More Frequent Extreme Floods Revealed by Observation-Constrained Projections

Current cascade-type models are the dominant approach for projecting future extreme floods, but they suffer from major limitations, including substantial bias, large uncertainty and coarse spatial resolution. To address these issues, we developed an observation-constrained framework that integrates flood estimates, based on historical data and regional flood frequency analysis, with changes in design floods from cascade-type model projections, enabling 1-km resolution projections across 28.2 million river pixels worldwide. Our analysis reveals that cascade-type models overestimate the historical 100-year flood by about 160% globally, while forcing-based corrections still exhibit considerable bias. Further, our observation-constrained approach reduces multi-model uncertainty by a median of 22.8% globally compared to cascade-type modeling. Under a high-emissions scenario, 83% of the global land mass shows increasing flood frequency. Globally, the historical 100-year flood is projected to have a median return period of about 36 years – more frequent than suggested by cascade-type model projections. Our results highlight the acceleration of flood risks, which may leave communities unprepared for intensifying climate impacts.