Professional Documents
Culture Documents
[Ecole Polytechnique]
On: 14 February 2007
Access Details: [subscription number 731975950]
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954
Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
1. Introduction
The structure and evolution of large synoptic systems, such as atmospheric lows
of temperate latitudes, and their main climatic characteristics (areas of formation
and motion, lifetime) have been studied for several decades. The physical processes
involved in these systems are also quite well understood. A practical application of
these properties is the weather forecast, which nowadays gives reliable results over
a period of ve days. The mechanisms involved in the formation of various cloud
types must also be reproduced as closely as possible to reality in climate models,
because the interaction between clouds and radiation plays an essential role in the
retroactions and equilibria of the climate.
Forecast models are nevertheless unable to reconstruct correctly specic, fast
evolving, meteorological situations. For example, some winter storms observed over
*e-mail: szantai@lmd.polytechnique.fr
Western Europe result from the evolution of fronts associated to a primary extra-
tropical low. In such situations, it is possible that small-scale processes have an
inuence on larger scale processes. From an observational point of view, the life
cycle of meteorological structures, i.e. their formation, their evolution and their
disappearance, is not well known; only snapshots are available at specic instants.
In the case of cyclogenesis, the rst images, necessary for a reliable forecast, are
missing. In practice, it is easier to obtain individual images than to track these
meteorological objects during their whole lifetime. These images are reconstructed
from synoptic measurements made from meteorological stations on land or on the
sea (from ships or buoys), from aeroplanes, radiosondes and satellites (which in turn
send real images).
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
The observation of the temporal evolution of clouds at diVerent space and time
scales could improve our understanding of the physical processes involved in their
formation, their persistence and their dissipation. Our method for the calculation
of trajectories from satellite images was developed for this purpose. Its use could
help to characterize clouds more realistically and thus be of benet to general
(meteorological and climatic) circulation models.
The Lagrangian approach of motion, which is implicit in the determination of
trajectories, is currently used in several elds of atmospheric sciences. Trajectories
of cyclones and hurricanes over the North Atlantic have been reconstructed since
1871 (Neumann et al. 1978). Trajectories of pollutants have also been determined.
These can be of natural origin: dust storms originating from the Sahara desert have
been observed over the Atlantic Ocean and the Caribbean (Carlson and Prospero
1972, Bertrand et al. 1975); or the result of chemical or radioactive pollution: the
fallout of the radioactive cloud emitted by the Chernobyl nuclear power plant in
1986 had a strong ecological impact. Its trajectories (there have been several emis-
sions) have been reconstructed from analysed data and a circulation model, and
conrmed by ground samplings of radioactive dust (containing 137Cs) (Strauss and
Gros 1986, Albergel et al. 1988). Balloons have also been used to investigate the
atmospheric circulation. Their trajectories were rst constructed at lower levels of
the troposphere (Angell and Pack 1960). During large-scale experiments like EOLE
(Morel and Bandeen 1973, Desbois 1976) and TWERLE (TWERLE team 1977),
trajectories at high level were also determined in selected regions.
Geostationary satellites have the advantag e to provide images at short time inter-
vals. The evolution of clouds can be observed on them with a far better temporal
resolution (h for Meteosat images) than the resolution provided by classical measure-
ments (generally 6 h or 12 h). The spatial resolution of these images is also well
adapted for a precise study of meteorological objects at synoptic scale such as
extratropical lows, or at mesoscale such as tropical convective systems. In the rst
studies, squall lines in tropical regions were identied and tracked visually on a
series of images (Aspliden et al. 1976, Desbois et al. 1988). Automatic methods were
also developed: convective cloud clusters are extracted with the use of a threshold
on the brightness temperature deduced from thermal infrared images. Basically, a
convective cluster can be tracked if the areas it covers on two consecutive images
overlap. With this type of method, the motion of convective clusters over North
America (Maddox 1980), in the region of Borneo (Williams and Houze 1989) and
over Africa (Arnaud et al. 1992, Laing and Fritsch 1993) has been reconstructed.
The motion of extratropical cyclones can also be reconstructed from a series of
images. A recently developed method reconstructs the motion of a singularity, in
Construction of cloud trajectories f rom satellite images 1701
this case the centre of a spiral, and its neighbouring points can be represented with
a two-dimensional linear model. The parameters dening the position of the centre
of rotation and characterizing its motion are estimated by a statistical method
(Maurizot et al. 1997 ).
Another recent development from satellite images is the determination of large-
scale atmospheric motions: water vapour transport (which can also be determined
from three-dimensional analyses (Pierrehumbert and Roca 1998)) has been recon-
structed from groups of trajectories. Such groups show the circulation of upper
tropospheric water vapour from the intertropical convergence zone and South Pacic
convergence zone to subtropical subsidence zones (Soden 1998).
This article is focused on the principle, possible applications and limits of a
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
method for the construction of cloud trajectories from a series of images taken by a
geostationary satellite. No scientic result will be presented, but possible applications
of this method will be proposed. Its aim is the tracking of small cloud structures, at
a scale about or smaller than 200 km. In 2, the method used for the determination
of cloud motion vectors from two images will be briey reviewed. In 3, this method
will be extended in a simple way to construct cloud trajectories and the quality of
the latter will be estimated. Section 4 will describe necessary improvements for the
construction of representative trajectories. In 5, a variation on this improved method
will enable the tracking of cold cloud elements, selected by a threshold on brightness
temperatures (BT). Several trajectory elds of structures tracked in the thermal
infrared (IR) and water vapour absorption (WV) channels will be presented, the
possible but limited use of trajectories in the visible (VIS) will also be mentioned.
