WATER RESOURCES RESEARCH, VOL. 35, NO.

7, PAGES 2259 –2268, JULY 1999

On the effect of digital elevation model accuracy on hydrology

and geomorphology
Jeffrey P. Walker and Garry R. Willgoose
Department of Civil, Surveying and Environmental Engineering, University of Newcastle
Callaghan, New South Wales, Australia

Abstract. This study compares published cartometric and photogrammetric digital

elevation models (DEMs) of various grid spacings with a ground truth data set, obtained
by ground survey, and studies the implications of these differences on key hydrologic
statistics. Inferred catchment sizes and stream networks from published DEMs were found
to be significantly different than those from the ground truth in most instances.
Furthermore, the width functions and cumulative area relationships determined from the
published DEMs were found to fall consistently outside the 90% confidence limits
determined from the ground truth for more than 60% of the relationship, suggesting that
these hydrologic properties are poorly estimated from published DEMs. However, the
slope-area relationships determined from published DEMs were found to be less sensitive
to catchment shape, size, and stream network, with the relationship falling outside the
90% confidence limits for less than 40% of the relationship for all catchments identified
from the published DEMs. A published relationship linking the horizontal resolution with
the vertical accuracy of the DEM was tested, predicting a horizontal resolution of about
10 m for the published DEMs tested.

1. Introduction paths tend to zigzag, not following the drainage lines, and are
hence systematically too long.
Recent developments in computational models to monitor Generally, raw elevation data in the form of stereophoto-
and predict hydrology, erosion, and landscape evolution by graphs or field surveys and the equipment necessary to process
Earth and water scientists rely heavily on the integrity of the these data are not readily available to potential users of a
digital elevation models (DEMs) available [Moore and Gray- DEM. Most users are therefore forced to rely on the DEMs
son, 1991; Moore et al., 1991; Lane et al., 1994]. Fryer et al. published by government agencies. The most common form of
[1994] suggested that Earth and water scientists are not fully DEM available in Australia, Organisation Européenne
aware of the limitations of DEMs as sources of spatial infor- d’Estudes Photogrammetriques Experimentales member
mation. This concern is not unwarranted, with few published countries, and the United States are those produced by digi-
reports on the accuracy of DEM-derived elevation and slope tizing the contours on existing topographical maps [Moore et
[Bolstad and Stowe, 1994; Robinson, 1994] or the effects intro- al., 1991; Robinson, 1994], known as “cartometric DEMs.” In
duced on geomorphic parameters; DEMs are created, distrib- many countries these contour-based DEMs are interpolated
uted, and used very frequently without any reference to the onto a grid. In addition to interpolation errors this process
magnitude of the error implied or to the methods applied to its introduces artifacts such as pits or depressions, which interfere
disclosure or correction [Felicı́simo, 1994]. with drainage algorithms based on flow in the steepest down-
A DEM consists of either (1) a two-dimensional array of slope direction. In the state of New South Wales (NSW),
numbers that represents the spatial distribution of elevations Australia, this pit removal incorporates a drainage enforce-
on a regular grid; (2) a set of x, y, and z coordinates for an ment algorithm developed by Hutchinson [1989], which utilizes
irregular network of points; or (3) contour strings stored in the the blue lines drawn on topographic maps to represent the
form of x, y coordinate pairs along each contour line of spec- permanently flowing or major intermittent streams [Moore et
ified elevation. DEMs on a regular grid are the most widely al., 1991].
used data structures because of their computational efficiency In recent years, there has been an increasing move toward
and lower storage requirements. Furthermore, grid DEMs are using automated digital correlation techniques to generate
used to calculate all other types of digital terrain models whose what are known as “photogrammetric DEMs” directly from
accuracy is limited by that of the original source data [Florin- stereoscopic imagery, especially where contour data are not
sky, 1998]. However, Moore et al. [1991] notes that grid DEMs available or are not accurate enough [Robinson, 1994]. These
have several disadvantages: (1) they cannot easily handle dis- methods create gridded DEMs directly with similar artifacts to
continuities in elevation; (2) the resolution of the mesh effects the cartometric DEMs, though studies of these errors for hy-
drologic purposes have not been performed. An additional
the results and computational efficiency; (3) grid spacing needs
drawback is that the points which are correlated (used to de-
to be based on the roughest terrain in the catchment, resulting
termine the elevations) are often the tops of trees, buildings,
in redundancy in smoother areas; and (4) the computed flow
etc., consequently requiring a significant amount of editing by
Copyright 1999 by the American Geophysical Union. hand to produce a realistic DEM [G. Burgess, personal com-
Paper number 1999WR900034. munication, 1995].
0043-1397/99/1999WR900034$09.00 Few studies have been undertaken on the effects of relative

