See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/222463571

A New Procedure for Gridding Elevation and Stream Line Data with Automatic
Removal of Spurious Pits

Article in Journal of Hydrology · April 1989

DOI: 10.1016/0022-1694(89)90073-5

CITATIONS READS
1,470 2,543

1 author:

M.F. Hutchinson
Australian National University
164 PUBLICATIONS 14,670 CITATIONS

SEE PROFILE

All content following this page was uploaded by M.F. Hutchinson on 19 April 2018.

The user has requested enhancement of the downloaded file.

Journal of Hydrology, 106 (1989) 211-27,2 211
Elsevier Science Publishers B.V., Amsterdam m Printed in The Netherlands

[1]

A N E W P R O C E D U R E FOR GRIDDING E L E V A T I O N A N D STREAM

LINE D A T A WITH AUTOMATIC REMOVAL OF S P U R I O U S PITS

M.F. HUTCHINSON*
CSIRO Division of Water Resources, G.P.O. Box 1666, Canberra, A.C.T. 2601 (Australia)
(Received March 28, 1988; accepted after revision July 1, 1988)

ABSTRACT

Hutchinson, M.F., 1989. A new procedure for gridding elevation and stream line data with
automatic removal of spurious pits. J. Hydrol., 106: 211-232.

A morphological approach to the interpolation of regular grid digital elevation models (DEMs)
from surface specific elevation data points and selected stream lines is described. The approach has
given rise to a computationally efficient interpolation procedure which couples the minimization
of a terrain specific roughness penalty with an automatic drainage enforcement algorithm. The
drainage enforcement algorithm removes spurious sinks or pits yielding DEMs which may be used
to advantage in hydrological process studies. The drainage enforcement algor~d,,,, ka= :!Jo been
found to significantly increase the accuracy of DEMs interpolated from sparse, but well chosen,
surface specific elevation data. Moreover, it facilitates the detection of errors in elevation data
that would not be detected by more conventional statistical means ~ud forms a sound physical basis
for cartographic generalization.

INTRODUCTION

Regular grid or raster digital elevation models (DEMs) have become the
basis for recent approaches to process raodelling of the earth's surface. Of
prime interest is the use of DEMs in hydrological modelling, as embodied in the
Syst~me Hydrologique Europ~en (Abbot et al., 1986a, b; Bathurst, 1986) and the
Deterministic Site Model of the Braunschweig Research Group (Bork and
Rohdenburg, 1986; Rohdenburg et al., 1986). At suitable levels of generalization
DEMs also have a major role to play in geographic information systems (Evans,
1980; Berry, 1985; Wiltshire et al., 1986), particularly in the modelling of erosion
(Knisel, 1980), in the classification of landforms and soils (Speight, 1974;
Heerdegen and Beran, 1982; Pennock et al., 1987), in the integration of
biophysical data with remotely sensed satellite data (Shelton and Estes, !981;
Jupp et al., 1986) and in the modelling of mesoscale and macroscale climatic
phenomena (Tesche and Bergstrom, 1978; Hutchinson and Bischof, 1983;

*Present address: Centre for Resource and Environmental Studies, The Australian National
University, G.P.O. Box 4, Canberra, A.C.T. 2601 (Australia)

0022-1694/89/$03.50 © 1989 Elsevier Science Publishers B.V.

212

Weiringa, 1986). They have even been useful at quite coarse levels of
generalization in detecting geological structures of significance for mineral
exploration (Harrington et al., 1982).
Raster DEMs can be calculated directly from stereophoto maps when these
are available (Kelly et al., 1977), or more recently from satellite imagery
(Konecny et al., 1987), but there remains a significant role for the interpolation
of DEMs fl'om scattered point elevation data, perhaps accompanied by stream
line data, particularly when the point data include surface specific points such
as peaks, pits, saddles and selected points on stream lines and ridge lines. These
data can be obtained by ground survey at minimal cost for the small catchment
areas often used in hydrologic studies and can also be digitized at moderate
cost for larger areas from existing topographic maps. At national or even
continental scales, existing data banks which hold trigonometric points
consisting of the principal peaks, as well as scattered ground control points
and digitized stream line data, may be sufficient, especially if the interpolation
technique is of high quality, to generate broad scale DEMs of sufficient
accuracy to be useful in natural resource assessment and global climatic
modelling (Trezise and Hutchinson, 1986; Hutchinson and Dowling, 1989).
This paper describes a morphological approach to the interpolation of
digital terrain data which attempts to take into acount the special nature of
terrain surfaces, and the surface specific points that can be used to sample
terrain, as well as potential hydrological applications of the interpolated
elevntinn grid It has given rise to a procedure, first outlined by Hutchinson
(1986), which can efficiently calculate raster DEMs witn sensible drainage
characteristics from large numbers of irregularly spaced elevation data points
and stream line data. The principal innovation of the procedure is a drainage
enforcement algorithm which automatically removes spurious sinks or pits
from the fitted grid, in recognition of the observation that sinks are rare in
nature tMark, 1984; Band, 1986; Goodchild and Mark, 1987). The approach
extends more conventional statistical approaches implicit in existing
assessments of optimum sampling strategies for digital terrain modelling
(Ayeni, 1982) and provides an alternative to an existing physically sound
approach to digital terrain modelling using phenomenon-based data structures
(Mark, 1979).
The drainage enforcement algorithm is coupled with an iterative finite
difference interpolation technique which is based on minimizing a terrain
specific, rotation invariant roughness penalty. This technique has its origins in
the minimum curvature interpolation method of Briggs (1974) but is more
computationally efficient. The roughness penalty has been tailored to yield
good results in conjunction with the drainage enforcement algorithm while
maintaining artifact-free behaviour of the fitted surface away from data points.
The finite difference nature of the technique facilitates the monitoring of the
surface characteristics of the fitted grid and permits the imposition of simple-
ordered chain constraints. These constraints are the means by which the
 213

