Abstract
Physical considerations and previous studies suggest that extremal dependence between ocean storm severity at two locations exhibits near asymptotic dependence at short inter-location distances, leading to asymptotic independence and perfect independence with increasing distance. We present a spatial conditional extremes (SCE) model for storm severity, characterising extremal spatial dependence of severe storms by distance and direction. The model is an extension of Shooter et al. 2019 (Environmetrics 30, e2562, 2019) and Wadsworth and Tawn (2019), incorporating piecewise linear representations for SCE model parameters with distance and direction; model variants including parametric representations of some SCE model parameters are also considered. The SCE residual process is assumed to follow the delta-Laplace form marginally, with distance-dependent parameter. Residual dependence of remote locations given conditioning location is characterised by a conditional Gaussian covariance dependent on the distances between remote locations, and distances of remote locations to the conditioning location. We apply the model using Bayesian inference to estimates extremal spatial dependence of storm peak significant wave height on a neighbourhood of 150 locations covering over 200,000 km2 in the North Sea.
Article PDF
Similar content being viewed by others
Avoid common mistakes on your manuscript.
References
Brown, B.M., Resnick, S.I.: Extreme values of independent stochastic processes. J. Appl. Probab. 14, 732–739 (1977)
Coles, S., Heffernan, J., Tawn, J.: Dependence measures for extreme value analyses. Extremes 2, 339–365 (1999)
Davison, A.C., Padoan, S.A., Ribatet, M.: Statistical modelling of spatial extremes. Stat. Sci. 27, 161–186 (2012)
Ewans, K.C., Jonathan, P.: The effect of directionality on northern North Sea extreme wave design criteria. J. Offshore Arct. Eng. 130, 041,604:1–041,604:8 (2008)
Ferreira, A., de Haan, L.: The generalized Pareto process; with a view towards application and simulation. Bernoulli 20(4), 1717–1737 (2014)
Ferro, C.A.T., Segers, J.: Inference for clusters of extreme values. J. R. Stat. Soc. B 65, 545–556 (2003)
de Fondeville, R., Davison, A.C.: Functional peaks-over-threshold analysis. arXiv:2002.02711 (2020)
Heffernan, J.E., Resnick, S.I.: Limit laws for random vectors with an extreme component. Ann. Appl. Probab. 17, 537–571 (2007)
Heffernan, J.E., Tawn, J.A.: A conditional approach for multivariate extreme values. J. R. Stat. Soc. B 66, 497–546 (2004)
Huser, R.G., Wadsworth, J.L.: Modeling spatial processes with unknown extremal dependence class. J. Am. Stat. Soc. 114, 434–444 (2019)
Keef, C., Papastathopoulos, I., Tawn, J.A.: Estimation of the conditional distribution of a vector variable given that one of its components is large: additional constraints for the Heffernan and Tawn model. J. Mult. Anal. 115, 396–404 (2013)
Kereszturi, M., Tawn, J., Jonathan, P.: Assessing extremal dependence of North Sea storm severity. Ocean Eng. 118, 242–259 (2016)
Ledford, A.W., Tawn, J.A.: Statistics for near independence in multivariate extreme values. Biometrika 83, 169–187 (1996)
Mardia, K.V., Kent, J.T., Bibby, J.M.: Multivariate Analysis. Academic Press, New York (1979)
Oceanweather: North European Storm Study User Group Extension and Reanalysis Archive. Oceanweather Inc. (2002)
Reich, B.J., Shaby, B.A.: A hierarchical max-stable spatial model for extreme precipitation. Ann. Appl. Stat. 6, 1430–1451 (2012)
Ribatet, M.: Spatial extremes: max-stable processes at work. J. Soc. Fr. Stat. 154, 156–177 (2013)
Roberts, G.O., Rosenthal, J.S.: Examples of adaptive MCMC. J. Comp. Graph. Stat. 18, 349–367 (2009)
Ross, E., Kereszturi, M., van Nee, M., Randell, D., Jonathan, P.: On the spatial dependence of extreme ocean storm seas. Ocean Eng. 145, 1–14 (2017a)
Ross, E., Randell, D., Ewans, K., Feld, G., Jonathan, P.: Efficient estimation of return value distributions from non-stationary marginal extreme value models using Bayesian inference. Ocean Eng. 142, 315–328 (2017b)
Schlather, M.: Models for stationary max-stable random fields. Extremes 5, 33–44 (2002)
Shooter, R.: Conditional Models for Spatial Extremes. PhD Thesis, Lancaster University, UK (2020)
Shooter, R., Ross, E., Tawn, J.A., Jonathan, P.: On spatial conditional extremes for ocean storm severity. Environmetrics 30, e2562 (2019)
Smith, R.L.: Max-stable processes and spatial extremes. Unpublished article, available electronically from www.stat.unc.edu/postscript/rs/spatex.pdf (1990)
Spiegelhalter, D.J., Best, N., Carlin, B.P., van der Linde, A.: Bayesian measures of model complexity and fit (with discussion). J. R. Stat. Soc. B 64, 1–34 (2002)
Tawn, J.A., Shooter, R., Towe, R., Lamb, R.: Modelling spatial extreme events with environmental applications. Spat. Stat. 28, 39–58 (2018)
Vallis, G.: Atmospheric and Oceanic Fluid Dynamics: Fundamentals and Large-Scale Circulation. Cambridge University Press, New York (2017)
Wadsworth, J., Tawn, J.: Dependence modelling for spatial extremes. Biometrika 99, 253–272 (2012)
Wadsworth, J.L., Tawn, J.A.: Higher-dimensional spatial extremes via single-site conditioning. In submission (2019)
Wadsworth, J.L., Tawn, J.A., Davison, A.C., Elton, D.M.: Modelling across extremal dependence classes. J. R. Stat. Soc. C 79, 149–175 (2017)
Young, I.R., Ribal, A.: Multiplatform evaluation of global trends in wind speed and wave height. Science 364, 548–552 (2019)
Acknowledgements
The authors would like to thank Jenny Wadsworth (Lancaster University) and Matthew Jones, David Randell and Ross Towe (Shell) for helpful discussions. We also thank two anonymous reviewers and an associate editor for comments on an earlier draft of the manuscript.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
R. Shooter would like to acknowledge financial support from Engineering and Physical Sciences Research Council grant EP/L015692/1 (STOR-i Centre for Doctoral Training) and Shell Research Ltd.
