Title: Leveraging Data-Driven AI to Decode the Mechanics of Compound Flooding

Abstract:

The prevention of compound flooding remains a formidable obstacle due to the nonlinear interplay among various hydrometeorological variables. Current forecasting methodologies, whether rooted in physics or data science, tend to prioritize temporal patterns, thereby neglecting the joint influence of multiple interacting factors on flood dynamics. To bridge this gap, we perform a comprehensive, data-driven investigation into compound flooding in South Florida—a region prone to such events—by synthesizing data on tidal conditions, precipitation, groundwater levels, and anthropogenic water management practices.

Our analysis yields three pivotal insights: First, models reliant solely on temporal dynamics are insufficient for capturing the multi-factor interactions characteristic of compound events. Second, in this porous coastal environment, subsurface saturation, indicated by groundwater levels, serves as the primary determinant of flood severity, frequently surpassing the impact of immediate rainfall intensity. Third, the spatial configuration of nearby monitoring stations within a specific effective radius offers essential causal context for flooding, whereas extending the temporal history provides diminishing benefits during extreme weather scenarios.

These results indicate that compound flooding is driven more by spatially coupled system states than by long-term temporal dependencies, thereby questioning the prevailing rain-centric and sequence-focused forecasting frameworks. By positioning data-driven models as instruments for scientific discovery rather than mere predictive tools, this research provides novel understanding of compound flooding mechanisms and guides the development of more physically robust early-warning systems for coastal regions. The associated dataset and code are accessible at https://github.com/AslanDing/SFBench.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC