Abstract
<title>Abstract</title> <p>Extreme rainfall in the West African Sahel has intensified in recent decades, rendering conventional stationary extreme value models inadequate. Existing nonstationary approaches rely on predefined climate covariates whose relationships with Sahel rainfall are uncertain and unstable. This study develops a Bayesian Gaussian Process–Generalized Extreme Value (GP–GEV) framework that models temporal evolution of GEV parameters as latent Gaussian processes, learning nonstationarity directly from data without requiring physical covariates. The framework is validated through simulation experiments across eight data-generating regimes and applied to annual maximum daily rainfall from ERA5 dataset. Results show that while stationary GEV models remain competitive under stationarity, the GP–GEV framework substantially outperforms parametric-trend alternatives when latent nonlinear variability is present, achieving superior predictive accuracy, improved return level estimation, and better calibrated uncertainty intervals for high return periods. Empirical analyses reveal strong spatial heterogeneity, with several regions showing pronounced increases in 50-year return levels relative to stationary benchmarks. These findings demonstrate that latent-process Bayesian models offer a flexible, statistically robust alternative to covariate-based methods in regions with uncertain climatic drivers, underscoring the importance of uncertainty-aware frameworks for resilient infrastructure design and climate adaptation policy in the Sahel.</p>