A Spatial Data Structure Integrating GIS and

Simulation in a Marine Environment

CHRISTOPHER M. GOLD
ALFONSO R. CONDAL
Dept. des Sciences Godsiques et dc Tldtection. Universit Laval.
Ste-Foy. Qubec. Canada

Keywords geographic information systems. Voronoi diagrams. interpolation. navigation. simulation

Introduction: Marine Data Structures

Traditional geographic information system (GIS) data structures have always been
some- thing of a problem in a marine context. as so much of the available information is
collected in point form, while the spatial structuring available within a land-based GIS is
usually based on vector line segments or arcs, or else on a raster-based image
approach (or some hybrid system converting back and forth between the two). it would
thus be very desirable to be able to work within spatial data structures that could both
handle objects that were unconnected and provide the necessary basis of spatial
adjacency or proximity. People have no difficulty navigating around isolated features, so
why should a computer?
Additionally. in a marine environment even more than in a land-based one. objects
are likely to change location, and it would be nice if this could be handled in a dynamic

Received 23 February 1994: accepted 2 June 1994.

The funding for this research was made possible by the foundation of an Industrial
Research Chair in Geomatics at Laval University. jointly funded by the Natural Sciences and
Engineering Research Council of Canada and the Association de llndustrie Forestire du
QuCbec. The authors are very grateful to Mr. Denis Nadeau for providing the sea-surface
temperature example. and to Mr. Weiping Yang and Mr. Franois Anton for considerable help
with the programming and diagram preparation.
Address correspondence to Dr. Christopher M. Gold. Dept. des Sciences Godsiques et de
Tldtection. Universit Laval. Ste-Foy. Qubec. Canada G 1 K
7P4.

s yst em . Finally, w h i l e in both terrestrial and marine systems we may be

representing both discrete o b j e cts and fields that vary continuously over

the map. in a marine context these fields are less often classified into a discrete polygon
tiling.
Our recent work has been concerned with the development of spatial data
structures with some of the desired properties mentioned above, as a potential
replacement for the usual terrestrial techniques. This matters because SO often the
operations we w ish to perform are severely restricted by the available tool set and,
even worse, we attempt to pose the problem in available. but inappropriate terms. The
spatial data structures used are based on the concepts of space used for the software
development and may not be appropriate elsewhere. Finally, we have questions of
theoretical computational efficiency that are relevant when we are concerned with timevarying situations-because the objects have changed, or because the operator wishes to
interact with the system in a timely (and maybe urgent!) fashion.
The techniques developed at Laval University are based on the idea of Voronoi
tesselations, used as a dynamic spatial data structure. These are described in various
articles (e.g.. Gold. 1990a. 199Oc. 1991. 1992 a . 1992b), and the purpose of this
article is to evaluate them for USC in a marine GIS. W e consider that a reasonable
wish list for a marine GIS would include the ability to handle nonconnected and
connected objects, as well as field-type data. that would be able to vary their relative
positions and values over time. At any appropriate moment the spatial structure
should respond to queries concerning values or spatial relationships, e.g., for navigation
purposes. Implementation is not feasible with static polygon structures. but it is
possible using dynamic Voronoi tesselations.

