Upload
others
View
8
Download
0
Embed Size (px)
Citation preview
Statistical Downscaling Prediction of Sea Surface Winds over the Global Ocean
CANGJIE SUN AND ADAM H. MONAHAN
School of Earth and Ocean Sciences, University of Victoria, Victoria, British Columbia, Canada
(Manuscript received 11 October 2012, in final form 7 March 2013)
ABSTRACT
The statistical prediction of local sea surface winds from large-scale, free-tropospheric fields is investigated
at a number of locations over the global ocean using a statistical downscaling model based on multiple linear
regression. The predictands (the mean and standard deviation of both vector wind components and wind
speed) calculated from ocean buoy observations on daily, weekly, andmonthly scales are regressed on upper-
level predictor fields from reanalysis products. It is found that in general the mean vector wind components
are more predictable than mean wind speed in the North Pacific andAtlantic, while in the tropical Pacific and
Atlantic the difference in predictive skill between mean vector wind components and wind speed is not
substantial. The predictability of wind speed relative to vector wind components is interpreted by an idealized
model of the wind speed probability density function, which indicates that in the midlatitudes the mean wind
speed is more sensitive to the vector wind standard deviations (which generally are not well predicted) than to
the mean vector winds. In the tropics, the mean wind speed is found to be more sensitive to the mean vector
winds. While the idealized probability model does a good job of characterizing month-to-month variations in
the mean wind speed in terms of the vector wind statistics, month-to-month variations in the standard de-
viation of speed are not well modeled. A series of Monte Carlo experiments demonstrates that the in-
consistency in the characterization of wind speed standard deviation is the result of differences of sampling
variability between the vector wind and wind speed statistics.
1. Introduction
Sea surface winds play a central role in influencing the
exchange of heat, momentum, and mass between the
ocean and the atmosphere (e.g., Garratt 1992; Bates
et al. 2001; Jones and Toba 2001; Donelan et al. 2002).
As well, sea surface winds represent a potentially sig-
nificant energy resource (e.g., Liu et al. 2008; Capps and
Zender 2009) and high sea surface winds represent
hazards to shipping (e.g., Sampe and Xie 2007). While
the output of global climate models represents the best
tool available for studying large-scale climate variabil-
ity, it is generally not directly relevant for inferences
about local climate. The coarse resolution and approx-
imate parameterizations of subgrid-scale processes both
limit the accuracy of the representation of local vari-
ability, especially in the planetary boundary layer. In
particular, local surface winds may be influenced by
local small-scale processes that are not resolved well in
climate models. The process of downscaling is designed
to relate local, small-scale variability to variability on
large scales. Dynamical downscaling approaches this
problem by nesting finely resolved, local dynamical
models within coarsely resolved, large-scale models. In
contrast, statistical downscaling (SD) is a complemen-
tary strategy employed to empirically downscale large-
scale variability through statistical methods. Although
dynamical downscaling has the merits of being physi-
cally based and not assuming a stationary climate, po-
tentially significant drawbacks such as possible errors
associated with imperfect parameterizations of key
processes (e.g., clouds and boundary layer processes),
systematic biases, coarse spatial resolution, and ex-
tremely high computational demands constrain the use
of pure dynamical downscaling. Although SD has the
weakness that it assumes statistically stationary re-
lationships between large-scale and local variables (as
represented by historical observations), it is inexpensive
and easy to implement.
This study considers the statistical relationships be-
tween local ocean winds and large-scale free-tropospheric
Corresponding author address:AdamMonahan, School of Earth
and Ocean Sciences, University of Victoria, P.O. Box 3065 STN
CSC, Victoria BC V8W 3V6, Canada.
E-mail: [email protected]
7938 JOURNAL OF CL IMATE VOLUME 26
DOI: 10.1175/JCLI-D-12-00722.1
� 2013 American Meteorological SocietyUnauthenticated | Downloaded 12/23/21 03:00 PM UTC
circulation based on buoy observations and reanalysis
products. Following earlier work by Monahan (2012a)
and Culver and Monahan (2013), SD is used to in-
vestigate the predictability of sea surface winds (both
wind speed and vector wind components) measured at
buoys over the global oceans on daily, weekly (10 days),
and monthly time scales. Along with the absolute pre-
dictive skills of surface-wind statistics, the predictability
of the statistics of wind speed relative to those of the
vector wind components is considered in this study.
A simple multiple linear regression will be used for
the SD in place of a more sophisticated analysis (e.g.,
stepwise linear regression) in order to minimize the
chances of overfitting the model. With appropriate cross-
validation, multiple linear regression produces statisti-
cally robust prediction models. In the present study, the
number of statistical degrees of freedom available for
building the statistical model can be fairly small (e.g.,
12 years of 3 months for a given season at a typical buoy).
The number of model parameters increases with the
complexity of a statisticalmodel, requiring larger datasets
for their robust estimation and to avoid overfitting the
model. The use of a relatively simple statistical model
reduces the potential risk of skill inflation due to model
overfitting.
This study has four primary goals: 1) Characterize the
predictive information that free-tropospheric large-
scale predictors carry for the statistics of local sea sur-
face winds across a range of wind climates. 2) Explore
the predictive skills of different surface-wind statistics
(mean and standard deviation) on different temporal
scales (daily, weekly, and monthly). 3) Investigate the
relationship between the predictability of mean wind
speed and that of the vector wind components presented
in Culver and Monahan (2013) over a wider range of
wind climate using an idealized probability distribution
model (IPM) of the wind speed probability density
function introduced by Monahan (2012a). 4) Assess the
modeling skill of the IPM for the submonthly standard
deviation of wind speeds across the global ocean.
Monahan (2012a) investigated the predictability of
surface winds in the subarctic northeast Pacific off of
western Canada, while Culver and Monahan (2013)
studied the statistical predictability of historical land
surface winds over central Canada. Other studies (e.g.,
Salameh et al. 2009; van der Kamp et al. 2012) have
investigated the predictability of various vector wind
components in regions of complex topography. In con-
trast to land surface winds, sea surface winds are less
influenced by stationary local features (e.g., topography
or fixed surface inhomogeneities). Therefore, the con-
nection between sea surface winds and upper-level
large-scale atmospheric fields is expected to be simpler
than that for surface winds over land. Furthermore, the
range of wind climates is much greater over oceans than
over land (because of the much weaker surface drag
over water). Thus, consideration of sea surface winds
allows for the analysis of surface wind SD in a relatively
idealized setting over a relatively large parameter range.
This study does not consider the temporal structure of
winds [which is considered in detail in Monahan
(2012b)]. The focus is on the ‘‘instantaneous’’ prediction
of surface wind statistics on various averaging time
scales from large-scale free tropospheric predictors on
the same time scale.
Section 2 describes the data used in this study, and
section 3 presents both the methodology to be used and
the prediction results of surface-wind statistics. Section 4
introduces the idealized model used to understand the
relationship between the predictability of the statistics
of the vector wind components and wind speed. A dis-
cussion and conclusions are presented in section 5.
