Identification of Dominant Flood Influencing Factors with Grid-Based Explainable Machine Learning

This study develops a high-resolution, grid-based explainable machine learning (XAI) framework to systematically identify dominant flood-influencing factors across Jeju Island, Korea, by integrating historical flood trace maps with multi-source spatial datasets. Flood occurrence was classified at a 100 m grid resolution using four state-of-the-art tree-based ensemble algorithms, enabling robust modeling of nonlinear interactions between hydro-meteorological, geomorphological, and infrastructural variables. Model performance was rigorously evaluated across multiple subregions to quantify spatial heterogeneity in predictive skill and controlling mechanisms. The models achieved moderate to high classification performance, with maximum recall and F1-scores reaching 0.81 and 0.75, respectively, demonstrating strong capability in detecting flood-prone conditions.
Explainability analyses based on feature-importance metrics consistently identified short- and long-duration extreme rainfall (3-hour and 12-hour maxima), 5-day antecedent precipitation, maximum wind speed, groundwater level, and proximity to detention facilities and river networks as the most influential predictors of flood occurrence. Notably, their relative contributions exhibited pronounced spatial variability. In inland and high-elevation basins, flood dynamics were primarily governed by rainfall persistence and subsurface hydrological responses, whereas in coastal and highly urbanized zones, flood occurrence was more strongly modulated by drainage connectivity and proximity to hydraulic infrastructure.
These spatially differentiated controls reflect the complex volcanic hydro-geomorphological setting of Jeju Island and highlight the limitations of uniform flood warning criteria. The findings underscore the necessity of region-specific, dynamically adaptive warning thresholds that explicitly account for local hydrological processes and infrastructure configurations.
Overall, this study demonstrates the methodological advantages of grid-based explainable machine learning for physically interpretable and spatially adaptive flood risk assessment. The proposed framework provides a transferable blueprint for data-driven disaster risk management in volcanic island environments and other hydrogeomorphologically complex regions under intensifying climate extremes.

Acknowledgements
The research for this paper was carried out under the KICT Research Program (Project no. 20260161-001, Development of Digital Urban Flood Control Technology for the Realization of Flood Safety City) funded by the Ministry of Science and ICT.