Principles of Voronoi Methods

Static point-Voronoi tesselations are well known in the literature, and algorithms have
been used for many years (see Aurenhammer, 199 1. for a summary). Less well known
are dynamic algorithms that allow point creation, deletion. and movement. and also
Voronoi tesselations of more complex objects- typically line segments as well as points.
These have been the subject of considerable research at Laval University, and they are
being used in terrestrial GIS applications, especially those concerned with managing the
history of map changes. We have also been involved in the evaluation of Voronoi fieldmodeling techniques in the marine context. This article will explore the state of current
dynamic Voronoi techniques, and their potential application in marine applications.
The basic point Voronoi tesselation in the Euclidean plane is a tiling of the mapping
plane such that each convex tile contains that portion of the plane closest to any one of
an arbitrarily distributed set of data points (Figure 1a). This figure shows the Voronoi
tesselation for each data point. as well as the dual Delaunay triangulation. [It should be
noted that one is not necessarily restricted to the plane: Augenbaum and Peskin ( 1985)
have developed methods for the sphere. and Tanemura et al. (1983). among others. have
developed methods for three dimensions. A large variety of algorithms exist to generate
this tesselation, from the simple point-by-point insertion methods (e.g.. Green & Sibson.
1978). to more elaborate sweep-line (e.g.. Fortune. 1987) and divide-and-conquer techniques (e.g.. Guibas & Stolfi. 1985). There are no particular resolution constraints.
except those of preserving large. high-precision coordinate values. and the limitations
of the precision of geometric calculations using finite-precision arithmetic. that are
standard for all vector GIS operations. Indeed, as they are intended to express local
spatial relation- ships, as far as possible minimizing the USC of global coordinates, they
are excellent for

preserving information
at
widely
differing densities.
They
have
traditionally been used for rainfall analysis, tree-crown modeling, and many
types of urban nearest-service anal- ysis [see Okabe et al. ( 1992) for many
examples]. It has also been shown that for many applications the static
Voronoi tcssclation may be used as a superset of the traditional topological
data structure (Gold et al.. 1994). The tesselation combines many o f the

properties of both field and object views of space, as on the one hand the
adjacency of Voronoi cells give a valuable definition of the proximity of
unconnected objects, and on the other hand the cells themselves may be
assigned the attributes (e.g.. rainfall. elevation)
of the observation at each data point. While this gives a discontinuous
model of the
techniques
continuous field being modeled, various area-stealing
(Sibson. I981 : Wat- son & Philip. 1987: Gold. 1989. I99Ob) can be us e d
to estimate intermediate
values on the surface, based on a form of the
simple weighted-average interpolation that guarantees surface continuity
and, if necessary, slope.
Algorithms for generating the simple point-Voronoi tesselation have
improved significantly in theoretical cflicicncy in recent years. Where the whole
structure may be constructed at once, randomized incremental algorithms
such as those of Boissonnat and Teillaud (1986). Guibas et al. ( 1992). and
Teillaud (1993) can create these diagrams in expected time
log n),
However, as a major motivation for this work
which is optimal.
concerned the maintaining of a map when one or more objects are
moving, an alternative technique was developed that maintained the
Voronoi spatial relationships while map objects were being inserted,
deleted, or displaced. This is achieved by deter- mining when the Voronoi
cell of a moving point gains or loses a neighboring cell, moving the point to
that location, and locally updating the topological structure accordingly.
For the case of all points moving simultaneously, Guibas et al. (1991) give
a rather complex theoretical efficiency based on Davenport-Schinzel
sequences, but in the case of one point being inserted at a time by
splitting it from the nearest preexisting point and then moving it to its
destination (see below), the expected time efficiency should again
approximate

Dynamic
Structures

log n).

Spatial

Data

The development of a dynamic algorithm from the static-point Voronoi

tesselation. per- mitting the points to be moved around without
destroying the bubble structure of the cells. allowed greater freedom
in the development of spatial data structures for time- varying maps.
While still generating the Voronoi tesselation. the algorithms are significantly different. with the emphasis on detecting when the Voronoi cell of
a moving point will change its set of neighbors. This has been described
for point sets by Bajaj and Bouma ( 1990) and by Roos ( 1990, l993),
and for points and line segments by Gold (1990a). In two dimensions
this involves switching the diagonals of adjacent Delaunay triangle pairs
as a point moves, in order to preserve the empty circumcircle criterion for
each triangle. Given methods to split an old generating point and its
cell into two. generating a new one capable of being moved to its desired

destination. and the inverse operation of merging two points and cells, a
dynamic data structure was obtained that could be used as a
replacement for the static point-tesselation. Apart from the ability both to
add and to delete points. however, it permitted points to be moved about
the map in sequence. while preserving their spatial relationships- i.e..
their Voronoi neighbors--at all times. This has obvious applications for
navigation, where one or more of the points represent ships. etc. A
further development includes the area-stealing model within this
navigation process, allowing interpolation of the seafloor, etc., from the
neighboring

Figure I. ((11 The Voronoi diagram of a set o f points. (b) The point Voronoi diagram o f
plus a line segment.

part (a).

