Sea Level Measurement <strong>and</strong> Interpretati<strong>on</strong><strong>sea</strong> <strong>level</strong>s. It begins with the classical method of AnnualExtremes, which first appeared in the early 1960s <strong>and</strong>c<strong>on</strong>tinued to be developed for some time thereafter.Following this, the Joint Probability Method, which wasdeveloped in the late 1970s, is c<strong>on</strong>sidered. This makesmore efficient use of data by incorporating our extensiveknowledge of the tides <strong>and</strong> storm surges, whichare the two main comp<strong>on</strong>ents of <strong>sea</strong> <strong>level</strong>, as a part ofthe estimati<strong>on</strong> procedure.More recent work <strong>on</strong> the Annual Exceedance Method isdiscussed, followed by a revisi<strong>on</strong> of the Joint ProbabilityMethod to correct its deficiencies in areas where the <strong>sea</strong><strong>level</strong> is dominated by the meteorological surge comp<strong>on</strong>ent.Finally, very recent work <strong>on</strong> the spatial estimati<strong>on</strong>of extremes is menti<strong>on</strong>ed. References are given at eachstage so that the reader can examine any of the methodsin greater depth. Although extreme high <strong>sea</strong> <strong>level</strong>sare c<strong>on</strong>sidered, results for extreme low <strong>sea</strong> <strong>level</strong>s can beobtained in an analogous way.2.8.2 The Annual Maximum Method (AMM)This is the classical general method of analysis ofextremes having been applied to <strong>sea</strong> <strong>level</strong> estimati<strong>on</strong>since 1963 (Lenn<strong>on</strong>, 1963; Suth<strong>on</strong>s, 1963). It is based<strong>on</strong> a result from probabilistic extreme value theorywhich states: if X 1 ,... X n is a sequence of independent<strong>and</strong> identically distributed r<strong>and</strong>om variables, thenmax(X 1 ,... X n ), suitably linearly normalized, c<strong>on</strong>vergesas n ∞ , to a r<strong>and</strong>om variable with a distributi<strong>on</strong>functi<strong>on</strong> which is <strong>on</strong>e of the so called extremevaluedistributi<strong>on</strong>s. The general case is known as theGeneralized Extreme Value (GEV) distributi<strong>on</strong>. Animportant special case is the Gumbel distributi<strong>on</strong>.The Annual Maximum Method takes the GEV to bethe distributi<strong>on</strong> functi<strong>on</strong> of the maximum <strong>sea</strong> <strong>level</strong> ina year. Therefore, for a place of interest, the annualmaximum for each year is extracted from hourly observati<strong>on</strong>s<strong>and</strong> is used as data to estimate the parametersof the distributi<strong>on</strong> that they follow. From the estimateddistributi<strong>on</strong> <strong>on</strong>e can obtain the <strong>sea</strong> <strong>level</strong> corresp<strong>on</strong>dingto a chosen ‘Return Period’. In practice, return periodsof 50, 100 <strong>and</strong> 1,000 years are comm<strong>on</strong>. The basicmethod assumes that there is no trend in the data, butit can be extended to deal with those cases where atrend is present.A recent extensi<strong>on</strong> of the annual maximum methodinvolves using probabilistic extreme value theory toobtain the asymptotic joint distributi<strong>on</strong> of a fixed number(r) of the largest independent extreme values, forexample the five largest in each year. Essentially theapproach is the same as above except that more relevantdata are included in the analysis thereby improvingthe estimati<strong>on</strong>. Care must be taken to ensure that thenumber of annual maxima ‘r’ is not excessive, such thatthe lower extremes fall outside the tail of the extremevalue distributi<strong>on</strong>.This method of estimating <strong>sea</strong> <strong>level</strong> extremes is highlyinefficient in its use of data, since it extracts very fewvalues from each yearly record. This is particularly importantwhen the <strong>sea</strong> <strong>level</strong> record is short, since it yieldsreturn <strong>level</strong> estimates with unacceptably large st<strong>and</strong>arderrors. In additi<strong>on</strong>, it makes no use of our knowledge ofthe <strong>sea</strong> <strong>level</strong> <strong>and</strong> storm surge processes. However, theadvantage of annual maxima methods is that they d<strong>on</strong>ot require knowledge of tide–surge interacti<strong>on</strong> whichcan sometimes be a significant feature of the data.C<strong>on</strong>sequently the methods are relatively straightforwardto apply.2.8.3 The Joint Probabilities Method (JPM)This