othod is
ersely it
n like a
in more
r which
cteristic
]though
hm, it is
is very
ingle of
t more
he most
e of the
hod, we
system
ic Eval-
(Mexico,
ence on
for 3-D
stereo
lings of
1ational
dy on
ble and
rence on
REMOVING STRUCTURAL CLUTTER
FROM LAND-COVER CLASSIFICATIONS
OF VERY HIGH SPATIAL RESOLUTION
IMAGES USING REFLEXIVE-MAPPING
TECHNIQUES
Stuart Barr and Mike Barnsley
Department of Geography
University of Wales Swansea
Singleton Park
Swansea, SA2 8PP, UK.
Commission VII, Working Group 4.
KEY WORDS: structural composition, clutter, reflex-
ive mapping, spatial resolution, urban land use.
ABSTRACT
A new generation of satellite sensors, with a spatial resol-
ution of 1m-5m, is due to be launched over the next few
years. The images generated by these devices should,
in theory, provide improved information on land cover
and land use, particularly in areas where the individual
parcels of land are small and their spatial arrangement
complex, such as towns and cities. In practice, however,
improvements in the ability to identify and delineate the
features of interest go hand-in-hand with an increase in
unwanted spatial detail. This produces ‘clutter’ in land-
cover classifications and, consequently, inhibits the po-
tential to infer information on land use through an ana-
lysis of the structural composition of the observed scene.
A structurally-based, reflexive-mapping procedure has
been developed to overcome this problem. It is used to
remove much of the clutter in an initial land-cover map
of an urban area produced from a 2m spatial resolu-
tion multispectral image. 'The resultant data set reveals
more clearly the underlying structural properties and re-
lations that are indicative of the principal categories of
urban land-use present in the scene.
INTRODUCTION
At present, the highest spatial resolution multispectral
image data-sets acquired by satellite sensors are those
provided by the SPOT HRV (20m) and Landsat Them-
atic Mapper (30m) instruments. In certain situations,
where a scene under investigation has a complex spa-
tial arrangement the resolutions of these sensors have
been found to be to insufficient to perform the accurate
and consistent classification of both land-coverand land-
use. (Ridley et al. 1997). For example, one particular
application often cited in this context, is the classifica-
tion of both land-cover and land-use within urban areas
(Blamire and Barnsley 1995, Barnsley and Barr 1997).
This has been attributed to the fact, that not only is the
spatial arrangement of urban areas complex, but also,
that the individual parcels of land-cover which comprise
them and which characterise the land-uses present tend
to be relatively small (Barnsley and Barr 1997).
In the near future though, the potential to accurately
classify spatially complex scenes, such as urban areas, is
likely to improve. This is because, a new series of op-
tical satellite sensors, which will produce digital image
data-sets with a spatial resolution of 1-5 m in panchro-
matic mode and 4-15 m in multispectral mode, are due
to be launched (Fritz 1996). In theory, data from these
new instruments should offer considerable benefits over
current coarser spatial resolution satellite sensor images
for the derivation of information from spatially complex
scenes such as urban areas. In particular, it should be-
come possible to accurately identify and delineate the
principal physical objects in such scenes (e.g., individual
buildings, roads and small parcels of vegetation) as com-
plete geometric entities rather than individually classified
pixels. In turn, this should facilitate the inference of fur-
ther information, such as land-use, on the basis of the
structural properties and relations exhibited by theses
derived land-cover objects (Barnsley and Barr 1997).
However, in relation to this point, it is important to
note that a number of previous studies have found that
the perceived benefits of increased spatial resolution have
not always yielded the improvements in classification ac-
curacy anticipated (Forster 1985, Martin et al. 1988).
This has been attributed to the fact that as the spatial
resolution of.the sensor increases individual scene ele-
ments (e.g., buildings, roads and small parcels of veget-
ation) begin to dominate the detected spectral response
of each pixel; as such, the spectral response of scene
as whole becomes more varied, with individual scene
elements such as buildings, roads and small parcels of
vegetation tending to become more spectrally hetero-
geneous (Gastellu-Etchegorry 1990, Barnsley and Barr
1996). Clearly, this phenomenon is likely to have an
impact on the potential to derive information on land-
cover and land-use from the images which are to be ac-
quired by these new very high spatial resolution satellite
sensors.
In this paper, we explore how the likely increase in spec-
tral heterogeneity which may result from increased spa-
tial resolution, is is likely to affect the derivation of land-
cover and land-use information from the images acquired
by these new satellite sensors. In particular, we demon-
strate that a spatially complex urban scene acquired at a
very spatial resolution results in the occurrence of large
amounts of structural "clutter" when classified to the
level of land-cover. In response to this, we introduce
a structurally-based reflexive mapping procedure which
can be used to correct for presence of structural clutter in
land-cover data-sets derived using traditional per-pixel
spectrally-based classification procedures. The suitabil-
ity of this structurally-based reflexive mapping proced-
ure is demonstrated by comparing the results obtained
against the land-cover composition of a digital map data-
set, derived from Ordnance Survey, 1:1,250 scale Land-
Line.93+ comprising of five land cover types, namely,
roads, buildings, areas of grass, areas of trees and water
bodies (Figure 1).
AIRBORNE SCANNER DATA AND LAND
COVER CLASSIFICATION
A multispectral image covering part of the London Bor-
ough of Bromley in the U.K., recorded by a Daedalus
AADS-1286 airborne scanner, is used in this study. This
image data-set had a nominal spatial resolution of 2m.
Although the Daedalus scanner collects data in eleven .
spectral channels, only four are used here, namely bands
2 (0.45-0.52um), 3 (0.52-0.60um), 5 (0.63-0.69um) and
7 (0.76-0.90um). These broadly correspond to the blue,
green, red and near infra-red channels of the Space Ima-
ging sensor, scheduled for launch later this year (Ap-
lin et al. 1997). Variations in detected radiance as a
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 315