Applications of the Dynamic Voronoi Method

W e have briefly reviewed the basic concepts o f dynamic Voronoi

data

structures, as described in the literature. To support the approach for use in a

marine GIS, it is perhaps best to illustrate some o f the advantages by using
several applications as examples. The key point for us is that while each
of the applications may be developed separartely, to our knowledge no
other underlying, methodology is capable of being applied to all o f them----that
is,
using the same basic data structure
and set
of
spatial
operations. This strongly suggests an underiyin g generality that may be
used for rapid development ot further appl icat ions as needed .

Interpolation of Precise, Arbitrarily Distributed Point Data

The area-stealing (Gold. 1989) or natural neighbor (Sibson. 198 I : W
a t s o n & Philip.
1987) method o f interpolation is based on one idea: If the query point (at
the
Iocation where it is desired to make an interpolation ecstimate) is
inserred into the Voronoi tesse- lation formed by the re a l data points,
then it will reduce (steal) s o m e
area from the adjacent Voronoi cells.
These areas are ideally suited as weights to be used for calculating the
weighted average of the heights of the neighboring data points. The areas
have zero value when a data point c a s e s to be a neighbor to the query
point (i-e.. their Voronoi cells no longer touch); and when the query point
coincides with a data point, all o f the stolen area will be taken from that
particular data points Voronoi cell. This produces one o f the few spatial
models that can guarantee that an interpolated surface both passes
through all data points and has no discontinuities or other artifacts it is not
dependent
on the data distribution-in dee d.
it has been used successfully
with extremely anisotropic data (Figure 2. from Caulfield et al. (1984)).
Sibson (I 98 I
showed that the interpolated surface resulting frorn
this is continuous within the convex hull o f the data, and has continuous
slope, except at the data points themselves. As described in Gold (1989), a
simple transformation of the weightings may produce a continuous surface
with continuous slopes everywhere, the slopes being zero at the data
points. in order to achieve more plausible slopes at the data points (e.g.,

the data are on an inclined plane), a slope estimate may be made at each
data point. and the weightings applied to this data-point function (evaluated
at the query-point location).
when

In geological applications. for example. such a modification is required, but in many

cases bathymetric modelin g is legally required not to be extrapolated beyond the local
data range- in which case slopes would be left as zero at data points (in Figure 2, for
example). As shown in Figure 3 (usin g terrestrial seismic data. with slope estimates at
data points), the method produces intuitively reasonable results even for the highly anisotropic data.
Oceanographic data are generally obtained along transects that follow different directions, generally with dense data along the transects by comparison with the transect
spacing. To obtain a synoptic view of these types of data requires a suitable interpolation
method. To our knowledge. Voronoi methods have not been used for oceanographic
work, except in the case of Figure 2. although simple Voronoi functions exist in various
commercial GIS. and Okabe et al. (1992) discuss many applications in geography. We
therefore wished to include an evaluation of the interpolation technique within this review
of the Voronoi data structure for marine GIS. as such contour mapping is a basic requirement of such a system.

Figure 2. Bathymetry around a caisson. (From Caulfield, et al., 1984.)

Figure 3. Contour map of subsurface seismic data.