procedure enforces drainage and incorporates the ordered descent and

breakline conditions implicit in stream line data.
The drainage enforcement algorithm eliminates one of the principal
weaknesses of elevation grids produced by general purpose interpolation
techniques which has limited their usefulness in hydrologic applications, par-
ticularly those which rely on the automatic calculation of channels, ridges and
basin catchment areas (Peucker and Douglas, 1975; Mark, 1984; Band, 1986;
Palacios-V~lez and Cuevas-Renaud, 1986). If spurious sinks are not removed
from the DEM then programs which delineate basins and simulate drainage
networks must take special action in order to produce acceptable results
(Marks et al., 1984; Yuan and Vanderpool, 1986). However, eliminating sinks
from an existing DEM is difficult. Manual methods can be time consuming and
subject to individual operator error. On the other hand, existing automatic
methods for digital filtering or smoothing of DEMs can remove most, but not
all, spurious sinks only at the expense of simultaneously oversmoothing well
defined surface features (Mark, 1984). Heerdegen and Beran (1982) have
similarly noted the tendency of one interpolation technique to produce grids
with nonsensical surface properties if one attempts to choose the grid spacing
fine enough to capture known complexities in the data. The technique
presented here in fact provides the "judicious filtering" called for by Band
(1986).
Moreover, the imposed drainage condition has been found in practice to be
a powerf!,~] condition which can significantly increase the accuracy, especially
in terms of their ~trainage propertms, of digital elevation models interpolated
from sparse, but ~::el] chosen, surface-specific data sets. The size of such data
sets can be at least an order of magnitude smaller than the number of points
normally requireG to describe elevation using digital contours. This can
minimize the expeT,se of obtaining hydrologically sound digital elevation
models in terms of the capture, correction and storage of primary elevation
data.
An important feature of the method is its sensitivity to data errors. The
drainage enforcement algorithm acts conservatively when attempting to
remove sinks by not imposing drainage conditions which would contradict the
elevation data by more than a user supplied elevation tolerance. Consequently,
errors in both elevation and position of input elevation data can often be
indicated by sinks in the final fitted grid and thence easily corrected. This is
useful when processing large data sets which almost inevitably contain errors.
The procedure routinely detects errors in this fashion which are small enough
to remain undetected as outliers by conventional statistical methods.
The interpolation procedure and its accompanying drainage enforcement
algorithm are described in detail below. A much analysed data set of Davis
(1973) is used to illustrate several points and to provide a useful comparison
with other interpolation methods. The paper concludes with an application to
a larger data set.
It should be noted that the procedure has been designed principally to
214

interpolate scattered surface-specific point elevation data and stream line data.
It can be applied to contour elevation data with acceptable results, but it is
possible to capitalize on the special nature of contour data to improve the
quality of interpolation. The extension of the present technique to the inter-
polation of contour elevation data will be the subject of a separate study.

THE INTERPOLATION ALGORITHM

The relative merits of various general purpose interpolation methods have

been discussed by a number of authors including Ripley (1981), Sibson (1981),
Hutchinson (1984), Laslett et al. (1987) and others. Most existing approaches to
the interpolation of irregularly distributed data can be broadly classified as
having either global or local character.
The two most popular global methods are the thin plate splines of Duchon
(1976) as taken up by Wahba and Wendelberger (1980) and others and the
method of Kriging advocated by Matheron (1965). These techniques offer
elegant solutions to the general interpolation problem. The interpolated
surface normally depends cn the data in a rotation invariant manner and can
be made to have reasonable continuity properties. Their chief disadvantage is
a computational cost proportional to n 3 where n is the number of data points,
making them prohibitively expensive for very large data sets.
Local methods are usually based on partitioning the region containing the
data points into small elements and fitting simple functions on each element so
that the functions are continuously differentiable across the boundaries of
adjoining elements. This significantly reduces the computational burden at the
expense of somewhat arbitrary restrictions on the form of the fitted functions.
Such techniques are also sensitive to the positions of data points, particularly
when the data points are very irregularly spaced, and spurious edge effects can
be generated. Recently developed local methods of Sibson (1981) and Watson
and Philip (1985), which employ Dirichlet tessellations of the plane, appear to
overcome at least some of these difficulties.
The iterative finite difference interpolation method adopted here has been
designed to have the efficiency of a local method without sacrificing the
continuity and rotation invariance of high quality global methods. Most impor-
tantly, because the fitted grid values are available at every stage during the
iteration, it is relatively straightforward to monitor the drainage characteris-
tics of the fitted grid and to impose appropriate ordered chain constraints to
enforce drainage. This is not easily achieved for either the global or local
methods described above. The interpolation problem is solved by minimizing a
discretized rotation invariant roughness penalty which is defined in terms of
first and second order partial derivatives of the fitted function. Similar
approaches have been adopted by Briggs (1974), Swain (1976) and Testud and
Chong (1983). An application of an early version of this technique to the
interpolation of bathymetric data occurs in Torgersen et al. (1983).
The iteration technique employs a simple nested grid strategy which
 215