2. Data
This study assesses the cross-validated statistical pre-
dictability of the statistics of historical sea surface wind
observations from a total of 52 moored ocean buoys (the
predictands) using free-tropospheric large-scale circula-
tion data from global reanalysis products (the predictors).
a. Buoy data
The buoys considered in this study are situated in the
tropical and North Pacific and Atlantic Oceans, with
data durations of between 8 and 28 years (Table 1). The
SouthernOcean and IndianOcean are not considered in
this study as we were not able to find buoy observations
in those locations of sufficient duration to establish ro-
bust statistical relationships with the flow aloft. The
wind and direction data from the 52 buoys were ob-
tained from four sources:
1) Prediction and Research Moored Array in the At-
lantic (PIRATA) project 10-min averaged data
(measured at 3–4m above mean sea level) from five
buoys in the tropical Atlantic (downloaded from
http://www.pmel.noaa.gov/pirata/);
2) Tropical Atmosphere Ocean (TAO)/Triangle Trans-
Ocean Buoy Network (TRITON) project 10-min
averaged wind data (measured at 3–4m above mean
sea level) from 12 buoys in the tropical Pacific
(downloaded from http://www.pmel.noaa.gov/tao/
disdel/disdel-pir.html);
3) National Data Buoy Center (NDBC) hourly reports
of 8-min averaged data (approximately 5m above
mean sea level) from 31 buoys off of the west and east
15 OCTOBER 2013 SUN AND MONAHAN 7939
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
coast of North America (downloaded from http://
www.ndbc.noaa.gov/); and
4) Japan Meteorological Agency (JMA) three-hourly
reports of 10-min averaged data (approximately 5m
above mean sea level) from four buoys in the north-
west Pacific; (downloaded from http://www.data.
kishou.go.jp/kaiyou/db/vessel_obs/data-report/html/
index_e.html).
TABLE 1. The location, duration, anemometer height, and data archive of the buoys considered in this study. Buoy identification (ID) is
given where applicable.
Buoy ID Latitude (8N) Longitude (8W) Duration
Anemometer
height (m) Source
41001 32.31 75.48 1973–2001 5 NDBC
41004 32.50 79.10 1978–98 5 NDBC
41006 29.30 77.40 1982–96 5 NDBC
41009 28.52 80.17 1988–98 5 NDBC
41010 28.91 78.47 1988–99 5 NDBC
42001 25.89 89.66 1975–2001 5 NDBC
42002 25.79 93.67 1973–2001 5 NDBC
42003 26.04 85.61 1976–2001 5 NDBC
42019 27.91 95.35 1990–98 5 NDBC
42020 26.97 96.70 1990–2001 5 NDBC
44004 38.48 70.43 1977–98 5 NDBC
44005 43.19 69.14 1978–98 5 NDBC
42007 30.09 88.77 1981–99 5 NDBC
51001 23.45 162.28 1981–2000 5 NDBC
51002 17.09 157.81 1984–98 5 NDBC
51003 19.35 160.62 1984–98 5 NDBC
51004 17.53 152.38 1984–98 5 NDBC
46002 42.59 130.47 1975–98 5 NDBC
46035 57.07 177.75 1985–97 5 NDBC
46005 46.10 131.00 1976–2000 5 NDBC
46041 47.35 124.73 1987–98 5 NDBC
46030 40.42 124.53 1984–98 5 NDBC
46028 35.74 121.88 1983–96 5 NDBC
46027 41.85 124.38 1983–2001 5 NDBC
46025 33.75 119.05 1982–98 5 NDBC
46026 37.76 122.83 1982–99 5 NDBC
46023 34.71 120.97 1982–99 5 NDBC
46042 36.79 122.40 1987–98 5 NDBC
46012 37.36 122.88 1980–98 5 NDBC
46013 38.24 123.30 1981–98 5 NDBC
— 12.00 38.00 2000–09 4 PIRATA
— 8.00 38.00 2001–09 4 PIRATA
— 4.00 38.00 2001–09 4 PIRATA
— 0.00 35.00 2001–08 4 PIRATA
— 15.00 38.00 2001–09 4 PIRATA
— 8.00 155.00 2000–08 4 TAO/TRITON
— 2.00 155.00 1998–2005 4 TAO/TRITON
— 22.00 155.00 1998–2005 4 TAO/TRITON
— 9.00 140.00 1998–2009 4 TAO/TRITON
— 2.00 140.00 2000–08 4 TAO/TRITON
— 2.00 125.00 2001–09 4 TAO/TRITON
— 22.00 125.00 1999–2008 4 TAO/TRITON
— 2.00 195.00 1997–2006 4 TAO/TRITON
— 5.00 195.00 1999–2006 4 TAO/TRITON
— 28.00 125.00 2001–09 4 TAO/TRITON
— 5.00 170.00 2001–09 4 TAO/TRITON
— 2.00 180.00 1999–2004 4 TAO/TRITON
22001 28.30 233.85 1978–2000 5 JMA
21001 39.50 214.50 1978–91 5 JMA
21003 25.60 224.00 1978–87 5 JMA
21004 29.00 225.00 1982–2000 5 JMA
7940 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
These data were then used to calculate means of
vector wind components (projections along 36 direc-
tions around the compass) and wind speed on daily,
weekly and monthly time scales. Standard deviations of
these quantities on subaveraging time scales were also
calculated. Other than removing missing data from the
buoy time series, no other preprocessing was carried out
on these datasets.
b. Global reanalysis products
Ten-meter and 850-hPa fields (zonal wind U, me-
ridional wind V, and temperature T) were obtained
from National Centers for Environmental Prediction
(NCEP)/Department of Energy (DOE) Reanalysis 2
data (downloaded from http://www.esrl.noaa.gov/psd/
data/gridded/data.ncep.reanalysis2.html). Wind speed
fields W were computed from the U and V fields. These
data are available 4 times daily from January 1979 to
December 2011 at a resolution of 2.58 3 2.58. The down-
scaling predictors were calculated from the 850-hPa data.
An analysis using predictors at other pressure levels
demonstrated that predictive information is largely in-
dependent of predictor pressure level throughout the free
troposphere (Sun 2012).
c. Construction of surface-wind predictands
The statistics of both wind speed and vector wind
components were predicted on three different time scales
(daily, weekly, and monthly) in this study. For all pre-
dictions, the same averaging time scale was used for
both the predictors and predictands. Predictions were
carried out separately in each calendar season [December–
February (DJF), March–May (MAM), June–August
(JJA), and September–November (SON)] to minimize
the influence of nonstationarities in the relationship be-
tween predictors and predictands resulting from the sea-
sonal cycle. An inspection of the seasonal variations
in surface wind data (not shown) demonstrated that
these calendar seasons characterize the dominant non-
stationarity in the data. A summary of the predictands is
as follows:
1) mean wind speed on the specified averaging time
scale w,
2) subaveraging time scale standard deviation of wind
speed sw,
3) mean vector wind components in the direction along
the basis vector ~e, ~u � ~e (These components are
considered at 108 increments around the compass.
By construction, projections separated by 1808 arethe same up to a sign.), and
4) subaveraging time scale standard deviation of the
vector wind components along ~e, s~u�~e.
Throughout this paper, an overbar will denote aver-
aging on daily, weekly, or monthly time scale. If the time
scale is not explicitly specified when discussing a given
result, it will hold on any time scale. In particular, w2 is
the mean of the square of the wind speed.
Two further statistics, m and s, are also calculated: 1)
the amplitude of the average vector wind m5ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2 1 y2
p
and 2) the isotropic vector wind standard deviation
s5ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi(1/2)(s2
u 1s2y)
p, where u and y are means of two
orthogonal vector wind components and su and sy are
standard deviations of two orthogonal vector wind
components (u and y generally denote arbitrary or-
thogonal components, unless otherwise explicitly spec-
ified). The quantities m and s arise in the context of the
idealized model of the wind speed probability density
function considered in section 4.
