Istanbul 2004
JS) is a project
1imed to the
| to meet the
is currently
h the follow-
ense, Hugin,
Consultants
partment of
nental Stud-
big compo-
UTS, made
| from Earth
illary infor-
iding socio-
1al scale fo-
1 developing
GUS in the
nd use map-
emely more
rse, the two
re both clas-
' classes are
ia only, and
patial analy-
rcel may be
> inferred by
and periur-
n two ways.
ap obtained
on routines.
iques is ap-
the land use
xternal data
wledge sys-
1 on a direct
st classifica-
ation. Thus,
ig based on
> classes, an
s first some
1 areas, and
e from land
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV , Part Bl. Istanbul 2004
[ Block | GUS Product | Sensors) [| Spatial res. | Temporal res. (requ.) |
| UTS Land Use Spot 5 1:15.000 2-4 year
Land change Spot 5 1:15.000 2-4 year
Hot-spot monitoring Spot 5/ Ikonos | 1:25.000 (5.000) 6 12 months
City volume model Airborne
Modelling tool
Sealing map Spot 5 1:15.000 2-4 year
Noise observatory Spot 5 1:15.000 2-4 year
Heating efficiency Airborne 10 years
REG Basic Land Use Landsat/Spot 1:50.000 2 year
Regional sealing Envisat 1:50.000 2 year
DEV Basic urban mapping dev. countries | Spot S/Ikonos | 1:15.000 (5.000)
Table 1: GUS current portfolio spectral, spatial and temporal specifications.
Land cover mapping in urban areas with the accuracy re-
quested nowadays by the users requires dealing with high
or very high spatial resolution satellite sensors. The cost
and the acquisition problems of these sensors at the mo-
ment allow usually to work on single date, single sensor
data sets. Therefore, co-registration problems are usually
not considered at this stage and the pre-processing steps are
devoted to remove image artifacts, distortions due to the
viewing geometry of the sensors, and atmospheric effects.
From this point of view, all data providers are quickly adapt-
ing to the market, which requires data with all these correc-
tions already. Moreover, models for the viewing geometry
of all these sensors have been or are planned to be included
in the most widespread COTS software.
The classification approach to these images aims at fully
exploiting their spatial resolution, and therefore to inte-
grate pixel-by-pixel classification with spatial analysis. Its
of course impossible to refer here of all the methods that
have been proposed. However, from the research view-
point, land cover classification using satellite data have
been recently discussed in the September 2003 Issue of the
IEEE Transactions on Geoscience and Remote Sensing,
devote to “Urban Remote Sensing by Satellite” (Gamba
et al., 2003) , which is therefore a good introduction to the
topic. Once land cover maps are obtained, land use can
be extracted, tough only to some extent, by means of au-
tomatic or semi-automatic ways. A final manual reclassifi-
cation is always required, at least at the moment, to reach
the accuracy required by the final user. Kernel-based re-
classification (Kontoes eft al., 2000) is based on the evalu-
ation of the patterns of vegetation and built areas, for in-
stance, to provide information about residential and indus-
trial areas. Similarly, a much more complex, graph-based
approach has been proposed (Barr and Barnsley, 2000) to
discriminate among building districts built in different cen-
turies, with different spatial patterns of man-made features.
It requires the extraction of each building, the characteri-
zation of a graph connecting it to its neighborhood, and a
strategy to compare graphs to match the given residential
model.
An example of knowledge-based integration of GIS data
into the classifier is found in Stefanov et al. (2001), where
ASTER data have been used to analyze a large urban area
in Arizona, and coordinated with many different layers of
321
information. Results are encouraging, and have been re-
cently proposed in a second paper on the same area using
LANDSAT data, which proves the robustness of the sys-
tem. Re-classification by knowledge-based classifiers is
also used, since it is currently available in COTS like E-
cognition by Definiens (Definiens, 2004) and Erdas Expert
Classifier by Leyca Geosystems (Erdas, 2004). They in-
corporate spectral and spatial information at different scales
and provide very good land cover classification accuracy at
an affordable price. For land use mapping (Kressler et al.
2001) problems arise from class definition and support the
above mentioned consideration that remote sensing data
alone are not able to provide all the information required
for accurate land use mapping.
A different approach to land use mapping tries and takes
into account directly spatial statistics, represented mainly
by texture measures, to provide a land use map, or at least
a map with classes other than “buildings”, “roads”, “trees”,
“meadows” and “water”. The basic idea is that if we incor-
porate texture measures or statistical measures as a sup-
plementary band, we may recognize the different textural
appearance of urban environments. The process is driven,
for instance, by studies that show that census data is corre-
lated with texture statistics in Landsat TM images (Chen,
2002). So, in Gong and Howarth (1990) the authors com-
bine edge-density image with the two principal compo-
nent (PC) bands to obtain a better overall accuracy with
SPOT imagery. They also observe that the edge-density
image eliminates the confusion between the rural and ur-
ban land use that have similar spectral characteristics. Sim-
ilarly, in Gong ef al. (1992) the authors compare gray level
co-occurrence matrix (GLCM), simple statistical transfor-
mations (SST) and texture statistics (TS) approaches for
SPOT image of urban area. Their results indicate that some
spatial features derived using GLCM and the SST methods
could improve the classification accuracies obtained by the
use of spectral images only. On the contrary, TS method
makes limited accuracy improvements.
Despite the large number of land cover classifiers, the need
of extracting information from very high resolution satel-
lite sensors in urban areas has not been fulfilled yet. Thus,
also land use mapping is still in the process to become a
“mature” application. At the moment the best approach re-
mains strictly connected with human intervention, mainly
