Resolving the cascade of uncertainty in global flood projections

Projections of river flooding are crucial for adaptation, but conventional impact modelling is plagued by a ‘cascade of uncertainty’ arising from global climate model (GCM) outputs, statistical corrections, and hydrologic model structures. To overcome this cascade, we developed a data-driven approach that uses machine learning (ML) models to directly predict 10-year flood magnitudes from spatially-varying summary statistics derived from 19 uncorrected CMIP6 GCMs. We find that ML models trained on uncorrected GCMs are 40% more accurate than hydrologic models driven with bias-corrected GCMs. ML rectifies spatially-variable GCM biases by adjusting the contribution of basin attributes, correcting heavy rainfall underestimation in wet river basins. Applying our ML models to over 4.7 million kilometres of global river channels, we find that precipitation intensification is more likely to increase river flood magnitude in dry and high-altitude regions. By 2100, only 34% of rivers are projected to experience larger floods under SSP5-8.5, with the largest increases found in dry climates and high-altitude, steep river segments.