Sei sulla pagina 1di 23

DK342X_book.

fm copy Page 59 Thursday, May 19, 2005 3:42 PM

3 Applications of
Hilbert-Huang Transform
to Ocean-Atmosphere
Remote Sensing
Research
Xiao-Hai Yan, Young-Heon Jo,
Brian Dzwonkowski, and Lide Jiang

CONTENTS

3.1 Introduction ....................................................................................................60


3.2 Analyses of TOPEX/Poseidon Sea Level Anomaly Interannual
Variation Using HHT and EOF .....................................................................62
3.3 Application of HHT to Ocean Color Remote Sensing
of the Delaware Bay ......................................................................................64
3.4 Mediterranean Outflow and Meddies (O & M)from Satellite
Multisensor Remote Sensing .........................................................................71
3.5 Conclusion......................................................................................................78
Acknowledgments....................................................................................................79
References................................................................................................................79

ABSTRACT

The Hilbert-Huang transform (HHT) is a newly developed method for analyzing non-
linear and nonstationary processes. Its application in oceanography and oceanatmo-
sphere remote sensing research is still in its infancy. In this chapter, we briefly introduce
the application of this method in oceanatmosphere remote sensing data analyses and
present a few examples of such applications.

59

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 60 Thursday, May 19, 2005 3:42 PM

60 The Hilbert-Huang Transform in Engineering

3.1 INTRODUCTION
Spectral analysis is a very useful tool to analyze a time series signal. However, this
method does not fully describe a data set that changes with time. The spectrum gives
us the frequencies that exist over the entire duration of the data set. On the other
hand, time–frequency analysis allows us to determine the frequencies at a particular
time. Hence, the fundamental idea of time–frequency analysis is to understand and
describe phenomena where the frequency content of a signal is changing in time.
Scientists traditionally use short Fourier transform by sliding the window along
the time axis to get a time–frequency distribution. Since it relies on the traditional
Fourier spectral analysis, one has to assume the data to be piecewise stationary.
Currently, the most famous time–frequency analysis method is wavelet transform.
The most common method used is Morlet wavelet, defined as Gaussian enveloped
sine and cosine wave groups with 5.5 waves (1). The problem with Morlet wavelet
is the leakage generated by the limited length of the basic wavelet function, which
makes the quantitative definition of the energy–frequency–time distribution difficult.
Once the basic wavelet is selected, one has to apply it to analysis of all the data (2).
Recently Huang et al. (3) introduced a new and potentially more robust method
for time–frequency analysis. This method, the empirical mode decomposition–Hil-
bert-Huang transform (EMD-HHT), is applicable to both nonstationary and nonlin-
ear signals. In real ocean and ocean atmosphere coupling, most processes are non-
linear and nonstationary. One example is that at the onset of El Nino: nonlinear
Kelvin waves carry warm water from the western Pacific to the east (4). This process
is exhibited as a nonlinear pattern in altimeter data. For this reason, we use the
EMD-HHT technique in our El Nino study and in many of our other studies.
Since the EMD-HHT is relatively new to the ocean remote sensing community,
a brief summary of the technique based on Huang et al. (3) is given in this section.
Basically, the EMD-HHT method requires two steps in analyzing the data. The first
step is to decompose time series data into a number of intrinsic mode functions
(IMFs). These functions must satisfy the following two conditions: (a) within the
entire data set, the total number of extrema (as a function of time) and the total
number of zero-crossings must either be equal or differ at most by one, and (b) at
any point, the mean value of the envelope defined by the local minima (as a function
of time) and the envelope defined by the local maxima (as a function of time) must
be zero. The second step is to apply the Hilbert transform to the decomposed IMFs
and construct the energy–frequency–time distribution, designated as the Hilbert
spectrum. The presentation of the final results of the time–frequency analysis is
similar to the wavelet transform method, which is a spectrogram (time–fre-
quency–energy plot). For clarity, a spectral analysis of the corresponding signal is
shown next to the spectrogram when we present our results in the next sections.
The decomposition of the time series data (i.e. H(t)) into IMFs uses separately
defined envelopes of local maxima and minima. Once the extrema are identified, all
the local maxima are connected by a cubic spline to form the upper envelope. The
procedure is repeated for the local minima to produce the lower envelope. Their
mean is designated as m1(t), and the difference between the time series data and
m1(t) is the first component, h1(t). One can repeat this procedure k times, until hk(t)

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 61 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 61

is an IMF. Then, hk(t) = c1(t) is the first IMF component of the data. c1(t) should
contain the finest scale or the shortest period component of the signal. Then, c1(t)
is separated from the data, and the process is repeated until either the component
cn(t) or the residue rn(t) becomes so small that it is less than a predetermined value,
or when the residue rn(t) becomes a monotonic function from which no IMF can be
extracted.
After all IMFs have been determined, one can check the original data with the
sum of the IMF components

()
H t = ∑ c (t ) + r (t ) .
i =1
1 n (3.1)

Thus, a decomposition of the data into n empirical modes and a residue rn(t) is
achieved. The rn(t) can be either the mean trend or a constant.
After IMFs of the data have been generated, the next step is to apply the Hilbert
transform to each IMF time series. For an arbitrary time series X(t), one can define
its Hilbert transform, Y(t), as:

()

( ) dτ.
X τ

1
Y t = (3.2)
π t−τ
−∞

With this definition, X(t) and Y(t) form a complex conjugate pair, so one has an
analytic signal Z(t) as

() ()
Z t = X t + iY t , () (3.3)

where its amplitude function a(t) is

() ()
a t =  X 2 t + Y 2 t  2 , () (3.4)

and its phase function (t) is

()
Y t 
θ t = arctan  .
()
()
(3.5)
 X t 

The instantaneous frequency is defined as

ω=
dθ t ( ). (3.6)
dt

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 62 Thursday, May 19, 2005 3:42 PM

