International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
in the final stage.
We stress however that the approach currently followed by
GUS service providers, which heavily relies on manual in-
terpretation (approximately half of the total production ef-
fort), is the most suitable to obtain the accuracies required
by the users. The methods that we have highlighted are in-
deed a way to reduce the final manual re-classification step
and therefore the time-to-market and the cost of land use
mapping products, and this is available and used by some
of GUS service providers using COTS software.
Among the classification approaches, it is our opinion that
morphological analysis, which is currently pursued by many
authors, is able to provide in the short term some kind of
improvement in this field. Similarly, methods aimed to in-
tegrate texture measures may be useful to provide semi-
land use class, i.e. something that is not a land use map,
but more than a land cover one. These features are already
used in standard software: what lacks now is a clear def-
inition of which nomenclature is possible to extract with
textures. More far in the future is, to our knowledge, the
possibility to integrate GIS information with remote sens-
ing data, at least at the European level.
Summing up, limitations of the present version of land use
mapping product are the limited use of spatial information
in the images to improve land use mapping, the lack in def-
inition of nomenclature, the problems in integrating GIS
layers and remote sensed data, the large percentage of the
work still done manually. Research lines that should be
addressed to improve them are therefore:
e criteria for the selection of simple spatial feature to
improve land use mapping;
e realization of simple procedures for incorporating GIS
data into classification tools exploiting their charac-
teristics;
e definition of the land use nomenclature that it is pos-
sible to extract from each sensor or, on the contrary,
of the requirements of sensors for extracting a given
nomenclature.
3.20 REG block
Among the REG block a particular interest is in sealing
mapping products. This point is confirmed by the realiza-
tion of a very recent symposium promoted by one of the
European Environmental Agency Technical Committees,
for the definition of what “sealing” really means or should
mean.
As a matter of fact, sealed area maps are of particular value
in relation to increasing urbanization, increases in surface
run off and increasing concern with the unpredictability of
weather patterns in the context of global warming. The
map of sealed areas offers a means of addressing issues
which are on the foreground on the political agenda and are
therefore matters for which positive remedies are sought
throughout the European context.
322
A first way to provide sealing maps with different seal-
ing factor comes from an accurate characterization of the
cover classes in the urban area of interest. For instance,
after determining the built up area with precision we may
compute the percentage of coverage to provide the sealing
map. Therefore, a first group of methods for the proposed
task is made by procedures starting from high resolution
data, typically SPOT or IRS-1 at 5 m spatial sampling, and
classify these images with very high precision with respect
to urban cover classes.
The largest part ofthese procedures adds one or more bands
to the original data. In Shaban and Dikshit (2001), for
instance, textural features extracted from grey level co-
occurrence matrix, grey level difference histogram and sum
and difference histogram are compared and used to im-
prove the urban classification accuracy. It was found that
the best results are obtained by combining spectral and tex-
tural features, without any advantage by a conventional
Principal Component Transform before the combination.
Moreover, usual separability criteria (like transformed di-
vergence) are not useful to select the best combination. A
similar approach is proposed in Chen ef al. (1997), where
a fractal measure is used to improve the classification. The
paper shows that the use of this information improves the
accuracy values for heterogeneous classes, slightly degrad-
ing homogeneous areas, e.g. water.
A different approach is presented in Zhang (1999), where
the textural measures are used to filter out the classification
results to improve the accuracy of the built up classes. The
homogeneity of the class map is computed ina 3 x 3 win-
dow and in the four diagonal directions, and then filtered
to discard uninteresting areas and improve by some sort
of majority voting the initial guess based only on spectral
characteristics.
The complementary approach to those in previous para-
graphs is to compute information about the sealing density
by means of a more direct approach. To this aim, we may
define two major methodologies. The first one, exempli-
fied in Karathanassi et a/. (2000), refers to the use of textu-
ral features to directly decompose the urban environment
into areas with different urban density. In this work the
classes are defined by setting up thresholds in built.up to
overall area ratios (« 0.3, low density, 0.3 to 0.7 medium
density, > 0.7 high density). It is found that significantly
larger window size than in [1] should be used, because we
are not looking for buildings, but for blocks. Instead of
3 x 3 co-occurrence measures, 11 x 11 or wider windows
are used. Large improvements were obtained with respect
to spectral features alone.
Finally, the so-called Vegetation - Impervious surface Soil
(VIS) model may be used to discriminate among different
degree of impervious surface. This is done for instance in
Phinn et al. (2002), where samples from these three classes
are extracted using a first simple classifier, and then a man-
ual analysis of the model allows finding end members for
a refined segmentation. The paper shows that this method
enables distinctive densities of commercial, industrial and
