The determination of cloud pattern motions from geosynchronous satellite image data

Pattern Recognition

PergamonPress 1970. Vol. 2, pp. 279-292. Printedin Great Britain

The Determination of Cloud Pattern Motions From Geosynchronous Satellite Image Data JOHN A. LEESE, CHARLES S. NOVAK and V. RAY TAYLOR National Environmental Satellite Center, Washington, D.C., U.S.A. (Received 17 February 1970) A l m t r a e t ~ l o u d motion can be determined from sequential pictures obtained from the geosynchronous Applications Technology Satellites, ATS-I and ATS-III. Experiments have been conducted using two automated techniques for computing the cloud motions over the time interval between two digitized pictures. One technique is a direct application of cross-correlation using the fast Fourier Transform (fFT). In the other, a relatively simple matching technique is used in the computer to recognize binary images of cloud fields on each member of a picture pair. Results obtained from the two techniques are similar. The computed directions are in good agreement with those determined by manual methods with indications that computed speeds are more accurately resolved than those determined manually.

1. I N T R O D U C T I O N THE Applications Technology Satellites (ATS) each contain cameras which obtain information about the earth's cloud cover. ATS-I and ATS-III are in geosynchronous orbits and so remain almost stationary over a given longitude at the Equator. ATS-I is located over the Pacific Ocean area and ATS-III over the Atlantic Ocean area. An example of the kind of pictures obtained is shown in Fig. 1. This picture was taken at 2124Z on November 5, 1969 by the spin-scan camera aboard ATS-I. It covers a significant portion of the Pacific Ocean with the United States and Mexico located in the upper right portion of the earth's disk. ATS spin-scan cameras can provide pictures of the earth's cloud cover at intervals of 24 rain during daylight hours. Characteristics of the ATS-I satellite and camera were described by MCQUAIN.(1~ ATS-III, with its multicolor spin-scan camera, was described by SUOMI and PARENT (2) and WARNECKEand SUNDERLIN. (3) Wind speed and direction are the most important parameters used in numerical weather prediction models. However, surface based measurements of the wind field are practically nonexistent over many areas of the ocean. The ability to determine cloud motion from ATS pictures was demonstrated shortly after this kind of information became available. The "loop" movies produced by Dr. Fujita at the University of Chicago exhibit this possibility in a very striking manner; cf. FUJITA, (4) FUJITA et al., tS~ IZAWA and FUJITA (6) have shown that the cloud motion vectors are very closely related to the wind field at particular altitudes, e.g. the wind speed and direction at an altitude of 5000 ft can be approximated by the motion vector determined from the low level cumuliform cloud patterns. Manual techniques are currently being used to determine cloud motion vectors from ATS pictures. This is done by viewing the loop movies and fixing the positions of cloud elements at the beginning and end ofthe picture sequence in which each appears. The tedious nature of this manual process makes highly desirable the development of an automated processing technique capable of determining cloud motion information on a near real time basis. 279

280

J.A. L~SE, C. S. NOVAKand V. R. TAYLOR

2. I N P U T DATA The ATS picture shown in Fig. 1 is available in digital form as an array consisting of approximately 2000 scan lines with 4096 samples per line, each digitized to one of 64 grayscale levels. Cloud displacements have been computed from sequential pairs of digitized pictures in this format. However, our approach is designed for a pair of image sectors mapped onto a standard projection. This has been done to simplify the program logic which otherwise would be required to deal with distortions arising from variations in perspective and relative coordinate systems. A mercator mapped image of the ATS picture in Fig. 1 is shown in Fig. 2. BRISTOR~7) has described the mapping procedures used at the National Environmental Satellite Center. The dimensions of the map shown in Fig. 2 are approximately 100 degrees of latitude by 100 degrees of longitude. Each degree of longitude consists of 25.6 samples. The number of samples per degree of latitude on a mercator projection varies from 25.6 at the Equator to 39 at 50° N or 50 ° S. The grid spacing currently used in Numerical Weather Prediction is about 2.5 degrees latitude. A comparable grid square on the mercator map of ATS data contains a matrix of 64 by 64 individual gray scale values. This matrix size was selected so that computed cloud motion vectors could be used to provide wind data on a scale compatible with the Numerical Weather Prediction grid spacing. 3. D E S C R I P T I O N O F T E C H N I Q U E S 3.1

