Professional Documents
Culture Documents
Research Article
1. Introduction
Although positional uncertainty and error assessment have been recognised as
fundamentally important in geographic analysis (Atkinson 1999), very few countries
have digital geospatial data standards dealing with these issues. Instead, map accu-
racy standards relating error to original source scale are often the only indication
of positional accuracy, even though such estimates are, at best, indirectly relevant
to digital data sources. Consequently, without adequate accuracy reporting stand-
ards, data users can unknowingly be provided with datasets that do not meet even
the less pertinent map accuracy standards (Van Niel and McVicar 2001). This can
be particularly problematic because users may assume that spatial data provided by
national or state mapping agencies are reliable.
When positional accuracy estimates of maps or digital geospatial data are calcu-
lated, they are often quanti ed by comparing positions of well-de ned points in a
test dataset to the positions of identical points in a reference dataset of higher
accuracy (e.g. FGDC 1998a, b, Van Niel and McVicar 2001). However, for point-
based assessment to be universally relevant to digital geospatial data, positional
error throughout lines and polygons must be directly related to the error of their
primitive points at identi able locations (Leung and Yan 1998). Errors of point, line,
and polygon GIS features could then be estimated from a point-based positional
Internationa l Journal of Geographica l Informatio n Science
ISSN 1365-881 6 print/ISSN 1362-308 7 online © 2002 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/13658810210137022
456 T . G. Van Niel and T . R. McV icar
error model, thus simplifying the de nition of accuracy standards of GIS datasets
(Leung and Yan 1998). This, in fact, is the basic assumption behind the point-based
methods adopted by the USA in their National Standard for Spatial Data Accuracy
(NSSDA) and those likely to be adopted by the international geospatial community.
However, the validity of the point-based testing method for estimating spatial
error of linear features has been questioned, partly because error distributions of
well-de ned points are thought to contain diVerent error distributions than the links
between them (Goodchild and Hunter 1997). Likewise, line segment endpoints have
been determined to contain disproportionatel y higher error terms than midpoints
(Shi 1998). Also, since linear networks often do not contain well-de ned points,
a simple comparison of features to either a linear reference of higher accuracy
(e.g. Goodchild and Hunter 1997, Tveite and Langaas 1999) or to internal digital
data geometry (e.g. Duckham and Drummond 2000, Veregin 2000 ) may be more
appropriate in many cases.
Considering these issues, assessment of point-based testing methods with respect
to alternate techniques is relevant to geospatial data users worldwide, especially
where positional accuracy standards are yet to be de ned. In this study, we compare
point-based positional accuracy results de ned from both US geospatial data stand-
ards and Australian map accuracy standards to results from a linear feature-based
testing method developed by Goodchild and Hunter (1997). A digital road network
previously determined to be non-compliant with Australian map accuracy standards
from a point-based assessment was used as a case study because of its ability
to provide:
1. a large range of positional errors facilitating comparison of point- and
line-based errors; and
2. speci cally relevant information for the anticipated de nition of digital
geospatial positional accuracy standards in Australia.
Speci cally, error distributions within the dataset and the eVect of the number
of points used in the calculation are examined. A brief discussion of US and
Australian spatial data accuracy standards and the theoretical basis of positional
errors determined from point- and line-based testing methods is provided in §2.
2. Background
2.1. Spatial data accuracy standards
The US Bureau of Budget (USBB) issued the United States National Map
Accuracy Standards (USNMAS) in 1941 (updated in 1947), which require that not
more than 10% of well-de ned points be in error by more than an allowable distance
calculated from source map scale (USBB 1947). The well-de ned points are compared
to identical points from an independent source of higher accuracy. These map
accuracy standards have been the only indicator of positional error in digital geospa-
tial datasets in the US until the Federal Geographic Data Committee (FGDC)
published the NSSDA, which outlines both new positioning accuracy standards and
accuracy reporting standards for digital geospatial data (FGDC 1998a, b).