calculates grids at successively finer resolutions, starting from an initial

coarse grid, and successively halving the grid spacing until the final user
specified grid resolution is obtained. For each grid resolution, the data points
ale simply allocated to the nearest grid point and values at grid points not
occupied by data points are calculated by Gauss-Seidel iteration with over-
relaxation (Young, 1971; Golub and Van Loan, 1983) subject to the specified
roughness penalty and ordered chain constraints. The iteration matrix used in
the Gauss-Seidel iteration depends on the roughness penalty and is easily seen
to be positive definite and symmetric for the roughness penalties considered
below, thereby guaranteeing convergence. Starting values for the first coarse
grid resolution are calculated from the average height of all data points while
starting values for each successive finer grid are linearly interpolated from the
preceding coarser grid. An empirically determined overrelaxation parameter of
1.6 has been found to give useful acceleration of convergence for the roughness
penalties considered and iteration terminates for each grid resolution when the
user specified maximum number of iterations (normally 25) has been reached.
The computational cost of the technique is then optimal in the sense that it is
essentially proportional to the number of interpolated grid points.

A TERRAIN SPECIFIC ROUGHNESS PENALTY

The roughness penalty which defines the interpolating function, f, is based

on two nonnegative, rotation invariant functionals. These are defined in terms
of first and second order partial derivatives of f by:

J,q) = j'(f + g)dxdy

and:

J2 (D f(]~x + 2~y + gy) dx dy

where the range of integration is the region occupied by the fitted grid.
Minimizing Jl (f) in its discretized form over all suitably continuous inter-
polating functions f gives rise to discretized minimum potential interpolation
while minimizing J2 (f) gives rise to the minimum curvature interpolation of
thin plate splines in the discrete form first devised by Briggs (1974). Testud and
Chong (1983) have discussed these roughness penalties in the context of data
smoothing.
Minimizing J2 (D is a good general purpose strategy which can give visually
pleasing results as seen in Fig. 1 where it has been applied to the elevation data
in table 6.4 of Davis (1973, p. 316). In terms of freedom from spurious surface
features, the result displayed in Fig. I is clearly superior to the results of four
different interpolation methods applied to the same data by Lodwick (1982) and
compares favourably on the same grounds with the results of the preferred
version of Kriging in fig. 4.17 of Ripley (1981, p. 64) and the local interpolation
method in fig. 5 of Watson and Philip (1985, p. 322). However, minimum
curvature interpolation of terrain is not ideal. Its tendency to maintain trends
216

~ _ / :ooo / 730 / ,o,

830 820

y 675

/ .~. N 861!

.o
\ 9,0 ~ 9,~ ?
860

Fig. 1. Discretized minimum curvature interpolation of the point elevation data of Davis (1973,
table 6.4). Elevations in feet.

away from data points can generate spurious overshoot and undershoot in
regions containing closely spaced data points with large variations in
elevation. Though not in evidence here, this phenomenon is well documented,
particularly in the context of univariate minimum curvature interpolation (De
Boor, 1978; Fritsch and Carlson, 1980). Moreover, a closer look at Fig. 1 reveals
more subtle shortcomings. The real pattern of stream lines and ridges, as
illustrated in Fig. 8 and in fig. 6.10 of Davis (1973, p. 323), is not accurately
represented in Fig. 1 despite the lack of any spurious sinks.
Consideration of the statistical nature of natural terrain by Mandelbrot
(1982), using the theory of fractals (see also Goodchild and Mark, 1987), has led
Frederiksen et al. (1985) to suggest that a more appropriate roughness penalty
should lie somewhere between J~ (f) and J2 (f). This would allow the fitted
 217

surface to follow more closely the sharp changes in terrain often associated
with ridges and streams.
The result of minimizing Jl (t9 is shown in Fig. 2. This discretized minimum
potential surface is less sensitive to local trends in the data, displaying
characteristically flatter appearance away from the data points, of which most
are represented as sharp local maxima and minima. It should be noted that
minimizing J~(f) is only well defined in its discretized form, since it can be
shown that the corresponding continuous interpolation problem does not have
a unique solution amongst continuous functions. Moreover, the discretized
solution is sensitive to the grid spacing, with local maxima and minima
becoming increasingly sharp as the grid spacing decreases. While this is un-
desirable in general, the important feature here is that the local minima occur
at data points on stream lines and the local maxima occur at data points on
ridges and peaks.
An empirically determined compromise between the minimum curvature
and minimum potential roughness penalties is illustrated in Fig. 3 which has
been obtained by minimizing the discretized form of:
J (]) = 0.5 h-2 J~ (f) + J~ (f)
over all suitably continuous interpolating functions f where h is the grid
spacing. The effect of this roughness penalty is to modify the usual finite
difference recurrence formula for minimizing J2(f) away from data points (see
Briggs, 1974; Testud and Chong, 1983) in a way which is independent of the grid
spacing h. It has been confirmed empirically that the resulting interpolated
surface is also insensitive to the grid spacing. This surface maintains trends
away from data points in a similar fashion to minimum curvature interpolation
but still identifies most of the points on stream lines as sinks. It also identifies
breaks in slope corresponding to data points on ridges more sharply. If the
sinks, which have been identified in Fig. 3 by circles, were to be linked in a
sensible fashion to form stream lines, as indicated by the dashed lines in Fig.
3, then the drainage pattern of this landscape would have been effectively
recovered from a very small data set. Moreover the surface specific nature of
the data points would have been automatically recovered without the need for
explicitly identifying their nature in the data. This is precisely what is
achieved by the drainage enforcement algorithm described below.