Cross-correlation

For our purposes, cross-correlation is defined as a measure of the relationship between the members in two sets of quantities. The technique is also commonly referred to as crosscovariance because the two sets differ by a factor which is a constant for many problems. The use of these techniques is the basis for many of the pattern analysis and recognition techniques. The essential features of two-dimensional cross-correlation as applied to this problem are shown graphically in Fig. 3. The input array (or matrix) representing a digitized picture consists of gray scale values over an area, denoted as G,. G t = {go~t)}n,nj

(I)

where : i and j are the indices of the rows and the columns of the picture matrix, G t . nl and nj are the total number of rows and columns of the picture matrix, G,. gii(t) is the gray scale value at the location (xl, y~) at the time, t, that the picture was taken. (Where x and y are the orthogonal coordinates of the picture and the sampling locations, (xi, y y s , are spaced at equal intervals.) The root-mean-square variation of the input array, G,, is : 1

~,

[go(t)-- G,]

(2)

where 1

-

nj

ni

~

~ go(t).

n i + nj j 7 1 i = 1

(2a)

FIG. 1. ATS-I Picture, November 5, 1969, 2124Z.

[facing p. 280

FIG. 2. Mercator m a p o f ATS-1 Picture 1 shown in Fig. 1.

FIG. 4. Superposition of two binary images.

Determination of cloud pattern motions

281

The cloud motion can be determined through the analysis of the correlation between two pictures covering the same area but taken at different times, to and tl. The corresponding picture matrices are denoted G,, and Gt,. The covariance between Gt,, and Gt, with respect to a displacement between these two pictures is defined as : l

nj - q ni

p

2 i=1 y Lgi÷P4+q(t°)-G"'][gifltl)-ffJt~] r cov(p,q) -- N j~l

(3)

where: p and q are the integers which determine the displacement through the shifting of indices. The actual displacement is (x i ÷ p - x i ) and (Yi+q-Y/) in the x-direction and the y-direction respectively. N = ( n i - p ) + O l j - q ) - 1. l h e correlation coefficient of G,,, and G,, with respect to the displacement, (p, q), is defined as the following: R(p, q) -

coy{p, q)

(4)

O't I)O't 1

where the covariance, cov(p,q), and the root-mean-square variations, at, and at,, are defined in equation (3) and equation (2) respectively. In order to compute the speed and direction of the cloud motion in the time interval between to and t~, the displacement that yields the maximum correlation coefficient needs to be determined. This is done by computing a number of correlation coefficients, R(p, q), with respect to different displacements. The range of the displacement is determined according to the following limits : -P

<~ p <~ P

-Q~q~Q as shown in (c) of Fig. 3. Let p', and q' be the displacement indices that yield the maximum correlation coefficient : i.e. R(p', q') = max[R(p, q)]. The speed and direction of the cloud motion can be then computed as the following : [(p,. Ax)2 +(q,. Ay)2]l 2 I~:1 =

At

p'. Ax 0 = arctanq'. Ay

(5a) (5b)

where : I~ I is the speed of the cloud motion. 0 is the angle of the direction of the cloud motion. Ax = (xi+ ~- x~) and Ay = (y2+ t - y j) are the sampling intervals of the input picture matrix. At = t~ - to is the time interval between the times that the two pictures were taken. Equation (4) is the basic definition for the pattern matching required to compute cloud displacement. The main consideration in the use of the formal cross-correlation techniques is whether the computations can be accomplished in a sufficiently short time for real time operations.

282

J.A.

LEESE, C. S. NOVAK a n d V. R. TAYLOR

to

tI

Gt

Gt

a. Input Arrays p

to

I

qlj li I

I

b, Cross-Correlation at Lag p,q Q

-p

t k~p~q I , /

p

c. Cross-Correlatlon Matrix

FIG. 3. Graphical depiction of two dimensional cross-correlation.