The NSSDA provides a common reporting system for direct comparison of
datasets as well as a relevant estimation of digital geospatial data accuracy (FGDC
1998a) . Like the USNMAS, the NSSDA compares well-de ned points, but uses root-
mean-square error (RMSE) to estimate positional accuracy. Error is reported in
ground distances at the 95% con dence level, meaning that 95% of the positions in
Positional accuracy estimates f rom a linear network 457
the dataset should have an error that is equal to or smaller than the reported error
distance. These standards also de ne that at least 20 well-de ned points should be
used in any calculation of accuracy (FGDC 1998b) .
Conversely, Australia currently has no digital geospatial data accuracy standards,
but rather relies on map accuracy indicators to de ne positional accuracy estimates.
The National Mapping Council of Australia (NMCA) standards of map accuracy
test whether well-de ned points are within allowable error distances calculated from
map scale (see table 1). As the map scale indicator is not always a reliable estimate
of spatial accuracy in Australia (e.g. Van Niel and McVicar 2001), the de nition of
digital geospatial data standards is likely to be considered in the near future.
Table 1. Allowable error distances calculated for common map scales by Australian and US
map accuracy standard speci cations.
global context. However, because these two techniques diVer in how they measure
the closeness of coordinates to the reference at a fundamental level, they are expressed
more precisely as diVerent categories of positional error.
The error calculated from well-de ned points results in what has traditionally
been de ned as absolute positional error. This error, henceforth termed ‘absolute’
positional error, measures closeness by matching primitives (i.e. points) of the features
being assessed to identical primitives in a reference dataset. That is, matching points
to identical points. When Goodchild and Hunter (1997) developed the buVering
technique for measuring linear positional accuracy, they subsequently created an
error estimate that is diVerent from the traditional absolute accuracy measure. This
error, henceforth termed ‘functional’ positional error, like a mean distance between
lines (MDL) calculation, measures closeness of features by de ning the shortest
distance between primitives (Matos and Goncalves 1998 ). That is, measuring the
distance between points along a line to the closest point along a reference line. Thus,
the shortest distance criterion will associate non-matching points (Matos and
Goncalves 1998). Absolute positional accuracy, as de ned in this paper, is more
restricted and thus identi able from functional positional accuracy because all of the
locations tested must match identical locations in the reference dataset.
Although positional error assessment of linear features based on buVering tech-
niques provides fundamentally diVerent output than the traditional well-de ned
point-based technique, this issue is seldom addressed in detail (for an exception, see
Matos and Goncalves 1998). This omission in the literature can increase the likeli-
hood of, for example, improper calculations of error propagation through a geospatial
model because users may believe that these error estimations are interchangeable.
On the contrary, these errors carry very diVerent meanings in an operational sense.
The necessary removal of the restriction on matching points in a linear buVering
assessment predetermines that the functional error at every point along a line will
be de ned by the closest distance between matching line features ( gure 1(a)).
Consequently, functional error at any one point can never be greater than its
associated absolute error and results in a signi cant bias between these diVerent
positional accuracy metrics.
This bias, in turn, is explained by at least two factors:
1. Consistently shifted measurements of absolute versus functional positional
error, henceforth called the absolute/functional oVset.
2. DiVerent error distributions at well-de ned points versus non-descript points
along the links between them, henceforth called the point/link oVset.
We suggest that the absolute/functional oVset generally plays a substantially larger
role in determining this overall bias in road networks than does the point/link oVset.
An example of these two components of the overall bias is described in gures 1(a)
and 1 (b), respectively.
Figure 1(a) illustrates the absolute/functional oVset between two lines, a grey
reference dataset and a black test dataset, and ve matching non well-de ned points
along these two lines. The matching lines and points are identical to each other
except for an arbitrary geographic shift included for demonstration purposes. This
example is purely theoretical, for in practice, non-descript points like the ones
illustrated, by de nition, cannot be matched.