THE D R A I N A G E E N F O R C E M E N T A L G O R I T H M

The drainage enforcement algorithm attempts to remove all sink points in

the fitted grid which have not been identified as such in input sink data. The
essence of the algorithm is to recognize that each spurious sink is surrounded
by a drainage divide containing at least one saddle point. If the sink is
associated with an elevation data point then the lowest such saddle, provided
it is not also associated with an elevation data point, identifies the region of the
grid which is modified in order to remove the spurious sink. If on the other hand
218

oI.... ./." ! .!

~Qo/ .,-. ~ , , , , ...>

° '-+ ° I .+ !_ ..
Fig. 2. Discretized minimum potential interpolation of the point elevation data of Davis (1973, table
6.4). Sinks are denoted by circles.

the lowest saddle point is associated with an elevation data point but the sink
is not, then the sink and its immediate neighbours are raised above the height
of the data point saddle. If neither the sink nor the lowest saddle are associated
with elevation data points then grid points in the neighbourhood of both the
sink and the lowest saddle are modified to ensure drainage. Finally, ifboth sink
point and saddle point are associated with elevation data points, then a choice
is made, depending on a user-supplied tolerance, between enforcing drainage
and maintaining fidelityto the data. This last situation arises when calculating
generalized (coarse resolution) DEMs.
The action of this drainage enforcement algorithm is conceptually similar to
the action taken by the basin delineation and drainage network simulation
 219
I I . . . . . ~ ! '" ! a

~,,-- __ 765 765 ~ "-~ , ~ . , ~

/ I ~

p.o \ Xr l ~ .,, ~ t /

- .=\a/ /

\ "~ / £ ~ - \ .} "+,,,o-.
F-"-..,. \ ~ \ ( ~'\ \

Fig. 3. Interpolation of the point elevation data of Davis (1973, table 6.4) by minimizing the
roughness penalty J(f). Sinks are denoted by circles. True stream lines are indicated by dashed
lines.

programs o? Marks et al. (1984) and Yuan and Vanderpool (1986) to overcome
the problem of spurious sinks. It is also related to the methods of cartographic
generalization suggested initially by Warntz (1975) and taken up by Pfaltz
(1976) and Wolf (1984). In fact the approach suggested here of maintaining
connected drainage patterns provides a more secure physical basis for carto-
graphic generalization than the partially lexicographic approach adopted by
these authv Cs.
The drainage enforcement algorithm proceeds concurrently with the
iterative interpolation algorithm described in the preceding section. For each
grid resolution, after the first very coarse resolution, the grid is periodica'ly
220

inspected (normally a~ter every five Gauss-Seidel iterations) for sinks and their
accompanying saddle points. These are found by comparing the height of each
grid point with the height f each of its eight immediate neighbours (cf.
Peucker and Douglas, 1975). A sink point is characterized by having an
elevation no higher than the elevation of each of its eight immediate
neighbours while a grid point is a saddle point if it has at least two neighbours
strictly higher than itself interleaved by neighbours no higher than itself when
moving in a clockwise or an anticlockwise direction through the eight
immediate neighbours surrounding the grid point. Saddles are associated with
sink points by searching in each of the two, three or four possible steepest
downhill directions away from each saddle point until a sink or an edge of the
grid is found This is illustrated in Fig. 4 where the saddle points associated
with the sink point S 1 are the points A, B, C, D and E.
Ordered chain conditions effecting drainage clearance are then applied to
the grid by inserting ordered chains which lead from each spurious sink point,
via the lowest associated saddle point, to a data point or existing ordered chain
on the other side of the saddle, provided this does not lead to an elevation
conflict exceeding a user supplied tolerance. The action of the ordered chains
is to impose linear descent, to within a small tolerance, between successive
elevation data points down the entire length of the chain.
Thus, in the example of Fig. 4, the lowest saddle associated with the sink S 1
is the point D. Since this saddle point is not associated with an elevation data
point, and it leads to the sink point S 2 which is strictly lower than S 1, an
ordered chain is inserted from S 1 to S 2 via D as shown in Fig. 5. Each ordered
chain in Fig. 5 is made up of two flow lines leading from the ~we~t saddle
associated with each sink point in Fig. 4. The procedure for detecting sinks and
saddles and inserting ordered chain conditions is reasonably efficient,
requiring less computer time than the basic interpolation algorithm, especially
since it is only enacted once every five Gauss-Seidel iterations.
The action of the drainage enforcement algorithm is modified in practice by
the systematic application of three user-supplied elevation tolerances. These
tolerances allow the strength of drainage enforcement to be adjusted in
relation to the accuracy and devsity of the input elevation data as well as the
level of generalization required. Their detailed action has undergone consider-
able development and testing with data sets of varying densities and accuracies
at a variety of scales. The aim has been to achieve the strongest possible
drainage enforcement without making serious errors in the placement of
drainage lines. The action of the tolerances is most critical when the input data
are limited in terms of accuracy or density, especially when calculating
generalized DEMs for large areas (Hutchinson and Dowling, 1989). Their
action naturally becomes less critical as the accuracy and density of the input
data improve. When the tolerances have been set appropriately, the sink
points not cleared by the program are normally those associated with genuine
sinks, with s~gnificant elevation errors in input data, or, with areas where the
input data are not of sufficient density to reliably resolve the drainage charac-
teristics of the fitted grid.
 221

• 262 .26!

o" J

0~ d

/
/
I

i#
~ Oe

i#
iI

Fig. 4. Example showing how the sacdle points A, B, C, D, E are associated with the sink point S 1
via flow lines which are indicated by dashed lines. Additional sink points are denoted by S 2, S 3,
$4. Data points are indicated by their height in metres.