If we rely on the so-called "lagged-product" method of equation (4), the time required to compute the cross-correlation matrix is almost prohibitive. For example, some test cases consisted of 60 x 60 data points for each input array. A cross-correlation matrix of 25 x 25 lags for a single time step requires approximately 10 min to compute on the IBM System 360 Model 50, a third generation medium sized computer. A larger computer system may be able to reduce this time to about one minute per matrix but this is still too long for any significant real time use, when one considers that it is possible to get 1000 or more input arrays for cloud motion computation on a single pair of ATS pictures. An algorithm for the computation of Fourier coefficients, which requires much less computation effort than previously required, was reported by COOLEY and TUKEY. (8) This method is known as the "fast Fourier Transform" (fFT). The use of this transform has produced major changes in computational techniques used in digital correlation, spectral analysis, filter simulation and related fields. For brevity let us examine how the Fourier Transform can be used for computing crosscorrelation coefficients in a one dimensional time series. Extension to two dimensions is a

Determination of cloud pattern motions

283

straightforward process. Given two time series f~(t) and f2(t), the cross-covariance function at lag r is defined by h(r) = lim

T~°° ~

1 ff

T

fl(t) . f 2 ( t + z ) dt.

It is assumed that each of the functions f~ ,f2 and h has a complex Fourier Transform from the time domain to the frequency domain which we represent by the correspondence fl(t) ~-~ F l (~o)

f2(t) ~ F2(oJ ) htr) ~ H(oJ). Conditions which must be met for the existence of these transforms are discussed in most standard references (e.g. BENDAT and PIERSOL,(9)) and will not be repeated here. For a finite series of data at equally spaced intervals the Fourier transform becomes FI(09) =

N- 1 ~

[ - 2ni~tl

f l ( t ) . e x p /~~ / ; ~

! -

N

N

~< e9 ~< ~-.

(6)

t=O

The transform is computed to obtain Fl(e~) and F2(~). The cross-covariance function in the frequency domain is given by H(eo) = F~'(cn). Fx(co).

(7)

Where F~'(~) is the complex conjugate of Fz(~O). An inverse transform is then made to obtain the cross-covariance in the time domain : o,=N/2 [ 2nitotl h(0 =o,=~N/2 H(~o) • e x p t T ].

(8)

The cross-correlation coefficient is computed by R(z) -

h(0

a/,.

ay~

The details of the computational technique for the fFT have been discussed in numerous places and will not be repeated here. References devoted to the f F T are BR1GHAM and MORROW,( t o) COCKRAN et al., (~ ~~ GENTLEMAN and SANDE(12) and SINGLETON.(13) C o m p u t e r program subroutines for using the fFT have been written by several groups. SINGLETON(~4~ listed an A L G O L procedure for computing the f F T with auxiliary m e m o r y and limited high-speed storage. The particular subroutine used in our analysis was obtained from the IBM Scientific Subroutine Package'[. t System 360 Scientific Subroutine Package (360A-CM-03X) Version II Programmer's Manual. This package contains four subroutines for the fFT : HARM--Performs discrete complex Fourier transforms on a three dimensional array. DHARM--Same as Harm but uses double precision arithmetic. RHARM--Finds the Fourier coefficientsof one dimensional real data. DRHARM Sameas RHARM but uses double precision arithmetic. Copies of this publication can be obtained through IBM branch offices as document number H20-0205-2.

284

J . A . LEESE, C. S. NOVAK and V. R. TAYLOR

The array size used with the fFT was 64 x 64 using an IBM System 360 Model 50 computer. The time needed to compute the cross-correlation matrix, automatically search for the maximum value and print out the results was approximately 30 sec. This compares with approximately 10 min for the lagged product method on the same computer. This factor of 20 is approximate only, since no concentrated effort was made to obtain an exact comparison. Similar tests using a CDC 6600 computer, reduced the time to 4 sec as compared with approximately 30 sec for the IBM System 360 Model 50 computer. This resulted in an additional reduction of the computer time by almost a factor of 8. The timing results show that the fFT used in conjunction with a large computer system makes the cross-correlation a valid technique for consideration for real time operational use in computing cloud motion from geosynchronous satellite data. It should be pointed out that the fFT can be built into a special purpose device which can be attached to a computer system. These devices can provide significant increases in speed above that which can be achieved with a computer program and a large computer. Several fFT special purpose devices are already available.