2259
2260 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY

eral conclusions. Thus this paper takes a different tack. He-

mantha and Willgoose [1996] and H. J. Perera and G. R. Will-
goose (A simple model of subsurface saturation excess runoff
generation based on geomorphology, 1, steady state, submitted
to Water Resources Research, 1998) showed that the proportion
of saturation area in a catchment dominated by saturation
excess, and therefore the rate of runoff generation, can be
formulated in terms of the wetness index [Beven and Kirkby,
1979], the cumulative area diagram [Tarboton et al., 1989], and
the slope-area diagram [Willgoose et al., 1991]. Moreover, Sur-
kan [1969] showed the unit hydrograph and the runoff routing
response can be directly related to the catchment width func-
tion. Accordingly, we have investigated the effect of elevation
errors on the following geomorphic relationships: (1) catch-
ment boundary and stream network; (2) slope-area relation-
ship; (3) cumulative area relationship; and (4) normalized
width function.
The slope-area relationship is a log-log plot of the slope for
a drainage link against the total catchment area upstream of
that link [Willgoose et al., 1991], while the cumulative area
relationship is a log-log plot of the percentage of pixels in the
catchment having an upstream catchment area which is greater
than a specified value [Tarboton et al., 1989]. The normalized
width function is a frequency histogram for the number of
Figure 1. Extract of 1:25,000 scale series topographic map drainage links that exist at a given distance from the catchment
showing the location of the site under investigation (not to outlet, scaled such that the distance from the outlet ranges
scale).
from zero to one and the area under the curve is one [Surkan,
1969].
A normalized width function is used to allow comparisons
elevation accuracy in published DEMs on hydrologic modeling between catchments of different sizes and shapes to determine
or geomorphic parameters used in hydrology. Zhang and if generalizations can be made regarding all catchments. In
Montgomery [1994] and Wolock and Price [1994] have investi- addition, with the area under the width function scaled to one,
gated the effect of grid size on the topographic index used by it may be considered as a probability distribution function,
TOPMODEL [Beven and Kirkby, 1979]. Zhang and Montgom- where it describes the probability of the stream network having
ery [1994] found that increasing the grid size resulted in an an “equivalent width” at a given distance from the outlet. The
increased mean topographic index because of increased con- width function is also related to the unit hydrograph used in
tributing area and decreased slopes, while Wolock and Price hydrology, as the distance axis may be transformed to time by
[1994] found that increasing grid size resulted in higher mini- the assumption of constant velocity.
mum, mean, variance, and skewness of the topographic index
distribution. Wolock and Price [1994] also found that the map
scale used to produce the cartometric DEM has an observable
2. Data and Methodology
but much smaller effect on the spatial distribution of topo- 2.1. Data Sources
graphic index than grid spacing. Zhang and Montgomery [1994] For an assessment to be made on the accuracy of a published
conclude that a 10 m grid size presents a rational compromise DEM, it is necessary to have a set of “ground truth” with a
between increasing resolution and data volume for simulating higher degree of accuracy than that of the data being investi-
geomorphic and hydrologic processes. Gyasi-Agyei et al. [1995] gated. The ground truth used in this research was a parcel of
have investigated the effect of vertical resolution of elevation land of approximately 1.4 km2, situated at Pokolbin, NSW,
data on geomorphic parameters by modifying the vertical res- Australia, in undulating terrain (Figure 1). Prominent features
olution between centimeters and meters. They proposed that of the study site include a large farm dam, covering of trees on
the vertical resolution for the DEM of a catchment should be the southern side of Thompsons Lane (mainly on the sections
satisfactory for extraction of the drainage network if the ratio with higher elevation), and vineyards on the northern side of
of average elevation change per pixel (pixel drop) to elevation Thompsons Lane. The catchment elevation ranges from 110 m
error is greater than unity. Once the ratio fell below unity, they Australian height datum (AHD) to 205 m AHD, with steeper
were unable to reliably extract the stream network. They also slopes in the southern portion of the site, flattening out to
found that the cumulative distributions of slope-area relation- gently sloping in the northern portion of the site.
ship and normalized width function were not very sensitive to The ground survey data for this site were obtained by elec-
the vertical resolution. It should be noted, however, that this tronic distance measurement (EDM) tacheometry with coor-
analysis was coarse in so much as the vertical resolution was dinate evaluation on the Australian map grid (AMG). Data
altered by factors of 10. were collected at all changes of grade, as well as at any other
To perform a comprehensive study of elevation errors on points of interest. Therefore linear interpolation between data
hydrology would require multiple simulations with a wide points to a fine-grid spacing could be performed with a high
range of rainfall intensities and durations [e.g., Zhang and degree of accuracy. Approximately 3000 data points were col-
Montgomery, 1994], and it would be difficult to draw any gen- lected with an average spacing of 17.7 m.
 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY 2261