The first user-zupplied tolerance is essentially a measure of the elevation

accuracy of the data. Elevation differences between data points not exceeding
this tolerance are judged to be insignificant with respect to dz~:,inage. Thus data
points which block drainage by no more than this tolerance ~,'e removed. When
data points are not sufficiently dense to accurately reso!.~,e drainage, this
tolerance may be increased to yield a generalized but sensible drainage pattern
at the expense of fidelity of the fitted surface to the elevation data. The first
tolerance is also used when searching for possible clearances for sinks, to
slightly favour those saddle points which are not associated with elevation
data points over saddle points which are associated with elevation data points.
This is based on the understanding that the height of the grid at a data point
saddle is more reliable than the height at a nondata point saddle.
The second and third elevation ~-olerances play a more technical role in
limiting various searching operations by the procedure in order to increase
efficiency and to prevent nonsensical drainage clearances, particularly when
data are sparse. The second tolerance is a measure of local relief which is most
naturally set to the contour interval when gridding contour data. It limits the
222

• 262 ~........~252

• 262

Fig. 5. The result of draina~-e enforcement applied to the example of Fig. 4. Piecewise linear lines
indicate inferred drainage lines.

height above each data point sink of data point saddles which may be
considered as possible exits from the sink. This can remove from consideration
certain data point saddles, even though they may be the lowest saddle
associated with a particular sink, in order to allow drainage clearance via a
higher non-data point saddle. Ifthe second tolerance is set to a large value.then
the procedure acts conservatively when attempting to remove sinks because
more data point saddles will be considered. The third elevation tolerance is
simply used as a finalcheck to prevent drainage clearances which would entail
very large chan~,~ to the grid. It is only active when the elevation data are very
sparse or contain large errors in elevation.
T w o additional features of the drainage enforcement algorithm merit
comment. The firstis that spurious sinks are sorted by elevation and cleared
in order of increasing elevation. This facilitates the searching operations
required to associate saddle points with sinks and improves the placement of
ordered chains which clear higher sinks, particularly in their lower reaches
where they normally join existing ordered chains. This is illustrated in Fig. 5
where the sink point S 3 has not been immediately cleared to the lower sink
point S 4, since this point was firstcleared to S I. A subsequent enactment of
 223

the drainage enforcement algorithm cleared S 3 to the ordered chain leading

from $4 to S 1. The second feature is that ordered chains clearing sinks are
always extended beyond any lower data points encountered until they meet an
existing ordered chain, a sink or an edge of the grid. This ensures that a
connected drainage pattern is obtained.
The result of applying the drainage enforcement algorithm in conjunction
with the roughness penalty J ( f ) to the data of Davis is shown in Fig. 6. The
inferred drainage pattern and associated contours are remarkably similar to
the actual drainage pattern and the contours shown in Fig. 8 and in fig. 6.10 of
Davis (1973, p. 323). The remaining sinks at the top and bottom of Fig. 6
illustrate the essentially non-local nature of the drainage enforcement
algorithm. They would be removed if additional data beyond the extremes of
the figure were available.

T H E I N C O R P O R A T I O N OF S T R E A M LINE D A T A

Drainage enforcement can also be obtained by incorporating stream line

data. This can be useful when more accurate placement of streams is required
than can be calculated automatically by the procedure. It can also be used to
remove sinks which would not otherwise be removed by the automatic drainage
enforcement algorithm. This is in fact the recommended way to correct
remaining drainage anomalies in elevation grids calculated by the procedure.
All elevation data points which conflict with strict descent down each stream
line are removed, with all conflicts which exceed the third user specified
elevation tolerance being flagged for possible correction of errors. The incor-
poration of actual stream line data also provides a more reliable way of
ensuring sensibly generalized elevation models.
The method of incorporating stream lines and their associated side
conditions into a grid is illustrated in Fig. 7. First, the data stream line, which
flows from the bottom to the top of the figure as indicated by the dashed line,
is approximated by straight line segments, indicated by heavy lines. These have
been generalized so that successive segments change in direction by at most
45°. Thus, the segments CX and XY, which were an initial representation of a
portion of the stream, have been replaced by the single segment CY. This
minimizes the number of straight line segments required to represent each
stream line and removes unnecessarily sharp bends from its representation on
the grid. Side conditions arc then ~dded, as indicated by the light line segments.
These normally lie at 45° to the prevailing direction of ihe stream line except
at bends, the point Y in this example, where the side condition represented by
the segment XY is obtained as the uphill extension from the lower stream line
segment YF. Note that each grid point can have a number of adjoining upper
stream segments but at most one adjoining lower stream segment.
The action of these straight line segments on the interpolation conditions
which arise from minimizing the prescribed roughness penalty is then as
follows. For points such as B, with one lower adjoining segment and at least
224

/ /
755
t
( ?) :.o.i/o
/
!

