YGON APPROACH
e product. Typically,
int of the product is
urce errors at each
lata processing flow
le error source in an
f the same scene to
ons are investigated
inary results from a
source data bases,
which includes the
| bias can be a
rough spatial data
ipon the nominal
ith boundaries and,
hat exists in reality.
;ontiguous classes
n a construct of
ter 1991). The
delineate classes in
are subjective. The
dependent on the
Is, the scale of
allocated classes.
that, in interpreting
areas yielded less
that classification
han positional error
maps with fixed
a binary (yes/no)
in a measure of the
eing at that location
ation between the
> a cartographic line
| dependent on the
classes. However,
rring’ of a boundary
on of the extent of
geographic detail present as this zone is less distorted
than the original boundary.
Gong and Chen (1992) deal with methods that may be
used to determine, represent and display boundary
uncertainties in categorical (area-class) maps. They state
that it is impossible to tell the most accurate realisation
and they suggest ways in which the most probable
boundary could be determined using curve-fitting
techniques and blending functions. They generated a
number of realisations of land use categories from
classification and subsequently manually digitised the
map. Other authors (Maffini et a/ 1989; Dutton 1992)
have investigated the positional uncertainty of boundaries
resulting from the manual digitising of land cover maps.
The problem that arises in these cases is the introduction
of an additional interpretational process within data
processing, i.e. classification and digitisation. This paper
suggests that a framework for determining both local and
boundary errors resulting from multiple realisations of the
same phenomenon be resolved prior to raster to vector
conversion within the spatial database. In this case, the
operators are responsible for determining the classes
without the need to digitise each determination. Using the
GIS overlay function, the level of agreement can be
assessed, most probable class boundary positions
derived and then a once only vectorisation of the
polygons for entry into the GIS data base carried out.
1.3 Accuracy Assessment Used in Remote Sensing
Present problems with accuracy assessment of thematic
maps are that there is no indication of the variation of
land use / land cover from the sampled data within each
class. Each location on the ground has been allocated to
a particular class and the assignment of the appropriate
map label for some locations is ambiguous (Gopal and
Woodcock 1993).
The importance of accuracy assessment for remotely
sensed data is well recognised particularly when the data
may be used in a GIS (Congalton & Green 1993). Allan
et al (1996) detail previous literature concerning the
methods employed in assessing the accuracy of remotely
sensed data. In summary, the error (confusion) matrix
(Aronoff 1982), determination of producers and
consumer's risk (errors of commission and omission)
using row and column marginals of the matrix (Story &
Congalton 1986) and compensation for chance
agreement in the classes - Kappa coefficient of
agreement (Rosenfield & Fitzpatrick-Lins 1986) have
been used. Sampling designs have been investigated by
a number of authors (Congalton 1988; Stehman 1992).
2. ERROR SOURCES IN REMOTELY SENSED DATA
In the production of a thematic map the user needs to
have a knowledge of the error in the position and
labelling of the derived classes. As this product (map)
may be only one layer used in the GIS, quantitative
measures of error are necessary at its source before
progressing to error propagation in the GIS processing
flow. Until these source errors are thoroughly examined
and measured the utility of remotely sensed data as an
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
appropriate and valuable information source within GIS is
restricted. Goodchild et al (1994) suggest that
consistency, or replicability, of processes and realisations
have rarely been executed in practice.
An accuracy assessment of errors in a thematic map
should provide details concerning their nature, frequency,
magnitude and source (Gopal & Woodcock 1993). A
conceptual framework has been described by Veregin
(1989) in which he sets out 'a hierarchy of needs for
modeling error in GIS operations. In a five level
hierarchy, level 1 is concerned with the identification of
error sources, level 2: error detection and measurement,
level 3: error propagation modelling, level 4: strategies
for error management, and level 5: strategies for error
reduction. A number of authors have set out the various
stages in the spatial data ‘life-cycle’ in which error may
be introduced (Aronoff 1989; Lunetta et al 1991; Collins
& Smith 1994). They are summarised as follows:
e Data acquisition: geometric aspects, sensor systems,
platforms, ground control.
e Data input (processing): geometric registration and
resampling.
e Data analysis: classification systems, data
generalisation.
e Data conversion: raster to vector.
Data output: positional and attribute errors.
Data usage and interpretation: insufficient
understanding and incorrect use of data.
Lunetta et a/ (1991) identify source errors (Veregin's
Level 1) for each stage. They point out that error
accumulates for each successive stage but also may be
introduced within any stage.
The proposed framework, shown in Figure 1, describes
an approach to determine the degree to which stages in
the GIS information processing flow contribute to the
overall error in the data layer. It integrates the
hierarchical approach proposed by Veregin (1989) within
the framework of GIS processing with the potential error
sources suggested by Lunetta et al (1991).
Specifically, the approach is to detect and measure the
uncertainties in two stages of the processing flow: data
processing and data analysis after having identified the
source errors. At the source level in the framework, it
acknowledges that error may accumulate from one stage
to the next but may also contribute separately at each
stage. Two stages in the data processing flow are
selected to examine their respective contributions to the
determination of class accuracy assessment.
At the next level in the framework, the detection and
measurement of error phase, operational constraints are
imposed on the interpreters to elicit quantitative
estimates of error and its spatial variability. These
constraints may be the adoption of the same image
classification technique and the division of the same
image into a predetermined number of classes. Polygons
in disagreement are formed based on the realisations
from each image interpreter. The polygon characteristics
can be measured and aggregated based on a threshold
established for areas, perimeters, shapes or a
177