3.2 Binary matching In using the basic definition of cross-correlation to compute the displacement of cloud patterns, one implicitly assumes that it is necessary to maintain all the gray scale information in the original mapped arrays of cloud data. Binary slices could be made of the cloud pattern for input to the cross-correlation but there would be little or no gain in processing time, which is a major consideration. BRISTORet al. tl 5~applied a binary matching technique to the cloud motion problem which was an adaptation of technique developed for metallurgical applications by MOORE.t16~ The results of their work were encouraging and provided insight on improvements which would make the technique more applicable to the cloud motion problem. Cloud motion may be derived by measuring the change in position of some feature of the cloud field which can be easily and quickly defined and recognized in sequential pictures. One such feature is the cloud edge. A simple way to define this edge is to assign to it a brightness isopleth which lies somewhere between the brighter values which are obviously within a cloud and darker values in the adjacent cloud free area. This chosen isopleth may represent actual cloud edges, or because of the limited resolution of the data acquisition system, may be a boundary where the cloud field becomes broken into elements which cannot be resolved. The cloud "edge" then, may be a transition zone between cloud-covered and cloud-free areas and as such may be defined as the area between two brightness isopleths. To isolate the cloud edges, the cloud field may be redefined in terms of a binary image with events filling the area between two isopleths and nonevents filling all the remaining area. The superposition of two such images is shown in Fig. 4. Both images are from ATS-III pictures taken approximately one hour apart and show the motion of the clouds associated with hurricane Abby. The first binary image, shown in white, is from the 1534Z ATS-III picture and the second image, in black, is from the 1637Z ATS-III picture. Areas of overlap are stippled and cloud free areas are solid gray. The displacement of the clouds can be seen as the distance between white and black edges of corresponding features. This displacement can be measured and converted to cloud motion in the computer by matching

Determination of cloud pattern motions

285

the two images. The binary image of the first picture is divided into sectors (Fig. 5). Each sector contains cloud edges which constitute a pattern to be found or matched as closely as possible with a sector of the same size from the second image. A first image sector is matched with every possible sector of the second image. This matching is done within a restricted area determined by previous experience to contain all reasonable displacements of the pattern being sought. The origin of the motion vector is at the center of the first image sector, and the terminus is at the center of the second image sector which most nearly matches it.

1

-

:

;~t

FIG. 5. Binary image divided into sectors•

3.3 Mapping correction Ideally, one would prefer to have the total location errors at zero in the mapped input data. However, operationally useful output can be achieved with relatively large total variations in the initial data provided they are systematic between the picture pairs. For example, a random error of 25 miles with 20 min time difference between the two pictures would produce an error of 75 miles per hr in the cloud motion vector. A systematic error of the same magnitude would result in a displacement of 25 miles in geographical location

286

J.A. LEESE, C. S. NOVAKand V. R. TAYLOR

of the cloud displacement vector. In Numerical Weather Prediction, an error of 75 miles per hr in cloud motion makes the results useless, while a displacement of 25 miles in geographical location would have little effect on the utility of the data. The example cited above shows that relatively large systematic errors between picture pairs have little effect on the results. However, random errors in location between the two pictures have the greatest potential for producing invalid cloud motion vectors. Our objective is to achieve an automated technique capable of determining the speed of cloud motion to an accuracy of + 5 miles per hr. Our experience with the mapped data, using the procedures described by BRISTOR,t7) shows that the relative location error between two pictures falls between + 16 miles in most cases. However, the tests on landmarks across the picture show that the error is practically constant across a given pair of pictures. Since this is basically a simple translation, the error can be computed from a landmark in the picture and deleted from the resultant cloud motion vector. In both the cross-correlation and binary matching techniques a set of locations along the peninsula of Baja California is used to compute the relative mapping errors between the two pictures which then are applied automatically as corrections to the cloud motion vectors over the remainder of the picture. 4. TEST RESULTS O N REAL DATA

4.1 Detection of linear displacement The component of motion to be measured is that part due to simple translation of the cloud patterns. If this component was the only one present the patterns would remain uniform with time. However, the growth of new cloud elements and the decay of older clouds introduce complications to the problem. Other factors which can change the appearance of cloud patterns with time are variations in camera system response and solar illumination. The above factors, all acting on the cloud patterns, serve to degrade the maximum value of the cross-correlation coefficient and the maximum percentage of hits in the binary matching technique. Examples of a cross-correlation matrix and a matrix of percentage hits are shown in Figs. 6 and 7. These are examples of cloud displacement when there is only one layer of clouds present. Our experience has shown that the peak value of the crosscorrelation coefficient normally is no larger than about 0.7. Very distinct landmarks, such as those along Baja California (Fig. 2), yield maximum cross-correlation coefficients of about 0.9. Figures 6 and 7 show the characteristic of a very distinct relative maximum in both the cross-correlation and binary matching matrices. It was expected that the rapid growth and decay of some cloud patterns might yield spurious peaks in the matrices. However, this has not occurred and has made possible the use of a simple search for the maximum value in the matrix as the location proportional to the cloud displacement vector, see equation (5). 4.2 Cloud motion vectors for single layers Locations in which low level cumuliform type clouds were present were picked by visual inspection from ATS-! mapped imagery (Fig. 2). The cloud motion vectors were

