September 1971 665
UDC 881.601.76: 651.676.6:6561.W8.761:661.w7.362.2
WIND ESTIMATION FROM GEOSTATIONARY-SATELLITE PICTURES
L. F. HUBERT and L. F. WHITNEY, JR.
National Environmental Satellite Service, NOAA, Washington, D.C.
ABSTRACT
The motion of properly selected clouds derived from photographs taken by the geostationary Advanced Tech-
nology Satellites provides estimates of the wind a t cloud level. A large number of low- and high-cloud motions near
rawin stations are compared to the balloon-derived winds to determine the levels at which these wind estimates might
be used for map analyses and to assess their accuracy a t those levels.
For this sample, velocities of lower clouds correspond best to winds a t 3,000 f t , while upper cloud velocities cor-
respond best to winds at 30,000 f t . The median vector deviation of the cloud velocities from observed winds is 9 and
17 kt a t 3,000 and 30,000 f t , respectively. Direction deviations are modest; therefore, these wind estimates are ' .
representative of the actual flow patterns in the lower and in the upper troposphere.
Causes contributing to deviations are (1) uncertainty of cloud height, (2) nonadvective cloud motion, (3) photo-'
graph-measurement errors, (4) tracking errors, and (5) unrepresentative rawinsonde observations.
The principal sources of cloud-velocity deviation from observed winds are ranked in order of their significance to
identify the critical areas needing improvement. It appears that cloud-height uncertainty is the most significant cause
of deviations, particularly about upper cloud heights. Work aimed a t reducing this uncertainty is just beginning.
To point out the unique problems involved, we discuss the technique of selecting and classifying cloud targets.
The procedure is a subjective skill dependent upon the analyst's meteorological judgment.
Despite the problems-many as yet unsolved-the single earth-synchronous satellite, with its immense areal
coverage and high frequency of coverage, provides an important new source of data from remote regions.
1. INTRODUCTION
Pictures received from geostationary satellites have pro-
vided meteorologists with a new type of upper air data.
Each of the present geostationary satellites is designated
an Advanced Technology Satellite (ATS). ATS 1 is
located over the Equator at 15OoW, while ATS 3 has been
positioned a t several different longitudes from 45"W t o
S5"W. Several pictures are taken during the lifetimes of
trackable cloud patches, hence cloud motions can be
measured. Certain clouds are advected by the ambient
winds ; consequently measurements of such motions (cloud
velocities) provide estimates of the winds. Wind estimates
made in this manner have the advantage of wide coverage.
The accuracy and representativeness of these wind esti-
mates are uncertain, however, because we derive the
cloud velocities by exploiting characteristics of cloud
organization only recently discovered and, as yet, only
p art1 y understood.
Section 2 describes the data used for this study and the
procedure for assessing accuracy of our wind estimates.
Section 3 presents an assessment of error of wind estimates
based on measurement of cloud velocities. Accuracy is
enhanced by a skillful selection of cloud targets, the
subject of section 4. Section 5 identifies the various
sources of deviation of estimated cloud velocities from
observations. Section 6 summarizes our results.
2. PROCEDURES AND DATA
animate cloud motion; the motion, in turn, is measured
on a projected display. The steps are:
1. Preparing a movie sequence by photographing, with a movie
camera, the individual images that were taken from the satellite a t
intervals up to 30 min.1
2. Projecting the movie sequence repeatedly onto a worksheet by
means of an endless loop.1
3. Selecting clouds as targets and marking their initial and final
positions on the worksheet, classifying cloud type.
4. Measuring, scaling, and recording the cloud velocity vectors.
a PROCEDURE FOR ESTIMATING ACCURACY
To measure the accuracy of wind estimates, we must
compare them with observed winds a t the level of the
cloud target. But cloud height is difficult to derive; sound-
ings of temperature and humidity do not provide adequate
information, and below -4OOc humidity is not reported.
Consequently, it was necessary to devise some other
means of specifying cloud height; one that does not rely
on radiosonde data.
Heights of target clouds were estimated from nearby
wind soundings. It was assumed that the minimum vec-
torial difference between cloud velocity and balloon veloc-
ity occurred at the cloud level. This level of minimum
velocity difference was designated as the "level of best
fit" (LBF). To find LBF, we plotted each cloud velocity
on the hodograph of a nearby upper air observation as
shown in figure 1. The interpolated level (within the
1 Two similar closed-circuit televislon systems have been develoued to replace the PROCEDURE FOR DERIVING CLOUD VELOCITIES
In principle, the satellite camera is stationary relative t o movie camera and projector. The equipment designed and built atstanford Research
Institute (SRI) is described by Serebreny et 81. (197Ob). The Electronic Anbation
reported in the literature. Basic procedures for select- and tracking clouds are un-
changed, however, and some of the measurements included in our statlstics were obtained
the earth SO that fixed earth features are photographed in
location on the image. Time-lapse movies are produced to
system developed a t the National Environmental Satemte Service (NESS) hss not been
by SRIon thekequlpment.
the Same position on all pictures-only clouds change
MONTHLY WEATHER REVIEW Vol. 99, No. 9 666
1509
110‘.