3. Results of downscaling predictions
a. Spatial correlation map
As described in the previous section, 850-hPaU, V, T,
and W averaged on daily, weekly, and monthly time
scales were used as the predictors in this study. The
strength and spatial scale of the statistical relationship
between surface wind statistics and upper-level large-
scale predictors can be assessed through inspection of
spatial correlation fields. Correlation fields of each of
mean zonal wind u and wind speed w with U, V,W, and
T during DJF for one buoy at Atlantic Ocean are dis-
played in Fig. 1. Both the monthly and daily time scales
are displayed. It can be seen that u is strongly correlated
with U on large scales: positive correlations are found
locally while negative correlation fields are found to the
north. The other predictor fields also show large-scale
correlation structure with u In contrast, the absolute
correlation values for w are much smaller than those for
u, particularly on the monthly time scale. The horizontal
scales of strong correlations increase with the averaging
time scale: on the daily time scale, the spatial scales of
the correlation fields are smaller than those on monthly
time scale (Sun 2012). This result is consistent with the
fact that on the synoptic scale, the influence exerted on
surface winds by large-scale circulation is more local
while on longer time scales large-scale teleconnection
patterns become more important.
b. Combined EOF analysis
It is evident from Fig. 1 that predictive information
for surface wind statistics is spatially distributed within
individual predictor fields, and that some of this in-
formation is common across these different fields. To
efficiently distribute the predictor variance among the
15 OCTOBER 2013 SUN AND MONAHAN 7941
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
smallest number of time series, a combined empirical
orthogonal function (EOF) analysis following that in
Monahan (2012a) is carried out. Predictions on daily,
weekly, and monthly time scales are made using 26, 16,
and 6 combined PCs, respectively, as predictors. The
numbers of predictors chosen for each time scale were
selected as theminimum number needed to explain over
85% of the total variance in the four large-scale pre-
dictor fields. The prediction results are not sensitive to
reasonable changes in the number of predictors included
in the model, or to reasonable changes in the EOF do-
main. To maintain the consistency across different time
scales, the EOF domains are the same on daily, weekly,
and monthly time scales at a particular site. As the
spatial scales of correlation maps are smaller at shorter
time scale, more PC predictors are included in the
downscaling model on daily time scales than monthly
time scales to reproduce the smaller structures of vari-
ability. As the number of statistical degrees of freedom
is larger on smaller averaging time scales, the SD model
can accommodate more predictors without overfitting.
A ‘‘leave one year out’’ cross-validation strategy is
employed in the multiple linear regression model to
prevent model overfitting. For example, predictions of
the first year were determined from a regression model
built with data from all other years. The predictions of
the second year were then obtained in a similar way
(only the second year’s data were withheld when esti-
mating themodel parameters).When the predictions for
all the years were obtained, the r2 value (i.e., square of
FIG. 1. (left) Correlationmaps of mean zonal wind at buoy 41001 with large-scale predictors at 850 hPa onmonthly time scales: (top)U,
(second row) V, (third row) W, and (bottom) T. The position of the buoy is indicated by the white dot. The white boxes in the panels
denote the domain used for the EOF decomposition of large-scale predictor fields. (left center) As in (left), but for monthly-mean wind
speed. (right center) As in (left), but on a daily time scale. (right) As in (left center), but on a daily time scale.
7942 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
correlation between predictions and observations) was
computed to measure the prediction skill.
c. Predictability of surface wind statistics: A casestudy of three representative buoys
Statistics of the vector wind components (both means
and standard deviations) in 36 directions around the
compass, along with the mean and standard deviation of
wind speed, were predicted at all buoys on daily, weekly,
and monthly time scales. Figure 2 shows the DJF pre-
dictive skills (r2) of each of the surface-wind statistics for
the monthly time scale at three representative buoys. It is
evident that the predictive skills of vector wind compo-
nents are generally anisotropic, as had been previously
noted for land surface winds by van der Kamp et al.
(2012) and Culver and Monahan (2013). The speed pre-
diction is isotropic by construction, as wind speed is
a scalar quantity. Previous studies have suggested that the
maximum prediction skill of vector wind components is
aligned with topographic features in mountainous areas
(van der Kamp et al. 2012), although vector prediction
anisotropy is also observed in regions with little topo-
graphic variability (Culver andMonahan 2013). Wewere
unable to determine any dominant factor determining
the magnitude or orientation of this anisotropy. For ex-
ample, for the buoys considered in this study, the maxi-
mum prediction skills were aligned both along and across
shore. Note also that at buoy 41001 the predictive skill of
the best predicted mean vector wind is much better than
that of mean wind speed, while at buoy 51002 the mean
wind speed is as well predicted as the best predictedmean
vector wind component. Buoy 21001 represents an in-
termediate case.
d. Wind statistics predictability distribution
Maps of the DJF prediction skills (correlation r2 values)
of the best predicted mean vector wind components
max[r2(~u � ~e)] (i.e., the vector wind component that has
the highest predictive skill among all the components in
36 directions), the mean wind speed w, the best pre-
dicted standard deviations of vector wind component
max[r2(s~u�~e)], and the standard deviation of wind speed
sw on monthly time scales at all the 52 buoys are shown
in Fig. 3. Several general results follow from these pre-
diction maps.
1) As was found in Monahan (2012a) and Culver and
Monahan (2013), the prediction skills of the best
predicted mean vector wind component are gener-
ally higher than those of the mean wind speed across
all the 52 buoys. There is no general relationship
between the predictability of the mean speed and the
worst predicted mean vector wind component (not
shown).
2) The buoys which have relatively high prediction
skills of mean wind speed are generally located in
tropical regions. Through the midlatitudes, the pre-
diction skills of mean wind speed are generally
considerably lower. There is no general relationship
between the predictability of speed and proximity to
land.
3) The subaveraging time scale standard deviations of
both vector wind components and wind speed are
generally poorly predicted at all geographic locations.
Corresponding maps for the other calendar seasons and
averaging time scales produce results consistent with
these general results (Sun 2012).
These results are also illustrated by scatterplots (across
52 buoys and 4 seasons) of the relative predictability of
the vector wind component and wind speed statistics
(Fig. 4). Each point in these plots represents the pre-
diction skill of the specified surface-wind statistics in
one season at one buoy on the specified time scale. In
general, we see that mean quantities are generally better
FIG. 2. Monthly-time-scale DJF r2 prediction skills at three representative buoys. Shown are vector wind means
(solid red line) and standard deviations (red dashed line) in 36 directions, the mean wind speed (blue line), and the
wind speed standard deviation (dashed blue line). The black circle denotes a reference prediction skill of r2 5 0.8.
15 OCTOBER 2013 SUN AND MONAHAN 7943
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
predicted than standard deviations, particularly on
shorter averaging time scales. Furthermore, the best
predicted vector wind component is almost always bet-
ter predicted than the mean wind speed. To investigate
the relative predictability of the statistics of vector wind
components and wind speed, we now turn to an ideal-
ized model of the wind speed probability distribution.
4. Interpretation of the relative predictability ofvector wind and wind speed statistics
a. A Gaussian model of the vector wind probabilitydensity function
The results of the previous section showed that the
prediction skills of mean wind speed are generally
smaller than those of the best predicted vector wind
components. This result was also obtained for sea sur-
face winds in the northeast subarctic Pacific (Monahan,
2012a) and for land surface winds across Canada (van
der Kamp et al. 2012; Culver and Monahan 2013).
Monahan (2012a) introduced an idealized probability
model of the wind speed probability distribution to in-
vestigate the reason for these differences in predictability.