Appendices
Appendix A: Grouped adaptive MCMC
Under the ABN model (with piecewise linear distance-direction representations for α and β) the parameter set Ω to be estimated is \(\{\{\{\alpha _{k\ell },\beta _{k\ell }\}_{\ell =1}^{n_{\theta }},\mu _{k},\sigma _{k},\delta _{k}\}_{k}^{n_{d}},R_{1},R_{2}\}\) with nP = 2ndnθ + 3nd + 2 parameters.
Prior specification
Impose uniform prior distributions for each parameter. Explicitly, \(\alpha _{k\ell }\sim \text {Unif}(0,1)\), \(\beta _{k\ell }\sim \text {Unif}(0,1)\), \(\mu _{k}\sim \text {Unif}(-1,1)\), \(\sigma _{k}\sim \text {Unif}(0,\sqrt {2}+0.1)\) and \(\delta _{k}\sim \text {Unif}(0.5,2.5)\) for all k = 1, 2, … , nd, ℓ = 1, 2, … , nθ, and \(R_{1}\sim \text {Unif}(0,100)\), \(R_{2}\sim \text {Unif}(0,2)\),
Starting solution
We obtain a random starting solution Ω(0) by sampling the elements of Ω from their prior distributions, verifying that the starting solution has a valid likelihood (satisfying the conditional quantile constraints if applied).
Metropolis-within-Gibbs for n MiG iterations
Writing \(\varOmega _{k}^{(i)}\) as the value of the k th parameter of Ω at the i th iteration, use adaptive random walk Metropolis-within-Gibbs scheme for nMiG (> 2nP) iterations. That is, for i = 2, … , nMiG, update each \(\varOmega _{k}^{(i)}\) in turn, proposing candidate value \(\varOmega _{k}^{(i)c}\) from distribution
Grouped adaptive Metropolis-within-Gibbs for n GA iterations
For i > nMiG, use a grouped adaptive random walk Metropolis-within-Gibbs scheme, updating groups \(\boldsymbol {\varOmega }_{G_{k}}^{(i)}=\{\{\alpha _{k\ell },\beta _{k\ell }\}_{\ell =1}^{n_{\theta }},\mu _{k},\sigma _{k}\}\) jointly for k = 1, 2, … , nd in turn, before updating \(\{\delta _{k}\}_{k=1}^{n_{d}}\), R1 and R2 separately as before. Propose candidate \(\boldsymbol {\varOmega }_{G_{k}}^{(i)c}\) from distribution
where β = 0.05, as suggested by Roberts and Rosenthal (2009), and Ci is the empirical variance-covariance matrix of the parameters \(\boldsymbol {\varOmega }_{G_{k}}^{(i)}\) from the previous i iterations.
Iteration to convergence
Throughout, a candidate state is accepted using the standard Metropolis-Hastings acceptance criterion. Since prior distributions for parameters are uniform, and proposals symmetric, this is just a likelihood ratio. Candidates lying outside their prior domains, or violating the conditional quantile constraints if applied, are rejected.
Under the ABP model (with parametric forms for α and β with distance only), the estimation scheme is simplified so that the full set of parameters is updated at the grouped adaptive stage. The resulting procedure is the same as the original scheme of Roberts and Rosenthal (2009). Uniform priors \(A_{1\ell }\sim \text {Unif}(1,20)\), \(A_{2\ell }\sim \text {Unif}(0.1,5)\), \(B_{1\ell }\sim \text {Unif}(0.1,1)\), \(B_{2\ell }\sim \text {Unif}(1,5)\) and \(B_{3\ell }\sim \text {Unif}(0,20)\), ℓ = 1, 2, … , nθ were applied. The prior distributions for all other parameters under the ABN and ABP models are the same.