The complete trajectory of a cloud, from its birth to its dissipation, can be constructed
with an extension of the method presented in 6. The representativeness and
signicance of the trajectories in the diVerent channels will also be discussed.
inside a larger window (the search window) is computed for all possible displacements
(which are determined by the size of both target and search windows). The displace-
ment in pixels of the tracked structure(s) is determined by the relative position of
the target and search window that corresponds to the maximal value of the correla-
tion. If the geographical position corresponding to the position of the vector on the
image disc is known, the velocity of the tracked structure(s) can be deduced from
the displacement in pixels and from the time interval between images. The determina-
tion of the maximal cross-correlation is a maximum likelihood estimator (Wu 1995).
Another method can be used to compute CMWs: the Euclidean distance (or sum
of squared diVerences) method. The same denitions of pixel sets (target and search
windows) are used, but in this case the displacement vector in pixels (DL , DC) at
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
A
LT/2 1 CT/2 1
SSD(L , C , DL , DC)= W (L , C , l, c)
0 0 0 0
l= LT/2 c= CT/2
where i is the pixel value on rst image (target window); i is the pixel value on
1 2
second image; LT , CT are the number of lines and columns of the target window;
and W is the weighting function, equal to 1 for most applications.
Instead of the maximum of the correlation. Both methods have been considered
as giving similar results (Sitbon et al. 1974), but recent studies have shown that
diVerent results can be expected in a minority of cases (Wu et al. 1996, Dew and
Holmlund 2000). The Euclidean distance method is computationally a little faster
and we have chosen it for the measurement of cloud motion and thus for the
construction of their trajectories.
A series of quality tests eliminates inconsistent CMWs, which do not correspond
to a realistic motion of clouds. Commonly used tests include:
Comparison to the maximal possible displacement: If the maximal possible displace-
ment (determined by the size of the target and search windows) is reached, the CMW
is rejected.
A temporal symmetry check (Bowen et al. 1979): For this test, three images (taken
at instants tDt, t and t+Dt) are used. Each vector, computed from the central and
last images (at t and t+Dt) is compared to the collocated vector, computed from
the rst and central images, and is eliminated if both vectors show too great a
diVerence. The comparison can be realized on the vector diVerence, on the diVerence
or ratio between their speed, or on the diVerence between their direction, with the
use of selected thresholds.
A spatial consistency check (Green et al. 1975): Each CMW is compared with its
closest neighbours. If it shows large diVerences in velocity, speed and/or direction
with one or several of its neighbours (or with their average, according to the test
used), the vector is rejected.
A comparison with a forecast wind from a general circulation model: The height
of a CMW can be estimated from satellite images (see 2.2). If the CMW at this
estimated level shows large diVerences from the forecast wind at the same (or closest)
level, it is rejected.
1704 A. Szantai et al.
Another test can suppress very small or zero CMWs: the absence of displacement
can correspond to the tracking of terrain features (which do not move) or to the
tracking of orographic clouds; in this case, clouds appear to be blocked by mountains,
although local measurements can show strong winds. The value of the correlation
coeYcient can also be used as a selection criterion: if it is too small, the corresponding
CMW is refused.
The manual elimination of inconsistent CMWs, i.e. by a human operator, tends
to be abandoned nowadays or at least reduced, as automatic quality tests have
improved and as the number of CMWs can be very large, especially in operational
applications (Holmlund 1996, Nieman et al. 1997 ).
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
R ln(Pcloud)
z
cloud
=
g
a
ln(Psurface)
P T (P)d(ln(P))
v
(3)
gravity; and T is the virtual temperature of air at pressure P (can be deduced from
v
a temperature and relative humidity prole.
T he estimation of the height of low-level clouds from the IR brightness temperature:
CMWs are close to the wind measured at the base of low-level clouds (Hasler et al.
1979). Thus, the previously described method which estimates the cloud top temper-
ature is inappropriate because it leads to an underestimated pressure level (or an
overestimated height) of the cloud. In recent methods, the IR histogram of the pixels
of the target window is tted by Hermite polynomials (Le Marshall et al. 1994 ) or
by Gaussian curves (Ottenbacher et al. 1997). From these functions, the temperature
of the cloud base can be estimated. With the help of a temperature prole (from
other types of measures), a more realistic pressure level can then be deduced.
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
to height computed along the trajectory. And although this parameter may be
insuYcient for a precise height estimation, it may nevertheless be useful to show the
evolution of clouds and structures with time.
centre of the Earth disk)*, this corresponds to a motion of 108 km and thus to a
velocity of 216 km h 1 if the two images used are h apart.
1. if the time interval between consecutive images is small (this is generally the
case for h), rotation can be neglected in comparison with translatory motion
in most regions;
2. the determination of the rotation angle increases greatly the computation time.
For each possible rotation angle, all the pixel values of the search window
have to be interpolated to the same grid as the target window before computing
the (translational) displacement vector.
*In fact this corresponds to the distance between the centre of two neighbouring pixels.
The real size of a pixel is 5 km, but a small part of the surface it covers is overlapping with
its neighbours.
1708 A. Szantai et al.
limitations of the tracking algorithm: the principle of most block matching algo-
rithms is to recognize a similar structure on two images according to a statistical
criterion: the minimization of the Euclidean distance or the maximization of the
correlation. Thus changes of this structure with time may not be recognized. An
inconsistent motion can be measured when clouds appear, evolve fast or disappear,
when two or several clouds of similar aspect are in the vicinity (the algorithm may
track the wrong cloud ), when several cloud elements inside the target window move
diVerently (the algorithm will measure some kind of averaged motion, or the motion
of one of these cloud elements), or when (a part of ) a large featureless cloud is tracked.