9OC
60’
FIGURE 1 .-Typical hodograph illustrating the method of deriving
level of best fit (LBF) of target clouds from the observed winds.
Heights are given in thousands of feet and, in parentheses, in
constant-pressure levels. Cloud velocities are represented by
vector arrows with 5-kt error circles centered on their endpoints.
troposphere) at which the hodograph vector was nearest
the endpoint of the cloud velocity vector was taken to be
the LBF. Five-kt circles are centered on each cloud
velocity vector in figure 1 to illustrate the effect of a 5-kt
wind error on the LBF. Depending upon the direction of
error, the lower level LBF in this instance could be any-
where in the layer from the surface to 10,500 f t , while
the upper level LBF could be anywhere in the layer from
34,000 t o 35,000 f t . Such height ranges may be greater
or less than this example, depending upon the structure
of wind in the vertical. Where the vertical wind shear is
small, a 5-kt circle will encompass a very deep layer and
in such cases improbable LBF’s can be derived. The wide
range of LBF’s is, in part, due to that wind-error effect.
The LBF probably often yields a good estimate of the
actual cloud height, but more important for our purposes,
the distribution of the heights indicates the range of
heights of the selected targets. Such distributions are a
measure of the extent to which the analyst has properly
classified the cloud type. Hence, hodographs are a useful
tool for this investigation. Unfortunately, they cannot be
applied by the operational ATS analyst.
In an operational situation, the ATS analyst does not
have hodographs and cannot determine the LBF. There-
fore, the cloud velocities must be used for map analysis
at some level that is representative of the cloud type
classification, rather than at the individual LBF. The
distribution of LBF’s indicates which levels are most
representative. Accuracy of cloud velocities as wind esti-
mates, evaluated at two such representative levels for
low and high clouds, is presented in section 3.
TABLE 1.-Location, dafe, and number of ATS wind estimates
Loeation
Number velocitles of cloud
Low-level . Righ-level clouds clouds
Date --
Paclfic Ocean April and August 1967 47 19
Paclfic and Atlantic Oceans April, August, and 64 __
117 unlted States April 1988 __
Qulf of Mexico and Cardbean June 1968 39 72
PaciBc Ocean (OARP’ data) November 1989 462 366
Total 812 664
(derived by SRI) November 1987
*Global Atmospheric Research Program
DATA USED
The data sample used in this investigation consists of
measurements of approximately 600 low-cloud motions
and about 560 high-cloud velocities. In some cases, more
than one cloud velocity was compared to a single rawin-
sonde observation. The maximum distance between target
clouds and their associated upper air stations was 150 n.mi.
The time between upper air sounding and the ATS
picture sequence ranged up to 6 hr in the Atlantic and
United States and up to 4 hr in the Pacific. ATS 3 picture
sequences usually are centered near 1800 GMT, a t which
time few soundings are made. In those cases, the brack-
eting 1200 and 0000 GMT soundings were used to inter-
polate an 1800 GMT sounding. ATS 1 sequences frequently
end near 2300 GMT, so the nearest 0000 GMT sounding was
usually less than 3 hr from time of the cloud velocities.
Table 1 shows dates and general areas over which cloud
velocities were derived for this study.
3. ACCURACY OF CLOUD VELOCITIES
AS WIND ESTIMATES
Statistics presented in this section measure two aspects
of accuracy. First, we assess the degree of success achieved
in judging cloud genera by examining the dispersion of
LBF’s. Second, we summarize deviations of cloud ve-
locities from nearby rawinsonde observations at two levels
which are shown to be most representative: one summary
for low clouds, a second for high clouds. Such deviations
measure accuracy of the cloud velocities as wind estimates
for two specific analysis levels.
These results represent the state of the art during the
past few years. The quality of this sample is not uniform
because the analysts’ skills varied. The critical importance
of skill will become clear in the next section.
DISTRIBUTION OF LEVELS OF BEST FIT
The heights of the LBF’s for low- and high-cloud targets
were tabulated a t 2,000-ft class intervals. Frequencies of
occurrence appear in figures 2 and 3. For low clouds (fig. 2)
September 1971 L. F. Hubert and L. F. Whitney, Jr. 667
HEIGHT (thousands of f l )
FIQURE 2.-Frequency, in percent of total low-cloud sample,
versus height of LBF. Heights are shown in 2,000-ft class intervals.