Assuming that fluctuations in the vector winds are iso-
tropic, uncorrelated, and Gaussian, Monahan (2012a)
showed that the mean wind speed can be modeled as
a function of the magnitude of the mean vector wind
m5ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2 1 y2
pand the isotropic standard deviation
s5ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi(1/2)(s2
u 1s2y)
p:
w(m,s)5sF�ms
�, (1)
where (u, y) and (su, sy) are the mean and standard
deviation of orthogonal vector wind components. The
expression for F is as follows (Rice 1945):
F(x)5
ffiffiffiffip
2
rexp
�x2
4
���11
x2
2
�I0
�x2
4
�1
x2
2I1
�x2
4
��,
(2)
where Ij(x) is the associated Bessel function of the first
kind of order j. It should be pointed out that no explicit
assumptions aremade about the temporal autocorrelation
structure of the wind components. To assess the perfor-
mance of this model, we compared the IPM modeled w
and the actual w on monthly time scales using 10-m sur-
face wind data from NCEP/DOE Reanalysis 2. We have
also assessed the IPM’s performance with buoy data; the
results are consistent with those obtained with the re-
analysis data (Sun 2012). For each of the four calendar
seasons, the following calculations were carried out:
1) at each grid point and for eachmonth in the record, we
computed m and s from monthly means and standard
deviations of the 10-m zonal wind and meridional
wind;
2) these values of m and s are used to compute monthly
w using the IPM; and
3) we calculated the correlation between the modeled
monthly w from the IPM and the monthly w com-
puted directly from NCEP/DOE Reanalysis 2 data.
The square of the correlation (r2), which describes
the fraction of variance held in common between the
two time series, provides a linear measure of the
model performance in modeling mean wind speed.
It was found (not shown) that the modeled mean wind
speed from the IPMhas a high correlation with themean
FIG. 3. Cross-validated DJF r2 predictive skills on the monthly
time scale. (top) Best predicted vector wind component; (second
row) mean wind speed; (third row) best predicted standard de-
viation of vector wind component; and (bottom) standard de-
viation of wind speed.
7944 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
FIG. 4. (top) The prediction skills (cross-validated r2) of the standard deviations of wind speed relative to those of the mean wind speed,
(middle) the best predicted standard deviations of vector wind components relative to the best predicted means of vector wind com-
ponents, and (bottom) the mean wind speed relative to the best predicted means of vector wind components. (left) The daily time scale
predictions, (center) the weekly time scale predictions, and (right) the monthly time scale predictions.
15 OCTOBER 2013 SUN AND MONAHAN 7945
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
wind speed from the reanalysis data on a global scale (r2
above 0.9 at all grid points). Themeanwind speedmodel
derived from the IPM is demonstrated to work well
across the global ocean. In later section, we will dem-
onstrate that this model is less successful in modeling
month-to-month variations of the submonthly wind speed
standard deviation.
b. Sensitivity of w to m and s
Having provided evidence that the IPM is able to char-
acterize the variability of w in terms of the variability of m
ands, we can use thismodel to investigate the sensitivity of
w to changes in these vector wind statistics. While w is
a function of m and s, the sensitivities of the mean wind
speed to m and s are functions of the ratio m/s alone:
›mw5›w
›m5
›hsF
�ms
�i›m
5s
sF 0�ms
�5F 0
�ms
�and (3)
›sw5›w
›s5
›hsF
�ms
�i›s
5F�ms
�2
m
sF 0�ms
�. (4)
For convenience we can define the bounded scalar
quantity u (Monahan 2012a; Culver andMonahan 2013):
u5 tan21�ms
�. (5)
The sensitivities of w to m and s as functions of u can be
computed numerically and are shown in Fig. 5a.
From the sensitivity plot, it is clear that in the low u
regime, w is more sensitive to the variability of s than of
m. In contrast, in the high u regime w is more sensitive
to the variability of m. For intermediate values of u, w
has similar sensitivity to both s and m. This result in-
dicates that in a high u regime, variability of w is de-
termined by variability of m, irrespective of the variability
of s. In contrast, in a low u regime, variations in w are
determined by those of s and are insensitive to changes
of m. These different regimes can be illustrated by con-
sidering the skill of modeled w by the IPM allowing m to
vary frommonth to month at each point while holding s
constant at its climatological value. The results of this
calculation for DJF are displayed in Fig. 6a. Consistent
results are obtained for other seasons. For comparison,
we also calculated monthly time scale u values at each
grid point using themonthly time scalem ands, and then
averaged these across all months to produce one cli-
matological u field (Fig. 6b). These maps demonstrate
the following:
1) The u regimes are geographically organized. In
general, high u values are found in the tropical Pacific
and Atlantic as well as the Indian Ocean. The low u
regime is in the subtropical and subpolar latitudes.
Intermediate u values are found predominantly in
the midlatitudes.
2) The model with fixed s represents month-to-month
variations inw well in some regions (e.g., the tropics)
and poorly elsewhere (e.g., subpolar and subtropical
regions).
3) The two maps in Figs. 6a and 6b match closely. They
clearly indicate that where high u dominates, themodel
with fixed s can successfully represent monthly-time-
scale variability of w, while where low u prevails, the
FIG. 5. (a) Sensitivity of w to m and s as functions of u and (b) sensitivity of sw to m and s as
functions of u.
7946 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
model with fixed s cannot accurately characterize
month-to-month variations in w. This result is consis-
tent with the sensitivity plots in Fig. 5a: for low u
regimes, w computed from the IPM with fixed clima-
tological s is not very accurate because variations of w
are more sensitive to those of s in this regime. In
contrast, in the high u regimes, the performance of the
model with fixed s remains good as w is primarily
dependent on m.Note that the maps in Figs. 6a and 6b
are similar but not identical, because there is not
expected to be an exactly linear relationship between
u and the modeling skill r2.
The results of this analysis suggest that the scalar
quantity u is a good measure of the dependence of w on
m and s for observed sea surface winds.
Maps of the DJF u distribution on each of the three
averaging time scales (daily, weekly, and monthly) with
the locations of all the 52 buoys superimposed are pre-
sented in Figs. 6b–d. It can be seen that u generally de-
creases as averaging time scales increase: on daily time
scales, the mid-to-high u regime (above 0.9) dominates
all the buoys. On weekly time scale, an intermediate u
regime (0.5–0.9) appears on the flanks of the surface
westerlies. On the monthly time scale, the low u regime
appears in subtropical and subpolar latitudes. As the
variability of extratropical sea surface winds is strongest
on the synoptic time scale of several days, subdaily
variability is smaller than the subweekly or submonthly
variability. In consequence, s is much smaller thanm on
the daily time scales, resulting in a broadly distributed
high u regime. It is noteworthy that in the tropical Pa-
cific and Atlantic, a high u regime generally dominates
most buoys on daily, weekly, and monthly time scales.
In the tropics the major forms of variability, such as
the Madden–Julian oscillation (MJO), have time scales
much longer than those of midlatitude synoptic eddies.
The relatively steady tropical trade winds result in m
values that are much larger than s, resulting in a high u
regime. The u field also displays seasonal variations (Sun
2012).
The distribution of the buoys considered provides
good coverage of the full u range on weekly andmonthly
time scales. In consequence, we have a representative
sample of buoys within each of the three u regimes with
which to establish statistical relationships.
c. Predictability of w relative to m and s
The results of the previous section suggest that the
predictability of w relative to that of m and s is a func-
tion of u. Scatterplots of SD predictive skills (correlation
r2 values) of w relative to those of m and s are shown
in Fig. 7 for daily, weekly, and monthly averaging time
scales. Corresponding values of u are indicated by color.
