the contradiction between the models. By including index im-
ages in the classification process, the topographic effect could
be significantly reduced, whereas for the visual interpretation
only the original images were available. The same considera-
tion is valid for the confusion between natural vegetation and
no vegetation, though areas without vegetation cover were
slightly underestimated in the automated approach.
Summing up, the confusion between the two data-models can
be explained by three major causes - the different methods
applied, the different representation of the data, and in a few
cases a different interpretation of the nomenclature. The last
aspect refers above all to the definition of natural areas and is
subject to ongoing discussion. A certain amount of confusion is
due to the limited accuracy of linear features in raster represen-
tation.
The major differences, though, result from the different meth-
ods applied. Whereas the automated approach has a tendency
of smoothing complex shapes, it is rather sensitive to local
variations of land-cover. The visual interpretation, on the con-
trary, is very accurate in the demarcation of single objects but
is less sensitive to changing patterns within a larger heteroge-
neous environment. These effects are obviously caused by the
different approaches towards generalisation. Visual interpreta-
tion usually defines the dominant class in an image as
“background” and “cuts out” the remaining classes. The post-
classification algorithm is always limited to the window size
and therefore does not consider dominant classes in a larger
environment, but reacts to any change within this local neigh-
bourhood. This effect can best be observed when analysing
heterogeneous agricultural areas. This class is strongly con-
fused with arable land, pastures and forest, which are all the-
matic neighbours of this class and typical candidates for
“background classes”. The total of this confusion comes up to
11% of the entire test area and therefore represents the major
disagreement between the two models.
5. SUMMARY AND CONCLUSIONS
The presented paper gives a contribution to the discussion of
land-use versus land-cover classification. A method is pre-
sented that examines the spatial composition of land-cover
types in a local neighbourhood and assigns land-use classes
based on a predefined set of rules. Though postclassification of
this kind will always have a generalising effect and therefore
leads to a loss of details, it is powerful in detecting heterogene-
ous land-use classes composed of a particular composition of
land-cover types.
The application of the method is not limited to a single test
region but is performed for the entire area of Austria. Compari-
son of the resulting land-use model with parts of the CORINE
land-cover map of Austria confirms the usefulness of the cho-
sen procedure for mapping land-use on a regional scale. Never-
theless, there exist obvious differences in the two models,
which are due to the different approaches towards generalisa-
tion. For the postclassification process, the size of the local
neighbourhood seems to be the crucial parameter. Though two
different window sizes were used in the application this might
not be sufficient for a reliable recognition of all land-use ob-
jects. Furthermore, the postclassification algorithm could be
improved by not only considering the frequency of land-cover
types, but also their spatial arrangement.
846
Unexpected contradictions between the models were found in
the alpine areas. Besides the different thematic interpretations
of a few land-cover types, the essential reason for the confusion
of classes seems to lie in the different illumination angles in
rugged terrain. Whereas in the manual approach no correction
was performed at all to overcome this problem, index images
were used in the automatic classification process, thus reducing
the topographic effect to a certain extent. Although image ra-
tioning does improve classification accuracy in alpine areas, it
might be valuable to perform topographic normalisation by
applying a Digital Terrain Model.
Although the presented approach needs further research, it
represents a valuable alternative to visual interpretation of
satellite imagery, as it is definitely less time-consuming and
therefore significantly reduces the costs of land-use mapping
on regional or national scale.
REFERENCES
EUR 12585 - CORINE land-cover project - Technical guide,
Luxembourg, 1993.
Ecker, R., Gsandtner, M., Jansa, J., 1991. Geocoding using
hybrid bundle adjustement and a sophisticated DTM. In: Pro-
ceedings of the 11th EARSel Symposium, Graz, Austria.
Ecker, R., Kalliany, R., Steinnocher, K., 1995. Fernerkund-
ungsdaten für die Planung eines Mobilfunknetzes. Oster-
reichische Zeitschrift für Vermessung und Geoinformation,
83(1), pp. 14-25.
Haralick, R.M., Shanmugam, K., Dinstein, L, 1973. Textural
features for image classification. IEEE Transactions on Sys-
tems, Man and Cybernetics, SMC-3(6), pp. 610-621.
Franklin, S.E., and Peddle, D.R., 1990. Classification of SPOT
HRV imagery and texture features. International Journal of
Remote Sensing, 11(3), pp. 551-556.
Fung, T., and Chang, K., 1994. Spatial composition of spectral
classes: a structural approach for image analysis of heterogene-
ous land-use and land-cover types. Photogramm. Engng. &
Rem. Sens., 60(2), pp. 173-180.
Gong, P. and Howarth, P.J., 1992. Land-use classification of
SPOT HRV data using a cover-frequency method. Interna-
tional Journal of Remote Sensing, 13(8), pp. 1459-1471.
Guo, L.J., and Moore, J.M., 1991. Post-classification process-
ing for thematic mapping based on remotely sensed image data.
In: Proc. Int. Conf. IEEE Geoscience and Remote Sensing So-
ciety, Espoo, Finland, pp. 2203-2206.
Sali, E., and Wolfson, H., 1992. Texture classification in aerial
photographs and satellite data. International Journal of Remote
Sensing, 13(18), pp. 3395-3408.
Steinnocher, K., Staufer, P., and Franzen, M., 1993. Land-
nutzungsdaten zur Modellierung zellularer Mobilfunknetze:
Der integrative Einsatz digitaler Bildverarbeitungstechniken
und geographischer Informationssysteme zur Erfassung ur-
baner Strukturen. In: Proceedings AGIT V, Salzburg, Austria,
(7Salzburger Geographische Materialien 20), pp. 307-318.
Webster, C.J., and Bracken I.J., 1992. Exploring the discrimi-
nation power of texture in urban image analysis. In: Interna-
tional Archives of Photogrammetry and Remote Sensing,
Washington DC, Vol. 29, Part B7, pp.942-948.
Wilkinson, G.G., 1993. The generalisation of satellite-derived
raster thematic maps for GIS input. Geo-Information-Systems,
6(5), pp. 24-29.
Zhang, Z., Shimoda, H., Fukue, K., Matsumae, Y., and Sakata,
T.: New classification algorithms using spatial information for
high resolution image data. In: International Archives of Pho-
togrammetry and Remote Sensing, Vol. 27, B 7, Commission
VII, 1988.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