The published DEM data were provided by the Land Infor-

mation Centre (LIC) at Bathurst, NSW, Australia. These data
were provided on the AMG coordinate system thus dispensing
with any georeferencing problems in making data comparisons.
The published DEM data included the following.
Cartometric DEMs are the 6.25 m, 12.5 m, and 25.0 m
gridded elevation data from contours on the 1:25,000 scale
topographical map series (10 m contour interval). Hutchinson
drainage enforcement was used, and an auxiliary contour
height was assigned to the shorelines of the large dams in the
area. Use was also made of spot heights shown on the contour
maps. In addition, 95 m and 105 m intermediate contours were
added by the LIC to help with interpolation in the flatter
region located in the southeast corner of the site.
Photogrammetric DEMs are 6.25 m and 12.5 m gridded
elevation data produced by automatic image correlation of
stereopairs of controlled photography using the Orthomax
software which is a subset of the ERDAS Imagine software
package (J. Perry, personal communication, 1995). No editing
of the data set had been performed; hence in timbered areas,
correlated points may be at treetop level or ground level.
Each of the above DEMs was produced independently from
the same source data using the same parameter settings so that
even at common points the elevations could be different.
The three data sources used in this study are shown as
perspective surface plots in Figure 2. The plot for the ground
truth DEM in Figure 2a shows a relatively smooth surface, with
a few fine features clearly visible. These features are the cut-
ting for Thompsons Lane, the large farm dam at the base of
one of the gullies, as well as isolated areas with abrupt changes
in elevation (see also Figure 1).
The cartometric DEM in Figure 2b shows a much smoother
surface than the ground truth, with the dam being the only fine
feature visible from those mentioned above. The gullies are
still clearly visible, and the overall shape of the site is main-
tained. However, the ridge near the dam appears to have a dip
in it which would allow for the catchment to drain across into
the next gully. This is not the case in the ground truth.
The photogrammetric DEM in Figure 2c, however, shows an
extremely rough surface with a significant amount of high-
frequency noise. This plot fails to allow identification of any of
the fine detail clearly visible in the previous two DEMs, and
even the gullies are hard to identify. However, it is possible to
identify that the same overall shape of the site has been main-
tained. This may be seen more clearly in Figure 2d, where the
DEM has been filtered with a 5 3 5 Gaussian filter to remove
the high-frequency noise, though there are still artifacts on the
flatter northern section of the site.
By an error propagation analysis of the EDM tacheometry
Figure 2. Surface plots of (a) ground truth digital elevation
equations, the original EDM data were found to have centi-
model (DEM), (b) cartometric DEM, (c) photogrammetric
meter accuracy with significant correlations existing only be- DEM, and (d) photogrammetric DEM after filtering with a
tween errors in easting and northing. From a Monte Carlo 5 3 5 Gaussian filter. Horizontal plot size is 1400 m 3 1800 m;
analysis of the Delaunay triangulation gridding procedure vertical exaggeration is 5.
[Sloan, 1993] the gridded EDM data were found to have
slightly larger errors in elevation than the original data, with
the majority of grid points having a standard deviation of less found to be more accurate than the photogrammetric DEMs,
than 5 cm. It was also found that errors in elevations and with rms errors in elevation of approximately 3.5 m and 4.5 m,
derived slopes for the gridded data could be described by an respectively, and maximum absolute errors in elevation of ap-
isotropic single-exponential correlation function, with correla- proximately 12 m and 28 m, respectively. Absolute rms errors
tion lengths directly proportional to the raw data spacing. The in DEM-derived slope were approximately 6% and 20% for
DEM-derived slopes from finite differences were found to the cartometric and photogrammetric DEMs, respectively,
have an average error of 0.5%. with maximum absolute DEM-derived slope errors of approx-
By comparison with EDM data the cartometric DEMs were imately 75% and 290%, respectively. The cartometric DEMs
2262 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY

The confidence limits on the ground truth were determined

from Monte Carlo simulations, which involved simulating 100
replicate ground truth data sets using error estimates for the
tacheometry, assuming each survey observation to be indepen-
dent. Data sets were simulated by adding a perturbation to the
x, y, and z coordinates of each original data point, using the
covariance matrix of that data point. The covariance matrices
were estimated from error theory [Mikhail and Ackerman,
1976] by differentiation of the observation equations [Fryer et
al., 1987] and evaluation using the raw observations. Given the
large correlation between the x, y, and z coordinates of the
individual data points, random perturbations to the data points
were estimated by sampling from a multivariate normal distri-
bution using a technique which maintained the correlation
structure between the perturbations of the x, y, and z coordi-
nates. As the published DEMs were provided on a regular grid
and the method used for extracting the inferred stream net-
work [O’Callaghan and Mark, 1984] and associated geomor-
phic parameters was based upon the data being on a regular
grid, the 100 simulated data sets were interpolated onto a grid
using linear interpolation triangles. The width function, slope-
area relationship, and cumulative area relationship were then
evaluated for each of the 100 replicate gridded elevation data
from which 90% confidence limits on the geomorphic param-
eters were estimated by ranking and selecting the 5th and 95th
values.
The location for the catchment outlet used in this study was
taken as the grid point in the original ground truth DEM which
drained the largest catchment area. However, the main stream
network did not pass through the same outlet position for the
100 replicates of the ground truth but moved either north or
south by one or two pixels. Therefore the specified outlet
position was the nearest grid point with the same easting. An
interesting feature of the adopted catchment outlet was that it
was only slightly downstream from a junction of two major
streams (Figure 3a), requiring special care to ensure that the
Figure 3. Catchment and stream network from (a) ground correct catchment was identified.
truth DEM, (b) cartometric DEM, and (c) photogrammetric Gyasi-Agyei et al. [1995] suggested that a DEM is adequate
DEM for initial outlet position with data on a 6.25 m grid for extracting the stream network, if the ratio of average pixel
spacing.
drop and vertical resolution is greater than unity. The average
pixel drop can be determined from the average slope and grid
spacing, and the vertical resolution can be considered as being
satisfy the U.S. Geological Survey (USGS) specifications for approximately equal to the standard deviation of relative er-
level 2 data, which is currently the highest level of accuracy at rors between grid points.
which DEMs are produced by the USGS [Moore et al., 1991].
average pixel drop a D x
< .1 (1)
2.2. Effects of DEM Errors on Hydromorphic Parameters vertical resolution s Dz
In assessing the effects of DEM error on geomorphic pa-
where a is the mean slope (m/m), D x is the grid point spacing
rameters, both qualitative and quantitative assessments were
(in meters), and s Dz is the standard deviation of relative error
undertaken. The qualitative assessment was made by compar-
in elevation (in meters).
ing the inferred stream networks, catchment boundary, and
The standard deviation of relative error in elevation can be
study site boundary from each of the published DEMs with
estimated using two approaches. In the first approach the stan-
that from the ground truth DEM. The threshold area used for
dard deviation of relative errors can be estimated by an au-
defining the stream networks in this qualitative analysis was
toregressive order 1 series [Bras and Rodriguez-Iturbe, 1985],
nominally set to 625 m2. The quantitative assessment of the
using the error analysis of the differences between the pub-
effects on geomorphic parameters was made by comparing
lished and the ground truth DEMs. The relationship used to
plots of the width function, slope-area relationship, and cumu-
describe this is
lative area relationship from the published DEMs with 90%
confidence limits derived from the ground truth DEM. The s Dz 5 s z Î1 2 r 21 (2)
quantitative analysis is independent of the threshold area cho-
sen above, as geomorphic parameters were estimated from all where s z is standard deviation of absolute error in elevation
grid cells (flow paths) contained in the catchment inferred (in meters) and r1 is the lag 1 correlation. The lag 1 correlation
from the DEM. is the correlation of absolute error in elevation at a distance of
 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY 2263