Determination of cloud pattern motions

LAGP--32

-16

0

287

+16

+32 LAGQ = + 3 2

+16

1

1

-16

i'

'

I

-3?. FIG. 6. Cross-correlation matrix. ~l+Cm]r++

n ! .........

~PC++74+;P l,' r

TAUt[ ~(+ t

]

30 I 33333

~Z22~

e o o e ~ t o • •

4n l 444~a

e~

t~

o • oe

~

5~ ! 5555b

~0 ! 6666~

~ o o o o o o o

• •

~

Q Q • ooo~

22~ ~2Z

• • o o • • •

70 l 77777

oeeQoooet

o o ~ o o o o t

3333333333 3333333333333

• . . •

22~ 333333333333 2~22 333333 333 22~ 33333 33 ~2Z 3333333 4~&& pP?~ 333333 4444444~ 2~ZZ 333333 44444 .444 2 2ZZ 333 444 4 44~4 2~ 3333 444~ 4~.4 22 333 444 555 44 2P 33 444 55~555 44~ 2 Z2Z 33 ~ 555555555 .44 Z~ ~22 33 44 555 555 4~ 222222222k 3 44 55 66 55 4. 2£22~2£ 33 44 ~ b~ 6 6 6 55 . 4 4

•

~2222~

•

• • • e • • • • •

• •

eoeoeooeo

b

6

6

5

44

ot

• • • • • • • o

• e • • • •

3

"

• •

oooeo~o~e

"

90 I 9999q

5 ~ 7777 6 5 4. 33 5506 7X77 6 5 4 33 333 ~44 556b 777 6 % 44 3 3 3 33333 4~ 55 ~ 6b ~ 44 3 3 ~ 333333333 .4~ b b b 6 6 6 55 4 33 333333333 4~. 5 555 ~ 333 22 3333333 646 5655%55 ~ 3 22P2 333333 t44~ 44 3 Z2?Z2 33333 444444 444 33 22~ 2 3333 .4~.4.44t4444 33 2ZZ 3333 4*44444 3333 22222 +33333 + 333333 +2~E 3333333333 3333 2P~ 3333 33333333 33333 Z~ 3333 33~333333333 ~Z~Z~ ~Z 33~333 33 333 222~22~22~22 ~2222 3333 ?~Z?ZZ Z Z Z Z Z 2 Z Z ~ Z 3333 +++~Z ?++?~+++2

2~

3

BO t ~8~

33 4~ 333 444

eo~oeooeo

ee~eoeo~o

oeeeoee~o

FIG. 7. Percentage-of-hits matrix.

•

• • " * " •

• • • •

•

oo~eoeeoo

oe

lO0 X 0

288

J.A. LEESE,C. S. NOVAK and V. R. TAYLOR

computed at each of these locations and the directions and speeds mapped by the computer. The map of cloud m o t i o n vectors is shown in Fig. 8; each vector represented by a five digit number. The first two digits give the direction from which the cloud m o v e d in tens of degrees and the last three the speed in knots. For example, 09025 means the cloud m o v e d from the East (090 degrees) at a speed of 025 knots. The barbs at each location show the direction in the manner customarily used by meteorologists.

CLOUD

I

ATS

rlOTION

VECTORS

DATE

i .g=°i L,

O'LO~O

l-

Dido

Nooq

ll/

51[;9

........

TIME I

,

"

1 'f,

I

',.

i

~

2 ~4Z - ,?.147Z

A'

II

'-

,

J {

~'

3b Dido

35N

i3"1 O/t5

OO'OgO

t

'i /

J~":'?-.. I

O~ Di~b

|

Dh 011 D3D~H:]b u= uac

--

5

.

.--

,~:. . . . . .

° ~'°i~. ! :

I

b.i',, -'oo ,.--rro~o

t , ?oon .