15 -I
Z N 564
HEIGHT (thousands of f t )
FIQURE 3.-Same as Figure 2 but for the high-cloud sample.
the mode falls in the lowest class interval. Taken at face
value, this suggests that wind estimates from low clouds
are representative of the layer from surface to 2,000 f t .
However, this modal level is not necessarily the elevation
where the deviations are minimized. Figure 4, discussed
later, shows that there is little to choose between various
low levels. For this sample it appears that the “deviation-
minimizing” level is 3,000 f t , but a different sample could
well favor another level.
Figure 3 illustrates the height distribution of the upper
cloud LBF’s. The principal mode and the median LBF
are quite near 30,000 ft. However, a significant fraction of
high-cloud targets lies below 20,000 ft, suggesting that,
despite the analysts’ efforts to track only cirrus, many
middle-cloud targets were selected.
Two conclusions are suggested by these figures: (1)
under operational conditions when no cloud height data
are available, the low-cloud motions are best applied to
3,000-ft to 5,000-ft analyses and the upper cloud motions
are best applied to the 30,000-ft analysis; (2) it is clear
that the height uncertainty is greatest for high-level
clouds.
VECTORIAL DEVIATIONS AT 3,000 AND 30,000 FT
From the viewpoint of the user, the accuracy of wind
estimates is measured by the deviations of cloud velo-
cities from those of actual winds at the level to which the
estimate has been assigned-namely, the 3,000- and
30,000-ft levels. From the user’s viewpoint, therefore,
wind estimates might be in error even if the cloud velocity
-s .m II
--- 3 .m f l
..... 1.m ft
N = 610
‘5 Y
L
30
FIQURE 4.-Magnitude of vectorial deviations of low-level cloud
velocities from observed winds At 1,000, 3,000, and 5,000 ft
versus cumulative frequencies.
’T / I 1
:u
VECTOP DEVIATION l i b 1
FIQURE 5.-Magnitude of vectorial deviations of high-level cloud
velocities from obqerved winds at 30,000 ft versus cumulative
frequency.
were precise. For that reason we tabulated, for each cloud
motion vector, the magnitude of vector deviation and the
magnitude of directional deviation from the observed
wind at those two “representative” levels.
Cumulative frequencies of deviation of low- and high-
cloud motions from observed winds are shown in figures 4
and 5. Curves of deviation at 1,000 and at 5,000 f t are
also shown on figure 4 to illustrate two points. First, with
a sample that consists of a wide variety of low clouds, the
cloud velocities correspond reasonably well to the flow a t
various low levels. Second, comparison of these curves
shows that the deviation-minimizing level is not necessarily
the modal level of the LBF’s; it tends toward the mean
LBF.
Figure 4 shows that half of the targets judged to be low
clouds deviated from the observed 3,000-ft wind by no
more than 9 kt, while figure 5 shows that half of the targets
judged to be cirrus deviated no mole than 17 k t from the
30,000-ft winds. This larger deviation of upper level cloud
velocities is partly the consequence of the greater disper-
sion in elevation of upper level targets that was shown in
k u r e 3.
440-818 0 - 71 - 2
668 MONTHLY WE ATHE R REV1 EW VOI. 99, No. 9
FIQIJRE &--Magnitude of directional deviations of low-level cloud
velocities from 3,000-ft observed wind and corresponding deviation
. of high-level cloud velocities from 30,000-ft observed wind,
versus cumulative frequencies.
DIRECTIONAL DEVIATIONS
Despite the size of the vectorial deviations, the direc-
tional patterns correspond reasonably well to those of the
real flow patterns at the two representative levels. Hence,
streamline analyses of cloud directions delineate cyclones
and anticyclones, troughs and ridges, and sometimes more
subtle flow configurations. This characteristic is illustrated
by the cumulative frequencies of directional deviations
given in figure 6. More than 75 percent of the high-level
cloud directions were within f40° of the 30,000-ft wind
direction, while 65 percent of the low-level directions were
within f4Q0 of the 3,000-ft wind direction.
4. SELECTING AND CLASSIFMNG TARGET CLOUDS
The intent of this section is not to present a complete
guide for the AT’S analyst, but to stress the type of clues
that must be employed to surmount some of the problems
inherent in these data. The selection procedure is subjec-
tive, depending upon the meteorological knowledge of the
analyst in a manner that is shown by the following outline.
The goals of the selection procedure are to classify cloud
genus and to discriminate between “passive tracers” and
those cloud targets that are unusable because they are
displaced by nonadvective mechanisms. Some targets can
be seen to move in directions totally inconsistent with the
wind at any level; therefore, tracking every visible target
would degrade the quality of wind estimates. As a conse-
quence of this selection process, only a fraction of all
visible clouds are tracked.
