The Ethiopian Rift Valley Lake Basin, a data-scarce region characterized by sparse and discontinuous rainfall observations. Reliable estimation of extreme rainfall remains a major challenge for hydrological design in data-limited regions. This study brings together satellite rainfall, extreme value analysis, and machine learning to derive spatially continuous Intensity–Duration–Frequency (IDF) relationships across the basin. Daily CHIRPS rainfall was bias-corrected using a hybrid cumulative distribution matching and LOESS approach, reducing basin-wide RMSE from 8.19 to 5.66 mm (≈31%) and improving correlation from 0.22 to 0.31. Annual maximum series from the corrected data were used to estimate rainfall intensities through L -moment-based Generalized Extreme Value modeling. These estimates were then linked to spatial and physiographic predictors using a Random Forest model, with grouped station-based validation to ensure robust spatial generalization. The approach explains about 84% of the variance in rainfall intensity and performs consistently across durations. Rainfall intensity decreases with increasing duration and increases with return period, while spatial patterns reveal localized high-intensity zones shaped by terrain variability. SHAP analysis identifies duration and event rarity as the dominant controls, with elevation and location providing additional influence. Uncertainty analysis shows that absolute variability decreases with duration, whereas relative uncertainty increases slightly, particularly in complex terrain. These results offer a practical basis for estimating reliable IDF relationships in data-scarce regions. • Bias correction reduces CHIRPS rainfall error by ∼31% across the basin. • Random Forest model captures ∼84% of rainfall intensity variability. • Duration and return period dominate rainfall intensity scaling. • Spatial uncertainty increases in complex terrain and northern regions.
Tesfaye et al. (Wed,) studied this question.