The SD predictions of each wind statistic on each of
three averaging time scales were done separately for
each of the four calendar seasons and each of the 52
buoys (resulting in a total of 208 points in each plot).
The following are observed: 1) On the daily time scale,
for which u is consistently large, the predictive skill of
w is strongly correlated with the predictive skill of m
across all stations and seasons. In contrast, the pre-
dictive skill of w has no strong relationship with that of
s (except for the smallest values of u). 2) On weekly
and monthly time scales, the points in the scatterplot
of r2(m) with r2(w) gather around the 1:1 line for high u
values, while the points are more broadly scattered for
low u values. In contrast, in the scatterplot of r2(s) with
r2(w), the data points gather around the 1:1 line for low
FIG. 6. (a) Modeling skill of DJF mean wind speed by the IPM
[Eq. (1)] with month-to-month variations in m but s held constant
(at its long-term average value). (b) Climatological DJF u distri-
butions onmonthly time scale, with positions of all 52 buoys. (c) As
in (b), but on a weekly time scale. (d) As in (b), but on a daily time
scale. Note that the color bar for (a) is between 0 and 1, while for
(b)–(d) it is between 0 and 1.5.
15 OCTOBER 2013 SUN AND MONAHAN 7947
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
u values and scatter away from the 1:1 line for high u
values.
The above results indicate that u is a good metric for
characterizing the statistical predictability of w relative
to that of m and s. In the high u regime, good predictions
of w require accurate predictions of m. On the other
hand, in low u regimes, how well w can be predicted
depends on the predictability of s. In a medium u re-
gime, the situation is complicated as w has comparably
strong and nonlinear dependence on m and s. While
these results have been demonstrated using a linear
statistical downscaling model, the sensitivity results of
the IPM suggest that they should hold irrespective of
how the predictions are made.
d. Predictability of w relative to that of the bestpredicted vector wind component
Having related the predictability of w to that of m and
s, we are on the way toward understanding the pre-
dictability of w relative to the vector wind components~u � ~e. In the low u regime, the predictive skill of w is
determined by that of s. Predictions of the isotropic
standard deviation s are not better than max[r2(s~u�~e)]; it
follows from Fig. 4 that s will generally not be as well
predicted as the max[r2(~u � ~e)]. It follows that in the low
u regime, in which w variations are dominated by those
of s, the predictability of w should be less than that of
the best predicted mean vector wind component. On the
FIG. 7. The correlation-based predictive skill of w relative to that of (left) m and (right) s on
(top) daily, (middle) weekly, and (bottom)monthly averaging time scales. The color of the data
points denotes the value of u. One-to-one lines are given in solid blue.
7948 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
other hand, in the high u regime, the predictive skill of
w is determined by that of m. To complete the connec-
tion between the predictability of speeds and of vector
wind components, we need to relate the predictability of
the amplitude of the mean vector wind m to the pre-
dictability of the vector wind components themselves.
Culver and Monahan (2013) provided the following
theoretical expression for the linear predictability of m
relative to that of the vector wind component aligned
along the long-term mean wind:
r2(m)5mean(u)2
std(u)21mean(u)2r2(u) , (6)
where r2(u) and r2(m) are the (correlation-based) pre-
dictabilities of the wind component u along the long-
term mean and of m by a single predictor x (assumed
to have aGaussian distribution), respectively. The result
in Eq. (6) is based on the assumption that variations
of u (e.g., from month to month) are isotropic, un-
correlated, and Gaussian. The quantities mean(u) and
std(u) are the mean and standard deviation of u over the
entire record, respectively. For convenience, we intro-
duce the quantity g5mean(u)/std(u), which gives us
r2(m)5
�1
g21 1
�21
r2(u) . (7)
From Eq. (7) it is clear that the predictability of m is
bounded above by that of the along-mean wind compo-
nent. Furthermore, r2(u) will itself be bounded above by
the predictability of the best predicted component (by
definition). Figure 8 displays the predictive skill of m
relative to that of u in high u regimes (u $ 1) in relation
to g on daily, weekly, and monthly time scales. It can be
seen that in general, when g � 1, the predictability of m
approaches that of u. On the other hand, when g de-
creases, the predictive skill of m becomes smaller than
that of u. As discussed in the previous section, a majority
of the buoys are in a high u regime on daily time scales
across all seasons and locations, while on weekly and
monthly time scales, many buoys are in a medium or
low u regime. As a result, fewer data points with u. 1
are displayed for the plots on weekly and monthly
time scales.
It should be emphasized that g increases with aver-
aging time scales. In Fig. 8 it can be observed that the
upper limit of g increases from 6 to 10 as the averaging
time scale increases from daily to monthly. On longer
averaging time scales, a smaller fraction of the variance
of the wind is retained in variations of u, and a larger
fraction is contained in subaveraging time scale vari-
ability (i.e., s). For an increasingly large fraction of
stations and seasons, as the averaging time scale is in-
creased the variability in the averaged vector wind u
becomes much smaller than the climatological mean
wind [i.e., std(u) � mean(u)].
In general, shorter averaging time scales are associ-
ated with larger values of u and smaller values of g, while
longer averaging time scales with smaller u and larger g.
Predictability of w is generally smaller than that of the
vector wind components on long averaging time scales
because of lower u values (and the fact that s has weak
FIG. 8. The predictive skill ofm relative to that of the along-mean
vector wind component u in high u regimes (u $ 1) in relationship
to g [Eq. (7); as indicated by the color of the data points].
15 OCTOBER 2013 SUN AND MONAHAN 7949
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
predictability). In contrast, predictability of w can be
limited on short averaging time scales when m is more
poorly predicted than u because of the small value of g.
It is possible that there may be an optimal averaging
time scale on which the relationship among u,m, andw is
balanced in such a way as to yield optimal predictability
of w relative to u.
e. Prediction of subaveraging time scale standarddeviation of wind speed with the IPM
The IPM introduced above has been shown to be able
to successfully model w in terms of m and s. An analytic
expression for the standard deviation of wind speed sw
can also be derived from the IPM. By definition,
s2w 5w2 2w2 , (8)
from which it follows that
sw5
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2s2 1m22
hsF
�ms
�i2r5sG
�ms
�. (9)
From Eq. (9), we have an exact expression for the wind
speed standard deviation sw in terms of m and s.
As with the modeled w, the sensitivities ›msw and
›ssw are functions of the scalar variable u, as illustrated
in Fig. 5b. In contrast to modeled w, the modeled sw is
most sensitive to variations in s over the entire u range.
In the same way that we assessed the ability of the IPM
to represent observed variability of w in terms of vari-
ability in m and s, we now consider a similar calculation
to test the performance of Eq. (9) in capturing observed
the month-to-month variability in the submonthly
standard deviation of wind speed:
1) at each grid point, we calculated monthly m and s
frommonthlymeans and standard deviations of 10-m
NCEP/DOEReanalysis 2 zonal wind andmeridional
wind;
2) from these, we computedmonthlysw usingEq. (9); and
3) we then calculated the correlation between the
modeled month-to-month variations in sw from Eq.
(9), and those directly computed from NCEP/DOE
Reanalysis 2 data [the squared value of the correla-
tion (r2) is used to assess the model performance in
modeling the month-to-month changes in the sub-
monthly standard deviation of wind speed].