62 The Hilbert-Huang Transform in Engineering

Once the instantaneous frequency (as function of time) of each IMF has been
generated, the final result of the time–frequency analysis is similar to the wavelet
transform method, time–frequency–energy plot, or spectrogram. The energy density
similar to Fourier transform can be generated by summing the spectrogram for the
whole time series along a constant frequency for each frequency.
From the definition of the phase function and instantaneous frequency, we can
clearly see that for a simple function such as a = sin(t), the Hilbert spectrum is
simply cos(t), and the phase function is a monotonic straight line. Hence, the
spectrogram is simply a horizontal line for a whole time along fixed frequency, and
its spectrum is simply a single peak at that particular frequency.
In the next sections, we describe a few examples of applications of the HHT in
ocean-atmosphere remote sensing data processing and research. These examples
include an analysis of TOPEX/Poseidon sea level anomaly interannual variation
using HHT and empirical orthogonal function (EOF), application of HHT to ocean
color remote sensing of the Delaware Bay, and Mediterranean outflow and Meddies
determined from satellite multisensor remote sensing.

3.2 ANALYSES OF TOPEX/POSEIDON SEA LEVEL ANOMALY


INTERANNUAL VARIATION USING HHT AND EOF
The measurement of sea surface height provides a unique opportunity to make
significant contributions to many of the science objectives of oceanatmosphere
remote sensing scientists working in the area of climate variability, oceanatmosphere
interactions, and upper ocean dynamics. Satellite altimeter data can be employed to
gain vital new understanding of the nature of the ocean’s circulation and ocean
atmosphere coupling, both in terms of mean state and variability and interrelation-
ships between this circulation, its variability, and global climate change.
We used TOPEX/Poseidon (T/P) monthly sea level anomaly (SLA) data from
January 1993 to August 2002, obtained from the University of Texas’ Center for
Space Research (UT/CSR). The deviation of the sea surface was removed from a
mean surface and was preceded by instrument corrections (ionosphere, wet and dry
troposphere, and electromagnetic bias) and geophysical corrections (tides and
inverted barometer). The availability of accurate altimeter data, such as that from
T/P, with the root-mean-square (RMS) error as small as 2 cm (5), allows one to
examine the annual and interannual sea surface variability due to El Niño Southern
Oscillation (ENSO) and global climate change. The original spatial range was 65°S
to 65°N, 180°E to 180°W; however, we limited the latitude range to 60°S to 60°N.
The resolution of the data set was 1° by 1°.
To isolate the signal of the interannual variation of SLA, the annual fluctuation
was removed by subtracting the overall monthly mean for a given month from the
individual monthly data for that respective month. According to calculation, the
interannual energy accounted for about 60% of the total energy. Furthermore, to
analyze the resulting time series, EOF method was applied. The first three EOFs,
EOF-1, EOF-2, and EOF-3, accounted for 35.4%, 12.6%, and 9.8% of the total
interannual energy, respectively.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 63 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 63

Interannual EOF-1 (35.4%) IMF-1 of EOF-1


0.02
60N
0.01
40N
0
20N
−0.01
EQ
−0.02
93 94 95 96 97 98 99 00 01 02
20S IMF-2 of EOF-1
0.1
40S
0.05
60S 0
60E 120E 180 120W 60W
−0.05
−0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02
−0.1
93 94 95 96 97 98 99 00 01 02
IMF-3 of EOF-1
0.2

0.3 0.1

0.25 0

0.2 −0.1
0.15 −0.2
93 94 95 96 97 98 99 00 01 02
0.1
Residual of EOF-1
0.08
0.05
0.06
0
0.04
−0.05
0.02
−0.1 0
−0.15 −0.02
93 94 95 96 97 98 99 00 01 02 93 94 95 96 97 98 99 00 01 02

FIGURE 3.1 The spatial (upper) and temporal (lower) interannual EOF-1 (left) and EMD
of the temporal mode (right) of the T/P-SLA.

Due to its important role in interannual variation, EOF1 was further analyzed
using the HHT. Empirical mode decomposition (EMD) was applied, and EOF1 was
decomposed into three IMFs plus a residue (Figure 3.1). The temporal mode of
EOF1 shows a peak in late ‘97, which correspond to the strong 1997–1998 El Niño.
To examine the effect of ENSO events on EOF1, the Southern Oscillation Index
(SOI) was also used as a reference by decomposing it using EMD. The IMF4 of the
SOI appeared to be a smoothed curve of the SOI, and Salisbury and Wimbush (6)
used this to predict future ENSO events.
Here a comparison was made between the EOF-1 and the SOI by calculating
the correlation coefficient between the IMF-3 of the EOF-1 and IMF-4 of the SOI.
The correlation coefficient was found to be as high as 0.85 (Figure 3.2). Considering
this, the EOF-1 can be considered to behave under the impact of the ENSO, and we
can further infer that the ENSO contributes to about one third of the interannual
SLA variation.
Moreover, we can reexamine the spatial mode of EOF1 (Figure 3.1) to see the
global impact of the ENSO events on human activities. Besides the coastal areas of
America, Australia, and Indonesia, which are most strongly and directly affected,
those of east Africa and Japan also are undergoing remarkable ENSO variation. The

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 64 Thursday, May 19, 2005 3:42 PM

64 The Hilbert-Huang Transform in Engineering

Correlation = 0.85
0.1

0.05

−0.05

−0.1

−0.15
93 94 95 96 97 98 99 00 01 02 03

FIGURE 3.2 Correlation between negative IMF-4 of SOI (red) and IMF-3 of EOF-1 (blue)
is 0.85. It suggests that the primary interannual EOF is impacted by ENSO.

impact even spread as far as Antarctica. On the other hand, coastal areas of the
Atlantic Ocean are less affected by the ENSO events.
In addition, the HHT spectrum of the EOF1 was investigated to extract the
frequency information (Figure 3.3). The frequency distribution of the EOF1 is
between 0 and 4 cycle/yr. The 2.5 to 4 cycle/yr range corresponds to the spectrum
of the IMF1, which behaves in a relatively random pattern. The 0.5 to 2 cycle/yr
range corresponds to the spectrum of IMF2, which has more energy within the
frequency range 0.5 to 0.8 cycle/yr. The 0.2 to 0.4 cycle/yr range has much higher
energy, which is shown as a dark red curve at the bottom part of the spectrum. This
curve corresponds to the frequency feature of IMF3, i.e., a period of 2.5 to 5 yr,
which can be considered as the typical frequency of ENSO events. The lowest
frequency part of the 0 to 0.1 cycle/yr range corresponds to the frequency of the
residue, which is the longterm trend. It has a period of tens of years or longer and,
therefore, requires much longer time series to analyze.

