CER EE DE es EE SL
function of sensor scan angle — present in all images
acquired by wide field-of-view sensors — were removed
using a simple normalization technique (Barnsley and
Barr 1993). The data were geometrically registered to
the U.K. national grid using a standard, least-squares
polynomial fit and nearest-neighbour resampling.
The image was classified into ten principal land-cover
types (concrete roads, tarmac roads, tile-roof buildings,
concrete-roof buildings, metal-roof buildings, bare soil,
healthy grass, senescent grass, trees and open water)
using a standard, maximum-likelihood classification al-
gorithm. The training-set statistics indicate that these
classes had an average transformed divergence value of
99.90. The overall classification accuracy was found to
be very high (94.77%), with a Kappa coefficient of 0.90.
To permit direct comparison between the classified im-
age and the digital land-cover map, the original ten land-
cover classes were re-mapped (at the pixel level) into
five composite classes. This involved collapsing the two
road classes and the three building classes into two new
classes which shall be referred to as roads and build-
ings, respectively. The resultant, simplified land-cover
classification is shown in Figure 2.
STRUCTURAL COMPARISON OF THE
LAND-COVER DATA DERIVED FROM THE
DIGITAL MAP AND MULTISPECTRAL
IMAGE
Structural Description
Quantitative characterisation of the structural composi-
tion of the digital map data and the classified image was
performed using the syntactic pattern-recognition sys-
tem developed by Barr and Barnsley (1997). In this
system, raster data measured on either a nominal (i.e.,
class type) or an ordinal (i.e., class number) scale are
analysed to identify and label the complete set of dis-
crete (i.e., non-overlapping) regions present within the
scene (where a region is taken to mean one or more con-
tiguous pixels with the same class label or number). The
boundaries of these regions are derived using a simple
contour-tracing algorithm (Gonzalez and Woods 1993)
and encoded using Freeman chain-codes (Freeman 1975).
These data are used to derive information on a series of
structural features of the regions, such as their morpho-
logical properties (e.g., size and shape) and spatial re-
lationships (e.g., adjacency, containment, distance and
direction). This information is represented in a graph-
theoretic data-model, known as XRAG (eXtended Rela-
tional Attribute Graph; Barr and Barnsley 1997). In this
study, for reasons of space, we will restrict our attention
to an analysis of one morphological property, area, and
one spatial relation, adjacency.
Structural Comparison
Tables 1, 2 and 3 provide summary statistics on the
structural composition of the digital map data and clas-
sified image, on a class-by-class basis. In previous stud-
les it has been suggested that the structural properties
of, and spatial relationships between, road and building
regions might provide a means by which the dominant
land use (e.g., residential, commercial, industrial) in dif-
ferent parts of an urban area could be inferred (Blamire
and Barnsley 1995, Barnsley and Barr 1997). For this
reason, emphasis will be placed on a comparison of the
structural composition of the regions for these two land-
RT
Map Regions | Image Regions
Label # % # %
All 912 | 100.00 | 14608 | 100.00
Road 8 0.88 3925 26.87
Building | 729 79.93 5187 35.51
Grass 130 14.25 3562 24.38
Tree 13 1.43 1930 13.21
Water 32 3.51 4 0.03
Table 1: Number and percentage of regions as a function
of land-cover type in the classified image and digital map
data sets.
cover classes in the map (Figure 1) and image (Figure
2) data. The land parcels in the digital map data have,
of course, been simplified and generalized with respect
to reality, as part of the normal cartographic produc-
tion process. We would therefore expect these data to
exhibit a much simpler structural composition than the
land-cover parcels identified in the remotely-sensed im-
age. The latter is expected to contain unwanted spatial
information or ‘clutter’; for example, small, isolated re-
gions corresponding to individual trees. While it may be
an accurate representation of the distribution of different
land-cover types within the scene, this level of detail is
often unhelpful because it reduces the clarity and inter-
pretability of the resultant land-cover map, and because
it unduly complicates (or possibly even prevents) infer-
ence of the dominant land use in an area by means of
structural pattern-recognition techniques. In short, the
digital map data provides a point of reference against
which we can assess the structural composition of the
classified image and, hence, determine the amount of
structural clutter that it contains. By the same token,
it permits an evaluation of reflexive-mapping techniques
as a means of removing structural clutter.
Table 1 clearly shows that the classified image contains
considerably more road and building regions than the
corresponding digital map data, whether expressed in
terms of the absolute number or percentage of regions
in the scene. The difference in the number of these re-
gions is reflected in their mean area in each of the two
data sets (Table 2), such that individual road regions
are generally much smaller in the classified image than
in the digital map data set, even though the total area
for this land cover class is greater in the image. This
suggests that the road network — a single, continuous,
morphologically-complex entity in the map data — has
been fragmented into a series small, disjoint regions in
the image, due partly to occlusion by overhanging trees
and tall buildings and partly because of the raster data-
collection process. A similar, although somewhat less
pronounced effect is also evident for the building class.
The differences in the morphological properties of the
building and road regions between the map and image
data impact on their observed spatial relations (Table
3). Thus, the number of adjacency relationships in-
volving road and building classes is much larger in the
image than in the map. This confirms the hypothesis that
the spatial structure of the land-cover data derived from
remotely-sensed images will generally be more complex
316 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
Fig
Ord