Adaptive distribution selection and bootstrap uncertainty in intensity–duration–frequency curves with limited records

Abstract Intensity–duration–frequency (IDF) curves are the basis of hydrological design in small catchments, but they remain difficult to construct in regions with short rainfall records. Conventional methods apply a single probability distribution to all storm durations and report point estimates without quantifying uncertainty, which can produce biased extrapolations and overconfident design values. We present an IDF framework that (i) segments events into short (5–120 min) and long (120–1440 min) duration regimes, (ii) selects the optimal distribution for each duration through a multi-criteria voting scheme combining the Anderson–Darling, AIC and Kolmogorov–Smirnov tests, and (iii) propagates parameter uncertainty by bootstrap resampling with 1000 iterations and regional pooling. Applied to 14 years of sub-daily rainfall records (2012–2025) from Goiânia, Brazil, the framework indicated regime shifts: Log-Pearson III prevails in 64.3% of the analysed durations; GEV is selected in 22.9% of durations, concentrated at the convective extremes; and Gumbel accounts for 12.7%, in the transition band. The duration-decay exponent ranges from 1.5–2.0 in the convective regime to 0.95 in the stratiform regime, marking the convective–stratiform transition. The GEV-based IDF model reached R² ≥ 0.98 in both segments, with bootstrap confidence bands of ± 13–21 mm h⁻¹ at the 5-min duration and ± 1–3 mm h⁻¹ at 1440 min, for return periods between 2 and 100 years. The framework supports evidence-based hydrological design and adaptation in tropical regions with limited data.