Appendix B: Conditional quantile constraints
We optionally restrict the space of feasible combinations of α and β to ensure that conditional quantiles from asymptotic independent models do not exceed those from asymptotic dependent models, as proposed by Keef et al. (2013). For any pair α and β corresponding to distance dk and direction θℓ, k = 1, 2, … , nd, ℓ = 1, 2, … , nθ on the distance-direction lattice, we require either
or
In the above, ν is a value of the conditioning variate (on standard Laplace scale) above the threshold level at which the SCE model is applied. Further z+(q) is the quantile of the distribution of standardised residuals from the conditional extremes model with non-exceedance probability q, and z+(q) is quantile of the distribution of standardised residuals from the conditional extremes model assuming asymptotic positive dependence (i.e. by imposing α = 1, β = 0) with non-exceedance probability q. In practice, as suggested by Keef et al. (2013), it is sufficient to satisfy the constraints above for q = 1 and ν equal to the maximum observed value of the conditioning variate.
Appendix C: Supporting diagnostic plots
Figure 10 supports the discussion in Section 4 around Fig. 8 regarding agreement between simulated spatial trajectories under the fitted SCE model, and observed trajectories. The figure shows 100 simulated trajectories for transects emanating to the North-East and to the North-West, from the central conditioning location, in black. The trajectories all correspond to a conditioning value x0 with non-exceedance probability in the interval (0.95, 0.96). Also shown are observed trajectories available (22 and 20 in number respectively for the left and right hand cases), for the same conditioning, in red.
Observed (black dashed) and simulated (red; grey in greyscale) trajectories from the fitted ABN model, for conditioning value x0 with non-exceedance probabilities in the interval (0.95,0.96]. Left: transect emanating from conditioning site to North-East; right: transect emanating from conditioning site to North-West. Numbers of observed trajectories are 22 (NE) and 20 (NW); 100 simulated trajectories shown in each case
Figure 11 illustrates the fitted conditional Gaussian correlation function of Eq. 6. Consider two remote locations, equidistant from a third conditioning location. For each of the curves in the figure, the x-coordinate of the black disc indicates the distance between the remote locations. The curve passing through each disc gives the value of conditional residual correlation as a function of the (common) distance of the remote locations from the conditioning site. Thus, the leftmost curve pertains to two remote locations approximately 30 km apart. When the distance of these locations to the conditioning site is large, the remote locations have large residual correlation since conditioning has effectively no effect: the residual process is essentially unconstrained, and the value of correlation determined by the powered exponential ρ. However, as distance to the conditioning site is reduced, residual correlation decreases, since more sample variation is explained by the SCE model (as opposed to the SCE conditional residual process). The minimum value of distance to conditioning site occurs when the three locations are collinear, with the conditioning site mid-way between remote locations. For this arrangement, conditioning induces a small negative residual correlation between remote locations. As the distance between remote locations increases, for different curves left to right, the maximum conditional residual correlation is reduced.
Illustration of the estimated Gaussian-scale residual correlation function. Consider a pair of remote locations with common distance to conditioning site. Distance between remote locations is indicated by x-coordinate of black disc for each curve. Curve through disc indicates the residual correlation between remote locations as a function of distance to the conditioning site
Deviance and DIC
The table below provides MCMC deviance statistics and DIC values to support the discussion in Section 4.
Model | nθ | Constraints | \(\overline {D(\varOmega )}\) | sd(D) | \(D(\bar {\varOmega })\) | DIC |
---|---|---|---|---|---|---|
ABN | 1 | Y | 1770 | 8 | 1740 | 1800 |
ABN | 6 | Y | 1550 | 11 | 1490 | 1620 |
ABP | 6 | Y | 1580 | 3 | 1570 | 1580 |
ABN | 1 | N | 1730 | 7 | 1760 | 1760 |
ABN | 6 | N | 1470 | 10 | 1422 | 1520 |
ABP | 6 | N | 1510 | 4 | 1500 | 1510 |
Columns in the table are as follows: nθ is the number of directional parameters in the model, ‘Constraints’ indicates whether the conditional quantile constraints of Keef et al. (2013) were imposed, \(\overline {D(\varOmega )}\) is the posterior mean deviance, sd(D) is its standard deviation and \(D(\bar {\varOmega })\) the deviance calculated using posterior mean parameters. In summary, there is evidence in favour of a directional model, since lower values of deviance and DIC are obtained for nθ = 6 compared with nθ = 1 both with and without application of conditional quantile constraints. DIC values for ABP are somewhat lower than for the corresponding ABN, but differences are small.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Shooter, R., Tawn, J., Ross, E. et al. Basin-wide spatial conditional extremes for severe ocean storms. Extremes 24, 241–265 (2021). https://doi.org/10.1007/s10687-020-00389-w
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10687-020-00389-w