Motion, displayed by means of animation, provides a
powerful aid to meteorological interpretation. Animation
enables one to examine pattern behavior-a significant
advantage over instantaneous views. Not only are bright-
ness, texture, and pattern of clouds disclosed, but so are
time changes and relative motions of cloud layers. Such
pattern behavior enables one to deduce a great deal more
about the vertical structure of the flow pattern, hence
about the synoptic regime, than is possible with “snap-
shot” pictures.
CkASSlFlCATlON OF CLOUD TARGETS
Typically, the analyst decides on cloud type by first
determining the synoptic situation. Cloud genus is then
deduced by judging how the cloud behavior fits the appro-
priate synoptic model. Where it is not possible to reoon-
cile a target’s behavior with the synbptic situation, it is
prudent to omit that target.
The known relationships between satellite pictures and
synoptic situation provide the starting point. Motion of
clouds aids the analyst in applying the following guide-
lines :
1. Locate cyclones, fronts, jet streams, squall lines, and inter-
tropical convergence zones by looking respectively for such indica-
tors as spirals, large bands, abrupt cloud edges, and lines or zones of
bright cloudiness.
2. Determine whether jet stream indicators are consistent with
the position of cyclones and fronts.
3. Identify cold air masses by looking for bright globular cloud
masses behind cold fronts near cyclonic centers.
4. Determine stages of cyclone development by the character of
spirals or waves.
5. Determine interaction between air masses and underlying
surfaces by looking, for example, for cloud types typical of cold
advection over warmer water or warm advection over colder water.
Once the synoptic situation is determined, observations
of cloud characteristics and cloud motion aid the analyst
in segregating cloud layers and in specifying cloud types.
The following guidelines should be considered when deter-
mining cloud level and cloud type:
1. Clouds a t different levels can be distinguished by their con-
trasting speed and direction.
2. Cloud patches moving in contrasting directions in the same
area frequently suggest, by considering climatology, which layer is
low cloud and which layer is high cloud. For example, tropical
easterlies are easily separated from upper level westerlies.
3. Cumulonimbi in a field of cumuli produce cirrus that usually
move differently than the low-level clouds.
4. Cirriform clouds usually move faster than low-level C ~O U ~P ,
particularly at midlatitudes.
5. High clouds in well-developed tropical cyclones move slowly
and both cyclonically and anticyclonically, whereas low and middle
clouds move rapidly in a cyclonic direction.
6 . Bright clouds with sharp edges usually are cumuliform and
are easily tracked.
7. Bright globular masses behind cold fronts and near cyclonic
centers are cumuliform-usually with some vertical development.
8. Large globular masses in the southeastern sector of anticyclones
are cumuliform-usually with little vertical development because
they are restricted by inversions.
9. Clouds influenced by coast lines and by ocean temperatures
are low-level clouds.
10. Thin clouds with diffuse edges tend to be cirriform and are
less easily tracked than cumuliform and middle clouds.
11. Clouds associated with jet streams are often cirriform. Beware
of the large masses of bright multilayered cloud associated with jet
streams, for the bright middle clouds will often obscure the thin
gray cirrus of jet stream level.
12. Large uniform cloud masses with few distinguishable elements
tend to be middle clouds that are difficult to track.
SELECTION Off BASSWE TRACERS
Selection of passive tracers is made concurrently with
the classification of clouds. Some guidelines are:
1. Follow the same point on cloud clusters and patches rather
than lines, bands, or areas of equal brightness.
2. Use only those clouds moving at speeds and in a manner that
is consistent with the synoptic situation. Beware of motions which
appear to move through a pattern of cloud, alternately suppressing
September 1971 L. F. Hubert and L. F. Whitney, Jr. 669
TABLE 2.-ATS Cloud velocity deviations f r h rawinsonde observa-
tions; magnitude of vectoi deviations for various cumulative per-
centages of sample i s shown. N i s the total number of measurements.
~~ ~
Low-Aoud velocities (N=612) Highcloud velocities (N=564) Cumulative percent of Total at Deviation Part due Total at Deviation Part due
sample 3,oOOft at LBF toheight 30,OOOft at LBF to height uncertainty uncertainty
Qt) et) Qt) Qt) (kt) (ktj
MI 9 3 6 17 5 12
76 13 6 7 26 10 15
86 18 8 8 31 l3 18
90 18 9 9 36 16 21
and enhancing brightness. This type of motion often conflicts with
motion of the individual cloud elements in the same layer and is
probably due to gravity waves. Upward motion in crests of such
waves enhances cloudiness, and downward motion in troughs sup-
presses cloudiness. These motions are frequently seen in invefsion-
dominated low clouds and at various upper levels near cold fronts.
As expected from theory, the orientation of waves and their direction
of motion bear no fixed relation to the ambient wind.
3. Use clouds which show the least change during the time-lapse
sequence.