3.3 APPLICATION OF HHT TO OCEAN COLOR REMOTE


SENSING OF THE DELAWARE BAY
The use of satellite-based ocean color data has become an integral part of oceano-
graphic studies, aiding in the exploration of a number of important topics. The ability
to collect ocean color data at high temporal and spatial resolutions, which only
satellite sensors can provide, has given the oceanographic community a rich data

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 65 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 65

HHT Spectrum of EOF-1

4
Hilbert Spectrum (cycles per year)

3.5

2.5

1.5

0.5

93 94 95 96 97 98 99 00 01 02
Year

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

FIGURE 3.3 The HHT spectrum of EOF-1. The dark red curve indicates high energy at
frequencies between 0.2 and 0.4 cycle/yr, which corresponds to the Hilbert spectrum of IMF-3
of interannual EOF-1. It indicates that ENSO events have a typical period of 2.5 to 5 yr.

source. However, use of this data source is dependent on the premise that there are
discernable relationships between reflectance exiting the water column and the
constituents in it. A primary constituent of study has been chlorophyll-a on account
of its connection to phytoplankton and thus to primary production and biomass (7).
As coastal management strategies begin to focus on large-scale ecosystem based
programs, the relationship between chlorophyll-a and primary production and bio-
mass has the potential to provide a cost-effective alternative to traditional ship-based
or point source sampling. Monitoring and assessing the health of ecosystem-size
regions will be much more feasible with satellite-based parameters due to the
frequent repeat periods and large spatial areas covered by satellite sensors. Thus,
the quantitative assessment of water constituents from satellite-based ocean color
data holds great promise for implementing dynamic management strategies. How-
ever, to interpret ocean color data appropriately, a full understanding of the period-
icity of chlorophyll concentrations in a given area is essential. The purpose of this
study is to examine the seasonal and interannual chlorophyll-a cycles from satellite
data of the coastal region at the mouth of the Delaware Bay.
Although quantifying chlorophyll concentration from satellite ocean color data
has been successful in the open ocean, coastal areas (case II water) can still present
problems. Water constituents associated with coastal regions, such as excessive
chlorophyll-a concentrations, suspended sediments, and colored dissolved organic

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 66 Thursday, May 19, 2005 3:42 PM

66 The Hilbert-Huang Transform in Engineering

material, create an optically complex area in which traditional empirical models


typically perform poorly (8). However, several studies have used an innovative neural
network (NN) methodology successfully to develop ocean color algorithms in Case
II water that outperform empirical methods (9, 10, 11, 12). NN based algorithms
are advantageous on account of their ability to model complex, nonlinear geophysical
transfer functions like those required to relate chlorophyll-a concentrations to the
reflectance measured by a satellite sensor (13). The NN requires training data
consisting of input data paired with the desired output, from which it “learns” the
geophysical relationship. The ability of a NN to generalize (i.e. accurately predict
the geophysical parameter) is assessed through a validation set of pair data that the
NN has never encountered. Given the success of NNs in coastal water algorithm
development, the chlorophyll-a data used in this study was produced from a NN
trained and validated using the Sea-Viewing Wide Field-of-View Sensor (SeaWiFS)
remotely sensed reflectance data paired with in situ chlorophyll data in the Delaware
Bay and its adjacent coastal water. These paired data points were collected during
a number of different days and seasons. The neural network showed significant
improvement over the current SeaWiFS processing algorithm (Ocean Color 4 [OC4])
and was used to reprocess the satellite data used in this study (14).
Over the course of the SeaWiFS project, daily high-resolution picture transmis-
sion (HRPT) images of the Delaware Bay and the adjacent coastal zone were
subjectively viewed to identify all the cloud-free images of the Delaware Bay and
the adjacent coastal zone from September 1997 to July 2003. This period covers
almost six years of data, from which 468 images were collected, with an average
of 6.6 images per month. The daily images were obtained at the L1A processing
level and converted to a remapped L2 remote sensing reflectance (Rrs) product. The
remapping used a standard Mercator projection with a fixed grid of 890 × 890 pixels.
Each image ranges from 34°N to 42°N latitude and from 68°W to 78°W longitude,
which covers approximately 9.58 × 105 km2. The Rrs product was then used as input
for the optimal NN model mentioned earlier to produce daily chlorophyll-a maps
of the Delaware Bay and the adjacent coastal zone. Since NNs are extremely good
at interpolation but poor at extrapolation, only a limited subset (the 121 × 121 pixel
Delaware Bay and adjacent coastal zone) of the entire SeaWiFS image was processed
using the NN. This ensured that the model would only be applied to the region
where it was specifically trained with in situ data. The SeaWiFS imagery was binned
by month and geometrically averaged to produce an image of the mean monthly
values of chlorophyll-a concentrations for each month. This organizational process
created spatially coherent patterns of chlorophyll-a values at evenly spaced time
scales that help reduce the effects of sensor and algorithm errors and minimize the
influence of high frequency variability on seasonal distributions of chlorophyll-a
concentrations (15). Furthermore, due to the fact that chlorophyll pigments tend to
be lognormally distributed, the geometric means were used in the various statistical
procedures applied to chlorophyll-a values in this study (15). A station at the mouth
of the Delaware Bay (38.916°N, –75.100°W) was used as the geographic point from
which a monthly times series of chlorophyll-a concentrations was produced. This
monthly time series is shown in Figure 3.4.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 67 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 67

25
Chlorophyll-a (mg/m3)
20

15

10

0
1998 1999 2000 2001 2002 2003

1
NAO Index

−1

−2

−3
1998 1999 2000 2001 2002 2003
Years

FIGURE 3.4 The time series of the monthly means of the chlorophyll-a concentrations (top
panel) and the North Atlantic Oscillation index (bottom panel) from September 1997 to July
2003.

