SHORT-TERM STATISTICS OF 10,000,000 WAVES OBSERVED BY BUOYS Merce Casas Prat1, Leo H. Holthuijsen2 and P. H. A. J. M.

van Gelder3
Time records of the surface elevation measured by four Waverider buoys in the Mediterranean Sea off the coast of Spain have been analysed to inspect the statistics of crest heights and wave heights. By concatenating the normalised records we obtained a long, quasi-stationary record of 10,000,000 waves, permitting a verification of the Rayleigh distribution and its theoretical variations at rather low levels of probability (wave heights up to 10 times the standard deviation of the surface elevation). The crest heights were almost perfectly Rayleigh distributed over the entire range of observation. The distribution of the wave heights is close to a Rayleigh distribution with scale factor 0.88 (rather than 1 as in the conventional Rayleigh distribution), but it is better approximated with a Weibull distribution with a shape factor 2.162 (rather than 2, as for a Rayleigh distribution). Supplementary observations with laser altimeters in the North Sea (10,000 waves) showed nearly identical results in the range of overlap (the normalised crest heights were slightly higher, showing a nonlinear behaviour). Introduction

The short-term statistics of wind-generated waves in deep water are usually based on the assumption that the sea surface elevation is a stationary, Gaussian process, resulting in the conventional Rayleigh distribution for wave heights and crest heights (Longuet-Higgins, 1952). However, field observations indicate that this Rayleigh distribution over-predicts the significant wave height by 7 - 8 % (e.g. Forristall, 1978; Holthuijsen, 2007) whereas the maximum crest height in a given duration is reasonably well predicted by the same theory with less than 2% error (Cartwright, 1958). With our rather long wave records of up to 10 million waves (obtained by concatenating normalised observed records of 20 minutes each), we can revisit the Rayleigh distribution at very low probabilities. Most of our data were measured with Waverider buoys in the Mediterranean Sea. To see the effect of using a fixed instrument, we supplemented our data with waves observed with laser altimeters from an offshore platform in the North Sea.
Observations

Most of our time series were measured by 4 Waverider buoys off the Catalan coast of Spain during a 15-year period (see Fig. 1 and Table 1). We are well aware of the fact that in the time records of the buoys the peaks of steep

Escola Tecnica Superior d' Enginyers de Camins , Canals i Ports, Technical University of Catalonia (UPC), Jordi Girona 1-3, 08034 Barcelona, Catalonia, Spain Faculty of Civil Engineering and Geosciences, Delft University of Technology, Stevinweg 1, 2628CN, Delft, the Netherlands Faculty of Civil Engineering and Geosciences, Delft University of Technology, Stevinweg 1, 2628CN, Delft, the Netherlands

560

COASTAL ENGINEERING 2008

561

waves tend to be flatter than they actually are and that buoys may be dragged through or swerve around the three-dimensional peaks of waves. However, most wave measurements at sea are made with buoys and understanding the statistics thus obtained is important.

Figure 1. The location of the four WAVERIDER buoys off the Catalan coast.

~T7
io>

55*

Figure 2. The location of the EDDA platform in the North Sea (from Allender et al. 1989).

562

COASTAL ENGINEERING 2008

Table 1 . Specifics of the buoy observations off the Catalan coast and of the altimeter observations in the North Sea. Roses 03 11.99 E 42 10.79 N 46 0.7 1/2.56 0.01 scalar buoy 20 2001-2006 Tordera 02 48.93 E 41 38.81 N 74 0.7 1/2.56 0.01 scalar buoy 20 2002-2006 Llobregat 02 08.48 E 41 16.69 N 45 0.7 1/2.56 0.01 scalar buoy 20 2001-2004 Tortosa 00 58.89 E 40 43.29 N 60 0.9 1/1.28 0.01 directional buoy 20 1991-1997 2001-2006 Edda 03 28E 56 28N 70

coordinates depth (m) diameter (m) sample interval (s) resolution (m) instrument record length (min) period (years)