The temporal and spatial scales of the tracked clouds can also inuence the
values of the measured CMW. A CMW measured with a large target window may
reect the motion of a large cloud structure, respectively of a group of smaller clouds,
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
whereas a small target window will produce a vector representing the motion of a
part of this structure, respectively of one (or several) cloud(s) of this group. This can
be the case for an expanding cloud.
images cover Europe, the central and western Mediterranean Sea and a part of the
north-eastern Atlantic Ocean (gure 1). The 16:00 image has a small defect, a shift
of a few lines and columns of pixels above 40 approximately. The trajectory compu-
tation starts on a regular grid, one starting point every 5 in latitude (between 40
and 60 N ) and longitude. No preliminary selection of the elements to track is
undertaken prior to the trajectory computation. Four synoptic meteorological
elements can be observed on these images:
1. clouds related to an extratropical low, located south of Greenland and Iceland,
moving westwards;
2. a large band of clouds oriented in a south-west/north-east direction oV the
west European coast, which bends in a west/east direction over Scandinavia
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Figure 1. Initial situation of the study area, 18 October 1989, 0:00 UTC, IR channel. (High
brightness temperatures are dark, low ones are light.)
Figure 2. Trajectories in the IR channel, 18 October 1989, between 0:00 and 23:00 UTC.
The target and search window, and the maximal displacement vector (24 pixels)
are drawn in the lower right corner. BT colours: black, BT <40C; dark blue,
40 BT <30C; light blue, 30 BT <20C; green, 20 BT <10C; red,
10 BT<0C; pink, BT>0C.
south-eastward motion of the band of clouds associated to the jet stream can be
clearly identied. But several trajectories (nos 19, 23, 31, 43, 44, 51, 53 and 54)
contain vectors with a large variation in direction (more than 90). Some trajectories
(nos 23, 31, 46, 53 and 54) include vectors with unrealistic speeds of more than
70 m s 1, which are rarely observed (i.e. in the core of jet streams or close to the
centre of tropical cyclones). Vectors with a large variation of speed and/or direction
are probably not associated to a physical motion of the tracked cloud elements.
Some trajectories ( beginning of 31, trajectory 41) contain no displacement
Construction of cloud trajectories f rom satellite images 1711
(At this stage, an unclassied vector undergoes comparisons with the previous vector.
These tests cannot be applied to the rst vector of a trajectory, which has no
predecessor.)
4. small vectors: in a pair of consecutive vectors, if one of them corresponds to
a small speed, below a chosen threshold (arbitrarily set to 3 m s 1), it does
not undergo the following tests;
5. vectors with a large change in speed: if the ratio between the smallest and the
largest vector of a pair is below a chosen threshold R (set to 0.5):
s
V
0< min <R (8)
V s
max
6. vectors with a large change in direction: when the two vectors are at an angle
larger than a chosen threshold DDIR set to 45:
max
|direction(v (t), v (t+Dt)|>DDIR (on trajectory i) (9)
i i max
7. vectors that satisfy both previous conditions (5 and 6) can be grouped together;
8. All the remaining vectors, which can be classied as vectors without problems.
Table 1 shows the results for the 18 October 1989 trajectories in the IR channel.
It appears that three quarters of the vectors are classied as without problems
(80% of the vectors, if the less severe thresholds of 0.33 for the speed ratio and 60
for the direction change are used). Computations of trajectories at other dates, in
other climatic regions or with other channels (in particular WV ) have shown that
this class remains the largest one. Noisy images can lower the percentage of this class.
Vector classes 5 6 7 8
Vectors with
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
inconsistent vector. But at the end of the rst extension of trajectory 23, two
consecutive correction attempts cannot extend this trajectory any more: this is the
physical end of the trajectory. Movie loops on a computer monitor show that the
cloud is dissipating at that time. The extension of trajectory 51 has the same direction
as the rst seven vectors of the original, but the vectors of the extension have a
smaller speed. A possible explanation is that new clouds tracked along the extension
are located at a lower level, where the wind is weaker.
We have identied on movie loops several causes that explain abrupt changes of
speed and direction along trajectories:
T he tracking of a similar structure (gure 4). On trajectory 31, the large inconsist-
ent vector is due to the tracking of the wrong cloud, whose aspect has become closer
to the aspect of the original cloud, which has evolved in the meantime.
T he dissipation and formation of clouds. In some cases, when the cloud dissipates,
the trajectory catches a cloud of similar aspect in the vicinity (extensions of trajectory
23). For a part of the trajectories, the formation of clouds cannot be detected because
the clouds already exist when the tracking starts. The formation of a cloud is
observed when non-zero vectors appear on a previously motionless trajectory
( beginning of trajectory 31 ).
Construction of cloud trajectories f rom satellite images 1713
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Figure 6. Trajectories in the IR channel, computed with a target window of 1616 pixels,
with target window, search window and maximal displacement vector (at lower right)
(cf. gure 2).
motion of clouds, the method has been improved. The determination of each CMV
is now divided in four steps:
1. the computation of the Euclidean distance surface and its minimal values;
2. a series of quality tests;
3. the correction of the vector if it is inconsistent;
4. stopping tests, to end the trajectory if necessary.
