Importantly, the first-derivative curve hardly fluctuates
beyond this threshold for any of the three land cover
types.
Having obtained the area threshold values for each of
the three principal land cover types, a clutter-removal
procedure was implemented as follows (i) any region of
the three principal land-cover types with an area smal-
ler than the threshold value for that class was re-labelled
and incorporated into an adjacent region, provided that
the other region belonged to one of the other two main
cover types and had an area greater than the threshold
value for that class; (ii) the boundary of the resulting re-
gion was re-determined, its structural features (e.g., area
and adjacency) re-calculated, and the XRAG model up-
dated accordingly. Where a clutter region was adjacent
to several possible non-clutter region, the re-assignment
was performed with respect to the smallest of these re-
gions. This procedure was run iteratively until no further
re-assignments needed to be performed.
RESULTS AND DISCUSSION
Figure 7 shows the results of the reflexive-mapping cor-
rection procedure applied to the classified land-cover im-
age presented in Figure 2, while Figures 8 and 9 show the
corrected road and building regions extracted from Fig-
ure 7. A comparison of Figures 3 and 8 reveals that most
of the spurious, small road regions have been removed
using this technique. As a result, the road network is
much closer to that shown in the digital map data (Fig-
ure 1), although the boundaries of the road regions are
somewhat ‘feathery’. Similarly, the reflexively-mapped
building regions (Figure 9) are considerably improved
with respect to the original land-cover classification (Fig-
ure 4).
Despite the considerable improvement in the visual ap-
pearance of the road and building regions after the struc-
tural clutter has been removed, Tables 4 and 5 suggest
that the structural composition of the reflexively-mapped
regions is still significantly different from their counter-
parts in the digital map (Tables 1, 2 and 3). In particu-
lar, the image still contains many more regions than the
digital map. Moreover, the total area of the road and
building regions has increased as a result of the correc-
tion process. In terms of the building class, this results
in a mean area closer to — albeit, slightly greater than
— the digital map data; however, the correspondence
between the mean area of the road regions between the
image and map data, as a result of the reflexive mapping
procedure, shows a much smaller improvement. Finally,
while the correction procedure reduces the spatial com-
plexity of the image with respect to the original land-
cover classification in terms of the number of adjacency
relationships (edges) between regions of different land-
cover type, the reflexively-mapped image still exhibits .
much greater structural complexity than the digital map
data (Table 5).
The fact that the reflexive-mapping procedure used here
to remove structural clutter does not produce a struc-
tural composition that is exactly comparable to that of
a digital map for the same study area is, perhaps, un-
surprising. After all, the map-making process involves
feature selection and abstraction, as well as various levels
of thematic and spatial generalization. A more appropri-
ate assessment of the clutter-removal procedures presen-
Corrected Regions Corrected Area
Label # T Y T
All 5142 100.00 | 1048576 | 203.92
Tarmac 725 14.00 289968 | 399.96
Built 1278 24.85 266608 | 208.61
Grass 1961 38.13 250228 | 127.60
Tree 1174 22.83 241736 | 205.91
Water 4 0.07 . 36 9.00
Table 4: Summary statistics on the number of regions
and the morphological property area as a function of
cover type in the reflexively-mapped land cover image.
Corrected Edges
Label # %
Al 20616 100.00
Tarmac 3934 19.08
Built 5455 26.46
Grass 6502 31.53
Tree 4720 22.89
Water 5 0.02
Table 5: Number and percentage of adjacency edges as
a function of cover type in the reflexively-mapped land
cover image.
ted here should be based on answers to two questions,
namely: (i) do they improve the visual appearance and
overall interpretability of the initial land-cover map; and,
(11) do they improve our ability to derive further inform-
ation from the land-cover data, such as the identification
of different types of urban land use based on an ana-
lysis of localised, structural patterns of discrete land-
cover parcels? On the basis of the evidence presented
in Figures 7-9, the answer to the first of these questions
would appear to be yes. The second question has not
been addressed directly in this paper and is the subject
of current research. Initial efforts in this direction in-
clude an examination of proximity graphs, derived as
a function of the distance between the centroids of the
building regions in the reflexively-mapped land-cover im-
age (e.g., Figure 10). Our aim, in this respect, is to see
whether it is possible to identify distinct and quantifiable
differences in the spatial pattern of buildings at different
locations within graph-space that correspond to partic-
ular categories of urban land use.
CONCLUSIONS
A method has been presented for identifying and remov-
ing structural clutter in land-cover maps derived from
very high spatial resolution images. The approach in-
volves an analysis of the morphological and structural
properties of, and relations between, discrete regions in
the original land-cover data. The procedure has been ap-
plied to image data of a spatially-complex urban scene.
The results show that considerable improvements can be
achieved in terms of the identification and representation
of roads and buildings. This has important implications
for future work on inferring urban land use from digital
remotely-sensed images.
320 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 °
Figu
