A Bayesian workflow for multi-station IDF curve development in data-limited island settings: application to Grenada

Abstract Small island states in the Caribbean face disproportionate climate risks, yet rainfall observations remain sparse, fragmented, and often insufficient for spatially representative intensity-duration-frequency (IDF) analysis. In Grenada, engineering practice has therefore relied largely on single-station IDF curves despite strong topographic controls on rainfall and substantial uncertainty associated with short, incomplete records. This study develops a Bayesian workflow for multi-station IDF curves and spatially distributed daily rainfall extremes in a data-limited island setting. A site-specific correlation diagnostic is introduced to evaluate the spatial structure and relative non-stationarity of daily rainfall, clarifying dependence patterns that are poorly resolved by conventional semivariogram-based approaches. Missing daily rainfall values are then imputed using ordinary kriging parameterized by non-stationary spatial correlation models (SCMs) inferred via Bayesian inference, after which rainfall extremes are estimated using Bayesian generalized extreme value (GEV) and generalized Pareto (GPD) formulations. The results show that daily rainfall dependence is controlled primarily by separation distance, with elevation exerting a secondary influence that is often moderated by temporal non-stationarity. Ordinary kriging using Bayesian-inferred spatial correlation models provided the strongest overall imputation performance among the deterministic, geostatistical, and machine-learning-based alternatives considered. Geostatistical interpolation of posterior return levels produced island-wide mean and standard deviation fields that were spatially coherent and broadly consistent with topographic rainfall gradients and Grenada’s known climatic zoning. The framework provides a transferable probabilistic basis for rainfall extreme estimation in Grenada and other data-limited regions.