Construction of cloud trajectories f rom satellite images 1715
Vectors
Total without Small Inconsistent
number problem vectors vectors
of vectors (8) (%) (4) (%) (2, 5, 6, 7) (%)
but it can save some trajectories when non-translationa l motion is also present (for
example, the expansion of a convective cloud). A DDIR value of 45 allows the
max
tracking of a cloud in a region of rotational motion (for example close to the centre
of rotation of extratropical lows). Similar smoothness constraints have been used
by Hodges to construct the trajectories of specic features (in particular maximal
values of relative vorticity) in meteorological analyses (Hodges 1994).
one used for the initial, non-corrected vector computation) exist and if they are not
too numerous. (The maximum number of local minima has been set to 32 and is
rarely reached. A larger number indicates that at least one image of a pair has been
degraded by random noise or has a marked texturefor example a eld of clouds
with a regular pattern.) The corrected CMV is given by the position of the local
minimum that is closest to the end of the vector obtained during the previous
iteration (gure 7). If this corrected vector is of better quality (i.e. it matches quality
tests on velocity and direction), the computation of the trajectory can be continued,
otherwise it is stopped. The rst two vectors of a trajectory undergo a modied
version of these tests: if the rst vector is small (below 5 m s 1), the test on speed
variation is not applied. This enables the saving of some trajectories, which start at
or closely after the formation instant of the tracked cloud(s).
been corrected, owing to an incorrect start of the trajectory (the rst vector cannot
be completely checked) or to the dissipation of the tracked clouds simultaneously
followed by the tracking of clouds of similar aspect. To avoid this, when three
consecutive corrected vectors are found, we interrupt the trajectory on the rst of
them. Finally, very short trajectories of three vectors or less are of little interest and
can easily be corrupted due to an inconsistent rst vector; these trajectories are
therefore removed.
The complete procedure has been applied to the 47 IR images and a new set of
trajectories has been computed. A rst attempt showed that all the trajectories but
one still remaining at 15:00 UTC are stopped at 16:00. This is caused by a shift of
several pixels on the 16:00 UTC image, conrmed by a visual examination of this
image and the surrounding ones. (The study of other cases has shown that the
simultaneous termination of several trajectories should draw ones attention in the
rst place onto a problem concerning the quality of at least one image. A meteorolo-
gical phenomenon of short duration (e.g. the outburst of a convective cloud system)
does not aVect trajectories on such a large scale.) After the removal of the 16:00
faulty image, trajectories have been computed again (gure 8); this time, several
trajectories persisted along the whole series of images. When compared to the
Figure 8. Corrected trajectories in the IR channel, 18 October 1989, between 0:00 and
23:00 UTC (cf. gure 2).
1718 A. Szantai et al.
uncorrected trajectory eld (gure 2), the new trajectories appear to be less numerous,
more regular and of shorter duration. Trajectories starting in initially cloudless areas
have disappeared. Trajectories 12 and 34 are (almost) identical to their originals.
The tracking has been improved for trajectories which had an inconsistent vector
shortly after their beginning (no. 46 ).
Figure 9. Trajectories in the WV channel, 18 October 1989, between 0:00 and 23:00 UTC,
superimposed onto the 0:00 UTC WV image. The longest living trajectory is 37.
Construction of cloud trajectories f rom satellite images 1719
cases, the trajectories are even identical. The use of the correction method and
interruption tests explains that the WV trajectory chart shows a more regular and
more realistic motion than the trajectory charts constructed with a simpler method
in the tropics by Soden (Soden 1998). In the latter method, each vector has only
undergone a self-symmetry test (see also 6.2), and non-self-symmetri c vectors have
been replaced by their predecessor. Only trajectories with less than 10% corrected
vectors were retained.
to a more selective tracking of cold, high-level clouds by selecting only the pixels
above a chosen BT threshold for the Euclidean distance computation and the
deduced cloud CMV (Schmetz and Nuret 1987). (This is realized in practice by
setting the weighting function W of equation (1) to 1 for the pixels of the target
window with a pixel count corresponding to a BT below the threshold value, and
by setting W to 0 for the pixel values above this threshold.) This selection reduces
the inuence of low-level clouds and the Earth surface, which in some cases slow
down CMVs, and in some cases enables the tracking of high-level clouds in situations
where the trajectory computation using all the pixels was not successful. Besides the
selection of cold pixels, the only diVerence is the introduction of a complementary
termination test on trajectories: when the number of cold pixels becomes too small,
below a chosen threshold (set in this study to 6.25% of all the possible pixels of the
target window); this corresponds generally to the dissipation of the cloud.
In gure 10, WV trajectories of high-level clouds, with a BT below 40 C are
represented. They are less numerous than trajectories computed without temperature
Figure 10. Trajectories of high-level clouds in the WV channel, after a selection of the pixels
with a BT below 40C; starting on 18 October 1989 at 0:00 UTC, with the end of
the longest living trajectory (46) on 19 October 1989 at 21:30 UTC.
1720 A. Szantai et al.
selection (18 and 22 trajectories respectively). Whereas some trajectories are identical
with and without cold pixel selection (possibly due to a large proportion of high-
level clouds inside the target windowfor example trajectory 37), in other cases the
tracked elements along trajectories with a cold pixel selection can move in a slightly
diVerent direction (trajectory 47) and with a diVerent, generally higher speed (traject-
ory 36). DiVerences between trajectories with the same origin can be related to
motions at diVerent levels. Trajectories with a cold pixel selection appear to be more
representative of the motion of active areas (for example convective areas).
channel (VIS) (0.41.1 mm), which is better adapted for the tracking of low-level
clouds and has a better spatial resolution (the side of a VIS pixel is 2.5 km). The
main limitation is that these images are in practice available only during daytime
(between 8:30 and 17:30 UTC for the studied area). The use of a solar correction
(dividing the pixel value associated to the albedo by a factor of cos h , where h is
0 0
the solar zenithal angle) enables better tracking, especially at the beginning of
trajectories.