1 0.001 laser altimeters 17min 4s 5-6/11/1985 21-23/12/1985

We subjected the buoy data to some rigorous tests. Of the original records, we accepted only those for which (in order of testing): 1. The record has the nominal length (see Table 1). 2. All absolute values of the vertical accelerations are </2g (the maximum acceleration in a Stokes wave, where g is gravitational acceleration). 3. After removing the linear trend in the record: no linear sections in the surface elevation occur (sequences of more than three consecutive data points with zero acceleration). 4. The significant wave height H =Aa (a is the standard deviation of the surface elevation in the record) is higher than 0.5 m (the size, shape and weight of the buoy may affect the measurement of smaller waves). No two consecutive data points are higher than 2.83// . A longer sequence is accepted as it may indicate the occurrence of a freak wave which we define as a wave with a height larger than 2.83// (this value is exceeded on average once per 100 storms, each with 2000 waves and a Benjamin-Feir index BFI - 0.8 ; see Discussion for definition). 6. Mean frequency / m 0 1 (based on the zero-th and first-order moment of the spectrum) is less than the Nyquist frequency fN = 1/2A/ divided by 2.2 (to avoid excessive sampling errors; At is the sample interval in the time record). 7. Low-frequency variance density is less than 0.004 m2/Hz (i.e., below 0.0065 Hz for the buoy data and 0.008 Hz for the altimeter data) to avoid records in which the buoy showed slow oscillations such as induced by a boat or flotsam hitting the buoy.

5.

COASTAL ENGINEERING 2008

563

The mean wave length (determined with the linear wave theory from the above mean frequency) is smaller than V2 the local water depth. We removed all records that did not pass these tests - we did not try to repair any records. In addition, we removed 16 unusual looking records. With these 9 criteria, we accepted 42,377 buoy records with a total of approximately 10 million waves. The highest individual wave height in this data set is 8.53 m and the highest significant wave height is 5.38 m. To find some indication of the differences with a fixed instrument, we supplemented our data with 92 time series obtained by altimeter observations during the WADIC project (Allender et. al, 1989; see acknowledgements) at the Phillips Edda platform (see Fig. 2). We did not censor these records - we only considered the requirement of shallow water (see above criteria), which caused 23 records to be rejected (including the ones with the highest significant wave height in November 1985). The remaining 69 records contained about 10,000 waves. The highest individual wave height in these records is 13.55 m and the highest significant wave height is 8.82 m.
Theory

8.

In the linear approach of Longuet-Higgins (1952), the crest height is Rayleigh distributed and, because the wave height is assumed to be twice the crest height, the wave height too is Rayleigh distributed:
P(nm) = Veres, e X p ( - X C
S

, ) = Pnml,Rayleish

( 1 )

and p(H) = y*Hex?(-KH2) = Pfi,Ra,e,gh C) in which the normalized crest height is fjcresl = rjmsl I yjm0 and the normalized wave height is H = HI yfm^ and m0 is the variance of the surface elevation. If the wave height is not assumed to be twice the crest height (instead the correlation between crest height and trough depth may be taken into account), the Rayleigh distribution for the wave height should be scaled: p(H) = rexp 2 4a v 8 a y with
2

(3) a depending on the spectrum, v= e.g.,

the
2

scaling -y2)v
2

factor

a =\-{y%n

in which the spectral width

^m0m2/mf-\

(Longuet-Higgins, 1980) and m0,mx and m2 are the zero-th, first- and secondorder moment of the variance density spectrum. Or, a2 = /2(\ - p) in which p is the correlation between the crest height and the trough depth of a wave (Naess, 1985). For large wave heights Vinje (1989) multiplies this distribution (with the same value for a ) with J%(l-p~]) and Tayfun (1990) multiplies this again

564

COASTAL ENGINEERING 2008

with

| l + (/72 - l ) / ( 4 p # 2 ) l . For weakly nonlinear waves, adaptations of the

Rayleigh distribution for the crest height have been suggested, e.g. by Mori and Janssen (2006) Pin**,, ) = [l + X (4| - 3) (l - , + X , )] Plc ,Ray,Clg>, which depends on the kurtosis XA of the surface elevation or, Tayfun (1994)
P(flcrest)
:

(4)

l + clAifjcresl(-fjles,-2)/j2

l{\-c^)p^msiMyleigh

(5)

which depends on the skewness/^ of the surface elevation ( q and c2 are constants). The cumulative distribution function of the maximum crest height or maximum wave height in a record of JV waves is readily obtained from the above, assuming that crest heights or wave heights are independent. The expressions are identical for the wave height and the crest height. Representing either of these as x, ^{xjnax<xmax} = P(xmax) = [P(x)\N
(

(6)

For the Rayleigh probability density function written as (7) 812 the expected value of the maximum is (Longuet-Higgins, 1952; Forristall, 1978)
s l

p(x) =

exp

where yF = 0.5772 is Euler's constant and 8 = 8a 2 and 8 = 2 for the wave height and the crest height respectively. Observed statistics We ranked the observed individual wave heights and crest heights and used the plotting position suggested by Goda (1988) for Rayleigh distributions Pr{x<x}~l-(i-a)/(N + b) (9) in which x is either the normalized crest height or normalized wave height and the values of the coefficients are a = 0.20 + 0.27 l-Jc and b = 0.20 + 0.23 / V c . with C = 2 for Rayleigh distributions. For the sake of clarity, we show the results for the wave heights that are nearest to multiples of /2yfm~0 and for crest heights that are nearest to multiples ofl/Ay[m~a. To show the full range, we added the highest value in each distribution. The results are shown in Fig. 3.