Dashed lines connecting light and dark circles illustrate absolute error (e.g. the
line segment between A and A² , denoted AA² ), while thin solid lines connecting light
Positional accuracy estimates f rom a linear network 459
Figure 1. Components of the potential error bias between positional error results determined
from (well-de ned) point-based testing methods to those determined from linear
buVering methods. In (a), absolute and functional positional error is demonstrated for
the thick grey line. In (b), absolute positional error is limited by the functional error
of its least accurate link. This error is exacerbated if one of the road intersections is
displaced. See article text for full details.
circles to dark squares illustrate functional error (e.g. the line segment between A
and A¾ , denoted AA¾ ). At any individual point, the functional error will always be
equal to or less than the absolute error, and in many instances, it can be much
smaller (e.g. CC¾ % CC² , and DD¾ % DD² ). The absolute/functional error oVset exists,
then, because the shortest distance between lines will not often represent identical
points within the lines themselves. Consequently, traditional point-based methods of
determining positional accuracy should report consistently higher errors than com-
parative results from a linear buVering method because they report diVerent metrics.
In gure 1(b), one potential explanation for the point/link oVset is demonstrated.
The error in this example is caused by the compounding of errors at the source of
well-de ned points, the road intersections. In cartographic terms, this displace-
ment of intersections is manifested as a loss of positional accuracy in favour of
logical consistency. This means that in digital datasets, although line intersections
460 T . G. Van Niel and T . R. McV icar
Figure 2. The DTDB roads data over the CIA. 1:100 000 scale mapsheets are displayed in
thick black lines and labelled by map name, 1:50 000 scale mapsheets are displayed in
thin black lines and labelled by map number. The roads are displayed with grey lines,
and the CIA area is shaded dark grey.
Figure 3. DGPS roads dataset at CIA in thin grey lines. Black points are the 83 road
intersections, the thick black line is the CIA boundary and the thick grey lines are
1:50 000 scale mapsheet boundaries introduced in gure 2. The labelled intersections
illustrate point and link sets de ned in §4.3
accuracy of the DGPS is expected to be within 1.5 m. The DGPS performance was
checked by retracing large portions of road at diVerent times during the day. The
mean diVerence between the 16 km portion of the traced and retraced road de ned
by 1161 DGPS positions was 0.96 m, with 99.8% of the measured distances within
2 m of each other. From this, the DGPS was shown to be performing within expected
speci cations.
Table 2. NSSDA and NCMA positional accuracy statistics for the DTDB roads dataset
generated using the point-based testing methodology.
*NSSDA 95% stands for the US National Standard for Spatial Accuracy estimate for the
95% con dence interval about the RMSE.
464 T . G. Van Niel and T . R. McV icar
example, a 170 m section of road in mapsheet 80271 was in error by more than
325 m, or over 13 times the NMCA allowable error.
These results show that the absolute error estimate of the DTDB dataset is much
greater than the functional error estimate. That is, 95% of the dataset is functionally
no farther than about 50 m from truth as determined by the whole line network,
whereas the 95th percentile of the absolute error is calculated to be approximatel y
105 m, or over twice the functional error. This analysis also con rmed that the
positional errors in the DTDB dataset are not spatially random, but contain a great
deal of spatial correlation ( gures 4(a-d )). Although large portions of the dataset are
above the NMCA standard 25 m ( gure 4(a)), the largest errors are consistently
found within one 1:50 000 scale mapsheet (mapsheet 80271) ( gures 2, 4(b), 4(c), and
4(d)). Directional distribution and location of errors between the test and reference
datasets showed no sign of either transformation gain or directional bias (Matos
and Goncalves 1998).
Figure 4. Linear accuracy results for diVerent buVer distances between the DTDB and DGPS
roads datasets over the CIA. Grey lines fall within the buVer, black lines fall outside
of the buVer. Results are displayed for (a) 25 m (NMCA allowable distance), (b) 50 m,
(c) 100 m, and (d ) 200 m.