*Various methods can be used to determine the brightness temperature of the cold pixels.
In this article, for each target window, this brightness temperature is derived from the average
numerical count of all the pixels that have a value below a threshold. This threshold is
determined by the average of the numerical count of all the pixels of the target window minus
the standard deviation of these numerical counts. Therefore the value of this threshold varies
with each correlation window.
Construction of cloud trajectories f rom satellite images 1721
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Figure 11. Consecutive target windows (inside squares) and their immediate surroundings
along trajectory 50, IR channel. (The 16:00 UTC image has not been used for the
construction of this trajectory.)
Figure 12. Brightness temperature of the coldest pixel (T ) and average BT of the cold
min
pixels (T ) along trajectory 50, and identied cloud structures. 1, cloud in lower left
cold
corner (dissipating); 2, at left centre; 3, high-level cloud at top, left and centre; 4, small
bright cloud just above centre; 5, several small cloud structures; 6, growing cloud
along lower edge, at left.
and n). The end of the direct vector gives the position of the centre of the target
window for the computation of the opposite vector. The search windows of both
vectors have the same centre, the origin of the direct vector. We have an excellent
tracking if the opposite vector has exactly the same magnitude and the opposite
direction of the direct vector.
Figure 13 shows the diVerence between both vectors in the IR channel. Isolated
non-self-symmetric vectors (with a pixel diVerence of more than two pixels in either
directionlines or columns between the direct and opposite vectors) can be
observed along several trajectories (nos 10, 11, 13, 14, 20, 22, 30, 37 and 50). More
than 20% of non-self-symmetric vectors are scattered along some trajectories (nos
3, 5, 8 and 15 ) or concentrated along the nal part of other trajectories (nos 26
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
and 46). An important diVerence between the direct and opposite vectors can be
explained by the presence of several features of similar aspect in the search window
or may be due to changes in shape or size of the tracked features. Therefore
interrupting the trajectory computation as soon as a non-self-symmetric vector is
found is a good way to improve the quality of trajectory elds, but at the cost of
having shorter trajectories. A vector diVerence of two pixels between direct and
opposite vector should be tolerated for this rather severe quality test, thus allowing
small deformations of the tracked features.
Table 3 presents the proportion of self-symmetric vectors represented in gure 13
and for two other situations (one in the IR channel, the other one in the WV channel)
during one day with a large number of trajectories covering the whole Meteosat
disk north of 35 S. The results for the 18 October 1989 IR trajectories validate
the use of a vector correction and trajectory interruption tests; the proportion of
non-self-symmetric vectors has been reduced thereafter.
The interruption of trajectories on a non-satisfactor y self-symmetry test leads to
a further reduction in the number of vectors by less than a half of the corrected set
Figure 13. DiVerence in pixels between direct and opposite vectors, coded in colour: black,
direct and return vector identical; blue, diVerence of one line or/and one column of
pixels; green, diVerence of two lines or/and two columns of pixels; red, diVerence of
more than two lines or two columns of pixels.
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Table 3. Number of trajectories and vectors, and diVerence in pixels between direct and opposite vectors (absolute value of the largest diVerence
between lines or columns of the sum of the two vectors) for three situations (18 October 1989, IR channel; 18 February 1996, IR and WV channels).
of trajectories (23% less vectors for the 18 February 1996 IR case, 42% less vectors
for the 18 October 1989 IR case) and to a reduction of the number of trajectories
by less than a quarter (16% less trajectories for the 18 February 1996 IR case, 25%
less trajectories for the 18 October 1989 IR case), with the benet of having more
consistent results. This supplementary process is needed for the computation of
reliable trajectories representing the motion of clouds during their whole lifetime.
trajectory of a cloud or cloud ensemble, from its birth or apparition to its dissipation.
For this purpose, complete trajectories are computed in four steps.
1. The direct trajectories, starting from selected positions (for example on a
regular grid) at a chosen time (t ), are computed (forward computation, from present
0
to future). All the previous quality tests and corrections, including the interruption
of trajectories on non-self-symmetric vectors, are applied except the suppression of
short trajectories (less than four vectors). The nal instant of a trajectory is t +Dt ,
0 1
where Dt is the duration of this trajectory.
1
2. The backward trajectories, starting at the same time from the same positions,
are computed backwards in time (from the present back to the past), with the same
conditions as the direct trajectories. The instant of termination of the corresponding
backward trajectory is t Dt , with a duration Dt .
0 2 2
3. Complementary quality tests are applied on the rst vector of the direct
and the corresponding backward trajectory: both vectors must have close speeds
and directions close to opposite. The same threshold values on speed (R =0.5) and
s
direction (DDIR =45) are used to compare the rst vector of a direct trajectory
max
and the opposite of the rst vector of the corresponding backward trajectory. If the
vectors pass the tests successfully, the direct and backward trajectories are joined to
form the complete, reconstructed trajectory (named b+d hereafter). This trajectory
is retained for the next step. When at least one of these complementary tests is not
satised, this can be caused: (a) by the dissipation of a cloud or cloud system ( long
backward trajectory with no, or a very short, non-signicant, direct trajectoryin
this case, only the backward trajectory is kept for the next step); ( b) by its formation
(long direct trajectory without backward trajectory) ; or (c) by the presence of two
clouds or cloud systems with a diVerent but consistent motion (long direct and
backward trajectory) .
