TRACKING FUZZY STORM CENTERS IN DOPPLER RADAR IMAGES

1 1 1 2
D. Cheng , R. E. Mercer , J. L. Barron and P. Joe

Dept. of Computer Science

1
King City Radar Station
2

The University of Western Ontario Atmospheric Environmental Services

London, Ontario, N6A 5B7 Toronto, Ontario M3H 5T4

ABSTRACT The method used to determine this point (center of

In this paper, we describe an automatic storm track- mass calculation) is sensitive to the properties of the
ing system to help with the forecasting of severe storms. storm (deformability, changing intensities). Although
The concepts fuzzy point, fuzzy vector, fuzzy length of the storm moves in one direction, the algorithm can
a fuzzy vector, and fuzzy angle between two non-zero actually record a movement in the opposite direction
fuzzy vectors are rst examined. We use a region split-
ting algorithm with dynamic thresholding to determine
storm masses in Doppler radar intensity images. We
represent the center of an hypothesized storm using a
fuzzy point. These fuzzy storm centers are tracked over
time using an incremental relaxation algorithm. The
algorithms are tested on actual radar images obtained
from the Atmospheric Environment Service radar sta-
tion at King City, Ontario, Canada. The algorithms
due solely to the somewhat arbitrary location of the
center of mass of the storm in the subsequent image.
This undesirable movement of the point representing
the storm aects the performance of the relaxation al-
gorithm. To obtain results that were close to reality
with the methodology described in Krezeski et al. , we
had to set thresholds in the relaxation algorithm at
levels that allowed incorrect tracks to be generated.
To diminish the eect of the arbitrariness of the
are capable of producing storm tracks which closely
match human perception.
1. INTRODUCTION
Because of the devastation caused by severe storms,
the forecasting of severe storm movement is one of the
most important tasks facing meteorologists. To help
with this task we have been developing an automatic
storm tracking system. Tracking past storm movement
is a prerequisite to forecasting future storm movement.
Tracking severe storms in data sets obtained from
an operational radar in southern Ontario poses inter-
esting problems. Storm systems move too great a dis-
tance and their properties change too much over time
storm location we have modied the representation of
the location of a storm to be a fuzzy point. A fuzzy point
is a circle whose inner region represents the uncertainty
of the location of a targeted point. Our current work [1]
uses fuzzy points, fuzzy vectors, fuzzy lengths of fuzzy
vectors, and fuzzy angles between two non-zero fuzzy
vectors in the relaxation framework to produce more
realistic storm tracks. Note that our interpretation of
a fuzzy point is not based on the theory of fuzzy sets
[6]; rather, it is in the spirit of fuzzy geometry [4, 5].
2. FUZZY POINT ALGEBRA
A fuzzy point P = hc; ri is a circle with center c =
(cx ; cy ) and radius r. It represents a region where a
targeted point can arbitrarily be located. We denote
(for instance, shapes deform and average intensities
change) to allow the use of correlation-based tracking
techniques. Other problems encountered include alias-
ing, storm path crossovers, and the birth and death of
storms.
Krezeski et al. [3] report on a relaxation algorithm
that we have developed to overcome these problems.
The results, although promising, showed some prob-
lems which can be attributed to the deformability and
changing intensity values in the storm. The represen-
the set of all fuzzy points as P. We dene the distance
 between two fuzzy points P1 = hc1; r1i and P2 =
hc2; r2i as:
 (P1; P2) = kc1 ? c2 k2 + jr1 ? r2j ;
where (P;  ) forms a metric space. If we consider P
as a subset of R2, then a fuzzy vector P??! 1P2 from fuzzy
point P1 to fuzzy point P2 is the set of all displace-
ment vectors from a point in P1 to a point in P2. The
tation of the location of a storm in an image (which fuzzy length of this fuzzy vector is then dened as the
is required by the relaxation algorithm) is a point. set of lengths of all displacement vectors in P??! 1P2 . This
set can be represented as the real interval [dmin ; dmax],
where dmin and dmax are respectively the least and
greatest distance between any two points in the circles
representing the fuzzy points P1 and P2. Any fuzzy
vector with fuzzy length [0; d]; d  0 is considered as
a zero fuzzy vector. The fuzzy angle subtended by a
non-zero??!fuzzy vector P??!2P3 relative to a non-zero fuzzy
 On the other hand, if sz is small, as indicated by
sz < Ts , then we consider z to be representative
of the intensity levels of most pixels in Sn and we
compare the value of z to the dynamic threshold
Tz. Tz is based on the mean, z0 , and standard
deviation, s0z , of the intensity levels of a subset of
pixels in the image with intensity levels greater
vector P1P2 is dened as the set of angles subtended than zmin = 16. It is computed as Tz = z0 +
by any displacement ??! vector in P??!
2P3 relative to a dis- ks0z , where k = ?0:5. If we have z  Tz, then
placement vector in P1 P2 having touching heads and the subnode Sn will be considered as a part of
tails, respectively. The angle between the two displace- a potential storm and will be marked for further
ment vectors, , can be determined by the dot-product. processing.
Like the fuzzy length of a fuzzy vector, the fuzzy an- The above splitting process continues recursively until
gle between two non-zero fuzzy vectors is a real inter- no subnode can be further split. Neighbouring subn-
val [min ; max ]  [0; ]. However, determining the two odes that are marked for further processing will be
endpoints min and max is not trivial. Currently, we grouped into a single region,
use an O(n) search algorithm to approximate the two
endpoints. R = Sk1 [ Sk2 [