Table 1. Results From Evaluation of the Empirical Relationship Describing the Adequacy of a DEM for Hydrological
Applications
Cartometric DEM Photogrammetric DEM

6.25 m Grid 12.5 m Grid 25.0 m Grid 6.25 m Grid 12.5 m Grid 25.0 m Grid
Resolution Resolution Resolution Resolution Resolution Resolution

Average slope, m/m 0.081 0.112 0.122 0.138 0.171 0.085

Average pixel drop, m 0.51 1.40 3.05 0.86 2.14 1.10
Correlation length, m 304 286 143 128 80 174
Lag 1 correlation 0.98 0.96 0.84 0.95 0.86 0.93
Standard deviation of absolute errors in elevation, m 2.57 2.89 2.73 4.00 2.97 2.58
Standard deviation of relative errors in elevation by AR(1) 0.51 0.81 1.48 1.25 1.52 0.94
series, m
Ratio of average drop to vertical resolution by AR(1) series 1.0 1.7 2.1 0.7 1.4 1.2
Standard deviation of slope errors, m/m 0.062 0.054 0.053 0.188 0.123 0.063
Standard deviation of relative errors in elevation by standard 0.39 0.68 1.33 1.18 1.54 0.79
deviation of slope, m
Ratio of average drop to vertical resolution by slope 1.3 2.1 2.3 0.7 1.4 1.4

one grid spacing away. Using the calculated correlation func- spacing, plots of inferred stream networks and catchment
tion and correlation length, the lag 1 correlation was estimated boundaries at the coarser-grid spacings were not significantly
(Table 1). different apart from a decrease in resolution. This suggests that
The second approach uses the standard deviation of errors grid spacings finer than 25 m do not have a significant effect on
in slope (m/m) which have been evaluated for each of the the ability to extract the inferred stream network and catch-
published DEMs. The standard deviation of relative errors in ment boundary.
elevation was evaluated by multiplying the standard deviation The stream network inferred from the photogrammetric
of slope with the respective grid spacing (Table 1). In this way DEM (Figure 3c) appears to be more realistic than that from
the empirical relationship for checking DEM adequacy was the ground truth DEM (Figure 3a). This is a result of the
tested for the DEMs used in this study. parallel streamlines in Figure 3a on the planar hillslopes and
the relatively low threshold area chosen for visual representa-
3. Results and Discussion tion of flow paths. The stream network in Figure 3c does not
posses so many parallel flow paths as a result of the high-
3.1. Qualitative Assessment of Effects on Geomorphic frequency noise in elevation data seen in Figure 2c.
Parameters
The catchment area identified from the cartometric DEMs
Plots of inferred stream networks and catchment boundaries (18 ha) in Figure 3b is significantly different from the ground
for the ground truth, cartometric, and photogrammetric DEMs truth (42 ha) in Figure 3a. These plots reveal that the catch-
are given in Figure 3 for 6.25 m grid spacing. In the ground ments and stream networks identified from the cartometric
truth DEM (Figure 3a) a small corner of the catchment under DEMs are similar in shape to that for the southern main
analysis is missed because of a lack of data in the upper stream identified in the ground truth, indicating that the two
reaches. However, this is not considered to have a significant main streams evident in the ground truth had not yet con-
effect on the comparison of results in the quantitative analysis, verged at the outlet in the cartometric DEMs. A field inspec-
as the plots of the catchments and stream networks for the tion of the outlet clearly indicated that the two main gullies
published DEMs are significantly different from those ob- converged before the specified outlet, and hence an assessment
tained for the ground truth.
was made of the cartometric DEMs for an outlet farther down-
Although Figure 3 only shows plots of inferred stream net-
stream.
works and catchment boundaries for DEMs with 6.25 m grid
Figure 4 contains a plot of the inferred catchment and
stream network from the cartometric DEM with 6.25 m grid
spacing for an outlet far enough downstream to capture the
convergence of the two streams. However, the second major
stream was still not captured in its entirety. It is evident from
these plots that the second major stream only extends back as
far as the large farm dam and that the hypothesis of the
cartometric DEMs draining water from the dam into the next
gully is correct.
From the plot of catchment boundary and stream network
for the photogrammetric DEM (Figure 3c) it may be seen that
the catchment area (42 ha) is of a similar size to that of the
ground truth (Figure 3a). However, the shape of the catchment
and the stream network within it are quite different. The major
Figure 4. Catchment and stream network from cartometric difference is that the two main streams evident in the ground
DEM for outlet chosen farther downstream from where the truth and converging near the outlet cannot be seen in the
two stream lines join. stream networks identified from the photogrammetric DEMs,
2264 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY

catchments from published DEMs shows that the stream net-

work and catchment boundary can be significantly in error.
However, a DEM grid spacing of less than 25 m does not
appear to have a significant effect on the adequacy of the DEM
to identify the catchment and stream network. This suggests
that the underlying data source used for deriving the DEMs is
the crucial factor in the observed differences.

3.2. Quantitative Assessment of Effects on Geomorphic

Parameters
The comparisons of geomorphic parameters presented here
are only for those locations of catchment outlet which gave the
Figure 5. Catchment and stream network from photogram- best qualitative catchment representation, that is, (1) the up-
metric DEM on 12.5 m grid, filtered with a 5 3 5 Gaussian stream outlet position for the cartometric DEM (Figure 6b)
filter, for initial outlet position. and (2) downstream outlet position for the photogrammetric
DEM (Figure 3c).
These plots are for a 6.25 m grid spacing; the other grid
with only one main stream being identified. The reason for this spacings gave similar results. The exception to this indepen-
is evident in Figures 2 and 3. In comparing Figures 3a and 3c,
the major difference in the stream network begins at the loca-
tion of the large farm dam. In Figure 3c (photogrammetric
DEM) the main northern stream has cut back across to the
main southern stream rather than continuing on in the same
manner as the ground truth in Figure 3a. Figures 2a and 2d
show that the dam has interfered with the elevations in the
DEM, with the upper section of the catchment draining across
to another gully within the catchment. This can be contrasted
with the cartometric DEMs which have drained the upper
section of the catchment across to a gully outside the identified
catchment.
The catchment and inferred stream networks for the photo-
grammetric DEM on a 12.5 m grid filtered with a 5 3 5
Gaussian filter were also extracted for the initial outlet position
and are presented in Figure 5. The catchment and inferred
stream network for the filtered data are similar to that for the
original data (Figure 3c), with the exception of the streamlines
becoming parallel. As the size of the filter increased, it was
observed that more parallel streamlines were formed as the
surface became smoother. These parallel streamlines are also
a feature of the stream networks from the cartometric DEMs.
The main reason for the differences in stream networks from
the ground truth in Figure 3 is the near coincidence of the
confluence of the two main streams and the selected outlet
position. To investigate how close the inferred stream network
and catchment boundary might be to the ground truth in the
absence of this coincidence, an outlet upstream of where the
two main streams merge was selected. Figure 6 shows the plots
of inferred catchment and stream network from the published
and ground truth DEMs with 6.25 m grid spacing. Plots from
coarser-grid spacings were similar.
The catchment area identified from the cartometric DEMs
(18 ha) in Figure 6b is approximately the same as that for the
ground truth (17 ha) in Figure 6a. Inferred catchment and
stream network shapes for the cartometric DEMs are essen-
tially the same as the ground truth apart from more parallel
streamlines. However, Figure 6c shows that the inferred catch-
ment area from the photogrammetric DEMs (39 ha) is signif-
icantly different from the ground truth (17 ha). The reason for
the inferred catchment area from the photogrammetric DEMs
being so large is that only one main stream was identified, Figure 6. Catchment and stream network from (a) ground
resulting in little loss of catchment area from the initial outlet truth DEM, (b) cartometric DEM, and (c) photogrammetric
position. DEM for outlet upstream from where the two stream lines join
This qualitative assessment of inferred stream networks and with data on a 6.25 m grid.
 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY 2265

dence of grid spacing was the cumulative area relationship

which was scale dependent, as previously identified by La Bar-
bara and Roth [1994]. Thus the only cumulative area diagrams
which could be directly compared with the 90% confidence
limits (derived for a 6.25 m grid spacing) were those from the
published DEMs on a 6.25 m grid. Cumulative area diagrams
at larger grid spacings maintained the same overall shape,
except that they moved to the right with increasing grid spacing.
The width functions for the cartometric and photogrammet-
ric DEMs are given in Figure 7. Though the 90% confidence
limits are quite wide, neither of the width functions from the
published DEMs are contained within these limits. The wide
confidence limits on the width function indicate that the width
function is very sensitive to even small errors in the elevations.
Although it would be expected that the width function would
be outside the 90% confidence limits for 10% of the relation-
ship on average, the width function for the cartometric and
photogrammetric DEMs is outside the limits for 60% and
70%, respectively. For the photogrammetric DEM this is be-
cause the inferred stream network is significantly different
from that of the ground truth. However, the inferred stream
network for the cartometric DEM is very similar to that of the
ground truth, yet it has not given any better results. An expla-
nation for this may be the parallel streamlines which can be
clearly seen in Figure 6b.
The slope-area relationship was estimated reasonably well
for both the cartometric and photogrammetric DEMs (Figure
8), lying outside the limits for 30% and 40% of the relation-
ship, respectively. Although not presented here, slope-area
Figure 8. Slope-area relationship from data on 6.25 m grid
spacing for (a) cartometric DEM with the upstream outlet
position and (b) photogrammetric DEM with the original out-
let position.