466 T . G. Van Niel and T . R. McV icar
Figure 5. Mean positional accuracy based on the diVerence between randomly combined
DTDB and DGPS road intersection points at CIA. The mean accuracy of the estimate
increases while the 90% con dence interval about the mean decreases as more points
are used.
the overall bias was calculated by comparing absolute error at the well-de ned
points versus functional error along the links, this bias contains two constituent
parts in unknown proportions. These are:
1. DiVerences between measuring absolute error versus functional error, or the
absolute/functional oVset.
2. DiVerences in error distributions at well-de ned points versus non-descript
points, or the point/link oVset.
The proportion of the point/link oVset could be inferred by eliminating the
absolute/functional oVset from the comparison of accuracy estimates. This was
accomplished by measuring the same type of error (i.e. functional error) at both
points and links. The absolute/functional oVset could then be isolated by subtracting
the point/link oVset from the overall bias.
Likewise, 83 associated road link sets were de ned for each of the roads datasets.
These were comprised of all the links that join the intersections in each point set.
For example, in gure 3, the DGPS link set for intersection A includes line segments
AB, AC, and AD. The mean diVerence between DTDB and DGPS links was also
calculated for each link set. In this study, the potential number of lines within a link
set was four.
The mean functional error calculated for each link set was then subtracted from
the mean absolute error calculated for each point set. This analysis, denoted the
‘absolute point/functional line assessment’ reveals the overall bias ( gure 6(a)).
Because the error of the link sets was subtracted from the error of their associated
point sets, instances where mean absolute error is higher than functional error are
represented by positive values.
Numerous non-descript points are used in the estimation of mean functional
error along links, whereas the mean absolute error determined for any point set only
uses a few points (a maximum of 5). This sample size mismatch makes it possible
for the mean absolute error estimate of a point set to be smaller than the mean
functional error estimate of its associated link set. This was never the case when the
sample size mismatch was eliminated. When absolute and functional errors were
compared at the same points (rather than the mean diVerence between point and
link sets), absolute error was always larger than functional error. For the 83 common
intersections, the minimum diVerence (calculated as absolute–functional) is 0.01 m,
while the maximum diVerence is 354.60 m. The median and mean are 8.31 m and
17.53 m, respectively, indicating that outliers have skewed the distribution in a
positive direction.
Figure 6. Histograms of (a) the overall error distribution bias containing both absolute/
functional and point/link oVset components, (b) the point/link oVset component of
the overall bias, and (c) the absolute/functional oVset component of the overall bias.
Positional accuracy estimates f rom a linear network 469
Figure 7(a) shows an example where the absolute error at the well-de ned point
is much larger than the functional error along most of its two adjoining roads. In
this case, the error is most likely related to either data processing or temporal
variance. That is, the road intersection has either been moved since the creation of
the DTDB data, or it was misallocated in the data processing procedure. Processing
error is likely the cause since this problem is typical of a straightening eVect, where
the end of one road is extended until it intersects with another road. This generally
results in inappropriate positioning and simpli cation of shape. The point-based
method is more sensitive to outliers such as these. Although these outlier errors are
not the typical case in this dataset, they are not extremely rare either. This example
clearly demonstrates where data production can exacerbate the diVerence between
absolute and functional error by maximising the error at the well-de ned point.
Conversely, gure 7(b) shows a rare example where the absolute error at one
well-de ned point underestimated the mean functional error of its adjoining roads.
The large positional inaccuracies seen here are related to a missing road intersection,
which in turn is very likely caused by either interpreter error or processing error.
The missing intersection was probably either not digitised by the interpreter or was
deleted in data processing, resulting in a shift of one road by over 100 metres east
of its DGPS location. Underestimates of error of this kind are rare for this dataset.
