Map Area Image Area
Label zZ Tr X T
All 1048576 | 1149.75 | 1048576 71.78
Road 74916 | 9364.50 271204 69.10
Building 123880 169.93 262716 50.65
Grass 743356 | 5718.12 265688 74.59
Tree 104152 | 8011.69 248932 | 128.98
Water 2272 71.00 36 9.00
Table 2: statistics on the morphological property
area for the regions of each land-cover type identified
in the classified image and digital map data sets.
Map Edges Image Edges
Label # % # %
All 1994 | 100.00 | 62268 | 100.00
Road 69 5.77 | 17622 28.30
Building 882 73.86 | 19056 30.60
Grass 987 82.66 | 16340 26.24
Tree 19 1.59 9244 14.84
Water 37 3.00 6 0.009
Table 3: Number and percentage of adjacency edges as
a function of land-cover type in the classified image and
digital map data sets.
than that of a corresponding digital map.
Figures 3 and 4 highlight the problems described above.
Clearly, the road and building regions identified in the
image bear only a partial resemblance — in terms of geo-
metrical, morphological and spatial structure — to their
counterparts in the digital map data (Figure 1). Thus,
while much of the real road network is correctly clas-
sified in the image, many other ‘road’ regions are also
identified: a similar effect is evident for the building
class. Some of this ‘clutter’ is an accurate representa-
tion of the complex spatial pattern of land cover in the
scene, identified because of the very high spatial resol-
ution of the image data. The remainder, however, rep-
resent errors of omission and commission in the initial
land-cover classification. Regardless of which is actually
the case, the presence of these ‘clutter’ regions complic-
ates the apparent structural composition of the scene.
What is required, then, is some means of identifying the
‘clutter’ regions and removing them through a process
of re-labelling (i.e., assigning them to another land-cover
category), such that the resultant data set exhibits a sim-
pler, realistic, more interpretable structural composition.
REMOVING STRUCTURAL CLUTTER
The process of removing structural clutter from a land-
cover image can be performed by means of a reflexive
mapping in which each of the pixels in a clutter region is
assigned an alternative land-cover label — one that res-
ults in a more credible structural composition for that
part of the image. It is possible to use per-pixel, mov-
ing window (i.e., kernel-based) techniques — such as
a simple, majority filter (Gurney and Townshend 1983)
— to perform this type of operation. However, Barns-
ley and Barr (1997) note that kernel-based techniques
have a number of inherent limitations, including the dif-
ficulty of selecting a priori the optimum kernel-size and
the fact that both the pixel and and the kernel are ar-
bitrary spatial constructs which bear little resemblance
to the geometric form of the principal spatial entities in
the corresponding scene. In view of this, we have de-
veloped an alternative approach, based on an analysis
of the morphological property area and the spatial re-
lation adjacency of multi-pixel land-cover regions iden-
tified in the image and represented in the XRAG data
structure.
The procedure makes two assumptions: first, that clut-
ter regions can be separated from non-clutter regions
through an analysis of their area; second, that clutter
regions are adjacent to at least one non-clutter region,
such that if the clutter region was re-assigned the land-
cover label of the non-clutter region it would result in
a more credible structural composition in that part of
the image. Thus, for example, a small road region which
was disjoint from the main road network, but spatially
adjacent to a building region, might reasonably be con-
sidered to form part of that building and would be re-
labelled accordingly. In examining the first of these two
assumptions, consider the frequency distribution of re-
Figure 5, for ex-
ample, presents data for the road class. It highlights the
very large number of single-pixel road regions and, more
gion area for each land-cover class.
generally, the inverse relationship between the number
and size (area) of the road regions. The relationship
is almost perfectly linear, in terms of a log-log plot, for
regions smaller than 50-100m?; beyond this point the re-
lationship becomes weaker and more erratic. A similar
relationship is evident in the data for the building and
vegetation (grass and tree combined) classes (not presen-
ted here). It seems reasonable to assume that most, if
not all, of the very small, often single-pixel, regions rep-
resent structural clutter, partly because the likelihood
that an individual pixel is misclassified is comparatively
high. By the same token, most of the larger regions
are likely to represent meaningful land-cover parcels, be-
cause the chances of misclassifying a large, multi-pixel
region in iis entirety are comparatively low. The prob-
lem is, however, determining objectively the threshold
that distinguishes clutter from non-clutter regions. The
solution adopted here is to compute the first derivative
of the region-size frequency distribution (Equation 1):
d m Zr Fo)
dz r-— ro
ZT — Zo
(1)
The rationale behind this approach is that the spatial
structure of land-cover parcels in urban area is typically
complex, such that the frequency distribution of parcel
size might be expected to be equally complex and irreg-
ular (e.g., the distribution for regions larger than 50m?
in Figure 5). We therefore seek the point at which the
region-size frequency distribution changes from being a
monotonic reduction in frequency as a function of area
(due to the presence of clutter regions), to a more com-
plex, irregular relationship. This point is identified as
the smallest area for which f (zo) = 0 (i.e., the first sta-
tionary point; Figure 6). Using this approach, the area
thresholds determined for the road, building and veget-
ation classes are 52m”, 35m” and 32m?, respectively.
318 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998