!I

J

"-,

I

J

",

'

'DO'ODD

1 Ob~

ao'ooo~fOJoSq

(

¢~ uc ~

-- ~

03

i

/~

' "', i

1

Ob05,L Dh 0~1

I

..,] leloib

e0~'~ "1

05"~33i

i

1

i./"

..........

_

OOe .....

OO'OOOi

Do ......

|/

i

----

~U

,

i I

OO'OO0

031

!Deo]-'~1

!

--i+ . . . . . . . . . . . . . . .

oqQ1-'i~

oqol-r~j2

--

i

15N

1!

aqaz-iTI/

l

I.

lllW

125~/-~ °~-n I i

'i

,

t i

r

FIG. 8. Map of cloud motion vectors--ATS-I November 5, 1969; Picture pair times 2124Z and 2147Z.

4.3 Comparison with manual results "Ground truth" information on the wind field is extremely rare over the geographical area shown in Fig. 8. In addition, we have only a very rough idea of the altitude at which

Determination of cloud pattern motions

289

the clouds occurred. A more meaningful comparison of our results can be made by using winds determined by manual techniques for the same geographical area. A comparison of the computer and manually determined directions is given in Chart 1. 36

32

28 C O M P U T E

24

20

R

D I

16

R

E C T

12

1

O N

8

C C

4

8

12

16

20

24

28

32

36

MANUAL DIRECTION CHART1. Comparison of computer and manually determined direction. The comparison of speeds is given in Chart 2. A line of perfect fit is drawn diagonally across each. The comparisons, for a four day period, are with the manual results taken directly from the operational analysis. One could be encouraged in that 81 per cent of the directions agreed within 30 degrees and 78 per cent of the speeds agreed within 10 knots. An important conclusion drawn from these comparisons is that weaknesses exist in both the manual and computer techniques. In the manual technique the speeds are clustered

290

J . A . LEESE, C. S. NOVAK and V. R. TAYLOR

95 90 85 80 75 70 65 60 C O M P

U T E

55 50

R

45

S P

40

1

35

1

E E D

1

30 3

25

10

20 3

15

5

/ C

5

¢/~

4 /

/

/

/ , / 1

10

12

12

I

2

8

10

1

2

5

2

I

10

15

20

25

10

C

3

1

/

/

30

35

40

45

MANUAL SPEED CHART2. Comparison of computer and manually determined speed.

Determinationof cloud pattern motions

291

around 15 and 20 knots (see Chart 2), while the computer technique gave estimates of between 5 and 25 knots. In this case, the computer estimates were more realistic. A relatively large number of calm or no motion situations (row labeled " C " in Tables 1 and 2) were derived by the computer when the manual method indicated motions were present. These cases were examined in detail and it was found they occurred when there were higher cloud layers present. The man examining the loop movies was able to discriminate between the low and high cloud layers moving at different directions and speeds. However, the cross-correlation technique integrated the motion of the two layers to yield a resultant vector near zero. Tests with cloud patterns synthesized into multiple layers have given some insight on this portion of the problem. Up to now, it has not yet been possible to demonstrate the feasibility of using computer techniques with multiple cloud layers.

5. CONSIDERATIONS FOR OPERATIONAL APPLICATIONS An automated technique is needed to compute valid cloud motion vectors from geosynchronous satellite data within the time constraints of a real time operational system. The experiments with cross-correlation and binary matching yielded similar results in the form of cloud motion vectors. Both techniques could automatically correct for any relative mapping errors by using landmarks. The binary matching technique has an edge in the processing time needed, by a factor of two or more. Comparison with manual results show that the automated and manual techniques are more complimentary than competitive, because each has different strengths and weaknesses. The strength of the automated techniques is in making the measurements after the locations of unique patterns have been identified ; this is where the manual techniques are the weakest. The major weakness of the automated techniques is in their inability to discriminate between motions when more than one cloud layer is present over the same geographical area. This is exactly the place where the man has a unique advantage over the automated techniques. The best operational solution to the problem at this time appears to be a combined man-machine interface where the manual interpreter has access to both the gray scale imagery and the automated techniques in a conversational mode with the computer.