[ Skn
if they are connected. We say that a subnode S 0 is
3. STORM HYPOTHESIS connected to a subnode S with a distance d if all of the
following conditions hold,
8
A region splitting algorithm with dynamic thresholding >
> l + d  u0y
is used to determine storm masses in a radar intensity < y uy ? d  ly0
image. This algorithm replaces the one in [3] that uses >
>
:
lx + d  u0x ;

Horowitz and Pavlidis' merge-split algorithm [2]. The ux ? d  lx0

merge-split process in [3] is determined by a threshold
on the dierence between the maximum and minimum
intensity levels of adjacent regions. We have observed
that this criterion is sensitive to outliers in the data.
Besides, in order to achieve good results, it requires the
threshold to be appropriately and manually chosen for
each image. In our new algorithm, we use the standard
deviation s of the average intensity level of a region to
where (ux ; uy ) and (lx ; ly ) denote the upper-left and
lower-right corners of S , respectively; (u0x ; u0y ) and
(lx0 ; ly0 ) denote the upper-left and lower-right corners
of S 0 , respectively; and d is a threshold currently set to
2 pixels.
3.1. Construction of Fuzzy Storm Centers
govern the splitting process. We nd that thresholding After region splitting we have a set of regions
using s is very robust. fR1; R2; : : : ; Rng in the Doppler radar intensity im-
A square image S is initially at level 1 of a quad- age which are hypothesized as storms. To represent
tree and is divided into four equally sized subnodes the location of each storm region, Rj , using a fuzzy
S1 ; S2 ; S3 and S4 which correspond to the north-western,
north-eastern, south-eastern and south-western regions
respectively.
For each subnode Sn , where n 2 + and  =
f1; 2; 3; 4g, we compute the mean, z, and the standard
deviation, sz , of the intensity levels of all pixels in Sn .
The splitting criterion of the subnode is based on the
value of sz . The value of sz indicates how well the value
of z represents the average intensity level of all pixels
point P = hc; ri, we rst compute the weighted av-
erages, x and y, of the centers of all subnodes form-
ing the region, in the x- and y-direction, respectively;
and also the corresponding standard deviations, sx and
sy . Then, the center of the fuzzy point P is taken as
c = (x; y); and the radius of the fuzzy point is deter-
mined by r = kr max(sx ; sy ), where kr is a parameter
to control the size of the fuzzy point. We currently
use kr = 0:5. We interpret the above construction as
in Sn . We currently set a threshold Ts on sz to 10. tting a quasi-circular Gaussian surface through the
storm region. Since the data points in the storm re-
 If sz is large, as indicated by sz  Ts , then the gion do not necessarily spread around the peak of a
intensity levels of most pixels in Sn dier signif- Gaussian surface, we could not employ a least-square
icantly from z and therefore, the subnode Sn is method to perform a Gaussian t. Since the fuzzy point
split into four smaller subnodes Sn1 ; Sn2; Sn3 and represents the uncertain location of the storm center,
Sn4 at the next level. we refer to the fuzzy point as the fuzzy storm center.
 4. INCREMENTAL RELAXATION
ALGORITHM
Once a set of storms has been hypothesized for the
radar intensity image sequence, the correctness of these
denote the supporting and contradictory evidence, re-
spectively. Let ns and nc denote the number of sup-