To achieve this goal we took a region in the Gulf of St. Lawrence (Figure 4a) where
sea surface temperature measurements had been obtained both by ship and by satellite.
The methodology for calculating the sea surface temperatures from the satellite data was
described by Condal and Le ( 1983). and the study is described in detail by Nadeau (
1994).
Figure 4c
shows the
Figure 4b shows the sea surface temperature image.
interpolated
surface using the Voronoi method at the (very few) ship locations. It is reasonable to say
that the Voronoi method produced a surface form comparable to the original image.
except where additional features. such as islands. were completely missed by the sparse
sampling. The surface form was also simple. in that it possessed no artifacts not attributable to the data-thus it appeal-cd to do as good a job as was feasible with the little
information available. This was important. as it gave a certain credibility to the interpolation of subsurface temperatures obtained from the sampling sites only-as obviously
no complete image would be available. Some experiments were made with the Kriging

5 0 0

Lat.

670

Long.

255

115

105

95

76

Figure 4. (a) index map of sea temperature study area. (b) Satellite s ea surface temperature
map. (c ) Temperature map interpolated from sample locations.

method. but with less success; the surface form was more irregular. This is not surprising,
as Kriging itself makes no statement about the exclusion of distant points from the
weighted-average calculation. In practice, however. some set of neighbors is selected and
the rest are rejected, and thus surface discontinuities are still possible when any data point
is accepted or rejected while its contribution is nonzero. Lam (1983) quotes Delfinier
(1976) as saying that choice of neighborhood will affect the continuity properties of the
estimates, which may lead to serious bias in interpretation. If the change of data points
from one neighborhood to the next is too abrupt, there may be discontinuities even though
the actual phenomenon is continuous.
While no full statistical techniques are available to test the validity of the results.
simple error checking for spurious values is easily performed by temporarily removing
each data point in turn. estimating the mean and standard deviation of the estimates used
in the weighted average. and comparing them with the true observation (see Gold. 1980).
Due to the dynamic nature of the Voronoi system. this approach may be used in an onship surveying operation (if the sampling rate is not too high), both to compare a new
observation with those previously accepted. and to detect regions of less reliability for
further sampling.

Topology of Complex Objects (Points Plus Line

Segments)
As described by Gold (1990a. 1992b). it is possible to maintain the Voronoi topological
structure while a point is moving across the map. A point is created by splitting it from
an existing point and moving it to the desired location. and a point is deleted by reversing
this operation. Line segments. which are the loci of moving points. may be generated in
the same way. by growing the segment behind the moving point. Line segments may be
deleted in the reverse process. The result is a dynamic Voronoi topological structure for
arbitrary points and line segments that can be extended to completed polygon sets with
islands. etc. (Figure 5) without any special treatment. if the line segments are assigned
height values (e.g.. bathymetric contours, ridge or valley lines. coastlines). then the
previously described area-stealing interpolation may be performed using these data objects. For line segments. slopes may be defined perpendicular to the line segment. Slope
or surface discontinuities may be defined across line segments, if required. As the Voronoi
tesselation is undefined for two objects at the same location. overlapping lines.
duplicate points. and zero-length line segments must be avoided.

Automatic Buffer Zone/Coastal

Beacon Control Zones

Boundary

Generation,

and

It is a standard operation in a terrestrial GIS to generate buffer zones around lakes, rivers.
etc. to represent environmentally protected areas. for example. The same function is used
in the marine context for national marine limits. and for restricted access areas. The
standard GIS treats this as a batch polygon overlay calculation. What is less understood
is that a buffer zone is a distance to the closest portion of the target set type of
operation, and as such may be resolved directly using the Voronoi tesselation, which by
definition is the partition of the map space into proximal regions associated with each
map object (e.g., the individual points or line segments used to construct more complex
features). Thus the generation of a buffer zone consists merely of the evaluation of each
Voronoi cell, and the drawing of a circular arc/straight linewithin the cell associated with
a point/line segment. This may be done dynamically, as the Voronoi tesselation is maintained dynamically, and may readily be applied to moving, not merely static. objects.

Figure 5. Voronoi diagram al points and line segments.

