A FRAMEWORK FOR THE ASSESSMENT OF LOCAL SPATIAL UNCERTAINTY USING A POLYGON APPROACH
R.C. Allan & G.P. Ellis
RMIT Centre for Remote Sensing & GIS
RMIT University
124 Latrobe Street, Melbourne, Victoria, Australia 3000
E-mail: rod.allan@rmit.edu.au
Commission lll, Working Group III/IV
KEY WORDS: Error, Classification, Accuracy, Recognition, Spatial Accuracy.
ABSTRACT:
For remotely sensed data to be effectively used in a GIS the user needs to know the reliability of the product. Typically,
in the production of a thematic layer in a land resource data base, an overall accuracy assessment of the product is
undertaken in which the user determines its fitness for use. However neither the magnitude of source errors at each
stage of data handling nor the within spatial variability is known from this assessment.
This paper proposes a methodology to elicit relative measures of error in the various stages of the data processing flow
and the extent of local spatial variability in the input data layer by identifying and then measuring the error source in an
iterative scheme. The process utilises overlays of independent realisations by image interpreters of the same scene to
create polygons in disagreement between the interpreters. Geometric characteristics of these polygons are investigated
to establish as to whether any changes in geometry are attributable to a particular source. Preliminary results from a
case study are discussed.
1. INTRODUCTION
An important but complex issue, when using Geographic
Information Systems (GIS) to integrate, analyse and
display spatial data, is the definition and quantification of
errors. With increasing emphasis placed on spatial
information processing, data are being used for purposes
they were never intended (Goodchild 1993). The resultant
products often have no indication as to their suitability for
use in the decision making process.
With the integration of disparate data sources required
for a GIS, error propagation and control throughout the
processes are not readily understood nor easily imposed.
1.1 Spatial Data Bases - A User's Perspective
For many applications, the effective use of spatial data
bases is dependent upon the data user who determines
its fitness for use (Chrisman 1994). A data user's own
perception of its worth for an application is based on
some a priori knowledge of the user about GIS and about
the data bases themselves (Coward & Heywood 1991).
Spatial data bases can represent multiple versions of
reality and the operation of a basic GIS function such as
generalisation for example, creates a less representative
version of reality. Indeed, for many users, these
operations on the data base are necessary to achieve the
required product.
The lack of any detailed knowledge of the extent to which
error is introduced and its magnitude, particularly at its
source, is one of the impediments in understanding error
propagation. Source errors enter in the data processing
flow at various stages and, importantly, not solely at the
176
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
data acquisition stage. With land resource data bases,
for example, the classification phase which includes the
operator's interpretative skills and bias can be a
significant source of error.
1.2 Local and Boundary Errors
The production of thematic maps through spatial data
processing is primarily based upon the nominal
categorisation of discrete classes with boundaries and,
by implication, is representative of what exists in reality.
In fact, the representation of contiguous classes
(polygons) is “nothing more than a construct of
cartographic convenience” (Trotter 1991). The
interpretive techniques employed to delineate classes in
the production of a thematic map are subjective. The
degree of subjectivity is largely dependent on the
heterogeneity of the image pixels, the scale of
representation and the number of allocated classes.
Cherrill and McClean (1995) found that, in interpreting
land cover change, smaller class areas yielded less
precise results. They also suggest that classification
(attribute) error is more significant than positional error
with land cover types.
The present problems in using maps with fixed
categorical attributes results in a binary (yes/no)
response to a spatial query rather than a measure of the
likelihood of a certain characteristic being at that location
(Lowell 1992). Boundary representation between the
classes, in this instance, would not be a cartographic line
but rather a transition zone of width dependent on the
spectral similarity of contiguous classes. However,
Goodchild (1994) asserts that the ‘blurring’ of a boundary
may give the user a false impression of the extent of
geographit
than the ol
Gong and
used to
uncertainti
that it is ii
and they
boundary
technique:
number c
classificat
map. Oth
have inve:
resulting f
The proble
of an ac
processin
suggests }
boundary
same phe
conversioi
operators
without th
GIS over
assessed,
derived z
polygons
1.3 Accur
Present p
maps are
land use
class. Ea
a particul
map labe
Woodcoc
The imp«
sensed ds
may be u
et al (1€
methods
sensed d
(Aronoff
consume
using rov
Congalto
agreemet
agreemel
been use
a number
2. ERRC
In the pr
have a
labelling
may be
measure:
progressi
flow. Unt
and mea