porting disparities and the number of contradictory dis-
parities, respectively. The four quantities are reset to
zero at the start of each iteration.
storms can be veried by tracking them over time. Our
tracking algorithm is based on Krezeski's temporal re-  To apply the temporal consistency constraint,
laxation algorithm with property coherence [3]. We for each adjacent disparity
????! ??????d!t to d of the form
have selected the size of a fuzzy storm center as a prop- dt = Sj ?1Sj or dt = Sj +1 Sj +2 , we compute the
erty to be coherent over time. compatibility at the kth iteration (k > 0) between
Let Sk be an hypothesized fuzzy storm center in the the two disparities as:
kth image. A disparity represented by a fuzzy vector
????!
Sj Sj +1 is constructed from Sj to Sj +1 if the inmum
(
Ck d; dt )=  
of the fuzzy length of the fuzzy vector is less than a pk?1(d) + pk?1(dt)
w1C (d; dt) + w2
threshold, Td , which is set to a default value of 10 pix- 2
els; and concurrently, the two fuzzy storm centers have
compatible sizes. We dene a property function, fs ,
to measure the size-compatibility of two fuzzy storm
centers S1 = hc1 ; r1i and S2 = hc2 ; r2i as:
( we add pk?1(d) to Ec and increment nc by 1.
jr1 ?r2 j if r > 0 or r > 0,
1 ? max(
(
fs S1 ; S2 )= r1 ;r2 ) 1 2
1 otherwise.
A size-compatibility threshold, Tsc , is set to 0.5. Note
that if Tsc is set to 1, then a disparity will be con-
structed between two fuzzy storm centers only when
they have exactly the same size. On the other hand, if
Tsc is set to 0, then the size-compatibility criterion is
eectively removed. We measure the partial compati-
where w1 and w2 are normalized weights that sum
to 1. We currently use w1 = 0:4 and w2 = 0:6.
If Ck (d; dt) > Tk (we use Tk = 0:6), then we add
pk?1(d) to Es and increment ns by 1. Otherwise,
 To apply the spatial consistency constraint, for
each disparity ds which has the same head storm
or tail storm as d, if pk?1(d)  pk?1(ds), then
we add pk?1(d) to Es and increment ns by 1.
Otherwise, we add pk?1(d) to Ec and increment
nc by 1.
 The certainty of the disparity d at the kth itera-
bility between two adjacent disparities using a weighted tion is modied by:
sum of three components: length compatibility Cd , an- pk (d) =
gle compatibility C , and size compatibility Cs. The (  
overall compatibility function is dened as: 1 1 + ws Es ?wc Ec if Es 6= 0 or Ec 6= 0,
2 ws Es +wc Ec
0 otherwise;
C = wd Cd + w C + ws Cs;
where ws is the weight of the supporting evidence
where wd ; w and ws are normalized weights such that and is computed as ws = nsn+snc ; and wc is the
wd + w + ws = 1. We currently use wd = 0:2; w = 0:2 weight of the contradictory evidence and is com-
and ws = 0:6. Two adjacent disparities are connected puted as wc = nsn+cnc .
together if their compatibility value is greater than a
threshold:  The iterative process stops at the iteration
kth
????! ??????! when the certainty of each disparity has con-
C Sj Sj +1 ; Sj +1Sj +2 > Tc ; verged to the desired level of condence, say to n
decimal places (we use n = 6):
where Tc is currently 0.2. When all qualied adjacent
disparities have been linked together, the certainty of " = j pk (d) ? pk?1(d) j < 10?n;
each disparity is rened iteratively by relaxation on
the overall compatibility among ????! its adjacent dispari-
ties. Consider a disparity d = Sj Sj +1 . The initial
certainty of the disparity, denoted as p0(d), is set to