Fi gure 6 shows the relation between the buffer zone and the Voronoi tesselation for a
simple collection of map objects. An even simpler proximal query concerns rhc generation
of control zones appropriate to various navigation beacons (in the absence of obstacles).
as these would be merely the Voronoi cells of the set of beacons. constructed on a separate
layer. Figure 6 also shows the relationship between the nearest object (point or half-line)
and the point-in-polygon problem. P and Q are readily determined to be inside the shaded
polygon. and R and S outside. merely by identifying the Voronoi cell containing them
(i.e.. the nearest object).

Dynamic
Maps

Modification of Topologically Structured

Because the Voronoi topology may be maintained during,the incremental map-building

process (and indeed the persistence of a complete topology is necessary to the navigation
process required for map building). a map may be built in an incremental fashion by
adding points and line segments in the digitizing process. It is not ncccssary. howcvcr.
that the points and line segments form a complete space-filling polygon set as with
traditional systems--the Voronoi cells themselves form the complete tiling necessary for
a continuous model of space. Thus the map may be edited at some later date without any
necessity to rebuild the whole structure. Points and line segments may be added and
deleted at will. and methods exist to detect collisions between the moving pen and the
existing map- in which case an intersection point may be generated. To take an example

from forestry, a forest stand map may be created (Figure 7a). At a later date a clearcut
may extract parts of several adjacent stands. The map updating process consists of adding
the lines forming the cut boundaries-the boundaries of the box in Figure 7b. Here these
edges have been added to the map and, where necessary, intersections generated with
the original boundaries. Each of the portions of the old and new boundaries
generates a Voronoi cell. Since the clear-cutting has eliminated the interior forest stand
boundaries. these may thus be removed from the map (Figure 7c). Ail of these are local

operations. which do not require the modification of the topological structure outside
the immediate area of interest-just the addition or removal of a Voronoi cell, and
the consequent modification of its immediate neighbors.
Because the digitizing paradigm is that of a moving point (the pen or cursor) interacting with its Voronoi neighbors. new and more user-friendly digitizing methods can be
developed (Gold. 1992b). These include interaction with the topology while the digitizing
is in progress. unlike the older batch construction from spaghetti files. This permits the
selection of objects to snap to before the action is performed. based on the
identification

Figure 6. Generating a buffer zone from the Voronoi diagram.

-2-24

Figure 7. (a) Schematic forest polygon map. (b) Cut boundary added to part ( a ) (c)
Interior boundaries deleted from part (b).

One interesting application of the Voronoi method to map digitizing is

found in Gold ct al. (1994). using a traditional static point-Voronoi
tesselation for the rapid digitizing of irregular-polygon data. Points at-e
digitized around the interior of each polygon (a f a s t intuitive operation)
and the point Voronoi tesselation constructed. On the basis of the spatial
structure, a single scan through the network suffices to extract all arcs
separating points with
differing
polygon
labels. This indicates that
traditional topology, in t h i s case at Ieast, may be considered to be a subset
of the Voronoi tesselation. Large time savings have been reported for
forest map digitizing (Nantel. 1993).

Because
an

the

creation

and

modilication

of

a map has been implemented

as

incremental

operation rather than as a batch polygon-buildin g operation, it is po ssi ble

to pr ese rv e a
log file of all commands executed by the program, along, with their date
T h i s may be played back at any time to re-create the map, stoppin g the
topological
construction at
any desired time. Thus there is no need to
preserve a copy of the map prior to updating, etc. _ with all the difticulties
in matching different snapshots of the map whenever some his- torical
comparison is required. All versions of the map may bc kept in one file and
any snapshot
may be
reconstructed as required. This also greatly
simplifies the problem of determining changes. and preserving an audit
trail of the persistence of some object. for legal or stock-taking purposes.
or to track changes. Indeed, it appears reasonable to distribute merely l o g
,files or their updates, rather than complete new charts, w h e n information is to be transferred. This is described in outline in Gold ( 1993).