4. The goal of the last step is to conrm that the trajectories previously con-
structed are of good quality. For each of them, the end position and the instant of
termination of the backward trajectory (t Dt ) determine the starting position and
0 2
instant of a new, recalculated trajectory, computed forward over the whole series of
images used for the direct and backward trajectory computation. In practice, two
favourable cases for comparisons can be distinguished:
the recalculated trajectory has a longer or the same duration (i.e. more
vectors or the same number of vectors) as the b+d trajectory;
the recalculated trajectory has a shorter duration than the b+d trajectory,
but still exists at the instant used as time reference (t ).
0
Construction of cloud trajectories f rom satellite images 1725
In both cases, if the recalculated trajectory is identical or stays very close to a portion
or the totality of the original b+d trajectory, then we can assume that the cloud
has been tracked correctly. This can be checked by comparing the distance in pixels
(d ) between the end position of the recalculated and the b+d trajectories to the
e
average length of both trajectories (l ), computed over their common duration. For
c
this purpose, the quality ratio R has been dened:
d
R=100 e (in %) (10)
l
c
A ratio of 10% has been set as a threshold for the identication of good trajectories
whereas a 5% ratio selects high quality trajectories.
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Figure 14. Recalculated IR trajectories between 17 February, 17:30 UTC and 19 February
1996, 17:30 UTC, superimposed onto the IR image of 18 February, 17:30 UTC. Red,
quality ratio R 5%; green, 5<R 10%; blue, R>10%.
Table 4. Comparison of b+d and recalculated trajectories and vectors for three situations.
Values for recalculated trajectories are indicated for three quality ratios (R<2 ,
R 10% and R 5%). Percentages refer to recalculated trajectories and vectors.
Number of trajectories
Recalculated Recalculated
Recalculated trajectories trajectories
trajectories R 10% R 5%
18 October 1989, IR 26 14 13
(100%) (53.8%) (50.0%)
18 February 1996, IR 1042 579 473
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Number of vectors
Recalculated Recalculated
b+d Recalculated trajectories trajectories
trajectories trajectories R 10% R 5%
been used in this study). During a long time interval, it is more likely that a cloud
will change its size or aspect, and in the case of extreme evolutions make the tracking
impossible.
5. A practical limitation is the large amount of data (satellite images) needed to
obtain long sequences and thus long trajectories, although this problem tends to be
reduced nowadays with the increasing storage capacities of recent computers (for
example, a half-hourly sequence over 48 h of full resolution Meteosat IR or WV
images represents a volume of approximately 600 Mb).
6. The presence of clouds at diVerent levels probably limits the tracking, but this
eVect has not been investigated yet. Basically, when clouds at two diVerent levels
have converging trajectories and one layer masks at least partly the second one, the
trajectory following the most outstanding cloud features is continued. Theoretically,
the motion of two cloud layers moving with a diVerent motion should lead to the
presence of several minima in the Euclidean distance surface, and potentially enable
the determination of two CMVs (although the measurement of each Euclidean
distance minimumi.e. each vector associated to one cloud layermight be slightly
modied by the presence of clouds from the other layer). Another way to extend the
trajectory with less visible features may consist of calculating CMVs using images
of another channel, where these features are prominent. For example, low-level clouds
and semi-transparen t thin cirrus (at a high level) can be tracked in the IR channel.
In cases of overlap of clouds from both types, thin cirrus may be tracked more easily
1728 A. Szantai et al.
indicates that these clouds are more likely located at a single level and that there is
no jump between clouds at diVerent heights. These observations could be conrmed
with the use of a precise height estimation method. Besides, such a method would
provide a description of the tracked structures as complete as possible.
The trajectory construction method is based on cloud motion wind calculation
techniques that have been validated for a long time. The comparison of trajectories
derived from satellite images with three-dimensional meteorological analyses should
provide a complementary validation, although the temporal resolution of analyses
(6 or 12 h) is not the same, and help for a height attribution of the tracked atmospheric
structures. Comparisons with the trajectories of balloons can also be undertaken
for case studies. A comparison of balloon trajectories launched for the INDOEX
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
experiment during the winter of 199899 with trajectories derived from Meteosat
5 images is under way at the Laboratoire de Meteorologie Dynamique.
The trajectory construction method described in this article can be applied for
meteorological and climatic studies. Possible applications include the study of the
evolution of large scale atmospheric structures, especially convergence (subsidence)
zones and divergence zones (in particular the intertropical convergence zone), and
at mesoscale the divergent motion of clouds generated on top of tropical convective
cloud clusters, or the motion of small cloud elements inside cyclones.
Acknowledgment
The Meteosat images were provided by the European Space Agency and by
EUMETSAT.
References
Albergel, A., Martin, D., Strauss, B., and Gros, J.-M., 1988, The Chernobyl accident:
modelling of dispersion over Europe of the radioactive plume and comparison with
air activity measurements. Atmospheric Environment, 22, 24312444.
Anandan, P., 1989, A computational framework and an algorithm for the measurement of
visual motion. International Journal of Computer V ision, 2, 283310.
Angell, J. K., and Pack, D. H., 1960, Analysis of some preliminary low-level constant level
balloon (tetroon) ights. Monthly Weather Review, 88, 235248.
Arnaud, Y., Desbois, M., and Maizi, J., 1992, Automatic tracking and characterization of
African convective systems on Meteosat pictures. Journal of Applied Meteorology,
31, 443453.