/ i°° i °.
750
765
620
791

i Bl2

862 908 682

_.¢:__L.Y.o ,o,,J ..o

Fig. 6, Interpolation of the point elevation data of Davis (1973, table 6.4) as in Fig. 3 and employing
the drainage enforcement algorithm, Sinks are denoted by circles. Piecewise linear lines indicate
inferred drainage lines.

one upper adjoining segment, the height at B is co:.strained so that descent is

(approximately) uniform from the lowest of A 1, A2 and A 3 down to C. For
points such as E, with one lower adjoining segment and no upper adjoining
segment, the height is similarly constrained between D (the uphill extension of
EF) and F. These actions maintain linear descent between elevation data
points down each stream line and ensure that each stream line acts as a
breakline for the interpolation conditions. This i n . t u r n ensures that each
stream line lies at he bottom of its accompanying valley. Side conditions are
not set for data poiz :ts beside streams whose elevations are more than the third
elevation tolerance below the height of the stream. These conflicts are also
flagged for possible correction of errors.
 225
+ ÷ + ~ + +

/!1_\

+ + X + +

~J, / ~ / " I

-~'A2 4A3 + + + +

Fig. 7. Example showing how a stream line is incorporated into a grid with associated side
conditions.

The stream lines in fig. 6.6 of Davis (1973, p. 312) were digitized and adde~
to the point elevation data to yield Fig. 8. The straight line segments describing
the stream lines and their associated side conditions are also shown. A modest
improvement to the modelled drainage pattern has been obtained in this case.
Stream line data are best used in practice to define the major drainage lines
associated with high order streams, leaving the drainage enforcement
algorithm to define the lower order stream lines to a level of detail that is
controlled mainly by the amount of available data and the specified resolution
of the DEM. This eliminates the need for digitizing the large number of lines
associated with low order streams which may not exist in mapped form.

A LARGER EXAMPLE

An application to a larger data set from an area in central Queensland,

Australia at 23°S and 143°E is now described. Figure 9 shows the result of
fitting a minimum curvature surface to this rather sparse data set. In contrast
to Fig. 1, in this case there are many remaining sinks which obscure the
underlying drainage pattern. The drainage enforcement algorithm, in combina-
tion with the terrain specific roughness penalty J(f), resolves all of the
apparent drainage anomalies as illustrated in Fig. 10 and the resulting
drainage pattern, though generalized because of the sparseness of the data, is
in good agreement with the actual drainage pattern shown in Fig. 11. The grid
was fitted to data extending beyond the limits of the figure so that no edge
anomalies are apparent.
226

.o) j' t
755

W r
I !

i e

800

765 765

830 °J,,'
[ 79i

812

8t2

130

873 875~

865

915~ 600

ego 0

Fig. 8. Interpolation of the point elevation data of Davis (1973, table 6.4) as in Fig. 6 and incorporat-
ing stream line data with associated side conditions.

DISCUSSION

An effective procedure for calculating digital elevation models with sensible

draiuage properties from comparatively small sets of surface specific point
elevation data and stream lines has been described. Digital elevation models
calculated by this procedure may be used to advantage in hydrological process
~tudies. The procedure embodies a morphological approach to digital elevation
modelling, since the interpolation algorithm and the accompanying drainage
enforcement algorithm are defined directly in terms of morphological
properties of the fitted surface. The success of the technique has also been
assessed primarily in terms of morphological properties of the fitted surface.
 227

~
......~0~° 26~

75

,8

254

2~f

~34
289

Fig. 9. Minimum curvature interpolation of scattered point elevation data in central Queensland,
Australia, 23°S, 143°E. Sinks are denoted by circles. Elevations in metres.

The drainage enforcement algorithm plays an essential role in the interpola-

tion process. Minimization of a variety of general roughness penalties as a
basis for interpolation has been seen to be inadequate without the imposition
of localized constraints as calculated by the drainage enforcement algorithm or
obtained from stream line data. Moreover, the drainage enforcement algorithm
provides a sound basis for cartographic generalization and displays significant
advantages over conventional statistical techniques for extracting informa-
tion from and detecting errors in spatially distributed elevation data.
228

259

Fig. 10. Interpolation of the point elevation data used in Fig. 9 by minimizing the roughness penalty
J(f) and employing the drainage enforcement algorithm. Piecewise linear lines indicate inferred
drainage lines.

The relative weighting of the minimum potential and minimum curvature

roughness penalties has been empirically determined to produce good results
in combination with the drainage enforcement algorithm when applied to
surface specific point elevation data. It is feasible that different weightings
would be better suited to different types of terrain and/or different data point
sampling strategies, however good results have been obtained with this
relative weighting for both arid and humid areas in Australia. The choice of
roughness penalty naturally becomes less critical as data density increases.
 229

~t
cr

Fig. 11. The drainage pattern for the area of Fig. 9 as digitized from a 1:250 000 topographic map.

The procedure can in principle be applied to data at any scale, the only limits
being the availability of sufficient data and practical limitations on the size of
the interpolated grid. The procedure has been incorporated by the Australian
Divisio~ of National Mapping into its production of a national digital topo-
graphic database at a nominal scale of 1:1 million (Trezise and Hutchinson,
1986). IL is also being used routinely by researchers in the hydrologic modelling
of small catchments (Moore et al., 1988).
The interpolation procedure is computationally optimal in the sense that
computer time is essentially proportional to the number c c interpolated grid
230

points. H o w e v e r , it is a n t i c i p a t e d t h a t a s u i t a b l e m u l t i g r i d t e c h n i q u e ( F u l t o n
et al., 1986) will yield a s i g n i f i c a n t r e d u c t i o n in t h e c o m p u t e r time r e q u i r e d .
W o r k is also in p r o g r e s s to e x t e n d t h e p r o c e d u r e to t h e o p t i m a l i n t e r p o l a t i o n
of c o n t o u r data.