4. Take care in tracking clouds that appear to penetrate vertical
shear layers. In these cases, try to track the upshear edge rather
than the center of mass. For example, in areas of active convection
the cloud area grows rapidly because of anvil growth. The origin
of the anvil (the brightest area a t rear of the growth area) moves
with the middle- and low-level wind. The leading edge of the anvil,
while advancing with the high-level wind, may be moving more
slowly than the wind because of evaporation. Thus the leading edge
of growing cirrus plumes should be avoided.
DISCUSSION OF SELECTION TECHNIQUES
The present shortcomings of our selection procedure
account for a significant part of the deviations in table 2.
The range of LBF’s shown in figure 3, for example, would
have been reduced appreciably had the analyst succeeded
in selecting cirrus targets exclusively. The errors shown in
table 2, due to height uncertainty, would have been
reduced.2
Misclassification of cirrus can be minimized by close
attention t o texture and brightness revealed on high-
quality pictures and by a cautious choice of targets. A
good practice is t o discard a target if its classification is
doubtful. But these improvements alone cannot eliminate
error due to height uncertainty. Cirrus occurs over a wide
range of altitudes;, consequently, vectors derived from
cirrus motions cannot correspond to any one representa-
tive level. Use of a single level will produce significant
error in regions where vertical shear is large.
Can cirrus altitudes be subclassified? We feel they might
be and are just beginning to examine the possibility. Cirrus
which flanks a jet stream is recognizable on satellite pic-
tures. Cirrus related to a polar jet probably forms at an
altitude different from cirrus associated with a subtropical
jet. High clouds behind a trough may be a t different heights
than cirrus ahead of the trough. Cirrus being carried away
from tropical disturbances may lie a t yet a different eleva-
* In a study sponsored by NESS, Serebreny et al. (1970~) found that deviations of
FIGURE 7.-Magnitude of vectorial deviations of low-level cloud
velocities from observed winds at their individual LBF’s, versus
cumulative frequencies.
tion. If these considerations enable one to assign wind
estimates to representative levels other than 30,000 f t , the
“height uncertainty” error will be reduced.
Wind estimates at low levels also can be improved by
this approach. An example of very small error due to height
uncertainty can be seen in a case study by Fujita (1969).
He derived many cloud velocities in tropical areas where
climatology insured that most targets we?e suppressed low-
level stratocumuli.
5. SOURCES OF ERROR
Five significant sources contribute to the error of wind
1. Uncertainty of target cloud height,
2. Nonadvective cloud motion,
3. Errors of measurement,
4. Errors in tracking cloud targets, and
5. Errors due t o nonrepresentative rawinsonde observations.
(While this does not affect the accuracy of our wind estimates, i t
does contribute to the deviations.)
Although data available at this time are inadequate to
measure the individual contribution of each source of error,
we believe that these are the principal sources, listed in
approximate order of decreasing importance.
estimates at 3,000 and 30,000 ft. These are:
UNCERTAINTY OF TARGET CLOUD HEIGHT
Cloud targets are tracked at a variety of elevations, but
for reasons already discussed, we assign the wind estimates
to specific representative levels. Deriving cloud velocities
at various elevations and assigning them to specific levels
for analysis is an error source, from the user’s viewpoint,
which we call uncertainty of cloud height. To the extent
that our LBF’s represent actual cloud heights, we can
deduce the influence of cloud height uncertainty. We shall
see that this factor accounts for over half of the error a t
3,000 and 30,000 ft-a result found by comparing the
deviations summarized in figures 4 and 5 with deviations
at the LBF.
Figure 7 presents cumulative frequencies of magnitude
of vectorial deviations between low-level cloud velocities
and observed winds at their individual LBF’s, while
figure 8 shows a similar curve for high-level cloud velocities.
high-level cloudvelocities from the 30,000-ft wind were reduced from 20 to 17 kt by
ebinating those targets whose LBF fell below 20,000 ft (eliminating, thereby, the lower About 70 percent of the low-level vectors deviated from
clouds that had been mistaken for cirms). observed winds (at LBF) by 5 k t or less. Slightly more than
670 MONTHLY WEATHER REVIEW
DISTANCE FROM PICTURE CENTER
PERCENT OF IMAGE RADIUS
Vol. 99, No. 9
FIQURE &-Same as figure 7 but for the high-level cloud sample.
50 percent of the high-level vectors deviated from the ob-
served wind 5 k t or less a t their LBF's These deviations
stem from the combined influence of all error sources with
the exception of cloud height uncertainty.
Figures 4 and 5 , on the other hand, summarize the
influence of all error sources, including cloud height uncer-
tainty. Hence, the differences between figures 4 and 7
and between 5 and 8 reveal the contribution of height
uncertainty to the total deviation; several such differences
are tabulated in table 2. For half of the cloud velocities
a t low levels, the deviations at LBF did not exceed 3 kt,
while half of the deviations a t 3,000 ft did not exceed 9 kt.
