Forecasting Global Rainfall in a Changing Climate: A Machine Learning Approach Using Köppen-Geiger Zones

Abstract Global society is facing growing risks from extreme rainfall events. Obtaining accurate sub-daily rainfall data to support flood-related engineering designs can mitigate such risks, yet it is difficult to obtain worldwide. Temporal scaling offers a method to infer sub-daily extremes from available daily observations. The scaling behaviour is described by the parameter $$\:\beta\:$$ ( $$\:{\beta\:}_{o\:}$$ for observed, $$\:{\beta\:}_{p}$$ for model-projected, and $$\:{\beta\:}_{f}$$ for future values). Using a global gridded precipitation dataset, this study presents a 0.1° resolution model for estimating extreme rainfall scaling from daily to sub-daily durations under present (1979–2020) and future (2071–2100, RCP4.5 and 8.5) climate conditions worldwide. We first assess the influence of geographical and climatic variables—latitude, longitude, altitude, distance to coast, and Köppen–Geiger class—on $$\:{\beta\:}_{o\:}$$ . Then we use four machine learning models to estimate $$\:{\beta\:}_{p}$$ with Random Forest achieving the best performance ( r 2 > 0.95 across all climate types) and greatly outperforming the baseline linear model ( r 2 = 0.13). The model was further applied to estimate 4-hour, 30-year return period rainfall intensities across eight global cities under present and future climate. Results show stronger time-scaling (lower $$\:{\beta\:}_{o\:}$$ ) at higher latitudes, with KG classification being a key predictor. Under future RCP8.5 climate scenarios, projected intensities rise by 20–62% at illustrative sites. This is the first global, high-resolution study using daily rainfall and geographic data to estimate sub-daily extremes, offering a practical tool for assessing flood risks and guiding infrastructure design in ungauged regions.