Hilbert-Huang transform (HHT) analysis provides an innovative way to study


the time–frequency characteristics of time series, due to its ability to handle nonlinear
and nonstationary data and to account for changes in frequency over time. This novel
technique has been applied to the time series of chlorophyll-a concentrations at the
mouth of the Delaware Bay, in order to better understand the forcing mechanisms
associated with seasonal and interannual variations. The initial chlorophyll-a time
series was decomposed into IMFs, to which the Hilbert transform can be applied.
The resulting Hilbert spectrum and its associated marginal spectrum are shown in
Figure 3.5.
The marginal spectrum is conceptually similar to a spectral density graph pro-
duced by Fourier analysis; however, it is produced by summing the instantaneous
energy associated with a constant frequency over the entire time period. Thus, the
marginal spectrum provides a measurement of the frequencies that contain the largest
amounts of energy. As expected, the dominant energy peak occurs at about 0.09,
which is the approximate annual forcing frequency and corresponds to period of
11.1 months. This result is supported by the fact that several annual physical forcing
factors, such as sea surface temperature, solar insolation, and stream flow, have been
shown to strongly influence chlorophyll-a concentrations in the Delaware Bay and
its adjacent coastal zone (8, 16). The fact that the chlorophyll-a periodicity, typically

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 68 Thursday, May 19, 2005 3:42 PM

68 The Hilbert-Huang Transform in Engineering

0.4

0.35
0.35

0.3 0.3
Frequency (cycles/month)

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05
0.05

0
1998 1999 2000 2001 2002 2003 0 10 20
Years Energy

FIGURE 3.5 The Hilbert spectrum (left panel) and marginal spectrum (right panel) for the
IMFs of the chlorophyll-a times series.

presumed to be annual, is not precisely 12 months is not unexpected, since the


biological processes are not simply controlled by physical processes. Examining the
Hilbert spectrum along the 0.09 frequency shows that there is generally high energy
along this signal, with the exception of two time periods. From January 1999 to
October 1999, the annual signal completely degenerates, and in the summer of 2001
there is a weakening of this signal. Without the ability of the HHT to ascertain
instantaneous frequency, these breakdowns in the annual periodicity would be dif-
ficult to determine.
In addition to the strong annual signal, there is a significant energy peak in the
lower frequencies. The marginal spectrum indicates an energy peak at the 0.02
frequency, which corresponds to a 4.2-yr period. However, looking at the Hilbert
spectrum shows that the energy at the 0.02 frequency is not constant; instead, it
appears to be roughly oscillating around the 0.02 frequency. This energy begins to
increase in frequency from mid-1999 to late 2001, when the signal briefly weakens
and starts to decrease in frequency. This interannual signal is primarily captured in
IMF 4 of the empirical mode decomposition (Figure 3.6). The relatively long period
of the mode suggests that a slowly varying basin-scale process could be a forcing
candidate.
One such basin-scale process is the North Atlantic Oscillation (NAO). NAO is
a measure of atmospheric conditions focused around the North Atlantic. The NAO
can have a positive or negative phase, as determined by the NAO index, which is
the surface sea-level pressure difference between the Subtropical (Azores) High and

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 69 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 69

0.5 2

Interannual variation of Chlorophyll-a


Interannual variation of NAO

0 0

−0.5 −2
1998 1999 2000 2001 2002 2003
Years

FIGURE 3.6 Interannual variation (IMF 4) for both the NAO index (blue line) and the
chlorophyll-a (green line) time series.

the Subpolar Low. Both positive and negative phases affect basin-wide variations in
the North Atlantic jet stream and storm tracks, as well as in large-scale modulations
of the typical patterns of zonal and meridional heat and moisture transport. These
alterations will consequently affect temperature and precipitation patterns and can
extend from eastern North America to western and central Europe (17). Furthermore,
recent studies have shown links between the NAO and hydrographic properties along
the northeast Atlantic shelf slope and the Gulf of Maine, and zooplankton and
chlorophyll-a concentrations within the Gulf of Maine (18). Thus, the interannual
frequency of the NAO was examined.
Monthly mean data of the NAO index from September 1997 to July 2003 were
obtained from the NOAA Climate Prediction Center. The NAO index data are shown
in Figure 3.4. This time series was also decomposed into its component IMFs in
order to isolate the NAO interannual mode. After decomposing the NAO index time
series, the fourth IMF contained an interannual signal that exhibited features very
similar to the fourth chlorophyll-a IMF (Figure 3.6).
To quantify this relationship, cross-correlation analysis was performed on the
IMF 4 of both the chlorophyll-a concentrations and the NAO index. The results are
shown in Figure 3.7. This study focused only on positive lag time (chlorophyll-a
time series following the NAO time series), because it would not be sensible to
assume that negative lags (chlorophyll-a time series leading the NAO time series)
were meaningful. The correlation analysis reveals that the NAO and chlorophyll-a

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 70 Thursday, May 19, 2005 3:42 PM

70 The Hilbert-Huang Transform in Engineering

Maximum correlation is r = −0.67 at tiag = 10


(95% significance level: rsig = 0.25)
1
Correlation coefficient (r)
95% significance level
0.8