COASTAL ENGINEERING 2008

565

P(ricresl) 1-10

P(H)

a = 0.878

buoy observations (this study) scaled Rayleigh distributions with scalingfactor a

'crest

I \Jm, Icrest ' "*0

10 H -*-*

12
i J HI, '

\"*0

Figure 3. The distributions of the normalized crest height and wave height as observed by the buoys in the present study, the best-fit Rayleigh distributions, theoretical and empirical distributions. Conventional Rayleigh distributions a = 1 added for reference.

As expected, because the wave height is not twice the crest height, the observed distribution of the wave heights deviates considerably from the conventional Rayleigh distribution ( a = 1). A Rayleigh distribution with a scale factor a = 0.878 does fit the data reasonably well but the observations show a gentle S-curve around this (best-fit) Rayleigh distribution with slightly higher values for the low wave heights. A fit over a lower range of normalised wave heights (e.g., from 0 to 8), would therefore give a larger scaling factor than over the complete range of the observations (from 0 to 10). This may explain (part of)

566

COASTAL ENGINEERING 2008

the discrepancy between the above scaling factor a = 0.878 and the slightly larger scaling factors of Forristall (1978) and Holthuijsen (2007). The data are better approximated with the Weibull distribution suggested by Forristall (1978) with one coefficient slightly adapted (Forristall's coefficient value a = 2.126 is replaced here by a = 2.162; for a Rayleigh distribution a = 2). In the linear theories the distribution of the wave heights deviates from the conventional Rayleigh distribution only in terms of a re-scaling (except the correction by Tayfun, 1990 which modifies the shape of the distribution). With the average observed values of the spectral width v and correlation p (determined per wave record and averaged over all records), we find values of a = 0.94 and a = 0.89 for the theories of Longuet-Higgins (1980) and Naess (1985) respectively. Obviously, the latter result agrees better with our observations ( a = 0.878) than the former. The observed distribution of the crest heights agrees almost perfectly with the conventional Rayleigh distribution (Fig. 3): the data cluster along a straight line with only slight deviations over the entire range of observation and the scaling factor of the best-fit Rayleigh distribution is an almost perfect with a = 1.0001 (fitted to the data shown in Fig. 3). In the nonlinear theories, the distribution of the crest heights deviates from the conventional Rayleigh distribution but the average skewness and kurtosis of our data are so close to those of a Gaussian population that the differences predicted by these theories are barely visible in Fig. 3. The above analysis was repeated for the observations with the laser altimeters of the WADIC experiment and the results are shown in Fig. 4. The best-fit Rayleigh distributions through the altimeter data have scales a = 1.048 and a = 0.869 for the crest heights and wave heights respectively. Obviously, the crest heights are slightly higher in the altimeter observations than in the buoy observations (confirming the dynamic effect of the buoys on the observations) but perhaps more remarkable: the wave heights in the altimeter observations are almost identical to the wave heights in the buoy observations, thus confirming the Weibull-type deviation from the Rayleigh distribution in the buoy data. For the buoy data, we also considered the maximum crest height and wave height in a sequence of N consecutive waves from the normalised records (the mean of n sequences, each obtained by concatenating observed normalised records of nominal duration). The results of this and the theoretical estimates are given in Fig. 5. As expected, because the wave height is not twice the crest height, the observed average of the maximum wave heights (upper set of data in Fig. 5) deviates considerably from the conventional Rayleigh distribution ( a = 1). But the line based on a scaled Rayleigh distribution with a = 0.863 fits the observations almost perfectly. This scaling is close to the scaling based on the observed wave height distribution in Fig. 3 in which the best-fit is achieved with

COASTAL ENGINEERING 2008

567

a = 0.878. The observed average of the maximum crest heights (lower set of data in Fig. 5) agrees very well with the conventional Rayleigh approach ( a = 1): the data cluster along the line based on this approach with only slight deviations over the entire range of observation and the scaling factor a = 0.991 is close to unity (fitted to the data shown in Fig. 5). Again, the deviations predicted by the nonlinear theories are so small that they are barely visible in Fig. 5.
P

