Hourly rainfall-flood relationships and risks of systematic flood underestimation in daily-scale analyses

Climate change is expected to intensify heavy rainfall with concomitant increases in flood hazards, yet most large-scale flood assessments remain based on daily data. While sub-daily rainfall variability differs fundamentally from daily statistics, its implications for flood generation and risk remain poorly understood at continental scales, largely due to the scarcity of long-term hourly streamflow observations. Consequently, flood hazards inferred from daily-scale analyses may be systematically underestimated. Here, a causally constrained deep learning model for hourly runoff reconstruction is developed, integrating multi-source hourly meteorological forcings with limited observed hourly streamflow and widely available daily discharge constraints. Using this model, we create a multi-decadal reconstruction of hourly runoff for nearly ten thousand basins across the continental United States. Building on these reconstructions, we provide a first assessment of how rainfall-flood relationships differ between hourly and daily timescales and investigate potential flood underestimation arising from daily-scale analyses. We show that sub-daily flood peaks can be masked when aggregated to daily resolution, and examine the temporal evolution and controlling mechanisms of this underestimation across events and catchments, with particular attention on catchment scale, intraday rainfall variability, storm duration, and aggregation timescales. This work highlights the importance of resolving sub-daily flood processes for flood risk assessment and early warning, and provides a foundation for ongoing extensions toward global hourly runoff reconstruction and large-scale sub-daily flood risk analysis under a changing climate.