0.6

0.4
Correlation coefficient

0.2

−0.2

−0.4

−0.6

−0.8

−1
0 10 20 30 40 50 60 70
Time lag
Chlorophyll-a follows NAO Index for lag > 0

FIGURE 3.7 Correlation coefficient (blue line) and the related 95 % confidence interval (red
dashed line). The maximum correlation coefficient with a positive lag (i.e. chlorophyll-a
interannual variation following NAO index interannual variation) is –0.67, corresponding to
a lag of 10 months.

concentrations have their highest correlation (correlation coefficient of –0.67) when


chlorophyll-a concentration lags (follows) the NAO by 10 months. The result is
significant at the 95% confidence level, as shown in Figure 3.7. Although a high
correlation coefficient does not guarantee a relationship between the two times series,
the fact that previous studies (mentioned earlier) have shown that the NAO can affect
the interannual time scales of biological and physical processes in oceanic regions
as distant as the Gulf of Maine suggests that this correlation is important.
This study examined the seasonal and interannual variability in satellite derived
chlorophyll-a data from the mouth of the Delaware Bay for a 6-yr period. By
applying the HHT to the chlorophyll-a time series, a more complete understanding
of its time–frequency characteristics was obtained. The resulting Hilbert spectrum
displayed a strong annual signal, with an unexpected weakening at two brief inter-
vals. In addition, the energy at interannual frequencies was compared to the NAO.
The fourth IMF of the NAO index and the fourth IMF of the chlorophyll-a concen-
trations displayed strong similarities and had a cross-correlation coefficient of –0.67

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 71 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 71

that was significant at the 95% confidence level. This relatively high correlation
coefficient occurred at a lag of 10 months, which suggests that the possible inter-
annual link between chlorophyll-a concentrations at the mouth of the Delaware Bay
and the NAO index may result from chlorophyll-a concentrations responding to
variations in the NAO index. Finally, the observations of both the weakening of the
seasonal signal and the suggested teleconnection of the interannual variation between
chlorophyll-a concentrations and the NAO are topics worthy of future investigation
that would have been difficult to identify with traditional time series analysis tech-
niques.

3.4 MEDITERRANEAN OUTFLOW AND MEDDIES (O & M)


FROM SATELLITE MULTISENSOR REMOTE SENSING
One of the most interesting and prominent features of the North Atlantic Ocean is
the salt tongues originating from an exchange flow between the Mediterranean Sea
and the Atlantic through the Strait of Gibraltar. The Mediterranean outflow through
the strait is heavier than Atlantic water due to its higher salt content. Evaporation
in the Mediterranean Sea raises water salinity to approximately 38.4 ppt, compared
to 36.4 ppt in the eastern North Atlantic. After leaving the strait under the incoming
lighter North Atlantic water, the Mediterranean outflow sinks and turns to the right,
due to the Coriolis force, following the continental slope of Spain and Portugal.
Eventually, the Mediterranean water leaves the coast and spreads out into the middle
North Atlantic, forming a tongue of salty water and generating clockwise eddies in
the process. These Mediterranean eddies, or Meddies, are rapidly rotating double
convex lenses that contain a warm, highly saline core of Mediterranean water. They
are typically 100 km in diameter, extend over 800 m vertically, and are located at
a depth of 1000 m.
Generally speaking, most of the remotely sensed oceanographic observations
are confined to either the sea surface or the top part of the upper mixed layer. This
limitation is not always true. Yan et al. (19, 20, 21) and Yan and Okubo (22)
developed methods to infer the upper ocean mixed layer depths from multisensor
satellite data. Stammer et al. (23) applied Geosat Exact Repeat Mission altimeter
data to investigate the possible surface signal of the Meddies. In this section, we
report a new method to study the O&M by satellite integration analyses of altim-
eter, scatterometer, SST, and XBT data with help of HHT. We computed the sea
level difference, namely ∆η′ = η′Total − ηUL
′ = η′Total − ( η′T + η′S + ηW
′ ). η′Total is the devi-
ation of the sea surface topography of the whole vertical water column, which can
be measured by TOPEX/Poseidon (T/P) altimetry. ηUL ′ is the sea surface variation
in the upper layer (400 m) due to thermal expansion (η′T ), calculated from XBT.
Salt expansion ( η′S ) and wind stress (ηW′ ) are computed from satellite scatterometer
data. The HHT was applied to help interpret ∆η′.
To estimate the sea level variation due to thermal expansion in the upper layer,
sea surface height anomaly was calculated using monthly mean XBT (η′T ) data from
the Joint Environmental Data Analysis (JEDA) Center by the integration of temper-
ature down to the 400 m depth from January 1993 to December 1999, i.e.,

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 72 Thursday, May 19, 2005 3:42 PM

72 The Hilbert-Huang Transform in Engineering

η′T = α
∫ ∆Tdz.
−400

∆T is the temperature difference between two different depths (dz), and α is the
thermal expansion coefficient, which is calculated as function of salinity (S) and
water pressure (P) based on empirical measurements (24);

1 ∆ρ
α=− .
ρ0 ∆T
( s , p)