{V crest)

P(H)

,^ Rayleigh a =\

o A

buoy observations (this study) altimeter observations (this study) scaled Rayleigh distributions with scaling factor a

H = HlJmn
Figure 4. The distributions of the normalised crest height and wave height as observed by the buoys and the altimeter in the present study (over the range of the altimeter observations) and the Rayleigh distributions that best fit the altimeter observations (compare with Fig. 3). Discussion

The above results for the buoy observations show that the observed distribution of the crest heights agrees exceptionally well with the expected (conventional, i.e., a = 1) Rayleigh distribution. This supports the linear

568

COASTAL ENGINEERING 2008

approach of Longuet-Higgins (1952) and it is also predicted by the nonlinear theories in the sense that these theories reduce to the linear theory for Gaussian sea states (which our composite buoy data set is, with kurtosis = 3.01 and skewness = 0.022). The observed wave height distribution is obviously not the conventional Rayleigh distribution. Here, the correlation between crest height and trough depth in a wave is important and, given the average observed correlation in the observed time series p = -0.57, the theory of Naess (1985) with a = 0.89 gives a better agreement with the observed value a = 0.878 than the theory of Longuet-Higgins (1980) with a = 0.94 (using the average observed spectral width v = 0.41). However, the slightly adapted Weibull distribution suggested by Forristall (1978) fits the observed distribution even better.

E\H.

K
10-

Rayleigh a = 1

a = 0.991 Mori and Yasiida

buoy observations (this study) for scded Rayleigh distributions with scahngfactor a
105 10s

[ 95% confidence interval

10J

10'

Figure 5. The expected values of the maximum wave height and crest height (normalised) observed by the buoys as a function of the number of waves in a sequence, and the predictions. For the crest heights observed by the laser altimeters the agreement with the conventional Rayleigh distribution is not as good as for the crest heights observed by the buoys but still reasonable if some scaling is allowed (Fig. 4): the shape of the observed distribution is close to that of a Rayleigh distribution with a scaling factor a = 1.048. This scaling may be indicative of some degree of nonlinearity in the waves, probably obscured in the buoy observations by the buoy dynamics (see section Observations). The distribution of the wave heights observed by the laser altimeter is remarkably similar to the distribution observed by the buoys (Fig. 4). In fact, they are almost identical except for the very highest values, suggesting that the statistics of wave heights observed by buoys,

COASTAL ENGINEERING 2008

569

in contrast to those of the crest heights, are not seriously affected by the buoy motion. It would be convenient if the theories could be used to predict the above statistics for arbitrary conditions, and preferably from the spectrum, which can be predicted from wind fields. A necessary condition then is that these statistics depend on the correlation, the spectral width, or the skewness or the kurtosis of the surface elevation. To inspect this, we ranked all records in ascending order of kurtosis and divided them into 5 groups (each containing 20% of the total number of records). Then we computed the corresponding conditional expected maximum in a sequence of N waves. The result for this kurtosis selection is shown in Fig. 6 in which the indicated values of the kurtosis are the averages for each group.

E\Hm
E

\f]cre.

Rayleigh a = 1. . - -c

a = 0.863 Rayleigh a = l-.

= 3.25 = 3.08 = 2.99

bvoy observations (this study) 10= 10' 10s 106

= 2.92 = 2 80

10'

Figure 6. The dependency of the average maximum crest height and wave height on the number of waves in a sequence and kurtosis (in groups in ascending order, lines in same order from top to bottom as in legend).

It is obvious that the observed maxima are well organised in the sense that, at each value of N, the observed values rise monotonically with rising kurtosis. We found similar well behaved dependencies for the spectral width, correlation and skewness. Mori and Janssen (2006) estimate the kurtosis of the sea surface elevation from the spectrum with the Benjamin-Feir index (Janssen, 2003, 2005) in /t4 =3 + (xlj3\BFI2, where the Benjamin-Feir index BF1 = Qpi, mOl fht

570

COASTAL ENGINEERING 2008

which the spectral width parameter is Qp =2m02 f fE2(f)df

(Goda, 1970)

with the mean wave number and in which the wave steepness stmOl estimated from the mean frequency with the dispersion relationship for
\oi

deep water km0l = (2nfm0i) Ig. This suggests that the kurtosis can be determined with the BFI. Unfortunately, we found no correlation between the kurtosis and the BFI thus computed (see Fig. 7). (It must be noted that the theory of Mori and Yasuda has been derived for unidirectional, i.e., long-crested waves. When applied to real, short-crested waves, the effects of their theory are greatly reduced. In fact, for young sea states with a typical directional spreading of 25 - 30, the spectral estimate of the kurtosis would be reduced by as much as a factor 5.)