method of analysis was introduced to exploit ourknowledge of the tide in short data sets to which theannual maxima method could not be applied (Pugh<strong>and</strong> Vassie, 1979). At any time, the observed <strong>sea</strong> <strong>level</strong>,after averaging out surface waves, has three comp<strong>on</strong>ents:mean <strong>sea</strong> <strong>level</strong>, tidal <strong>level</strong> <strong>and</strong> meteorologicallyinduced <strong>sea</strong> <strong>level</strong>. The latter is usually referred to as astorm surge. Using st<strong>and</strong>ard methods, the first two ofthese comp<strong>on</strong>ents can be removed from the <strong>sea</strong> <strong>level</strong>sequence leaving the surge sequence, which is just thetime-series of n<strong>on</strong>-tidal residuals. For simplicity these areassumed to be stati<strong>on</strong>ary. Because the tidal sequenceis deterministic, the probability distributi<strong>on</strong> for all tidal<strong>level</strong>s can be generated from tidal predicti<strong>on</strong>s. Thisdistributi<strong>on</strong> can be accurately approximated using 18.6years of predicti<strong>on</strong>s.The probability distributi<strong>on</strong> of hourly <strong>sea</strong> <strong>level</strong>s can beobtained either directly using an empirical estimate orby combining the tidal <strong>and</strong> surge probability densityfuncti<strong>on</strong>s (pdf). The latter is preferable, as it smoothes<strong>and</strong> extrapolates the former. However the nature ofthe combinati<strong>on</strong> of the pdf’s depends <strong>on</strong> whetherthere is dependence between the tide <strong>and</strong> surgesequences. Initially, c<strong>on</strong>sider the case in which theyare independent.By combining the pdf’s of tide <strong>and</strong> surge, the distributi<strong>on</strong>functi<strong>on</strong> of hourly (instantaneous) <strong>sea</strong> <strong>level</strong>s is obtained.From this, the distributi<strong>on</strong> functi<strong>on</strong> of the annual maximais required. If hourly values were independent, whichis approximately the case where the tide dominates theregime, then this is straightforward.The method has been widely applied. It makes betteruse of the data <strong>and</strong> of our extensive knowledge of thetides, <strong>and</strong> accounts for surges that could have occurred<strong>on</strong> high tide but by chance did not. Most successfulapplicati<strong>on</strong>s have been to sites which have several yearsof hourly records (>10 years) <strong>and</strong> where the site is tidallydominant, i.e. where the tidal range is large in comparis<strong>on</strong>to the surge amplitude. Least successful applicati<strong>on</strong>shave been to sites with both short lengths of data <strong>and</strong>where the site is surge dominant.8IOC <str<strong>on</strong>g>Manual</str<strong>on</strong>g>s <strong>and</strong> Guides No 14 vol IV
Sea Level Measurement <strong>and</strong> Interpretati<strong>on</strong>2.8.4 The Revised Joint Probabilities Method(RJPM)Particular emphasis was given to two principalimprovements that make the revised method morewidely applicable than the original joint probabilitiesmethod (Tawn et al., 1989). It was principally directedat sites where the storm surge was resp<strong>on</strong>siblefor a respectable proporti<strong>on</strong> of the <strong>sea</strong> <strong>level</strong> <strong>and</strong> toimprove the estimati<strong>on</strong> procedure for sites whereless than 10 years of data were available.The first issue was that of c<strong>on</strong>verting the hourly distributi<strong>on</strong>into annual return periods. It is clear thateach hourly value of <strong>sea</strong> <strong>level</strong> is not independent ofits predecessor or successor. Of the 8,760 hourly valuesin a year, it is necessary to determine the effectivenumber of independent observati<strong>on</strong>s per year. Thiswas d<strong>on</strong>e through an Extremal Index which is derivedfrom the mean overtopping time of a <strong>level</strong> for eachindependent storm which exceeds that <strong>level</strong>. In factthe Extremal Index can be shown to be a c<strong>on</strong>stantin the regi<strong>on</strong> of the extremes. Because large valuestend to cluster as storms, it should be expected thatthe Extremal Index >1; for example, in the NorthSea, it is 1.4. This effectively reduces the number ofindependent observati<strong>on</strong>s from 8,760 to 8,760/1.4.If the site is tidally dominant then the Extremal Indexis c<strong>on</strong>siderably smaller than if the site is surge dominant.The