Aspliden, C. I., Tourre, Y., and Sabine, J. B., 1976, Some climatological aspects of west
African disturbance lines during GATE. Monthly Weather Review, 104, 10291035.
Bannehr, L., Rohn, M., and Warnecke, G., 1996, A functional analytic method to derive
displacement vector elds from satellite image sequences. International Journal of
Remote Sensing, 17, 383392.
Bertrand, J., Baudet, J., and Drochon, A., 1975, Importance des aerosols naturels en Afrique
de louest. Journal de Recherches Atmospheriques, 8, 845860.
Bowen, R. A., Fusco, L., Morgan, J., and Roska, K. O., 1979, Operational production
of cloud motion vectors (satellite winds) from Meteosat image data. Proceedings of
T echnical Conference on Use of Data f rom Meteorological Satellites, L annion, France,
1721 September 1979 (Noordwijk: ESA c/o ESTEC), pp. 6575.
Buche, G., Ottenbacher, A., and Fischer, H., 1990, Wind vectors evaluated from structures
within consecutive Meteosat Water vapour images. Proceedings of 8th MET EOSAT
Scientic Users Meeting, Norrkoping, Sweden, 2831 August 1990 (Darmstadt:
EUMETSAT), pp. 8796.
Carlson, T. N., and Prospero, J. M., 1972, The large-scale movement of Saharan air outbreaks
over the northern equatorial Atlantic. Journal of Applied Meteorology, 11, 283297.
1730 A. Szantai et al.
Cayla, F. R., and Tomassini, C., 1978, Determination de la temperature des cirrus semi-
transparents. L a Meteorologie, 15, 6367.
Cote, S., and Tatnall, A. R. L., 1995, A neural network-based method for tracking features
from satellite sensor images. International Journal of Remote Sensing, 16, 36953701.
Desbois, M., 1976, Les resultats de lexperience EOLE. L a Meteorologie, 6eme serie, 5168.
Desbois, M., Kayiranga, T., Gnamien, B., Guessous, S., and Picon, L., 1988, Characterization
of some elements of the Sahelian climate and their interannual variations for July
1983, 1984 and 1985 from the analysis of METEOSAT ISCCP data. Journal of Climate,
1, 867904.
Dew, G., and Holmlund, K., 2000, Investigations of cross-correlation and Euclidean distance
target matching techniques in the MPEF environment. Proceedings of Fifth Interna-
tional W inds Workshop, 28 February 3 March 2000, L orne, Australia (Darmstadt:
EUMETSAT), pp. 235243.
Durst, C. S., 1955, Variation of wind with time and distance. Geophysical Memoires, 12, 331.
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Endlich, R. M., Wolf, D. E., Hall, D. J., and Brain, A. E., 1971, Use of a pattern recognition
technique for determining cloud motions from sequences of satellites photographs.
Journal of Applied Meteorology, 10, 105 117.
Eriksson, A., 1988, A new method of wind vector computation from METEOSAT water
vapor imagery. Proceedings of 7th MET EOSAT Scientic Users Meeting, Madrid,
2730 September 1988 (Darmstadt: EUMETSAT), pp. 321328.
Gandin, L. S., 1965, Objective Analysis of Meteorological Fields (Jerusalem: Israel Program
for Scientic Translation).
Green, R., Hughes, G., Novak, C., and Schreitz, R., 1975, The automatic extraction of
wind estimates from VISSR data. Technical Memorandum NESS 164, US Department
of Commerce, Washington, DC, pp. 94110.
Hasler, A. F., 1981, Stereographic observations from geosynchronous satellites: an important
new tool for the atmospheric sciences. Bulletin of the American Meteorological Society,
62, 194212.
Hasler, A. F., Palaniappan, K., Kambhammetu, C., Black, P., Uhlhorn, E., and Chesters,
D., 1998, High resolution wind elds within the inner core and eye of a mature tropical
cyclone from GOES 1-min images. Bulletin of the American Meteorological Society,
79, 24832495.
Hasler, A. F., Skillman, W. C., Shenk, W. E., and Steranka, J., 1979, In situ aircraft
verication of the quality of satellite cloud winds over oceanic regions. Journal of
Applied Meteorology, 18, 14811489.
Hasler, A. F., Strong, J., Woodward, R. H., and Pierce, H., 1991, Automatic analysis of
stereoscopic satellite image pairs for determination of cloud-top height and structure.
Journal of Applied Meteorology, 30, 257 281.
Hayden, C. M., and Purser, R. J., 1995, Recursive lter objective analysis of meteorological
elds: application to NESDIS operational processing. Journal of Applied Meteorology,
34, 315.
Hodges, K. I., 1994, A general method for tracking analysis and its application for meteorolo-
gical data. Monthly Weather Review, 122, 25732586.
Holmlund, K., 1996, Normalised quality indicators for EUMETSAT cloud motion winds.
Proceedings of T hird International W inds Workshop, 1012 June 1996, Ascona,
Switzerland (Darmstadt: EUMETSAT), pp. 155164.
Holmlund, K., 1998, The utilization of statistical properties of satellite-derived atmospheric
motion vectors to derive quality indicators. Weather and Forecasting, 13, 10931104.
Horn, B. K. P., and Schunck, B. G., 1981, Determining optical ow. Articial Intelligence,
17, 185203.
Hubert, L. F., 1979, Wind derivation of geostationary satellites. WMO T echnical Note,
166, 3359.
