Tr
INTEGRATION OF SPECTRAL AND SPATIAL CLASSIFICATION METHODS
FOR BUILDING A LAND-USE MODEL OF AUSTRIA
K. Steinnocher
Department for Environmental Planning
Austrian Research Centre Seibersdorf
2444 Seibersdorf, Austria
Commission 4, Working Group 6
KEY WORDS: Spatial Classification, Land-Use, Land-Cover, Environment, Generalisation, Landsat
ABSTRACT:
The first part of this paper deals with methodological aspects of land-cover and land-use classification. Due to the high resolution of
today's remote sensors, the single pixel represents land-cover rather than land-use. The outcome of per-pixel classifications will
therefore not meet the requirements of a land-use map if the spatial composition of cover types is not considered as well. This can be
achieved by applying a spatial postclassification method. We present an algorithm, which analyses the spatial composition of land-
cover types in the local neighbourhood of each pixel. The assignment of land-use classes is based on the comparison of the actual
composition of land-cover types with a predefined rule-set. In the second part the application of the method for building an Aus-
trian-wide land-use model is discussed. To asses the quality of both the method and the derived land-use model, the results are com-
pared to parts of the CORINE land-cover map of Austria.
1. INTRODUCTION
Information on spatial distribution of land-use represents an
essential input to environmental modelling. For analyses on a
national scale no appropriate land-use model of Austria has
been available up to today. Existing land-use maps do not meet
the requirements for an Austrian-wide representation, because
they either are not up-to-date or cover only selected regions of
Austria. Most of these maps result from photogrammetric or
terrestrial observations, which are both cost intensive and time
consuming. Satellite remote sensing represents a valuable al-
ternative to the traditional methods, by offering up-to-date
information of large areas for reasonable costs.
The first attempt to use remote sensing data on a national scale
in Austria was the development of the CORINE land-cover
data-base, a European-wide project initiated by the European
Commission (EUR, 1993). This model is derived from visual
interpretation of analogue satellite images and ancillary data
such as aerial photographs or topographic maps. Due to the
high expenditure of time needed for visual interpretation the
Austrian land-cover map will not be finished before 1997.
Nevertheless there exists an urgent need for such a data-set
today. This paper presents a method which allows a semi-
automated mapping of Level II land-use classes from high
resolution satellite imagery. The method is then applied for the
derivation of an Austrian-wide land-use model.
2. LAND-COVER VERSUS LAND-USE
Per pixel classification of image data acquired by sensors such
as Landsat TM or SPOT HRV is not always adequate for map-
ping land-use on a regional scale. This is due to the high geo-
metric resolution of the single pixel, which rather represents a
single land-cover type than certain land-use classes composed
of different cover types. For mapping heterogeneous land-use
types the context between a single pixel and its neighbours
seems to be the crucial point. This becomes apparent when
looking at different land-use types of built-up environments,
e.g. low density urban areas. Per pixel analysis of these areas
841
will result in a composition of different cover types such as
roofs, pavement, vegetation, bare soil, etc. thus producing a
*salt and pepper' pattern rather than the desired land-use class.
Therefore the spectral classification is not sufficient unless the
spatial composition of the cover types is considered as well. To
solve this problem various attempts have been made to include
spatial variation in the classification process.
Textural characteristics can be used to describe the spatial
variation of radiance within an image. Haralick (1973) pro-
posed various methods to derive textural measures from digital
images. When incorporated in multispectral data sets these
texture bands can significantly improve the accuracy of land-
use classification (Franklin and Peddle 1990, Sali and Wolfson
1992, Webster and Bracken 1992). Another approach applies a
two step process. First, a per pixel classification is performed
resulting in a land-cover layer. Second, a postclassification
algorithm analyses the spatial composition of the land-cover
types and assigns the land-use classes in question. The context
between land-cover and land-use classes can either be estab-
lished by training areas and statistical measures (Zhang et al.,
1988, Guo and Moore, 1991, Barnsley and Barr, 1992, Gong
and Howard, 1992) or be defined by rules (Steinnocher et al.
1993, Fung and Chan 1994).
The method presented in this study follows the second ap-
proach. A spatial postclassification algorithm, applied to the
result of a per-pixel land-cover classification, assigns the re-
quested land-use classes using a set of pre-defined rules. For a
better understanding, classes resulting from the spectral classi-
fication will be called primary classes, the final land-use classes
will be called secondary classes.
The algorithm works within a local neighborhood which is
defined by a moving window. Within this window a standard-
ized histogram of primary classes is calculated, representing the
spatial composition, i.e. the frequency of primary classes found
in the local neighborhood. The histogram is then compared to a
set of rules which represent the expected frequency of primary
classes for each secondary class. As soon as a rule is found to
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996