relationships for the other catchment outlet locations gave

qualitatively similar results. Hence the slope-area relationship
appears to be less sensitive to the accuracy of the DEM, as the
identified catchments were significantly different from that of
the ground truth in some instances. The increased number of
spikes for the photogrammetric DEM is due to the high-
frequency noise in the elevations and would be less intense if
the elevations were filtered, but the slope scatter is clearly
sensitive to elevation errors.
The cumulative area relationships are given in Figure 9,
showing that the 90% confidence limits are very tight, and the
cumulative area relationships determined for the published
DEMs are far from lying within these limits. The cumulative
area relationship which fits the limits best is that from the
photogrammetric DEM; however, neither of the fits are satis-
factory. This is counterintuitive, given that the cartometric
DEM has produced a closer representation of the catchment
and stream network than that from the photogrammetric
DEM. An explanation for this may again be the parallel
streamlines inferred from the cartometric DEMs. The narrow
confidence limits which can be seen on the cumulative area
diagram suggest that it is relatively insensitive to small eleva-
tion errors.
These geomorphic relationships were investigated for all
Figure 7. Width function from data on 6.25 m grid spacing outlet locations and gave qualitatively worse results than those
for (a) cartometric DEM with the upstream outlet position and presented here, the exception being the slope-area relationship
(b) photogrammetric DEM with the original outlet position. which was independent of the catchment outlet location or the
2266 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY

the trend is weak for the cartometric DEMs, and this may
indicate that the trend is coincidental, with differences being
sample effects.
The correlation length of errors in elevation decreased with
increasing grid spacing for both the cartometric and photo-
grammetric DEMs. Furthermore, the shorter correlation
length of errors in elevation for the photogrammetric DEMs
relative to the cartometric DEMs is consistent with the large
standard deviations in DEM-derived slope for the photogram-
metric DEM (19%) in comparison to the cartometric DEM
(6%) and the relatively noisy surface. Filtering of the photo-
grammetric DEM on a 12.5 m grid spacing with a 5 3 5
Gaussian filter reduced the standard deviation of slope error to
the same order as that for the cartometric DEM, with a cor-
responding increase in correlation length. The lag 1 correlation
decreases with increasing grid spacing for both the cartometric
and photogrammetric DEMs as expected.
The average DEM-derived slope for the unfiltered photo-
grammetric DEMs is greater than that for the cartometric
DEMs at corresponding grid spacings. This is expected, given
the large amount of high-frequency noise observed in the pho-
togrammetric DEMs. Filtering of the photogrammetric DEM
with a 5 3 5 Gaussian filter reduced the average slope to less
than that for the cartometric DEM at a corresponding grid
spacing. This would suggest that more smoothing has been
performed to the photogrammetric DEM than to the carto-
metric DEM.
The standard deviation of absolute errors in elevation show
Figure 9. Cumulative area relationship from data on 6.25 m
grid spacing for (a) cartometric DEM with the upstream outlet no consistent trends with grid spacing, and the standard devi-
position and (b) photogrammetric DEM with the original out- ations of absolute errors in elevation for the photogrammetric
let position. DEMs are significantly greater than those for the cartometric
DEMs, as expected. Filtering of the photogrammetric DEM
with the 5 3 5 Gaussian filter reduced the standard deviation
identified catchment boundary and stream network, indicating of absolute errors in elevation to the same level as for the
its independence of catchment shape and stream network. cartometric DEMs.
This quantitative analysis shows that published DEMs can The standard deviation of relative errors in elevation, how-
be used to determine the slope-area relationship with greater ever, shows a consistent trend with grid spacing for both the
reliability than the width function and cumulative area rela- cartometric and photogrammetric DEMs, with standard devi-
tionship, which cannot be determined with high confidence. ation increasing for increased grid spacing using both ap-
Since the width function characterizes the catchment routing proaches. This increase in standard deviation with grid spacing
behavior (through the unit hydrograph), this study suggests is to be expected as the correlation between errors in elevation
that the catchment routing behavior would be poorly estimated of adjacent grid points decreases with increased separation.
from published DEMs. Furthermore, the poor fit for the cu- The coefficient of variation of slope estimates decreases with
mulative area diagrams suggests that the region of saturation decreasing resolutions as the slopes are better estimated by
predicted for the saturation excess runoff generation mecha- lower resolution DEMs for both DEM types.
nism would be poorly estimated from the published DEMs The results from an evaluation of the empirical relationship
[Hemantha and Willgoose, 1996]. presented by Gyasi-Agyei et al. [1995] gave similar results for
3.3. Evaluation of DEM Accuracy for Hydrologic Purposes both approaches and show that the ratio of average pixel drop
by Empirical Relationship to vertical resolution is greater than or equal to unity in all but
Table 1 lists the data required for evaluating the empirical one case, that is, the photogrammetric DEM on a 6.25 m grid.
relationship of Gyasi-Agyei et al. [1995] along with the results Thus this relationship would indicate that the cartometric
and estimates of the standard deviation of relative errors in DEMs and photogrammetric DEM on a 12.5 m grid could be
elevation. used to extract the catchment and drainage network ade-
The average slope for the study site increased as the grid quately. However, the plots of catchment and stream network
spacing increased for both the cartometric and photogrammet- in Figures 3 to 6 would indicate otherwise, suggesting that the
ric DEMs. However, the standard deviation of errors in slope empirical relationship of Gyasi-Agyei et al. [1995] was either
decreased as the grid spacing increased for both DEM types. too loose or not a sufficient condition for adequacy but was
This decrease in standard deviation of slope is to be expected only a necessary condition.
for the photogrammetric DEM given the amount of high- The empirical relationship of Gyasi-Agyei et al. [1995] in (1)
frequency noise but is more difficult to explain for the carto- can be rearranged to identify the optimal horizontal resolution
metric DEM given the smoothness of the surface. However, D x of a DEM as
 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY 2267

