Estimation of ocean environmental return values is critical to the safety and reliability of marine and coastal structures. For ocean waves and storm severity, return values are typically estimated by extreme value analysis of time series of measured or hindcast sea state significant wave height HS. For a single location, this analysis is complicated by the serial dependence of HS in time and its non-stationarity with respect to multiple covariates, particularly direction and season. Here, we report a non-stationary extreme value analysis of storm peak significant wave height HSsp, assumed temporally independent given covariates, incorporating directional and seasonal effects using a spline-based methodology incorporating an ensemble of models for different extreme value thresholds. Quantile regression is used to estimate suitable thresholds. For each threshold, a Poisson process is used to estimate the rate of occurrence of threshold exceedances, and a generalised Pareto model characterises the magnitude of threshold exceedances. Covariate effects are incorporated at each stage using penalised tensor products of B-splines to give smooth model parameter variation as a function of covariates. Optimal smoothing penalties are selected using cross-validation, and uncertainty is quantified using bias-corrected and accelerated bootstrap resampling. We use the model to estimate environmental return values for a location in the Makassar Strait, in the South China Sea. Return values distributions for HSsp are estimated by simulation under the threshold ensemble model. Return values for HS are then estimated by simulating intra-storm trajectories of HS consistent with the characteristics of the simulated storm peak events using a matching procedure. Return values for maximum individual crest elevation C are estimated by marginalisation using a pre-specified conditional distribution for C given HS and other sea state parameters. Model validation is performed by comparing confidence intervals for cumulative distribution functions of HSsp and HS for the period of the data with empirical sample-based estimates. © 2015 John Wiley & Sons, Ltd.