The results of this calculation for DJF are presented
in Fig. 9a. In contrast to the highly accurate represen-
tation of w by the IPM on a global scale, the model fails
to represent month-to-month changes in sw over most
of the ocean. It can be seen that in the midlatitude
and high-latitude regions (of both the Northern and
Southern Hemispheres), the r2 prediction skill of the
model is generally below 0.5. In the tropical regions,
the IPM generally performs better although its perfor-
mance is still poor in a number of places. The failure
in reproducing month-to-month variations of the sub-
monthly standard deviation of wind speed from Eq. (9)
is perplexing: we should be able to obtain knowledge of
these statistics, as long as we have the correct wind speed
probability density function pw(w). The fact that we can
simulate w but not sw with our model leads us to re-
examine the three assumptions on which the IPM is
based: that the vector wind fluctuations are Gaussian,
uncorrelated, and isotropic.While these approximations
are reasonable for modeling first-order statistics (mean
wind speed), they may not be good approximations for
modeling the second-order statistics. Monahan (2006)
demonstrated that the along- and across-wind compo-
nents u and y are close to being uncorrelated and have
nearly isotropic fluctuations on a global scale (with some
exceptions in monsoon and ITCZ regions). However,
the skewness and kurtosis of the along-mean vector wind
components can differ substantially from zero (Monahan
2006). Therefore, we will investigate the influence of the
non-Gaussianity of vector wind components onmodeling
the standard deviation of wind speed.
FIG. 9. (a) The modeling skill of DJF submonthly time scale sw
from the IPM [Eq. (9)]. (b) As in (a), but with the non-Gaussian
vector wind model obtained from Eq. (11).
7950 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
1) WIND SPEED PDF FROM NON-GAUSSIAN
VECTOR WINDS
We first decompose the vector winds into components
along and across the time-mean vector wind. In this section,
these will be denoted by u and y respectively. Following
Monahan (2006), non-Gaussian surface wind components
are included in the model through a Gram–Charlier expan-
sion (Johnson et al. 1994) of the probability density func-
tion (PDF) of the along-mean wind component as follows:
pu(u)51ffiffiffiffiffiffi2p
psexp
"2(u2 u)2
2s2
#
3h11
n
6He3
�u2 u
s
�1
k
24He4
�u2 u
s
�i, (10)
where He3(x) 5 x3 2 3x, He4(x) 5 x4 2 6x2 1 3,
n5 skew(u)5mean[(u2 u)3]/std(u)3 (the monthly
along-mean wind skewness), and k5 kurt(u)5fmean[(u2 u)4]/std(u)4g2 3 (the monthly along-
mean wind kurtosis). The cross-mean wind compo-
nent is modeled as Gaussian (as is broadly consistent
with observations; Monahan 2006, 2007). Note that
pu(u) defined in this way is not strictly nonnegative,
and therefore is not necessarily a proper PDF. Nev-
ertheless, the resulting function has the correct mo-
ments and is a useful model of the PDF of u so long
as realizations of this random variable are not re-
quired (e.g., Johnson et al. 1994). A more compli-
cated expression for the wind speed PDF can then be
obtained:
pw(w)5
h12
n
6He3
�us
�1
k
24He4
�us
�iI0
�wus2
�1
1
2
�ws
�hnHe2
�us
�2
k
3He3
�us
�iI1
�wus2
�
11
8
�ws
�2h22n
�us
�1 kHe2
�us
�ihI0
�wus2
�1 I2
�wus2
�i1
1
24
�ws
�3hn2 k
�us
�ih3I1
�wus2
�
1 I3
�wus2
�i1
k
192
�ws
�4h3I0
�wus2
�1 4I2
�wus2
�1 I4
�wus2
�io�w
s2
�exp
�2w21 u2
2s2
�, (11)
where Ij is the associated Bessel function of the first kind
of order j (Monahan 2006). As with previous analyses,
we use this model to simulate month-to-month varia-
tions in sw given observedmonth-to-month variations in
u,s, n, and k. The performance in capturingsw is slightly
improved by including month-to-month changes in
the skewness and kurtosis of u, but model performance
remains much poorer than for w at most locations.
(Fig. 9b). Evidently, non-Gaussianity of the vector wind
components is not the primary cause degrading the
IPM’s performance in characterizing variability in sw.
Why should it be the case that the model performs well
for modeling monthly-mean wind speed but largely fails
when modeling submonthly wind speed standard de-
viation? As we will now show, a contributing factor is
related to differences in sampling variability of these
statistics.
2) SAMPLING VARIABILITY IN MONTH-TO-MONTH
FLUCTUATIONS OF m, s, w, AND sw
Having demonstrated that the model assumption of
Gaussian-distributed vector winds is not the primary
cause of the difficulty in modeling month-to-month
variations in sw, we now ask the question: might the
poor simulation of sw result from different sampling
variability of m, s, and sw? At any location, within each
month, the wind fluctuations have on the order of 15–30
statistical degrees of freedom as the surface winds gen-
erally have an autocorrelation time scale on the order of
one to two days (Monahan 2012b). As a result, there will
be nonnegligible sampling variability in all surface wind
statistics, which may be different from one statistic to
another. To assess the influence of sampling variability,
a series of Monte Carlo experiments were conducted to
examine how the potential sampling errors in m and s
influence the Gaussian model’s performance in model-
ing w and sw.
By construction in these idealized calculations the
vector winds were Gaussian, uncorrelated, and isotropic
so the wind speed population statistics are exactly re-
lated to those of the vector winds by Eqs. (1) and (9).
The strength of this analysis is that it is a ‘‘perfect
model’’ calculation—we know exactly what is true about
the underlying relationship between the statistics of
vector winds and wind speed, and within this can in-
vestigate the role of sampling variability.
In our first experiment, we let
u5m0(11 rmd1) and (12)
s5s0(11 rsd2) , (13)
where d1 and d2 are random numbers with a uniform
distribution on (2½, ½). We will interpret m0 and s0 as
15 OCTOBER 2013 SUN AND MONAHAN 7951
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
the climatological mean and standard deviation of the
vector winds, while the random numbers d1 and d2 de-
scribe month-to-month fluctuations in these statistics
with strengths scaled by rm and rs, respectively. Note that
fluctuations of u and s represent true month-to-month
variability in the vector wind statistics. That is, these are
the signal that we are interested in capturing with our
model.
We then generated N 5 120 realizations of u and s
each representing a separatemonth.Within eachmonth,
we randomly sampled M days (the number of inde-
pendent wind realizations within each month) of vector
wind components (u, y) from the Gaussian model with
the mean (u, 0) and isotropic standard deviation s for
that month. For each month, we computed the sample
mean and standard deviation of wind speed from the
sample u and y, as well as the sample m and s. The
sample m and s were then used to compute thew and sw
from theGaussianmodel using Eqs. (1) and (9).We then
correlated the observed and modeled wind speed mo-
ments, to characterize the modeling skills of w and sw.
This procedure was repeated for different values of rmand rs, withm ranging from 1 to 10m s21 and s kept fixed
at 3m s21 to assess sampling variability under different u
regimes. An ensemble of 300 estimates of the modeling
skill was computed.
We consider bothM5 500 andM5 20. The second of
these is closer to the real number of statistical degrees of
freedom within any month, while the first is considered
to illustrate how sampling variability changes as sample
size increases. The modeling skills of w and sw are
plotted as functions of u in Fig. 10, from which the fol-
lowing can be observed:
(i) The modeling skill of w is generally high with little
sensitivity to the values of rm and rs.Consistent with
the results presented earlier, the model is able to
reproduce month-to-month variability in w for
different sizes of the true signal strength.
(ii) Themodeling skill of sw can be quite poor for small
values of rs (the true month-to-month variability
of s). When rs 5 0, the modeling skill of sw is
FIG. 10. Monte Carlo experiment derived modeling skills r2 of w and sw by the Gaussian models
[Eqs. (1) and (9)] for different values of rs and rm, and for (left) M 5 500 and (right) M 5 20.