SUMMARY Cloud motion vectors as an estimate of the wind field are now extracted from geosynchronous satellite image data by manual techniques which are very tedious and time consuming. We have tested two techniques for automatically computing the cloud pattern displacements between a pair of mapped image sectors. The two-dimensional crosscorrelation technique makes use of the full range of 64 gray-scale values. In the binary matching technique, the gray scale image is redefined in terms of events which are cloud edges and non-events which constitutes the remainder of the image. Cloud displacements were detected equally well by either automatic technique when there was only one layer of clouds present in the image sector. Neither technique was successful in discriminating the motion vectors when there were two or more layers of clouds present.

292

J.A. LEESE, C. S. NOVAKand V. R. TAYLOR

C o m p a r i s o n s with the cloud m o t i o n vectors d e t e r m i n e d by m a n u a l techniques showed little differences in the directions for the vectors. However, the c o m p u t e r techniques were able to m e a s u r e the displacements to a finer resolution t h a n could be d o n e by the m a n u a l techniques. It is c o n c l u d e d that for single layer cloud p a t t e r n s the a u t o m a t e d techniques can do at least as well as the m a n u a l techniques. The inability of the a u t o m a t e d techniques to d i s c r i m i n a t e a m o n g multiple cloud layers present in the same image sector precludes their i m p l e m e n t a t i o n as a completely a u t o m a t e d operation. It is c o n c l u d e d that the best o p e r a t i o n a l solution at this time is a c o m b i n e d m a n - m a c h i n e interface system in which the m a n has access to the time-lapse display of the gray scale imagery a n d can c o n t r o l a n d override the cloud displacements c o m p u t e d by the a u t o m a t e d technique.

REFERENCES 1. R. H. McQvAIN,ATS-I camera experiment successful, Bull. Am. Met. Soc. 48, 74-79 (1967). 2. V. E. Suoui and R. J. PARENT,A color view of planet earth, Bull. Am. Met. Soc. 49, 74-75 (1968). 3. G. WARNECKEand W. S. SUNDERLIN,The first color picture of the Earth taken from ATS-III satellite, Bull. Am. Met. Soc. 49, 75-83 (1968). 4. T. FtL~ITA,Present status of cloud velocity computations from the ATS-I and ATS-III satellites, Proc. o f the Space Research IX; COSPAR, Plenary Meeting, l lth, Tokyo, Japan, pp. 557-570. North-Holland, Amsterdam (1969). 5. T. FWITA,K. WATANABEand T. IZAWA,Formation and structure of anticyclones caused by large-scale cross equatorial flows determined by ATS-I photographs, J. Appl. Met. g, 649-667 (1969). 6. T. IZAWAand T. FUJITA,Relationship between observed winds and cloud velocitiesdetermined from pictures obtained by the ESSA 3, ESSA 5 and ATS-I satellites, Proc. o f the Space Research IX; COSPAR, Plenary Meeting, l lth, Tokyo, Japan, pp. 571-579. North-Holland, Amsterdam (1969). 7. C. L. BRISTOR,The earth location of geostationary satellite imagery, Pattern Recognition 2, 269-277 (1970). 8. J. W. COOLEYand J. W. TUKEY,An algorithm for the machine computation of complex Fourier series, Math. Comput. 19, 297 301 (1965). 9. J. S. BENDATand A. G. PIERSOL,Measurement and Analysis of Random Data, p. 390, John Wiley, New York (1966). 10. E. O. BRIGHAMand E. R. MORROW,The fast Fourier transform, IEEE Spectrum 63-70 (Dec. 1967). I I. W. T. COCHRANet al., What is the fast Fourier transform? Proc. IEEE 55, 1664-1674 (1967). 12. W. M. GENTLEMANand G. SANDE,Fast Fourier transforms for fun and profit, Proc. Fall Joint Computation Conf., AFIPS 29, 563-578. Spartan, Washington, D.C. (1966). 13. R. C. SINGLETON,On computing the fast Fourier transform, Comm. A.C.M. 10, 647-654 (1967). 14. R. C. SINGLETON,A method for computing the fast Fourier transform with auxiliary memory and limited high speed storage, IEEE Trans. on Audio and Electroacoustics AU-l 5, 91-98 (1967). 15. C. L. BRISTOR,M. FRANKELand E. KENDALL,Some advances in automatic determination of cloud motion and growth from digitized ATS-I picture pairs, manuscript submitted to University of Wisconsin for publication in Weather Motions from Space ATS-I, University of Wisconsin Press, Madison (1970). 16. G. A. MOORE,Application of computers to quantitative analysis of microstructures, National Bureau of Standards Report No. 9428 (1966).