immediate advantages of this modificati<strong>on</strong>are: firstly, that no assumpti<strong>on</strong> about the localdependence of the process is required; sec<strong>on</strong>dly, thatthe c<strong>on</strong>versi<strong>on</strong> from the hourly distributi<strong>on</strong> to annualmaxima is invariant to sampling frequency.The sec<strong>on</strong>d modificati<strong>on</strong> enabled probabilities for<strong>level</strong>s bey<strong>on</strong>d the existing range of the surge datato be obtained, in additi<strong>on</strong> to providing smoothingfor the tail of the empirical distributi<strong>on</strong>. The methodis based <strong>on</strong> the idea of using a fixed number ofindependent extreme surge values from each year toestimate probabilities of extreme surges. The procedureinvolves two important steps. Firstly, the identificati<strong>on</strong>of independent extreme surges. Sec<strong>on</strong>dly,the selecti<strong>on</strong> of a suitable number of independentextreme surges from each year of data, perhaps fiveper year. Using these surge data, estimates can bemade of the parameters of the distributi<strong>on</strong> of theannual maximum surge (Smith, 1986).When interacti<strong>on</strong> is present, the <strong>level</strong> of the tideaffects the distributi<strong>on</strong> of the surge. In particular, thetail of the surge pdf depends <strong>on</strong> the corresp<strong>on</strong>dingtidal <strong>level</strong>. Thus the c<strong>on</strong>voluti<strong>on</strong> of tide <strong>and</strong> surgecan be adapted so that the surge parameters arefuncti<strong>on</strong>s of tidal <strong>level</strong>. This formulati<strong>on</strong> also enablesstatistical tests of independence to be performed.2.8.5 The Exceedance Probability Method(EPM)An alternative method of obtaining extreme <strong>sea</strong><strong>level</strong> estimates from short data sets is called theexceedance probability method (EPM) (Middlet<strong>on</strong>et al., 1986; Ham<strong>on</strong> et al., 1989). The EPM, like theRJPM, involves combining the tide <strong>and</strong> surge distributi<strong>on</strong>s<strong>and</strong> accounting for dependence in the <strong>sea</strong><strong>level</strong> sequence. The approach differs in the way thatit h<strong>and</strong>les extreme surges. The EPM uses results forc<strong>on</strong>tinuous time processes <strong>and</strong> makes assumpti<strong>on</strong>sabout the joint distributi<strong>on</strong> of the surge <strong>and</strong> its derivative.Improvement is achieved by allowing flexibilityin the surge tail through the use of a c<strong>on</strong>taminatednormal distributi<strong>on</strong>.2.8.6 Spatial Estimati<strong>on</strong> of ExtremesExtreme <strong>sea</strong> <strong>level</strong>s al<strong>on</strong>g a coastline are typically generatedby the same physical mechanisms, so the parametersthat describe the distributi<strong>on</strong> are likely to bespatially coherent. Models that describe the separatec<strong>on</strong>stituents of the <strong>sea</strong> <strong>level</strong> are best suited to exploitingthis spatial coherence, as the individual parametersshould change smoothly al<strong>on</strong>g a coastline.The joint distributi<strong>on</strong> of annual maxima over severaldata sites can be modelled using a multivariateextreme-value distributi<strong>on</strong> (Tawn, 1992). Changes ineach of the parameters of the distributi<strong>on</strong>, over sites,can be modelled to be c<strong>on</strong>sistent with the propertiesof the underlying generating process identified fromthe RJPM. The main advantage of the spatial methodis that it can utilize data sites with extensive <strong>sea</strong> <strong>level</strong>records <strong>and</strong> augment these with data from sites withshorter records of a few years.Using the ideas for extremes of dependent sequences,this can be related to the distributi<strong>on</strong> functi<strong>on</strong>of hourly surge <strong>level</strong>s, <strong>and</strong> then the empirical surgedensity functi<strong>on</strong> can be replaced by the adjusteddensity. Using the adjusted density functi<strong>on</strong>, thec<strong>on</strong>voluti<strong>on</strong> can be performed to combine the tidal<strong>and</strong> surge distributi<strong>on</strong>s to obtain the hourly <strong>sea</strong> <strong>level</strong>distributi<strong>on</strong> <strong>and</strong> hence the return periods can becalculated for different <strong>level</strong>s.IOC <str<strong>on</strong>g>Manual</str<strong>on</strong>g>s <strong>and</strong> Guides No 14 vol IV9