Therefore, an error up to, but not exceeding, 6 k t can be
attributed to the cloud height uncertainty.
The last column for each cloud type in table 2 shows
that the largest part of wind-estimate errors a t the two
representative levels is due to cloud height uncertainty.
In all portions of the sample, the error due to height
uncertaintly is greater by a factor of two for the 30,000-ft
wind estimate than for the 3,000-ft estimate. This is not
surprising, in view of the great range of high-cloud heights.
NONADVECTIVE CLOUD MOTION
We have no way of measuring the influence of non-
advective motions. The deviations illustrated in figures 7
and 8 suggest that the analysts selected a few improper
targets and tracked nonadvective cloud motions. We have
found that with increased experience ATS analysts become
more selective. This suggests that they detect and reject
a greater proportion of nonadvected features. It is not
surprising that our sample, in part derived by inexperi-
enced analysts, includes some improper targets.
ERRORS OF MEASUREMENT
Misregistration: The largest measurement errors &re the
result of imprecise registration of the first and last pictures
of the sequence. Apparent motion of fixed points on the
earth's surface are thereby added to the real motion of
cloud targets. Picture-to-picture registration can be
achieved by matching earth features, but precision is
limited by resolution of the system. Even when land-
marks are matched to that precision, distortion that
varies from picture to picture puts portions of the image
out of register.
DISTANCE FROM SUBPOINT-DEGREES OF GREAT CIRCLE ARC
FIGURE 9.-Scale adjustment factors for radial vectors, for tangent-
vectors and for vectors 45' from radials, as a function of distance
from picture center. This factor is the ratio of image scale at
the picture center to the image scale at various distances from
the center.
On the basis of analyzing hundreds of ATS pictures,
Fujita (19704 estimated that good-quality ATS3 pictures
can be registered within 10 km a t the subpoint. The error
becomes larger elsewhere on the image because the scale
varies across the image disk. If, for example, 1 mm on
the image corresponds to 10 km a t the subpoint, 1 mm
near the horizon might correspond to 20 km or more,
depending on scale variability.
Variability of scale: Image scale is a function both of
distance from the satellite subpoint (center of the image
disk) and the orientation of the vector being measured.
Scale varies inversely with distance from the subpoint
and directly with the angle between the vector and the
image radial. Line segments parallel to the radials are
foreshortened because the earth's curvature presents those
segments to the camera partly end-on. The same segment
lying perpendicular to the radial will not be foreshortened.
The only scale variation of tangential segments is caused
by increase of slant range to the satellite resulting from
increased subp oin t distance.
Figure 9 shows the factors required to adjust radial and
tangential components. The ordinate is labeled in terms
of the factor necessary to adjust the image scale in areas
distant from the subpoint to the image scale at the sub-
point. The curve for the radial component shows the large-
scale change that is the combined effect of foreshortening
and increased distance from the subpoint, while the curve
for the tangential component corresponds to the change
due to increasing slant range only.3
The effect of scale variability on misregistration error
cannot be assessed, because the direction of misregktration
is unknown. We can state only the upper and lower
limits after we have estimated the magnitude of misregis-
tration at the subpoint. A given misregistration will
9 These nonlinear scale changes are, of course, taken into account when cloud displace-
ments measured on the image are translated intc distances over the earth's surface. Each
cloud motion in the image plane is resolved into its radial and tangantid components and
the appmpriste scale factors are applied.
September 1971 L. F. Hubert and L. F. Whitney, Jr. .. 671
RADIAL DISTANCE FROM SUBPOINT
(Degrees of Great Circle Arc) 81 3
ob l b 2b Jo 4b 5b 6b 7b do Pb lob
RADIAL DISTANCE FROM SUBPOINT
(X of Picture Radius)
FIGURE 10.-Magnitude of the vectorial error of cloud motions
resulting from a 10-km error of registration a t the subpoint, as a
function of distance from picture center. Each shaded area
represents errors corresponding to different time periods of the
animated sequence. The range of errors (hatched areas) for each
time interval corresponds to the range of scale factors (fig. 9)
determined by the orientation of the error vector.
produce larger errors in that part of the field of view
where it has caused a radial displacement than where it
has caused a tangential displacement.
If, for instance, misregistration caused the subpoints
of the first and last pictures to be displaced east-west
by 10 km, then at 50 degrees of latitude north and south
of the subpoint the error would be only 10.5 km, while
at the same distance east or west of the subpoint, the
error would amount to 16.7 km. Fortunately, the in-
fluence of misregistration is minimized by proper selection
of the time interval for animation.