7952 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
substantially poorer than that with rs 5 0.45. For
instance, for M 5 500, when rm 5 0.45, rs 5 0 (Fig.
10a), and the r2 modeling skill of sw is about 0.6. In
contrast, when rm 5 0, rs 5 0.45 (Fig. 10c), and the
modeling skill of sw is close to 0.95. The value of rm(the true month-to-month variability in m) does not
substantially influence the modeling skill of sw (not
shown).
(iii) The modeling skill of sw increases with M. Under
the same set of rm and rs values, the modeling skill
of sw is better withM5 500 than withM5 20. For
M5 20, themodeling skill of sw is low even when rsis relatively large.
When rs 5 0, the population vector wind standard
deviation remains the same from month to month so
all fluctuations of s are produced by sampling fluctu-
ations alone. In this case fluctuations in sw modeled
by Eq. (9) differ significantly from the true fluctua-
tions in the wind speed submonthly standard de-
viation. As rs increases, the real month-to-month
fluctuations of s (the signal) increase in size relative to
those of the sampling fluctuations (the noise). Thus,
the signal-to-noise ratio (SNR) increases, and the
model does a better job of simulating month-to-month
changes in sw.
Similarly it is observed that the modeling skill of sw
increases asM increases for specified rs and rm values.
By increasing M, while the signal stays the same,
the noise is reduced so the SNR increases and model
performance is better. For small values of the SNR,
the IPM has difficulties modeling the month-to-
month variability in the wind speed standard deviation
even when the observed vector wind components
are Gaussian, uncorrelated, and isotropic without
approximation.
This analysis demonstrates that the skill of the IPM in
simulating month-to-month variations of sw is de-
termined by a SNR that is related to the size of the true
month-to-month fluctuations of s (characterized by rs)
and the number of independent wind realizations M
within the month. Sampling fluctuations in sw are dis-
tinct from those of m and s, so variability in sw is only
well predicted when the signal of true month-to-month
variability is sufficiently large relative to the sampling
noise. We will now develop a quantitative measure of
the SNR and use this to interpret the modeling results of
Fig. 9a.
This first two steps of this analysis are similar to those
of the previous Monte Carlo experiment, sampling
a broader range of values of rs andM (rs ranges from 0 to
1.5; rm is set to 0.3; and M ranges from 1 to 1000). We
define the signal-to-noise ratio as
SNR5
24 std(~s)
mean(~s)(rs,M)2
std(~s)
mean(~s)(rs 5 0,M)
35
std(~s)
mean(~s)(rs 5 0,M)
5
264 C(~s)(rs,M)
C(~s)(rs 5 0,M)
3752 1, (14)
where C(~s)5 std(~s)/mean(~s), mean(~s) is the ensemble
mean value of the sampled vector wind standard de-
viation over all 120 months, and std(~s) is the corre-
sponding standard deviation. As discussed above, it is
expected that the size of the variability in ~s will depend
on the true signal strength rs and the number of degrees
of freedom M. The SNR defined by Eq. (14) charac-
terizes the month-to-month fluctuation of s (the signal)
in the dataset relative to the sampling variability (noise)
given by [std(~s)/mean(~s)](rs 5 0,M). The SNR can be
computed fromMonte Carlo simulations and compared
with the modeling skill of sw.
Consistent with the qualitative analysis described
earlier, the modeling skill of sw is determined by the
SNR as shown in Fig. 11a: as the SNR increases, sw is
better modeled. To obtain a model r2 skill better than
0.9, the SNR has to achieve a value above 3. The re-
lationship among M, rs, and SNR is illustrated in Fig.
11b. Consistent with the previous analysis, the signal-to-
noise ratio increases with both M and rs. When rs 5 0.2,
the number of independent realizationsM has to exceed
1000 to get a signal-to-noise ratio of 3. When rs 5 0.9,
a SNR of 3 can be obtained with M below 50.
These results indicate that for real sea surface wind
data with a typical value of M smaller than 30, the
month-to-month fluctuation of s has to be relatively
large to result in a sufficiently high SNR, to obtain
a good modeling skill of sw with the IPM. We will now
estimate the SNR from the NCEP/DOE Reanalysis 2
surface wind dataset and compare it with the distribu-
tion of modeling skill of sw previously shown in Fig. 9a.
To estimate SNR, we need to have the estimates of
submonthly ~s and the value of M at each grid point.
The number of independent wind realizations within
a month can be estimated as follows:
M5N
Te
, (15)
where N is the duration of a month and Te is the auto-
correlation time scale. In computing Te, the autocorre-
lation function of the vector wind components was
modeled as a decaying exponential:
15 OCTOBER 2013 SUN AND MONAHAN 7953
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
af(t)5 exp
�2t
Te
�, (16)
where t is the time lag. The value ofTewas then obtained
from the observed autocorrelation function by linear
regression. In fact, the autocorrelation structures are not
generally exponential and in many locations over the
ocean the vector wind autocorrelation structure is an-
isotropic (Monahan 2012b). For this calculation, Te was
estimated from the zonal wind component, for which the
autocorrelation time scale is generally the largest.
Using the estimated value of M, we calculated val-
ues of [std(~s)/mean(~s)](rs 5 0,M) at each grid point
from the Monte Carlo simulation. The value of
[std(~s)/mean(~s)](rs,M) was estimated from the ob-
served month-to-month values in the standard deviation
of the vector winds. From these, the field of SNR was
computed (Fig. 12). Comparison of the SNR map with
that of the IPM modeling skill of sw (Fig. 9a) demon-
strates that the agreement between these two fields is
generally good. In the tropical regions, SNR is generally
high. Correspondingly, in Fig. 9a, the modeling skill of
sw in these regions is relatively good. In contrast, ex-
tratropical regions have smaller SNR values, which
correspond with the poor modeling skill of sw found in
these regions. Although the two maps do not match
perfectly, their high degree of correspondence indicates
that the Gaussian model’s performance in modeling sw
is strongly related to sampling variability asmeasured by
the SNR given by Eq. (14). It follows that while the IPM
provides a useful tool for describing variability of w in
terms of that ofm ands, it will not generally be useful for
doing so with sw, and presumably for other wind speed
statistics comparably sensitive to sampling fluctuations.
5. Summary of results
This study has investigated the predictability of local
sea surface wind statistics from those of large-scale free-
tropospheric flow fields. A statistical downscaling (SD)
model based on multiple linear regression was used to
predict the means and standard deviations of observed
vector wind components and wind speed at 52 ocean
buoys on daily, weekly, and monthly time scales. A
summary of our general results is as follows:
1) The predictive skill of the best predictedmean vector
wind component is generally higher than that of the
mean wind speed. Furthermore, the mean quantities
are generally better predicted than the subaveraging
FIG. 11. (a) The r2 modeling skill of sw as a function of the signal-to-noise ratio [Eq. (14)] and (b) the SNR as
a function of rs and M.
FIG. 12. Spatial distribution of the monthly-time scale DJF SNR as
defined by Eq. (14).
7954 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
time scale standard deviations for both vector wind
components and wind speed.
2) An idealized model of the wind speed probability
density function introduced in Monahan (2012a) was
used to investigate the relationship between w and
the statistics of vector wind components. This model
indicates that the predictability of w relative to the
magnitude of the mean vector wind m and the vector
wind standard deviation s can be characterized by
the scalar quantity u5 tan21(m/s), which is depen-
dent on season, geographic location, and averaging
time scale. The quantity u characterizes the local
wind climate: specifically, whether the local vector
winds are sustained or highly variable. Consistent
with the results of the IPM, the predictability of the
observed w was found to be determined by that of s
for low values of u and by m for high values of u.
