classification) and for extraction of buildings and roads, panchromatic band is needed
(Jensen and Cowen, 1999). Hyperspectral imagery data have been successfully used
for urban land use/cover classification (Benediktsson et al., 1995; Hoffbeck and
Landgrebe, 1996; Platt and Goetz, 2004; Thenkabail et al., 2004a,b), extraction of
impervious surfaces (Weng et al., 2008), vegetation mapping (McGwire et al., 2000;
Schmidt et al., 2004; Pu et al., 2008), and water mapping (Bagheri and Yu, 2008;
Moses et al., 2009). A large number of spectral bands provide the potential to derive
detailed information on the nature and properties of different surface materials on the
ground, but it also means a difficulty in image processing and a large data redundancy
due to high correlation among the adjacent bands. Increase of spectral bands may
improve classification accuracy, only when those bands are useful in discriminating
the classes (Thenkabail et al., 2004b).
Traditional classification methods of LULC based on detailed fieldwork suffered
two major common drawbacks: confusion between land use and land cover and lack
of uniformity or comparability in classification schemes, leaving behind a sheer
difficulty for comparing land use patterns over time or between areas (Mather, 1986).
The use of aerial photographs and satellite images after the late 1960s does not solve
these problems, since these techniques are based on the formal expression of land use
rather than on the actual activity itself (Mather, 1986). In fact, many land use types
cannot be identified from the air. As a result, mapping of Earth’s surface tends to
present a mixture of land use and land cover data with an emphasis on the latter (Lo,
1986). This problem is reflected in the title of the classification developed in the
United States for the mapping of the country at a scale of 1 : 100,000 or 1 : 250,000
commencing in 1974 (Anderson et al., 1976). Moreover, this U.S. Geological Survey
(USGS) Land Use/Land Cover Classification System (so-called Anderson scheme)
has been designed as a resource-oriented one. Therefore, eight out of nine in the firstlevel categories relate to nonurban areas. The success of most land use or land cover
mapping efforts has typically been measured by the ability to match remote sensing
spectral signatures to the Anderson scheme, which, in the urban areas, is mainly land
use (Ridd, 1995). The confusion between land use and land cover contributes to the
low classification accuracy (Foody, 2002). In addition, the spatial scale and categorical scale is not explicitly linked in the classification scheme. The former refers to the
manner in which image information content is determined by spatial resolution and
the way the spatial resolution is handled in the image processing, while the latter refers
to the level of detail in classification categories (Ju et al., 2005). This disconnection
leads to the problem of simply lumping classes into more general classes in multiscale
LULC classifications, which may cause a great lost of categorical information.
Since most classifications are conducted at a single spatial and categorical scale,
there remains an important issue of matching an appropriate categorical scale of the
Anderson scheme with the spatial resolution of the satellite image used (Welch, 1982;
Jensen and Cowen, 1999). However, the nature of some applications requires LULC
classification to be conducted at multiple spatial and/or categorical scales because a
single scale cannot delineate all classes due to contrasting sizes, shapes, and internal
variations of different landscape patches (Wu and David, 2002; Raptis et al., 2003).
This is especially true for complex, heterogeneous landscapes, such as urban
64
ON THE ISSUE OF SCALE IN URBAN REMOTE SENSING
(Jensen and Cowen, 1999). Hyperspectral imagery data have been successfully used
for urban land use/cover classification (Benediktsson et al., 1995; Hoffbeck and
Landgrebe, 1996; Platt and Goetz, 2004; Thenkabail et al., 2004a,b), extraction of
impervious surfaces (Weng et al., 2008), vegetation mapping (McGwire et al., 2000;
Schmidt et al., 2004; Pu et al., 2008), and water mapping (Bagheri and Yu, 2008;
Moses et al., 2009). A large number of spectral bands provide the potential to derive
detailed information on the nature and properties of different surface materials on the
ground, but it also means a difficulty in image processing and a large data redundancy
due to high correlation among the adjacent bands. Increase of spectral bands may
improve classification accuracy, only when those bands are useful in discriminating
the classes (Thenkabail et al., 2004b).
Traditional classification methods of LULC based on detailed fieldwork suffered
two major common drawbacks: confusion between land use and land cover and lack
of uniformity or comparability in classification schemes, leaving behind a sheer
difficulty for comparing land use patterns over time or between areas (Mather, 1986).
The use of aerial photographs and satellite images after the late 1960s does not solve
these problems, since these techniques are based on the formal expression of land use
rather than on the actual activity itself (Mather, 1986). In fact, many land use types
cannot be identified from the air. As a result, mapping of Earth’s surface tends to
present a mixture of land use and land cover data with an emphasis on the latter (Lo,
1986). This problem is reflected in the title of the classification developed in the
United States for the mapping of the country at a scale of 1 : 100,000 or 1 : 250,000
commencing in 1974 (Anderson et al., 1976). Moreover, this U.S. Geological Survey
(USGS) Land Use/Land Cover Classification System (so-called Anderson scheme)
has been designed as a resource-oriented one. Therefore, eight out of nine in the firstlevel categories relate to nonurban areas. The success of most land use or land cover
mapping efforts has typically been measured by the ability to match remote sensing
spectral signatures to the Anderson scheme, which, in the urban areas, is mainly land
use (Ridd, 1995). The confusion between land use and land cover contributes to the
low classification accuracy (Foody, 2002). In addition, the spatial scale and categorical scale is not explicitly linked in the classification scheme. The former refers to the
manner in which image information content is determined by spatial resolution and
the way the spatial resolution is handled in the image processing, while the latter refers
to the level of detail in classification categories (Ju et al., 2005). This disconnection
leads to the problem of simply lumping classes into more general classes in multiscale
LULC classifications, which may cause a great lost of categorical information.
Since most classifications are conducted at a single spatial and categorical scale,
there remains an important issue of matching an appropriate categorical scale of the
Anderson scheme with the spatial resolution of the satellite image used (Welch, 1982;
Jensen and Cowen, 1999). However, the nature of some applications requires LULC
classification to be conducted at multiple spatial and/or categorical scales because a
single scale cannot delineate all classes due to contrasting sizes, shapes, and internal
variations of different landscape patches (Wu and David, 2002; Raptis et al., 2003).
This is especially true for complex, heterogeneous landscapes, such as urban
64
ON THE ISSUE OF SCALE IN URBAN REMOTE SENSING