The spatial resolution of XBT data is 73° × 61° longitude/latitude, which is inter-
polated to 1° × 1° longitude/latitude data for this study. It is possible that errors
occurred due to interpolation of the time rates of change of the integrated upper
ocean heat storage anomaly (5 W/m2, on average); this is discussed by White and
Tai (25). We also checked interpolation error using 1° × 1° longitude/latitude opti-
mum interpolation sea surface temperature (OISST). We compared interpolated sea
surface temperature (SST) from XBT data and OISST by calculating correlation
coefficients. Correlation coefficients between OISST and SST from XBT were over
95% with RMS 0.2°C, except for two small areas, and correlation coefficient 80%
with RMS 1.2°C. We conclude that interpolation error of XBTs is less than 1°C at
the sea surface.
Because of scarcity of temporal and spatial salinity data, we estimated the effect
of η′S in the upper layer indirectly. Two different correlation and RMS calculations
were made: the first correlation is between η′T and OISST, and the second correlation
is between η′Total and OISST. Good temporal and spatial agreement between SST
and η′Total or η′T suggests that a robust regression between fields may have some
physical significance with respect to thermal expansion, but a low correlation with
high RMS may have some other variability in the mixed layer or below the mixed
layer due to the salinity. It appears that the difference is due to a change in ocean
salinity, which is reflected in the T/P sea level measurements but not in the η′T
measurements. The first correlation coefficients were all over 90% with smaller RMS
than the second ones, but the second ones were also over 60% to 80%, with larger
RMS than the first ones. This result is caused by a warm and salty core below the
mixed layer. Using Levitus 94 (26), the vertical structure of the salinity was examined
to see the salinity distribution. The whole upper layer in our study domain has a
uniform salinity distribution above 1000 m. We conclude that the anomaly of η′S is
spatially uniform with only small variability.
Wind effect on the η′Total signal was considered using scatterometer data from
European Remote Sensing Satellite-1/2 (ERS-1/2) from January 1993 to December
1999. First we calculated the correlations between wind stress curl and the η′Total for
the temporal variation. Negative correlation would be expected; rising sea level is
associated with negative wind stress curl. Correlation coefficients showed a relation
of about –30% in our study area, with southern areas having larger RMS than

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 73 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 73

northern areas. Second, we considered the magnitude of sea level variation due to
wind stress based on a linear barotropic vorticity equation, i.e.,

d η − g′
≈ ∇x τ,
dt f0ρg

where η is the sea level height, ρ is the water density, g′ is the reduced gravity, f0
is the Coriolis parameter, and τ is the wind stress. We also estimated ηW ′ by using
NASA’s scatterometer: NSCAT (August 1996 to June 1997) and Quikscat (June
1999 to present). The mean sea level variation due to wind stress was around 1 cm.
The result of ∆η′ = η′T P − ηUL′ is apparently some variation of the vertical water
column, which is a response according to the different incoming or outgoing water
mass under the upper layer. Sea level variation due to salinity (temperature) change
using mass conservation with 800 m thick is around 60 cm (32 cm), with 1-ppt
differences from background fluid (2°C), derived from the salt expansion coefficient
7.5 × 10–4/°C and thermal expansion coefficient 2.0 × 10–4/°C, respectively. The
bottom pressure (deep current) variation is negligible at this region (26, 27). How-
ever, since the Meddy was a weakly stratified result of extensive salt fingering (28),
there were regions of high stratification above and below the lens, where the back-
ground isopycnal surface becomes increasingly broader as it moves above the Meddy
toward the sea surface. Because of this isopycnal compensation, the O&M are not
revealed in the η′Total signal. This is why we cannot detect a Meddy with the altimeter
observation alone.
We computed the absolute differences of the sea surface height anomaly of ηUL ′
and η′Total to examine the trajectories of the O&M from January 1993 to December
1999. Figure 3.8 is an example of such trajectories from July 1998 to June 1999.
One can see a strong signature (|∆η′|) toward west and south from July to November.
Generally, southward Meddies are formed near 36°N by separation of the frictional
boundary layer at sharp corners (29) and in the Canary Basin (30, 31, 32). The
southward travelling Meddies can be explained by the strongest low frequency zonal
motions driven by baroclinic instability (33) and the influence of the neighboring
mesoscale features (cyclonic vortices or Azores Current meanders) in the regions
(34). The northward O&M over 36° to 40°N are also shown from August to October.
Furthermore, a strong signature over 40°N in February is considered to be the
influence of the returning Gulf Stream. From March to June the weak O&M signa-
tures are found because of the weak salinity deviation in 1000 m depth examined
from climatological data.
To investigate the reasons for the change in the direction of propagation of
Meddies, the stream function (ψ) was computed by using the T/P altimeter. This
representation of the stream function permits observations of the interactions
between the sea surface gradient and the Meddies’ propagation. The computation
of the sea surface height anomaly η′ in terms of the usual dynamic variables is
straightforward, if the flow is assumed to be quasigeostrophic: The stream function
(ψ) is defined as

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 74 Thursday, May 19, 2005 3:42 PM

74 The Hilbert-Huang Transform in Engineering

Jly 1998 Jan 1999


45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 30W 25W 20W 15W 10W 05W
Aug 1998 Feb 1999
45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 30W 25W 20W 15W 10W 05W
Sep 1998 Mar 1999
45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 30W 25W 20W 15W 10W 05W
Oct 1998 Apr 1999
45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 30W 25W 20W 15W 10W 05W
Nov 1998 May 1999
45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 30W 25W 20W 15W 10W 05W
Dec 1998 Jun 1999
45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 30W 25W 20W 15W 10W 05W

0 1 2 3 4 5 6 7 8

(
FIGURE 3.8 Calculation of the ∆η′ ∆η′ = η′Total − ηUL
′ )
from July 1998 to June 1999.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 75 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 75

Jan 1994 Apr 1994


45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 0W 30W 25W 20W 15W 10W 05W 0W

Jly 1994 Oct 1994


45 45

40 40

35 35

30 30
30W 25W 20W 15W 10W 05W 0W 30W 25W 20W 15W 10W 05W 0W

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

FIGURE 3.9 Comparisons of Meddy trajectories with stream functions (Equation 3.7) for
January, April, July, and October 1994.

