencing the flood causes, necessitating localized analysis for specific sites. (3) The local enhancement, data augmentation, and random strategies improved model performance, with data augmentation proving more effective for stronger models and having limited impact on weaker ones. Model performance requires an appropriate alignment between data complexity and model complexity. En-XGBoost provided support for capturing surface flood-prone features.