Robot
Objects

Navigation

between

The map-building process involves the control of a moving point within

the Voronoi tesselation representing the current map objects and their
spatial relationships. This mov- ing point cot-responds to the cursor or pen.
When the pen attempts to cross a preexisting line, the Voronoi structure is
capable of detecting this in advance (set Gold. 1992b) and generating an
intersection. The same mechanism may be employed for simulated navigation among obstacles in the map as. for example. a boat may be directed
through a marked channel. Because Voronoi cell adjacencies define local
neighbors, the small number of neighborin g objects to the moving boat
may be tested for possible collision- for example. boat A in Figure 8.
This is also closely related to two other techniques: buffer zone
generation, as described previously; and the use of the Voronoi tesselation
as the medial axis transform of a polygon (as in Figure 6) or some other set

of objects. The medial axis transform. or skeleton, defines the path

farthest from the adjacent object (Blum. 1967). Ognicwicz and Ilg (1990),
and others, have used this in various object- recognition problems, where
they refer to the endoskeleton and exoskeleton. Thus the same data
structure used for map or chart generation may be used to select the best
route or to detect potential collisions on the course specified (using the
local neighbors).

C . M Gold a nd A. R . C on

226
da l

Robot Navigation within a Bathymetric Channel

The collision-detection method just described. using the moving-point model. may be
modified based on the area-stealing interpolation described earlier. Since the moving point
has a Voronoi ceil. it has neighbors. and it has stolen area from each of them. It thus
can be used. as before. to estimate the elevation (or depth) at that location-for example,
boat B in Figure 8. The slope may also be determined. and thus the simulated boat
may be navigated to the deepest channel or. if preferable. controlled by the landmarks
(map objects) of the previous section but with some warning tolerance of available
depth. (These could be combined with tidal variations if so desired.) Using this
structure. contoured bathymetric models might not be necessary. as individual depths
would be calculated as required.

Fluid
Simulation

Flow

While still a subject of ongoing,research, there are some types of fluid flow modeling
that may be used with the Voronoi tcchniques. These are known as the free Lagrange
methods

Figure 8. Schematic navigation system.

References
Augenbaum. J. M.. and C. S. Peskin. 1 985. On the construction of the Voronoi mesh
o n 3 sphere.
J. Comp. Phys.5 9 177- 192.
Aurenhammer. F. IYY I . Voronoi diagrams--a survey of a fundamental geometric
data structure.
ACM Comp. Surv. 23:345--405.
Bajaj. C. L.. and W. J. Bouma. I 9 9 0 Dynamic Voronoi diagrams and Delaunay
triangulation.\.
PI-W. Second C a na d ia n Conf. in Computational G e o m e t r y Ottawa. August. pp. 273277. Blum. H. 1967. A transformation for extracting new descriptors of shape.
In W. Wharcn-Dunn
(cd.). Models for t h e perception of speech and visual form Cambridge. MA: MIT
Press. pp.
153-171.
Boissonat. J.-D.. and M. Teillaud 1986. A hierarchical representation of objects:
The Delaunay tree. P r o c . Second ACM S ym p . on Computational Geometry
Yorktown Heights. NY: ACM. pp. 260-268.
Caulfield, D. D., D. J. Kcnway. and C. M. Gold. 1984. Stereo side-scan as a
complement to echosounding for high resolution bathymetric studies. Report, USN/SEG
Symposium on Three
Dimensional Data, Diamond Head. MS.
Condal. A. R.. and H. V. Le. 1 983. Computer monitoring of sea-ice temperature by
use of NOAA
satellite data. P r o c Eighth Canadian S y mp . o n Remote Sensing Montreal. May.
pp. I45150.
Delfinier. P. 1976. Linear estimation of non-slationary spatial phenomena. In M.
Guarascio. M.
David. and C. Huijbregts (eds. ). A d van c ed geostatistics in the mining industrie
Dordrecht . Holland: Rcidcl. pp. 49-68.
Fortune. S. 1987. A sweepline algorithm for Voronoi diagrams. Algorithmica 2: 153174.

Fritts M. J.. W. P. Crowley. and H. E. Trease. 1985. The free-Lagrange m e t h o d

Lecture notes in physics. vol. 238. New York: Springer-Verlag.