The intersection shown in gure 7(c) demonstrates a case where the absolute
error at this one well-de ned point and the mean functional error along its two
adjoining road links are very similar. Much of this absolute error is probably caused
by either interpreter or data processing error, but could also be symptomatic of
imprecise measurement. Of note is the dominance of the least accurate adjoining
470 T . G. Van Niel and T . R. McV icar
Figure 7. Examples of well-de ned point and linear link errors (a)–(e) for the DTDB dataset
and their eVect on error estimation of absolute and functional positional error. DTDB
roads and road intersections are displayed in black, while DGPS roads and road
intersections are displayed in grey. Positions of these examples within the CIA are
illustrated in ( f ).
Positional accuracy estimates f rom a linear network 471
roads’ functional error term in the magnitude of the well-de ned points’ absolute
error term. Although the magnitude of error in this example is uncommon, this
relationship between absolute and functional error is typical of this dataset.
As seen in gures 7(d) and 7(e), this limitation of road intersection accuracy
probably increases the absolute error at these well-de ned points slightly. Typically,
the error at the road intersection usually either matches the error of the least accurate
road ( gure 7(d)), or combines to result in an error that is greater than either of the
adjoining roads ( gure 7(e)). In the special case where one road link is very close to
its true position (e.g. gure 7(d)), not only is the absolute error term at the well-
de ned point dominated by the functional error term of the least accurate road, but
these absolute and functional error terms will also be almost equal. When the
functional error terms of both adjoining roads are signi cantly far from the truth,
then both functional error terms combine to increase the absolute error term at the
well-de ned point. For example, although the functional error along the two roads
in gure 7(e) is approximately 11 m and 24 m, the diVerence between the two road
intersections (absolute error) is approximately 38 m. Due to the nature of these
overestimates, it seems likely that they would be found ubiquitously in digital road
networks, at least in Australia given common data sources, e.g. aerial photographs ,
and similar data processing steps.
5. Conclusions
From this study there were four noteworthy ndings with respect to the de nition
and interpretation of geospatial positional accuracy standards:
1. Absolute and functional positional errors were not interchangeable.
2. Absolute error at well-de ned points provided a relevant measure of the
absolute error of its links.
3. Displacement of road intersections caused the error of well-de ned points to
be slightly higher than non-descript points.
4. The NSSDA standard of 20 well-de ned points provided a reasonable
estimation of absolute error.
Overall absolute error estimation at well-de ned points was found to be greater
than two times the overall functional error estimation along linear features for the
DTDB dataset. This quanti cation of bias helped to determine that absolute and
functional positional errors are fundamentally diVerent, and although not totally
unrelated, they are not interchangeable. It was suggested that this bias was caused
by two constituent components:
1. DiVerences in the errors measured due to dissimilar closeness criterion (the
absolute/functional oVset).
2. DiVerences due to dissimilar error distributions at well-de ned points and the
links between them (the point/link oVset).
In this DTDB dataset, it was found that the absolute/functional oVset was much
greater than the point/link oVset, meaning that most of the diVerences between
point-based and line-based errors are due to the fact that they measure diVerent
types of positional error. Although the point/link oVset was small, it was determined
to be positive, revealing that well-de ned points for this dataset were biased towards
higher errors than were links. Importantly, the magnitude of this point/link oVset
shows that the diVerences in error distributions at points versus along links is only
472 T . G. Van Niel and T . R. McV icar
minor, suggesting that absolute error determined at well-de ned points is very much
similar to the absolute error of the links themselves (although this could not be
measured directly).
The assessment of the number of points used in the calculation of absolute
positional error at well-de ned points illustrated that the NSSDA standard of 20
points resulted in a reasonable estimation for the DTDB dataset. The con dence
intervals calculated for this analysis show that the sensitivity to outliers was suY-
ciently suppressed when the NSSDA standard of 20 points was used in the estimation
of absolute error for this dataset. Also, since this particular dataset contained a large
range of positional errors, we expect this standard of 20 points to be suYcient for
similar datasets across Australia.
Although this study suggests that the traditional point-based method provides a
suYcient estimation of positional accuracy for the entire road network, this relation-
ship might not be the same for more complex linear networks (e.g. a river network).
