Diagnosing region-specific flood influencing factors in a volcanic island using explainable machine learning

This original paper applies a grid-based, explainable machine learning framework to identify dominant region-specific factors influencing flood occurrence on Jeju Island, a hydrologically heterogeneous volcanic island in South Korea. The study integrates historical flood-trace records from 23 rainfall-induced flood events (2006–2022) with multi-layer spatial datasets—encompassing rainfall, hydrogeology, land use, socio-economic, and proximity features—at a 100-meter grid resolution.

The key findings indicate that flood occurrence on Jeju Island is driven by a combination of hydrometeorological variables (notably short- and long-duration rainfall), groundwater levels, and spatial connectivity factors such as proximity to rivers and detention facilities. The relative importance of these factors varies by region: while rainfall metrics consistently have the highest influence, subsurface and connectivity-related variables can be critical in some sub-basins, especially in urban areas or zones with limited drainage. Conventional surface runoff indicators had lower importance due to the island’s permeable geology. The authors recommend moving beyond rainfall-only warnings toward analytical frameworks that incorporate groundwater, topography, exposure, and spatial connectivity for more effective, region-specific flood risk management—highlighting that factors underrepresented in conventional assessments may substantially affect flood occurrence under heterogeneous conditions.