Scalable Pluvial Flood Risk Assessment: A Data-Driven Framework Integrating Machine Learning (ML) and Discrete Global Grid Systems (DGGS H3)

The escalating frequency of extreme rainfall events necessitates scalable and dynamic tools for urban flood risk assessment. The primary objective of this study is to demonstrate and evaluate how a machine-learning–derived building-level pluvial flood susceptibility index can be operationalised through multi-resolution aggregation using the H3 Hexagonal Discrete Global Grid System (DGGS). The study builds on the building-level Pluvial Flood Index ( PFI b ) developed in the InzureFlood project, which derives susceptibility from historical insurance damage patterns and geospatial predictors. In the proposed methodology, the analysis domain is partitioned using H3, and building-level scores are aggregated within hexagonal cells to produce a scalable hexagon-level index ( PFI h ). Results indicate that this grid-based approach substantially improves computational efficiency, reducing spatial query operations by approximately 98% compared to traditional geometry-based workflows. The resulting maps show spatial patterns that are qualitatively consistent with expected topographic controls, such as depressions and flow paths. However, quantitative analysis of resolution effects reveals a critical trade-off: while overall susceptibility levels are preserved, aggregation increasingly smooths local extremes. A Jaccard Similarity Index of 0.14 between street-level (R13) and neighborhood-level (R10) hotspots indicates that most fine-scale susceptibility hotspots are not visible at intermediate scales. These findings imply that coarser resolutions may support regional strategic screening, whereas high-resolution grids are necessary for safety-critical tasks such as emergency response and local mitigation. The framework provides a scalable and updateable representation that supports consistent multi-scale communication from micro-scale screening to macro-scale planning.