1.4

Figure 7. The kurtosis computed from the observed surface elevation (one data point corresponds to one buoy record) as a function of the Benjamin-Feir index ( BFI) computed from the spectrum.

Conclusions

Our analysis of 10,000,000 waves observed with buoys in the Mediterranean Sea shows that the crest heights are almost perfectly Rayleigh distributed without scaling ( a = 1.0001). Our analysis of 10,000 waves observed with laser altimeters in the North Sea shows an almost equally good correspondence for the crest heights but with a scaling factor of a = 1.048, indicative of some degree of nonlinearity (not seen by the buoys, probably because of the buoy dynamics). The distribution of the observed wave heights in both data sets is close to a

COASTAL ENGINEERING 2008

571

Rayleigh distribution with a scale factor a = 0.873 (average of the buoy data and the altimeter data). However, it is closer to a Weibull distribution with a shape parameter that is slightly larger than for a Rayleigh distribution (2.162 instead of 2), confirming a similar conclusion of Forristall (1978) for waves in five hurricanes in the Gulf of Mexico. The maximum crest height observed by a buoy in a sequence of waves agrees correspondingly with the Rayleigh distribution of the linear approach of Longuet-Higgins (1952) and the maximum wave height agrees similarly with the scaled Rayleigh distribution. Grouping the buoy data shows that the statistics correlate well with skewness, kurtosis, spectral width and the correlation between crest height and trough depth, suggesting that these statistics can be predicted with spectral parameters (presently under investigation).
Acknowledgements

We are grateful to the Xarxa dTnstruments Oceanografics i Meteorologies de la Generalitat de Catalunya (XIOM) and the participants of the WADIC experiment for their permission to use their data. We also want to thank Joan Pau Sierra Pedrico of UPC, Jesus Gomez Aguar of UPC and Steven Barstow of FUGRO-OCEANOR for their kind and very helpful assistance in accessing the data.
References

Allender, J., Audunson, T., Barstow, S. F., Bjerken, S., Krogstad, H. E., Steinbakke, P , Vartdal, L., Borgman, L. E. and Graham, C. 1989. The WADIC project: a comprehensive field evaluation of directional wave instrumentation. Ocean Engineering, Vol. 16, 5/6, 505-536. Cartwright, D. E. 1958. On estimating the mean energy of sea waves from the highest waves in a record. Proceedings of the Royal Society of London, Vol. 247, A, 22-48. Forristall, G. Z. 1978. On the statistical distribution of wave heights in a storm. Journal of Geophysical Research, Vol. 83, C5, 2353-2358. Goda, Y. 1970. Numerical experiments on wave statistics with spectral simulation. Port and Harbour Research Institute, Vol. 9, 3 Goda, Y. 1988. On the methodology of selecting design wave height. Proceedings of the 21s' International Conference on Coastal (Malaga), New York, ASCE, pp. 899-913. Holthuijsen, L. H. 2007. Waves in oceanic and coastal waters. Cambridge University Press. Janssen, P. A. 2003. Nonlinear four-wave interactions and freak waves. Journal of Physical Oceanography,. Vol. 33. 4. 863-884. Longuet-Higgins, M. S. 1952. On the statistical distribution of the heights of sea waves. Journal of Maritime Research. Vol. XI, 3, 245-267.

572

COASTAL ENGINEERING 2008

Longuet-Higgins, M. S. 1980. On the distribution of the heights of sea waves: some effects of nonlinearity and finite band width. Journal of Geophysical Research. Vol. 85, C3, 1519-1523. Mori, N. and Janssen, P. A. 2006. On kurtosis and occurrence probability of freak waves. Journal of Physical Oceanography, Vol. 36, 7, 1471-1483. Naess, A. 1985. On the distribution of crest to trough wave heights. Ocean Engineering, Vol. 105, 3, 221-234. Tayfun, M. A. 1990. Distribution of large wave heights. Journal of Waterways, Ports and Coastal Engineering, ASCE, Vol. 116, 6, 686-707. Tayfun, M. A. 1994. Distributions of envelope and phase in weakly nonlinear random -waves. Journal of Engineering Mechanics, Vol. 120, 5, 1009-1025. Vinje, T. 1989. The statistical distribution of wave heights in a random seaway. Applied Ocean Research, Vol. 11,3, 143-152.