s Dz Since the width function characterizes the catchment routing

Dx . (3) behavior, given by the unit hydrograph, this investigation sug-
a
gests that the catchment routing behavior would be poorly
If the horizontal resolution is higher (i.e., D x is smaller) than estimated from published DEMs. Furthermore, the poor fit for
the right-hand side, then the drainage network cannot be re- the cumulative area diagrams suggests that the region of sat-
liably extracted. A horizontal resolution equal to the right- uration predicted for the saturation excess runoff generation
hand side is therefore the finest DEM resolution from which mechanism will also be poorly estimated from the published
the drainage network can be extracted. If a high-resolution DEMs.
DEM is used to analyze drainage, only those characteristics of In contrast, the slope-area relationship was found to be less
the network with a horizontal length scale greater than D x are sensitive to both catchment shape and size, as well as stream
reliable, any finer detail is an artifact of the noise in the DEM. network, with all DEMs and catchments providing a relation-
D 2x might also be used to determine the threshold area for ship which fell outside the 90% confidence limits for less than
extraction of the drainage network. 40% of the relationship. This suggests that the slope-area re-
It is one thing to be able to extract a drainage network from lationship can be estimated from existing DEM sources more
a DEM; it is another thing entirely for this network to reflect reliably than other geomorphic relationships.
the underlying physics rather than DEM noise. For our carto- It is believed that these results are indicative of the errors
metric and photogrammetric DEMs the critical resolutions which are likely to exist when extracting hydrogeomorphic
were 5 to 10 m and 10 m, respectively. These values should be parameters from published DEMs around the world, since
viewed with some caution since Gyasi-Agyei et al. [1995] only many are derived by interpolating from 1:25,000 (or there
determined the threshold in (1) to within a factor of 10. about) scale topographical maps using the blue lines for drain-
Furthermore, the resolution is likely to vary across the catch- age enforcement [Florinsky, 1998]. The conclusions with regard
ment with varying slope. to photogrammetric DEMs may be considered indicative of
the loss of accuracy that might be obtained, but they are by no
means definitive, as digital photogrammetry software is rapidly
4. Conclusions improving. However, the characteristic noisy surface with short
We have shown that the catchment sizes and stream net- correlation length, leading to low slope accuracy and very ir-
works determined by published DEMs may be significantly regular networks, appears in our experience to be a problem
different from the ground truth. This is especially the case if for all existing digital photogrammetry packages, and in all
looking at a small catchment, where there is a localized error cases some form of filtering is required to remove this noise. At
in the elevation directing a major streamline in a wrong direc- this stage we cannot say definitively what the effect of the noise
tion. An example of where this occurred in the investigated and filtering might be on hydrologic characteristics.
study site is near the large farm dam. However, it has been Visually, the cartometric DEMs investigated were as smooth
shown that the published DEMs should be adequate to de- as the DEM from the ground survey data, with some loss of
scribe the catchment and stream network, providing there are fine detail, while the unfiltered photogrammetric DEMs pre-
no localized effects on elevations. It was also found that the sented a rough surface. Therefore photogrammetric DEMs
grid spacing of the DEM had no significant effect on the ability require some sort of filtering and editing before use. The use of
of the DEM to identify the catchment and the overall stream 5 3 5 Gaussian filter was shown to improve the visual appear-
network, with only the finer detail changing, provided the same ance of the surface; however, research into a large number of
base data is used for gridding. Therefore this study indicates DEMs for different terrains and different grid sizes with var-
that published cartometric and photogrammetric DEMs may ious filters is required before any recommendation may be
be used for determination of catchments and stream networks made about postprocessing requirements.
with caution by comparing the catchment and major stream In order to make concrete conclusions regarding the appli-
network defined from the DEMs with that observed from a site cability of results to DEMs around the world, it would be
inspection. necessary for case studies of this type to be made in diverse
The maximum horizontal resolution for which the details of locations. This work sets out a framework for such an analysis
the drainage network are reliable is related to both the vertical and justifies the need for both high-precision ground truth data
accuracy of the DEM and the slope. In order to predict this sets and further comparisons by highlighting the potential prob-
maximum horizontal resolution for a DEM, it is necessary to lems with published DEMs generated using current methods.
estimate the vertical accuracy. If the vertical accuracy is con-
sistent throughout the DEM, independent of elevation and
Acknowledgments. The provision of ground survey data by Ian
slope, then the horizontal resolution will be governed by the Marshall and Associates, Registered Surveyors and Subdivisional Plan-
topography in the flattest regions of the catchment. ners, Cessnock, and published DEMs by the Land Information Centre,
We have shown that both the width function and cumulative Bathurst, are gratefully acknowledged. Comments made by George
area relationship were very sensitive to elevation errors and to Kuczera are also gratefully acknowledged.
the source of the elevation data. The width function had rel-
atively wide confidence limits. Moreover, even though one of
the catchments identified from the cartometric DEMs was very References
similar to that of the ground truth, it still lay outside the 90% Beven, K. J., and M. J. Kirkby, A physically based, variable contribut-
confidence limits for more than 60% of the relationships. This ing area model of basin hydrology, Hydrol. Sci. Bull., 24, 43– 69, 1979.
suggests that these hydrologic properties will be poorly esti- Bolstad, P. V., and T. Stowe, An evaluation of DEM accuracy: Eleva-
tion, slope, and aspect, Photogramm. Eng. Remote Sens., 60(11),
mated from published DEMs. For the cartometric DEMs this 1327–1332, 1994.
may have been a result of the extensive parallel streamlines on Bras, R. L., and I. Rodriguez-Iturbe, Random Functions and Hydrology,
hillslopes. 559 pp., Addison-Wesley, Reading, Mass., 1985.
2268 WALKER AND WILLGOOSE: EFFECT OF DIGITAL ELEVATION MODEL ACCURACY