ψ= g η′Total , (3.7)
f 0

where ψ is the surface quasigeostrophic stream function (35). This is illustrated in


Figure 3.9. Due to the seasonal migration of the tropical water toward the north and
polar water toward the south, we can see that the seasonal signals divide into a
northern part and a southern part at 37°N. We compared the Meddies measured by
the floats and the stream functions in all the months in the experimental periods and
found that the Meddies preferred to travel toward the low streamlines and that they
stayed within the northmost and southmost streamlines in April and October, respec-
tively.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 76 Thursday, May 19, 2005 3:42 PM

76 The Hilbert-Huang Transform in Engineering

3rd EMD mode using HHT at A(30W, 30N) with other places
5
B(20W, 30N)

−5
93 94 95 96 97 98 99 00
5
C(10W, 30N)

−5
93 94 95 96 97 98 99 00
5
D(10W, 37N)

−5
93 94 95 96 97 98 99 00
5
E(10W, 42N)

−5
93 94 95 96 97 98 99 00
5
F(20W, 42N)

−5
93 94 95 96 97 98 99 00
5
G(30W, 42N)

−5
93 94 95 96 97 98 99 00
5
H(30W, 37N)

−5
93 94 95 96 97 98 99 00

FIGURE 3.10 A comparison of the third EMD modes using HHT are shown at given
locations. The solid curve is for location A (refer to Figure 3.11 for the location) in all panels
for comparison. The dotted curve is the individual EMD modes for those locations.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 77 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 77

Annual mean of |η′| with two experiments in 1994


45 5
AMUSE
4.5
SEMAPHORE

G F E 4

3.5
40

2.5
H D

35
1.5

0.5

30 A B C 0
30W 25W 20W 15W 10W 05W

FIGURE 3.11 Comparisons of annual mean |∆η′| signal from two float experiments in 1994.
The gray scale and contours show the annual mean of the |∆η′| signal estimated from our
method. All Meddies were discovered during the AMUSE experiment (stars) and SEMA-
PHORE experiment (circles) in 1994.

To find out the dominant signal of the sea surface interaction with the Meddies,
the HHT was applied to their EMD modes. We chose eight places to investigate the
dominant signal in our study area. Figure 3.10 showed that Place Group 1 (B, C,
and D, shown in Figure 3.11) and Group 2 (H and G, shown in Figure 3.11) had a
similar signal. However, Group 3 (E and F, shown in Figure 3.11) had slightly
different signals from Groups 1 and 2. Consequently, there were seasonal fluctuations
between location A and Group 3. To compute the dominant signal of the power for
locations A and E, HHT was also employed; this is shown in Figure 3.12. In both,
the dominant frequency was around f = 0.082 (1 year). However, location A had a
lower frequency, f = 0.03 (33.3 months), which seems to indicate that wind stress
with 33.3-month period produces sea surface forcing. The surface variations produce
the baroclinic instability on the Meddies, which is related to southward translation.
The southward translation of the Meddies due to the baroclinic effects were consis-
tent with that discussed by Müler and Siedler (36) and by Käse and Zenk (37). If
the current is vertically sheared in a stratified fluid, baroclinic instability can occur.
Figure 3.12 also demonstrates that the comparison between field observations from
two float experiments in 1994 and our computation was excellent.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 78 Thursday, May 19, 2005 3:42 PM

78 The Hilbert-Huang Transform in Engineering

Frequency-Spectrum-Energy of the streamfuction at A Energy


0.45 0.45

0.4 0.4

0.35 0.35

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0 10 20 30 40 50 60 70 80 0 20 40

Frequency-Spectrum-Energy of the streamfuction at E Energy

0.4 0.4

0.35 0.35

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0
0 10 20 30 40 50 60 70 80 0 20 40

FIGURE 3.12 Time–frequency–energy spectrum of the stream functions at locations A and


E. Refer to Figure 3.11 for location A and E. One can see the high energy at the 33-month
period at location A, but not at location E.

3.5 CONCLUSION
Three examples of applications of the HHT method in ocean-atmosphere remote
sensing research were illustrated in this chapter. These examples show that the HHT
method is indeed a potentially very useful and powerful tool for ocean engineering
and science studies.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 79 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 79

ACKNOWLEDGMENTS
This research was supported partially by the National Aeronautics and Space Admin-
istration (NASA) through Grant NAG5-12745 and NGT5-40024, by the Office of
Naval Research (ONR) through Grant N00014-03-1-0337, and by the National
Oceanic and Atmospheric Administration (NOAA) through Grants NA17EC2449
and NA96RG0029.

REFERENCES
1. Chan, Y. T. (1995). Wavelet Basics. Boston, MA: Academic.
2. Huang, N., Long, S., Shen, Z. (1996). The mechanism for frequency downshift in
nonlinear wave evolution. Adv. Appl. Mech. 32:59–111.
3. Huang, N., Shen, Z., Long, S., Wu, M., Shin, H., Zheng, Q., Yen, N.-C, Tung, C.,
Liu H. (1998). The empirical mode decomposition and Hilbert spectrum for nonlinear
and nonstationary time seriesanalysis. Proc. Royal Soc. Mathematical, Phys. Eng.
Science 454:903–995.
4. Yan, X.-H., Zheng, Q., Ho, C.-R., Tai, C.-K., Cheney, R. (1994). Development of the
pattern recognition and the spatial integration filtering methods for analyzing satellite
altimeter data. Remote Sens. Environ. 48:147–158.
5. Cheney, R., Miller, L., Agreen, R., Doyle, N., Lillibridge, J. (1994). TOPEX/POSEI-
DON: 2 cm solution. J. Geophys. Res. 99:24555–24563.
6. Salisbury, J. I., Wimbush, M. (2002). Using modern time series analysis techniques
to predict ENSO events from the SOI times series. Nonlinear Proc. Geophys.
9:341–345.
7. Kiddon, J. A., Paul, J. F., Buffum, H. W., Strobel, C. S., Hale, S. S., Cobb, D., Brown,
B. S. (2003). Ecological condition of U.S. mid-Atlantic estuaries. 1997–1998. Marine
Pollution Bulletin 46(10):1224–1244.
8. Keiner, L. E., Yan, X-H. (1998). A neural network model for estimating sea surface
chlorophyll and sediments from thematic mapper imagery. Remote Sens. Environ.
66(2):153–165.
9. Zhang, Y., Pulliainen, J., Koponen, S., Hallikainen, M. (2002). Application of an
empirical neural network to surface water quality estimation in the Gulf of Finland
using combined optical data and microwave data. Remote Sens. of Environ.
81:327–336.
10. Baruah, P. J., Oki., K., Nishimura, H. (2000). A neural network model for estimating
surface chlorophyll and sediment content at Lake Kasumi Gaura of Japan. In: Con-
ference Processing of GIS Development. Taipei, Taiwan, December.
11. Buckton, D., O’Mongain, E., Danaher, S. (1999). The use of neural networks for the
estimation of oceanic constituents based on the MERIS instrument. Int. J. Remote
Sensing 20(9):1841–1851.
12. Keiner, L. E., Yan, X.-H. (1997). Empirical orthogonal function analysis of sea surface
temperature patterns in Delaware Bay. IEEE Trans. Geoscience Remote Sensing
25(5):1299–1306.
13. Zhang, T., Fell, F., Liu Z.-S., Preusker, R., Fischer, J., He, M.-X. (2003). Evalating
the performance of artificial neural network techniques for pigment retrieval from
ocean color in Case I Water. J. Geophys. Res. 108(C9):3286–3298.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 80 Thursday, May 19, 2005 3:42 PM