Kamachi, M., 1989, Advective surface velocities derived from sequential images for rotational
ow eld: limitations and applications of maximum cross-correlation method with
rotational registration. Journal of Geophysical Research (Oceans), 94, 18 22718 233.
Laing, A. G., and Fritsch, J. M., 1993, Mesoscale convective complexes in Africa. Monthly
Weather Review, 121, 22542263.
Laurent, H., 1993, Wind extraction from METEOSAT water vapor channel image data.
Journal of Applied Meteorology, 32, 11241133.
Construction of cloud trajectories f rom satellite images 1731
Leese, J. A., Novak, C. S., and Clark, B. B., 1971, An automated technique for obtaining
cloud motion from geosynchronous satellite data using cross correlation. Journal of
Applied Meteorology, 10, 118132.
Le Marshall, J. F., Pescod, N. R., Seaman, B., Mills, G., and Stewart, P., 1994, An
operational system for generating cloud drift winds in the Australian region and their
impact on numerical weather prediction. Weather and Forecasting, 9, 361370.
Lunnon, R. W., and Lowe, D. A., 1992, Spatial scale dependency of errors in satellite cloud
track winds. Advanced Space Research, 12, 127131.
Lunnon, R. W., Lowe, D. A., Barnes, J. A., and Dharssi, I., 1992, Study of cirrus cloud
winds. Technical Report 35, Forecasting Research Division, Meteorological OYce,
Bracknell, Berkshire, UK.
Maddox, R. A., 1980, Mesoscale convective complexes. Bulletin of American Meteorological
Society, 61, 13741387.
Downloaded By: [Ecole Polytechnique] At: 14:34 14 February 2007
Maurizot, M., Bouthemy, P., and Delyon, B., 1997, Direct localization, motion characteriza-
tion and tracking of centers of atmospheric disturbances in Meteosat image sequences.
Proceedings of 1997 Meteorological Satellite Data Users Conference, Brussels, Belgium,
29 September3 October 1997 (Darmstadt: EUMETSAT), pp. 347354.
Menzel, W. P., Hayden, C. M., Nieman, S. J., Velden, C. S., and Wanzong, S., 1996,
Improvements in the quality assessment of automated satellite-derived cloud and water
vapor motion vectors. Proceedings of T hird International W inds Workshop, Ascona,
Switzerland, 1012 June 1996 (Darmstadt: EUMETSAT), pp. 197205.
Menzel, W. P., Smith, W. L., and Stewart, T. R., 1983, Improved cloud motion wind vector
and altitude assignment using VAS. Journal of Climate and Applied Meteorology,
22, 377384.
Menzel, W. P., Wanzong, S., Nieman, S., and Schmetz, J., 1993, Assigning heights to cloud
motion winds. Proceedings of Second International W ind Workshop, T okyo, Japan,
1315 December 1993 (Darmstadt: EUMETSAT), pp. 105115.
Morel, P., and Bandeen, W., 1973, The EOLE experiment: early results and current objectives.
Bulletin of American Meteorological Society, 54, 298306.
Neumann, C. J., Cry, G. W., Caso, E. L., and Jarvinen, B. R., 1978, Tropical cyclones of
the North Atlantic Ocean, 18711977. National Climatic Center, Asheville, NC and
National Hurricane Center, Coral Gables, FL.
Nieman, S. J., Menzel, W. P., Hayden, C. M., Gray, D., Wanzong, S. T., Velden, C. S.,
and Daniels, J., 1997, Fully automated cloud drift winds in NESDIS operations.
Bulletin of the American Meteorological Society, 78, 11211133.
Nieman, S. J., Schmetz, J., and Menzel, W. P., 1993, A comparison of several techniques to
assign heights to cloud tracers. Journal of Applied Meteorology, 32, 15591568.
Ondrejka, R. J., and Conover, J. H., 1966, Note on the stereo interpretation of Nimbus II
APT photography. Monthly Weather Review, 94, 611614.
Ottenbacher, A., Tomassini, M., Holmlund, K., and Schmetz, J., 1997, Low-level cloud
motion winds from Meteosat high-resolution visible imagery. Weather and Forecasting,
12, 175184.
Pierrehumbert, R. T., and Roca, R., 1998, Evidence for control of Atlantic subtropical
humidity by large scale advection. Geophysical Research L etters, 25, 45374540.
Purdom, J. F., 1996, Detailed cloud motions from satellite imagery taken at thirty second,
one and three minutes intervals. Proceedings of T hird International W inds Workshop,
1012 June 1996, Ascona, Switzerland (Darmstadt: EUMETSAT), pp. 137146.
Saunders, R. W., 1988, Cloud top temperature/height: a high-resolution imagery product
from AVHRR data. Meteorological Magazine, 117, 211221.
Schmetz, J., and Holmlund, K., 1992, Operational cloud motion winds from Meteosat and
the use of cirrus clouds as tracers. Advanced Space Research, 12, 95104.
Schmetz, J., Holmlund, K., Hoffmann, J., Strauss, B., Mason, B., Gaertner, V., Koch,
A., and Van de Berg, L., 1993, Operational cloud motion winds from Meteosat
infrared images. Journal of Applied Meteorology, 32, 12061225.
Schmetz, J., and Nuret, M., 1987, Automatic tracking of high-level clouds in METEOSAT
infrared images with a radiance windowing technique. ESA Journal, 11, 275286.
Sitbon, P., Berroir, A., and Chaume, D., 1974, Determination des champs de vents a partir
des photographies de satellites geostationnaires. 2eme partie: examen de procedes
1732 Construction of cloud trajectories f rom satellite images