Felicı́simo, A. M., Parametric statistical method for error detection in Moore, I. D., R. B. Grayson, and A. R. Ladson, Digital terrain mod-
digital elevation models, J. Photogramm. Remote Sens., 49(4), 29 –33, elling: A review of hydrological, geomorphological and biological
1994. applications, Hydrol. Processes, 5, 3–30, 1991.
Florinski, I. V., Combined analysis of digital terrain models and re- O’Callaghan, J. F., and D. M. Mark, The extraction of drainage net-
motely sensed data in landscape investigations, Prog. Phys. Geogr., works from digital elevation data, Comput. Vision, Graphics, Image
22(1), 33– 60, 1998. Process., 28, 323–344, 1983.
Fryer, J. G., M. H. Elfick, R. C. Brinker, and P. R. Wolf, Elementary Robinson, G. J., The accuracy of digital elevation models derived from
Surveying, 7th ed. (SI adaption), 472 pp., HarperCollins, New York, digitised contour data, Photogramm. Rec., 14(83), 805– 814, 1994.
1987. Sloan, S. W., A fast algorithm for generating constrained delaunay
Fryer, J. G., J. H. Chandler, and M. A. R. Cooper, Short communi- triangulations, Comput. Struct., 47(3), 441– 450, 1993.
cation on the accuracy of heighting from aerial photographs and Surkan, A. J., Synthetic hydrographs: Effects of network geometry,
maps: Implications to process modellers, Earth Surf. Processes Land- Water Resour. Res., 5(1), 112–128, 1969.
forms, 19, 577–583, 1994. Tarboton, D. G., R. L. Bras, and I. Rodriguez-Iturbe, The analysis of
Gyasi-Agyei, Y., G. Willgoose, and F. P. De Troch, Effects of vertical river basins and channel networks using digital terrain data, Tech.
resolution and map scale of digital elevation models on geomorpho- Rep. 326, Ralph M. Parsons Lab., Dep. of Civ. Eng., Mass. Inst. of
logical parameters used in hydrology, Hydrol. Processes, 9, 363–382, Technol., Cambridge, 1989.
1995. Willgoose, G., R. L. Bras, and I. Rodriguez-Iturbe, A physical expla-
Hemantha, J., and G. Willgoose, Scaling and scale-invariant behaviour nation of an observed link area-slope relationship, Water Resour.
in a geomorphically based conceptual model for runoff generation, Research, 27(7), 1697–1702, 1991.
Eos Trans. AGU, 77(46), F233, 1996. Wolock, D. M., and C. V. Price, Effect of digital elevation model map
Hutchinson, M. F., A new procedure for gridding elevation and stream scale and data resolution on a topography-based watershed model,
line data with automatic removal of spurous pits, J. Hydrol., 106, Water Resour. Res., 30(11), 3041–3052, 1994.
211–232, 1989. Zhang, W., and D. Montgomery, Digital elevation model grid size,
La Barbara, P., and G. Roth, Invariance and scaling properties in the landscape representation, and hydrologic simulations, Water Resour.
distribution of contributing area and energy in drainage basins, Research, 30(4), 1019 –1028, 1994.
Hydrol. Processes, 8, 125–135, 1994.
Lane, S. N., J. H. Chandler, and K. S. Richards, Developments in J.P. Walker and G.R. Willgoose, Department of Civil, Surveying and
monitoring and modelling small-scale river bed topography, Earth Environmental Engineering, University of Newcastle, Callaghan, NSW
Surf. Processes Landforms, 19, 349 –368, 1994. 2308, Australia. ([email protected].; [email protected].
Mikhail, E. M., and F. Ackerman, Observations and Least Squares, IEP, edu.au.)
New York, 497 pp., 1976.
Moore, I. D., and R. B. Grayson, Terrain-based catchment partitioning
and runoff prediction using vector elevation data, Water Resour. (Received August 7, 1998; revised January 29, 1999;
Res., 27(6), 1177–1191, 1991. accepted February 1, 1999.)