80 The Hilbert-Huang Transform in Engineering

14. Dzwonkowski, B. (2003). Development and application of a neural network based


ocean color algorithm in coastal waters. Masters thesis, University of Delaware,
Newark, DE.
15. Barnard, A. H., Stegmann, P. M., Yoder, J. A. (1997). Seasonal surface ocean vari-
ability in the South Atlantic Bight derived from CZCS and AVHRR imagery. Conti-
nental Shelf Res. 17:1181–1206.
16. Pennock, J. R. (1985). Chlorophyll distributions in the Delaware estuary: Regulation
by light-limitation. Estuarine, Coastal and Shelf Science 21:711–725.
17. Hurrell, J. W. (1995). Decadel trends in the North Atlantic Oscillation regional
temperatures and precipitation. www.cgd.ucar.edu/cas/papers/science1995/sci.htm.
18. Thomas, A. C., Townsend, D. W., Weatherbee, R. (2003). Satellite-measured phy-
toplankton variability in the Gulf of Marine. Continential Shelf Res. 23:971–989.
19. Yan, X.-H., Schubel, J. R., Pritchard, D. W. (1990). Oceanic upper mixed layer depth
determination by the use of satellite data. Remote Sens. Environ. 32 (1):55–74.
20. Yan, X.-H., Okubo, A., Shubel, J. R., Pritchard, D. W. (1991). An analytical mixed
layer remote sensing model. Deep Sea Res. 38:267–287.
21. Yan, X.-H., Niller, P. P, Stewart, R. H. (1991). Construction and accuracy analysis
of images of the daily-mean mixed layer depth. Int. J. Remote Sensing
12(12):2573–2584.
22. Yan, X.-H., Okubo, A. (1992). Three-dimensional analytical model for the mixed
layer depth. J. Geophy. Res., 97(C12):20201–20226.
23. Stammer D.,. Hinrichsen, H. H, Käse, R. H. (1991). Can Meddies be detected by
satellite altimetry? J. Geophys. Res. 96:7005–7014.
24. Wilson, W., Bradley, D. (1996). Technical Report NOLTR. 66–103.
25. White, W. B., Tai, C.-K. (1995). Inferring interannual changes in global upper ocean
heat storage from TOPEX altimetry J. Geophys. Res. 15:24943–24954.
26. Levitus, S., Boyer, T. P. (1994). World Ocean Atlas, Vol. 4, Temperature, NOAA Atlas
NESDIS 4, U.S. Govt. Print. Off., Washington, D.C., 117.
27. Iorga, M. C., Lozier, M. S. (1999). Signature of the Mediterranean outflow from a
North Atlantic climatology, 1. Salinity and density fields. J. Geophys. Res.
104:25985–26009.
28. Hebert, D. L. (1988). A Mediterranean salt lens. Ph.D. dissertation, Dalhousie Uni-
versity, Halifax, Nova Scotia.
29. D’Asaro, E. A. (1993). Generation of submesoscale vortices: A new mechanism. J.
Geophys. Res. 93:6685–6693.
30. Armi, L., Stommel, H. (1983). Four views of a portion of the North Atlantic sub-
tropical gyre. J. Phys. Oceanogr. 13:828–857.
31. Armi, L., Zenk, W. (1984). Large lenses of highly saline Mediterranean salt lens. J.
Phys. Oceanogr. 14:1560–1576.
32. Richardson, P. L., Tychensky, A. (1998). Meddy trajectories in the Canary Basin
measured during the SEMAPHORE experiment, 1993–1995. J. Geophys. Res.
103:25029–25045.
33. Spall, M. A., Richardson, P. L., Price, J. (1993). Advection and eddy mixing in the
Mediterranean salt tongue. J. Mar. Res. 51:797–818.
34. Tychensky, A., Carton, X. (1998). Hydrological and dynamical characterization of
Meddies in the Azores region: A paradigm for baroclinic vortex dynamics. J. Geophys.
Res. 103:25061–25079.
35. Douglas, B. C., Cheney, R. E, Agreen, R. W. (1983). Eddy energy of the northwest
Atlantic and Gulf of Mexico determined from GEOS 3 altimetry. J. Geophys. Res.
88:9595–9603.

© 2005 by Taylor & Francis Group, LLC


DK342X_book.fm copy Page 81 Thursday, May 19, 2005 3:42 PM

Applications of Hilbert-Huang Transform 81

36. Müler, T. J., Siedler, G. (1992). Multi-year current time series in the eastern North
Atlantic Ocean, J. Mar. Res. 50:63–98.
37. Käse, R. H., Zenk W. (1996). Structure of the Mediterranean water and Meddy
characteristics in the northeastern Atlantic. In: Ocean Circulation and Climate:
Observation and Modeling the Global Ocean, W. Krauss, Ed. Berlin: Gebrüder
Bornträger, Berlin, pp. 365–395.

© 2005 by Taylor & Francis Group, LLC

Potrebbero piacerti anche