fs (Sj ; Sj +1 ). During each iteration, we apply both spa-
tial and temporal consistency constraints to compute
the supporting and contradictory evidence of the dis-
for each disparity d or the maximum number of
iterations has been reached: k ! Tk , where Tk is
currently set to 20.
Once the relaxation process has converged, we con-
struct a set of all longest tracks such that each dis-
parity using the compatibility values. Let Es and Ec parity has a nal certainty over a threshold Tp (we use
Figure 1: Image 9 of the 77-series. Tracks indicate that
the storms at the current image are moving from left
to right.
Tp = 0:85). We choose a subset of these tracks, with
the condition that no storm lies upon more than one
chosen track.
We have changed the implementation of the algo-
rithm from the processing of a complete image sequence
of known length to an incremental process. Given n
images the hypothesized storm disparities are relaxed.
When an (n + 1)th image is added, the relaxation is
Figure 2: Image 8 of the 18-series. Tracks indicate that
some storms are moving from left to right; and some
are moving from top to bottom.
to plot the tracks, with one or more applications of 3-
point averaging of the fuzzy storm centers before the B-
spline approximation is performed. By using the fuzzy
storm center concept to obtain the tracks and the B-
spline approximation with 3-point averaging for their
display, we can now obtain storm tracks that are long
and smooth and which closely match human perception
of a \motion picture" of a storm image sequence.
restarted using the results for the rst n images plus the
hypothesized storms of the (n + 1)th image. We have
observed empirically that the result is always the same
as if all n +1 images had been initially relaxed. The dif-
ference in computation speed between the two methods
6. REFERENCES
[1] D. Cheng. \Tracking fuzzy storm centers in Doppler
radar images". Master's thesis, The University of
is insignicant since relaxation converges within ten it-
erations in either mode. The incremental algorithm
allows us to view the current storm tracks as the data
become available.
5. EXPERIMENTAL RESULTS
Due to space limitations, we only show test results
of our algorithms on two images from two image se-
quences: the 77-series and the 18-series. The storms
are moving from left to right in the 77-series and the
results obtained from our algorithms agree with this
fact, as shown in Figure 1. Storm movements are more
complicated in the 18-series. Some of the storms are
moving from left to right, and some are moving from
top to bottom. Our algorithms are capable of track-
ing these storm movements, as shown in Figure 2. To
ensure the display of smooth tracks, we use B-splines
Western Ontario, Dept. of Computer Science, 1996.
[2] S. L. Horowitz and T. Pavlidis. \Picture segmenta-
tion by a tree traversal algorithm". Journal of the
ACM, 23(2):368{388, 1976.
[3] D. Krezeski, R. E. Mercer, J. L. Barron, P. Joe,
and H. Zhang. \Storm tracking in Doppler radar
images". In Proc. International Conf. on Image
Processing (ICIP), volume III, pages 226{230, 1994.
[4] A. Rosenfeld. \Fuzzy geometry: An overview". In
Proc. First IEEE Conf. Fuzzy Systems (San Diego),
pages 113{117, 1992.
[5] A. Rosenfeld. \Fuzzy plane geometry: Trian-
gles". Pattern Recognition Letters, 15(12):1261{
1264, 1994.
[6] L. A. Zadeh. \Fuzzy sets". Information and Con-
trol, 8:338{353, 1965.