ACKNOWLEDGMENTS

T h e A u s t r a l i a n D i v i s i o n of N a t i o n a l M a p p i n g ( N A T M A P ) p r o v i d e d t h e
e l e v a t i o n d a t a used to p r o d u c e Figs 9, 10. I w i s h to t h a n k H.A. Nix, T.I. D o w l i n g
a n d J.G. S p e i g h t from t h e former C S I R O Division o f W a t e r a n d L a n d
R e s o u r c e s , A.L. C l a r k e a n d P.H. Trezise from N A T M A P a n d K. K u b i c for t h e i r
s u p p o r t , s t i m u l a t i n g f e e d b a c k a n d p a t i e n c e d u r i n g t h e c o n s i d e r a b l e p e r i o d it
h a s t a k e n to d e v e l o p t h e d i g i t a l e l e v a t i o n m o d e l l i n g p r o c e d u r e d e s c r i b e d
above.

REFERENCES

Abbott, M.B., Bathurst, J.C., Cunge, J.A., Connell, P.E. and Rasmussen, J., 1986a. An introduction
to the European Hydrological System -- Syst~me Hydrologique Europden, "SHE", 1: History
and philosophy of a physically-based, distributed modelling system. J.Hydrol., 87: 45--59.
Abbott, M.B., Bathurst, J.C., Cunge, J.A., Connell, P.E. and Rasmussen, J., 1986b. An introduction
to the European Hydrological System -- Syst~me Hydrologique Europ~en, "SHE", 2: Structure
of a physically-based, distributed modelling system. J. Hydrol., 87: 61-77.
Ayeni, O.O., 1982. Optimum sampling for digital terrain models: a trend to,Cards automation.
Photogramm. Eng. Remote Sensing, 48(11): 1687-1694.
Band, L.E., 1986. Topographic partition of watersheds with digital elevation models. Water Resour.
Res., 22(1)'. 1524.
Bathurst, J.C., 1986. Physically.based, distributed modelling of an upland catchment using the
Systbme Hydrologique Europden. J. Hydrol., 87: 79-102.
Berry, J.K., 1985. Computer-assisted map analysis: fundamental techniques. Computer Graphics
Conference, 1985, Natl. Comput. Graphics Assoc., Fairfax, Va.
Bork, H..R. and Rohdenburg, H., 1956. Transferable parameterization methods for distributed
hydrological and agroecological models. Catena, 13: 99117.
Briggs, I.C., 1974, Machine contouring using minimum curvature. Geophysics, 39: 3948.
Davis, J.C., 1973. Statistics and Data Analysis in Geology. Wiley, New York, N.Y.
De Boor, C., ,978. A Practical Guide to Splines. Springer, New York, N.Y.
Duchon, J., 176, Interpolation des fonctions de deaux variables suivant le principe de la flexion
des plaques minces. R.A.I.R.O. Anal. Numer., 10: 5-12.
Evans, I.S., 1980. An integrated system of terrain analysis and slope mapping. Z. Geomorphol. N.F.
Suppl.-Bd. 36: 274-295.
Frederiksen, P., Jacobi, O. and Kubik, K., 1985. A review of current trends in terrain modeling
I.T.C. Journal, 1985: 101106.
Fritsch, F.N. and Carlson, R.E., 1980. Monotone piecewise cubic interpc!ation. SIAM J. Numer.
Anal., 17(2): 238-246.
Fulton, S.R., Ciesielski, P.E. and Schubert, W.H., 1986. Multigrid methods for elliptic problems: a
review. Mon. Weather Rev., 114: 943-959.
Golub, G.H. and Van Loan, C.F., 1983. Matrix Computations. Johns Hopkins University,
Baltimore.
Goodchild, M.F. and Mark, D.M., 1987. The fractal nature of geographic phenomena. Ann. Assoc.
Am. Geogr., 77(2): 265-278.
 231