3) The subaveraging time scale variabilitys was generally
found to be poorly predicted by large-scale predictors.
Therefore, in the low u regimes, the predictive skill of
w (which is determined by that of s) is generally lower
than that of the best predicted vector wind component.
On the other hand, in the high u regimes, the predictive
skill of w is determined by that of m. The predictive
skill of m relative to that of the best predicted vector
wind u is determined by the quantity g5mean(u)/
std(u). When g � 1, the predictive skill of m relative
to that of the vector winds is greatest. For smaller
values of g, the predictive skill of m is much lower
than that of the vector wind components. Correspond-
ingly, the predictive skill of w is bounded above by
that of the best predicted vector wind component, and
can be much lower, even in high u regimes.
4) The IPM generally fails to capture month-to-month
variations of the subaveraging time scale standard
deviation of wind speed, sw, in terms of variations in
m and s. With a series of Monte Carlo experiments,
we demonstrated that this can be understood to be
a result of differences in sampling variability be-
tween the vector wind statistics and the wind speed
statistics. The amplitude of the real month-to-month
fluctuation in the vector wind standard deviation
relative to that associated with sampling variability
(characterized by a signal-to-noise ratio) accounts
for the mismatch between modeled and real sw.
The first three of these conclusions are consistent with
earlier results by Monahan (2012a) and Culver and
Monahan (2013). However, these earlier studies con-
sidered the SD predictive skill of surface winds in
a much more limited range of u values. The u range
considered in the present study allows the conclusions of
these earlier studies to be generalized to a much broader
range of wind climates. The focus of this study has been
the features of wind speed and vector wind predictability
that are common across many locations and seasons. In
consequence, no detailed analysis of predictability at
any individual buoy has been carried out. Such detailed
analyses—such as of the buoy in the Gulf of Mexico for
which the standard deviations are considerably better
predicted than the means (Fig. 3)—represent an inter-
esting direction of future study.
In the idealized setting of the Monte Carlo experi-
ments, in which the population statistics of the vector
wind are specified, the IPM fails to model variations of
the monthly standard deviation of wind speed despite
performing well in modeling monthly means of wind
speed. It can be expected thatmonth-to-month variations
of other higher-order statistics (e.g., 95th percentile,
skewness, and kurtosis) of wind speed would not be well
modeled by the IPM either. The failure of the more
general wind speed PDF model derived from non-
Gaussian vector winds to represent variations in sw in-
dicates that this difficulty will persist irrespective of
how the PDF of vector winds is modeled. As well, this
difficulty cannot be circumvented by considering longer
averaging time scales. On longer time scales, while the
number of statistical degrees of freedomM will increase,
the true variability of the subaveraging time scale stan-
dard deviation (the signal) will decrease. In general, we
cannot expect that the IPM will do a better job modeling
changes in the higher-order statistics of w on seasonal or
annual averaging time scales. It follows that the general
approach of using variations in the vector wind statistics
to model variations in these higher-order wind speed
statistics will be compromised by this strong sensitivity to
differences in sampling variability.
The results of this study demonstrate that the direct
SD predictive skill of mean sea surface wind speeds is
generally low outside of the tropics. The potential exists
that the SD prediction skills for the quantities that have
been considered in this study could be improved by
considering other sets of predictors or other SD tech-
niques. A more detailed investigation of alternative SD
approaches is an interesting direction of future study.
Furthermore, previous studies have shown that the an-
isotropy in the predictability of land surface winds can
be related to topographic features (van der Kamp et al.
2012; Salameh et al. 2009) although this is not always the
case (Culver and Monahan 2013). Unlike the land sur-
face, the sea surface is more homogeneous, and those
heterogeneities that are present (such as sea surface
temperature fronts) tend not to be fixed in place. The
control on the strength of this anisotropy, and the ori-
entation of the best predicted vector wind, are not well
understood. A detailed examination of the anisotropy in
15 OCTOBER 2013 SUN AND MONAHAN 7955
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC
the predictability of sea surface winds is another in-
teresting direction of future study.
Acknowledgments. The authors would like to thank
Aaron Culver, Andrew J. Weaver, Bill Merryfield, and
Julie Zhou for their comments, as well as those of three
anonymous reviewers. This work was funded by the
Natural Sciences and Research Council of Canada’s
Collaborative Research and Training Experience Pro-
gram in Interdisciplinary Climate Science.
REFERENCES
Bates, N. R., and L. Merlivat, 2001: The influence of short-term
wind variability on air–sea CO2 exchange.Geophys. Res. Lett.,
28, 3281–3284.
Capps, S. B., and C. S. Zender, 2009: Global ocean wind power
sensitivity to surface layer stability. Geophys. Res. Lett., 36,L09801, doi:10.1029/2008GL037063.
Culver, A. M. R., and A. H. Monahan, 2013: The statistical pre-
dictability of surface winds over western and central Canada.
J. Climate, in press.
Donelan, M., W. Drennan, E. Saltzman, and R. Wanninkhof, Eds.,
2002: Gas Transfer at Water Surfaces. Geophys. Mongr., Vol.
127, Amer. Geophys. Union, 383 pp.
Garratt, J., 1992: The Atmospheric Boundary Layer. Cambridge
University Press, 316 pp.
Johnson, N., S. Kotz, and N. Balakrishnan, 1994: Continuous
Univariate Distributions. Vol. 1. Wiley, 756 pp.
Jones, I. S., and Y. Toba, Eds., 2001: Wind Stress over the Ocean.
Cambridge University Press, 307 pp.
Liu, W. T., W. Tang, and X. Xie, 2008: Wind power distribution
over the ocean. Geophys. Res. Lett., 35, L13808, doi:10.1029/2008GL034172.
Monahan, A. H., 2006: The probability distribution of sea surface
wind speeds. Part I: Theory and SeaWinds observations.
J. Climate, 19, 497–520.——, 2007: Empirical models of the probability distribution of sea
surface wind speeds. J. Climate, 20, 5798–5814.
——, 2012a: Can we see the wind? Statistical downscaling of his-
torical sea surface winds in the subarctic northeast Pacific.
J. Climate, 25, 1511–1528.
——, 2012b: The temporal autocorrelation structure of sea surface
winds. J. Climate, 25, 6684–6700.Rice, S.O., 1945: Mathematical analysis of random noise (part 2).
Bell Syst. Tech. J., 24, 46–156.
Salameh, T., P. Drobinski,M.Vrac, and P. Naveau, 2009: Statistical
downscaling of near-surface wind over complex terrain in
southernFrance.Meteor.Atmos. Phys., 103, 253–265, doi:10.1007/
s00703-008-0330-7.
Sampe, T., and S.-P. Xie, 2007: Mapping high sea winds from
space: A global climatology. Bull. Amer. Meteor. Soc., 88,1965–1978.
Sun, C., 2012: Statistical downscaling of sea surface winds over the
global oceans. M.S. thesis, School of Earth andOcean Sciences,
University of Victoria, 111 pp.
van der Kamp, D., C. Curry, and A. Monahan, 2012: Statistical
downscaling of historical monthly mean winds over a coastal
region of complex terrain. II: Predicting wind components.
Climate Dyn., 38, 1301–1311.
7956 JOURNAL OF CL IMATE VOLUME 26
Unauthenticated | Downloaded 12/23/21 03:00 PM UTC