Also, because many linear networks do not contain numerous well-de ned points,
we recommend that an allowance be made for the calculation of both absolute and
functional error terms in the de nition of Australian geospatial standards. Most
importantly, GIS data users must realize that the positional accuracy estimates
resulting from point- and line-based techniques should be interpreted diVerently.
This may be particularly true for determining error propagation when combining
diVerent GIS datasets.
Acknowledgments
This research was funded by the Cooperative Research Centre for Sustainable
Rice Production and CSIRO Land and Water. Thanks to Arun Tiwari, Janelle
Dufty, and Greg Robertson of Coleambally Irrigation Cooperative Limited for access
to GIS data, Warrick Dawes, Susan Cuddy, Brent Henderson, and Kimberly Van
Niel for their helpful editing suggestions, GeoV Beecher and Brian Dunn of NSW
Department of Agriculture for use of their GPS, and to Dr Gary Hunter and Karin
Reinke, University of Melbourne for free access to the linear feature accuracy testing
code (www.sli.unimelb.edu.au/people/gjh_notes/linetest.htm). Thanks also to the
anonymous reviewers, whose input improved the text. The point-based positional
accuracy extension to ArcView used in this study is also available on the internet
(http://gis.esri.com/arcscripts/details.cfm?CFGRIDKEY 5 C03BD2F2-680F-11D4-
943200508B0CB41 9).
References
Atkinson, P. M., 1999, Geographical information science: geostatistics and uncertainty.
Progress in Physical Geography, 23, 134–142.
Duckham, M., and Drummond, J., 2000, Assessment of error in digital vector data using
fractal geometry. International Journal of Geographical Information Science, 14, 67–84.
FGDC, 1998a, Geospatial positioning accuracy standards. Part 1: Reporting methodology,
Federal Geographic Data Committee, FGDC-STD-007.1-1998.
FGDC, 1998b, Geospatial positioning accuracy standards. Part 3: National standard for
spatial data accuracy, Federal Geographic Data Committee, FGDC-STD-007.3-1998.
Goodchild, M. F., and Hunter, G. J., 1997, A simple positional accuracy measure for linear
features. International Journal of Geographical Information Science, 11, 299– 306.
Humphreys, L., Van Der Lely, A., Muirhead, W., and Hoey, D., 1994, The development of
on-farm restrictions to minimise recharge from rice in New South Wales. Australian
Journal of Soil and Water Conservation, 7, 11–20.
Positional accuracy estimates f rom a linear network 473
Hunter, G. J., 1991, Glossary of terms for land and Geographic Information Systems,
Australasian Urban and Regional Information Systems Association (AURISA),
Technical Monograph No. 2 (Revised).
Leung, Y., and Yan, J., 1998, A locational error model for spatial features. International
Journal of Geographical Information Science, 12, 607–620.
Matos, J., and Goncalves, A., 1998, Measurement and analysis of positional errors, In
Proceedings of the 8th International Symposium on Spatial Data Handling, Vancouver,
Canada (International Geographical Union: Geographic Information Science Study
Group, Burnaby, B.C., Canada), 151–160.
Shi, W., 1998, A generic statistical approach for modelling error of geometric features in GIS.
International Journal of Geographic Information Science, 12, 131–143.
Tveite, H., and Langaas, S., 1999, An accuracy assessment method for geographical line data
sets based on buVering. International Journal of Geographical Information Science,
13, 27–47.
USBB, 1947, United States national map accuracy standards, US Bureau of the Budget,
Washington, DC.
Van Niel, T. G., and McVicar, T. R., 2001, Assessing positional accuracy and its eVects on
rice crop area measurement: an application at Coleambally Irrigation Area. Australian
Journal of Experimental Agriculture, 41, 557–566.
Veregin, H., 2000, Quantifying positional error induced by line simpli cation. International
Journal of Geographical Information Science, 14, 113–130.