Harrington, H.J., Simpson, C.J. and Moore, R.F., 1982. Analysis of continental structures using a
digital terrain model (DTM) of Australia. Bur. Miner. Resour. J. Aust. Geol. Geophys., 7: 68-72.
Heerdegen, R.G. and Beran, M.A., 1982. Quantifying source areas through land surface curvature
and shape. J. Hydrol., 57: 359-373.
~~utchinson, M.F., 1984. A summary of some surface fitting and contouring programs for noisy data.
CSIRO Div. Math. Stat. Consult. Rep., ACT 84/6, 24 pp.
Hutchinson, M.F., 1986. Fitting surfaces to terrain data. CSIRO Div. Water Land Resour., Res. Rep.
1983-85: 63-64.
Hutchinson, M.F. and Bischof, R.J., 1983. A new method for estimating the spatial distribution of
mean seasonal and annual rainfall applied to the Hunter Valley, New South Wales. Aust.
Meteorol. Mag., 31: 179-184.
Hutchinson, M.F and Dowling, T.I., 1989. A new digital elevation model for Australia. CRES Pap.
Australian National University, Canberra, A.C.T.
Jupp, D.B.L., Hutchinson, M.F., Hutton, P.G. and Kalma, J.D.K., 1986. Models and data integra-
tion in thermal data: a comparison between sensors and scales. Proc. First Aust. AVHRR Conf.,
Perth, W. A., CSIRO Div. Groundwater Res., pp. 210-220.
Kelly, R.E., McConnell, P.R.H. and Mildberger, S.J., 1977. The Gestalt photomapping system.
Pho~ogramm. Eng. Remote Sensing, 43(11): 1407-1417.
Knisel, W.G. (Editor), 1980. CREAMS - - A field scale model for chemicals, runoff and erosion from
agricultural management systems. USDA Conserv. Res. Rep. 26, 643 pp.
Konecny, G., Lohmann, P., Engel, H. and Kruck, E., 1987. Evaluation of SPOT imagery ol~
analytical photogrammetric instruments. Photogramm. Eng. Remote Sensing, 53(9): 1223-1230.
Lasi¢~t, G.M., McBratney, A.B., Pahl, P.J. and Hutchinson, M.F., 1987. Comparison of several
spatial prediction methods for soil pH. J. Soil Sci., 38: 325-341.
Lodwick, G.D., 1982. A comparison of four digital terrain modelling packages. Proc. Harvard
Computer Graphics Week, Harvard University.
Mandelbrot, B.B., 1982. The Fractal Geometry of Nature. Freeman, San Francisco, Calif.
Mark, D.M., 1979. Phenomenon-based data-structuring and digital terrain modelling. Geo-proc., 1:
27-36.
Mark, D.M., 1984. Automa.ed detection of drainage networks from digital elevation models.
Cartographica, 21: 168-178.
Marks, D., Dozier, J. and Frew, J., 1984. Automated basin delineation from digital elevation data.
Geo-proc., 2: 299-311.
Matheron, G., 1965. Les Variables Regionalis~es et Leur Estimation. Masson, Paris.
Moore, I.D., O'Loughlin, E.M. and Burch, G.J., 198~. A contour-based topographic model for
hydrological and ecological applications. Earth Surf. Processes Landform 13(4): 305320,
Palacios-V~lez, O.L. and Cuevas.Renaud, B., 1986. Automated river-course, ridge and basin de-
lineation from digital elevation data. J. Hydrol., 86: 299-314.
Pennock, D.J., Zebarth, B.J. and de Jong, E., 1987. Landform classification and soil distribution in
hummocky terrain, Saskatchewan, Canada. Geoderma, 40: 297315.
Peucker, T.K. and Douglas, D.H., 1975. Detection of surface.specific points by local parallel
processing of discrete te~:ain elevation data.Cornput. Graph. Image Proc., 4: 375387.
Pfaltz, J.L., 1976. Surface networks. Geogr. Ann., 8(1): 77-93.
Ripley, B., 1981. Spatial Stm,~stics. Wiley, New York, NY.
Rohdenburg, H., Diekkrfiger, B. and Bork, H.-R., 1986. Deterministic hydrological site and
catchment models for the ar:~iysis of agroecosystems. Catena, 13:119 137.
Shelton, R.L. and Estes, J.E., 1981. Remote sensing and geographic information systems: an
unrealized potential. Geo-proc., 1: 395-420.
Sibson, R., 1981. A brief description of natural neighbourhood interpolation. In: V. Barnett
(Editor), Interpreting Multivariate Data, Wiley, Chichester, pp. 2L53.
Speight, J.G., 1974. A parametric approach to landform regions. Inst. Brit. Geogr. Spec. Publ., 7:
213-230.
Swain, C.J., 1976. A FORTRAN IV program for interpolating irregularly spaced data using the
difference equations for minimum curvature. Comput. Geosci., 1: 231-240.
 232

Tesche, T.W. and Bergstrom, R.W., 1978. Use of digital terrain data in meteorological and air
quality modelling. Photogramm. Eng. Remote Sensing, 44(12): 1549-1559.
Testud, J. and Chong, M., 1983. Three-dimensional wind field analysis from dual-doppler radar
data. Part I: Filtering and differentiating the raw data. J. Clim. Appl. Meteorol., 22: 1204-1215.
Torgersen, T., Hutchinson, M.F., Searle, D.E. and Nix, H.A., 1983. General bathymetry of the Gulf
of Carpentaria and the quaternary physiography of Lake Carpentaria. Palaeogeogr., Palaeocli-
matol., Palaeoecol., 41: 207-225.
Trezise, P.H. and Hutchinson, M.F., 1986. The development of a national digital elevation model.
Proc. 6th Aust. Cartogr. Conf. (Austra Carto Two), Melbourne, 1985, Aust. Inst. Cartogr.
Wahba, G. and Wendelberger, J., 1980. Some new mathematical methods for variational objective
analysis using splines and cross-validation. Mon. Weather Rev., 108: 1122-1143.
Warntz, W., 1975. Stream ordering and contour mapping. J. Hydrol., 25: 209-227.
Watson, D.F. and Philip, G.M., 1985. A refinement of inverse distance weighted interpolation.
Geo-proc., 2: 315-327.
Weiringa, J., 1986. Roughness-dependent geographical interpolation of surface wind speed
averages. Q. J. R. Meteorol. Soc., 112: 867-889.
Wiltshire, S.E., Morris, D.G. and Beran, M.A., 1986. Digital data capture and automated overlay
analysis for basin characteristic calculation. Cartogr. J., 23: 60-65.
Wolf, G.W., 1984. A mathematical model of cartographic generalization. Geo-proc., 2: 271-286.
Young, D.M., 1971. Iterative Solution of Large Linear Systems. Academic Press, New York, N.Y.
Yuan, L.-P. and Vanderpool, N.L., 1986. Drainage network simulation. Comput. Geosci., 12(5):
653-665.

View publication stats