Eflect ?f time interval: In addition to scale and mis-
registration considerations, measurement accuracy of
velocities depends on yet another factor, the total dis-
placement distance of the target cloud. Where the
displacement is large compared to the misregistration
error, neither speed nor direction will be much affected;
but where the cloud displacement is less than the mis-
registration error, the computed velocity might have speed
errors exceeding 100 percent or direction errors of 180
degrees. It is clear that reliable computation of velocity
by our procedures can be accomplished only where the
target displacements are relatively large, that is, where
the time interval of the sequence is long enough to permit
clouds to move a few tens of kilometer^.^
The time interval ultimately is limited by cloud life-
times; nonetheless, it is possible to use 2- to 3-hr sequences.
4 This relation of displacement to error is critical for all techniques that measure cloud
displacement. For example, computer procedures have been developed (Leese et al. 1971)
that match cloud patterns to detect displacement of cloud fields, from picture pairs half
to threequarters of an hour apart. Since many cloud displacements during this interval
will he no greater than 6 to 10 n.mi., clouds must he mapped (earth-located) with an error
substantially less than 2 n.mi ., for reliable velocities- flgure comparable to the resolution
limitation of the present ATS pictures.
Fujita (1970b) examined various types of cloud targets
and demonstrated that a significant number of targets
existed for this length of time. Nearly all of our sample
was derived from sequences upward of 2 hr. The advantage
of using long sequences is illustrated in figure 10.
Based on the assumption of a 10-km misregistration
at the subpoint, the diagram illustrates the magnitude
of vectorial error of cloud velocity that would resdt
from three different time periods. The upper limit of each
curved area corresponds to the error produced if the
misregistration is radially oriented while the lower limit
corresponds to the error vector that is tangentially
oriented. It is clear that the influence of a 10-km regis-
tration error at the subpoint is tolerable if the animation
period is 2 hr or longer.
Eject of orbital inclination: Two images can be registered
over their entire field of view only if the camera is fixed
relative to the earth. Any inclination of the orbital plane
produces a north-south motion of the satellite and pre-
cludes whole-image registration. Cloud velocities can be
derived, nonetheless, by measuring their apparent motion
and subtracting the apparent motion of fixed points on
the earth. Because the number and distribution of visible
landmarks is inadequate t o permit direct measurement
of apparent earth motions, these motions must be com-
puted from the geometry of the problem. This requires a
very accurate specification of the position and orientation
of the spin-axis of the spacecraft.
Accuracy suffered in many of our cloud velocity
determinations because the orbital plane of ATS 1 was
inclined. Corrections were applied, but inadequate
information concerning the satellite parameters and other
pertinent geometrical data at the time the velocities were
derived made the corrections themselves somewhe t
imperfect. Consequently, the misregistration in about
half our sample was somewhat greater than 10 km,
possibly averaging 15 to 20 km.
Errors in tracking: Measurement errors affect both
high- and low-level cloud vectors to the same degree.
Tracking errors, however, appear to affect the two levels
to different degrees. “Tracking error” refers to failure t o
track the same cloud feature throughout the sequence.
Cumulus clusters frequently appear as bright spots.
Cirrus targets, on the other hand, are likely t o be filmy
and poorly defined, suggesting that the tracking error will
be greater for the upper level. Columns 3 and 6 of table 2
show deviations of high-level cloud velocities at their
individual LBF’s to be nearly double those a t low levels
for all portions of the sample. This tends to confirm our
impression that high clouds are more dif6cult to track
than low-level clouds.
Discrepancies due to nonrepresentative rawinsondes:
Some of the deviations between our wind estimates and
rawinsondes are due t o errors in the sounding data. Klein
(1968) showed that the root-mean-square errors of
rawinsonde measurements vary from 5 to 17 kt, depend-
ing on wind speeds. The average speeds encountered in
this study suggested that the appropriate rawinsonde
error is a t the lower end of that range.
672 MONTHLY WEATHER REV1 EW Vol. 98, No. 9
Further discrepancies are caused by wind changes
between the time of upper air sounding and the time of
ATS picture sequence, and change of the flow between the
balloon location (downwind from islands) and the cloud
target location (sometimes upwind from the islands).
We cannot measure the contribution t o wind estimate
discrepancies of nonrepresentative rawinsondes, but cer-
tainly some discrepancies were so produced. Therefore,
the “errors” presented here somewhat overestimate the
true deviation of our estimates from actual winds a t
3,000 and 30,000 ft.
Generally, two levels of cloud motion are easily seen-
one at cumulus levels and one a t cirrus levels. To compare
cloud velocities with balloon velocities, we had to devise a
means of specifying cloud heights. This specification was
made by assuming that the minimum vectorial deviation
between cloud motion and balloon motion occurred a t the
cloud level-a level designated as the “level of best fit’,
The distributions of heights of LBF for low clouds and
for high clouds indicated that low-cloud velocities corre-
sponded most closely to winds a t 3,000 f t and that high-
cloud velocities corresponded most closely to minds a t
30,000 f t . Until better height information is available,
cloud velocities of this type should be regarded as mind
estimates a t those two levels for map-analysis purposes.
