Bayesian geoadditive modelling of climate extremes with nonparametric spatially varying temporal effects

Abstract

Non-stationary modelling of climate extremes has attracted significant attention in recent years. Generalized extreme value (GEV) distribution is the standard approach for modelling block extremes. The non-stationary form of GEV distribution with location and scale parameters linearly regressing to time has been widely used for single-site time series of climate extremes. In the present paper, such a model is extended to be geoadditive for regional climate extremes with nonparametric spatially varying temporal effects. The model is implemented through the Bayesian hierarchical approach and is applied to the annual minimum surface air temperature in Mainland China as a demonstration. The five bivariate functions representing the regional regression parameters in the model are approximated by low-rank thin plate regression splines. All the necessary software tools for producing a thin plate regression spline basis and Markov Chain Monte Carlo sampling can be seamlessly integrated and called through the R software environment. Results show that the model can capture the spatial inhomogeneity and temporal non-stationarity features of regional climate extremes. Expressions for return levels with different return periods and their trends are also derived accordingly. Trends in different return levels, especially those in the 'extreme state' (e.g., 50-year return level) that cannot be inferred by usual observational trend analysis, can also be derived through samples drawn from posterior predictive distributions. The regional trend in a 50-year return level is quite different to that in a 2-year return level. More attention should be paid to trends in the extreme state of climate under climate change.

Spatio-temporal modelling of lightning climatologies for complex terrain

“…GAMs allow the inclusion of expert knowledge through the choice of the covariates. GAMs have been used to compile climatologies of extremes (ChavezDemoulin and Davison, 2005;Yang et al, 2016) and full precipitation distributions (Rust et al, 2013;Stauffer et al, 2016). A further benefit of GAMs is that all the parameters of a distribution, e.g.…”

Spatio-temporal Bayes Regression with INLA in Statistical Downscaling Modeling for Estimating West Java Rainfall

“…The Bayes method is one solution in representing these complex phenomena, because the complexity of the model can be represented by designing a hierarchical structure for data and its parameters. Several studies have focused on Bayes modeling with MCMC inference for spatio-temporal data, including [13] analyzed annual minimum temperatures for the past 6 decades in Mindland, China, and assumed the data to have GEV distribution, data is assumed to be stationary and arranged hierarchically. [4] used a dynamic linear model based on GEV distribution on monthly maximum wind speed data.…”

Non‐stationary climate changes in summer high‐temperature extremes in Shanghai since the late 19th century

“…Yee and Stephenson (2007) developed a nonparametric method within the vector generalized additive models (VGAM) framework that implements splines to fit a non-stationary GEV model. Yang et al (2016) studied the non-stationary changes in spatial and temporal extremes in China using two-dimensional smoothing splines for a GEV model within the Bayesian framework. The generalized additive models for location scale and shape (GAMLSS) framework provides a flexible framework with more general distributions other than the exponential family and nonparametric and semi-parametric solutions (Rigby and Stasinopoulos, 2005), and more smoothing and model selection suggestions have been proposed based on GAMLSS (Wood et al, 2016).…”