A measure of errors introduced by using cloud velocities
for map analysis has been obtained by summarizing the
deviations of cloud vectors from observed winds at nearby
rawinsonde stations, a t their representative levels. Table 2
lists such deviations for various parts of the sample and
shows that upper level estimates have double the devia-
tion of those a t low levels.
Despite the vectorial deviations listed in table 2, the
directional deviations are modest. Hence, streamline
analyses made from cloud velocities provide realistic flow
patterns a t the two representative levels. This character-
istic, together with the prodigious areal coverage provided
by a single geosta,tionary satellite, shows the importance
of this new source of data.
The selection technique is a subjective procedure that
depends on a correct deduction of the synoptic situation.
Experience has shown that meteorological judgment
enables one to select “passive tracers” and classify their
cloud types. Animated picture sequences are of great help
in meteorological interpretation, in distinguishing between
different layers of clouds, and in identifying nonadvective
motion.
An examination of error sources revealed that, apart
from nonadvective cloud motions, measurement errors
axise chiefly from misregistration of pictures and image
distortions which are different from picture to picture.
Inclination of the orbital plane also creates registration
problems because the entire field of view cannot be
(LBF) .
registered when the satellite moves relative to the emth.
The corrections that must be applied depend upon very
precise knowledge of the satellite position and orientation.
Lack of such data has degraded accuracy of the sample
used here.
Accuracy of computed velocity also depends on the
total length of the cloud trajectory, because misregistra-
tion error has little effect if it is small relative to target
displacement. The influence of misregistration can there-
fore be minimized by using sequences long enough to
produce large cloud displacements. With the registration
accuracy that is achieved with good-quality pictures, a
sequence 2 hr or longer will reduce this type of error to
Nonrepresentative rawinsondes have undoubtedly con-
tributed to the discrepancies between wind estimates and
observed winds. For that reason, our various statistics
somewhat overestimate the actual error of o w estimates.
I n conclusion, the accuracies reported here represent
the state of the art as it existed diuing the past few years.
The greatest source of discrepancy between cloud velocities
and rawinsondes at the two representative levels is the
uncertainty in estimating the height of high-level clouds.
It follows that the greatest opportunity for improvement
is in refining the estimate of high-cloud heights. Work
along this line has already commenced. Moreover, infrared
images will be obtained by the next generation of geo-
synchronous satellites. Derived cloud-top temperatures
will aid in estimating their elevation.
2-5 kt.
REFERENCES
Fujita, Tetsuya Theodore, University of Chicago, Ill., IWOa (per-
sonal communication).
Fujita, Tetsuya Theodore, “Basic Problems on Cloud Identification
Related to the Design of SMS-GOES Spin Scan Radiometers,”
Satellite & Mesometeorological Research Project Research Paper
No. 84, Department of the Geophysical Sciences, University of
Chicago, Ill., Mar. 1970b, 33 pp.
Fujita, Tetsuya Theodore, Watanabe, Kazuo, and Ieawa, Tatsuo,
“Formation and Structure of Equatorial Anticyclones Caused by
Large-Scale Cross Equatorial Flows Determined by ATS-1
Photographs,” Journal of Applied Meteorology, Vol. 8, No. 4,
Klein, G. L., “A Comparison of the Wind Accuracy Obtainable
From a Standard Radiosonde and a Transponder Radiosonde
(Instrument Research and Development Report No. 6),”
C. M. R.R. 3/68, Canadian Meteorological Research Reports,
Meteorological Branch, Toronto, Ontario, Mar. 29, 1968, 47 pp.
Leese, John A., Novak, Charles S., and Clark, Bruce B., lrAn Auto-
mated Technique for Obtaining Cloud Motion From Geosyn-
chronous Satellite Data Using Cross Correlation,” Journal of
Applied Meteorology, Vol. 10, No. 1, Feb. 1971, pp. 118-132.
Serebreny, Sidney M., Wiegman, E. J., and Hadfield, R. E.,
“Further Comparison of Cloud Motion Vectors With Rawinsonde
Observations,” Final Report, ESSA Contract NO. E-210-69(N),
Stanford Research Institute, Menlo Park, Calif., Aug. 21, 197Qa,
Serebreny, Sidney M., Wiegman, E. J., Hadfield, R. E., and Eva-,
W. E., “Electronic System for Utilization of Satellite Cloud
Pictures,” Bulletin of the American Meteorological Society, v01. 51,
No. 9, Sept. P97(Pb, pp. 848-855.
Aug. 1969, pp. 649-667.
60 PP.
[Received August $1, 1970; revised January ldj 19711