16.81 to 12.24 to 5.61 with the increase in geographical scale. For any LST model,
LST was consistently underestimated over the low-temperature areas, while in the
high-temperature areas the distribution of residuals varied with scale. LST was
predominantly overestimated at the block scale. As the scale increased, LST became
generally underestimated. At the tract scale, LST was equally over- and underestimated. In the downtown area where high temperatures existed, predicted LST
values may be either over- or underestimated at any scale, implying the heterogeneity
in the thermal landscape. The spatial patterns of residuals appear analogous at the
block group and tract levels.
4.5.2 Population Estimation
Many remotely sensed images collected from different sensors have been utilized to
estimate population. With various spatial resolutions, they are especially applicable
at a certain scale for the study. For instance, high-spatial-resolution aerial photography is useful for population estimation at the microscale (Lo and Welch, 1977;
Lo, 1986; Cowen et al. 1995), while low-spatial-resolution data, such as those from
the Defense Meteorological Satellite Program Operational Linescan System
(DMSP-OLS), are suitable for modeling at the global or regional scale. However,
if a medium scale such as at a city level is concerned, images with medium spatial
resolution, such as those obtained from Landsat TM/ETM+ and Terra ASTER
sensors, should be considered. Research has proved that such data are efficient and
effective in predicting population in city or county levels (Harvey, 2002a,b; Qiu
et al., 2003, Li and Weng, 2005; Lu et al., 2006). Lo (1986) summarized several
approaches commonly used in population estimation with remotely sensed data:
counting the dwelling units, using per-pixel spectral reflectance, measuring urban
areas, and using land use information. The application of these methods may be, in
fact, considered to respond to different analytical scales, with the first two
applicable at small areas (1 km
2 or less) and the last two for larger (or regional)
and medium scales, respectively (Harvey, 2002a).
Because remotely sensed images are scale dependent, population models derived
from such data are also subject to the impact of scale. Lo (2001) used DMSP-OLS
nighttime light data to model the Chinese population and population densities at
three different spatial scales: province, county, and city. Either allometric growth or
linear regression models were found to be promising in estimating population at all
three levels, but the best models were obtained at the city level. Qiu et al. (2003)
carried out a biscale study of the decennium urban population growth from 1990 to
2000 in the north Dallas–Fort Worth metropolis using models developed with
remote sensing and GIS techniques. Both models yielded comparable results with
that obtained from a more complex commercial demographics model at the city as
well as the census tract levels, yet the GIS model remained robust to the scale
change because of its insensitiveness to the spatial scale. The remote sensing model
was attenuated when moved to the census tract from the city level. Liang et al.
(2007) estimated the residential population of Indianapolis at multiple scales of
census units (block, block group, and tract) using remote sensing–derived
SCALE DEPENDENCY OF URBAN PHENOMENA
71
LST was consistently underestimated over the low-temperature areas, while in the
high-temperature areas the distribution of residuals varied with scale. LST was
predominantly overestimated at the block scale. As the scale increased, LST became
generally underestimated. At the tract scale, LST was equally over- and underestimated. In the downtown area where high temperatures existed, predicted LST
values may be either over- or underestimated at any scale, implying the heterogeneity
in the thermal landscape. The spatial patterns of residuals appear analogous at the
block group and tract levels.
4.5.2 Population Estimation
Many remotely sensed images collected from different sensors have been utilized to
estimate population. With various spatial resolutions, they are especially applicable
at a certain scale for the study. For instance, high-spatial-resolution aerial photography is useful for population estimation at the microscale (Lo and Welch, 1977;
Lo, 1986; Cowen et al. 1995), while low-spatial-resolution data, such as those from
the Defense Meteorological Satellite Program Operational Linescan System
(DMSP-OLS), are suitable for modeling at the global or regional scale. However,
if a medium scale such as at a city level is concerned, images with medium spatial
resolution, such as those obtained from Landsat TM/ETM+ and Terra ASTER
sensors, should be considered. Research has proved that such data are efficient and
effective in predicting population in city or county levels (Harvey, 2002a,b; Qiu
et al., 2003, Li and Weng, 2005; Lu et al., 2006). Lo (1986) summarized several
approaches commonly used in population estimation with remotely sensed data:
counting the dwelling units, using per-pixel spectral reflectance, measuring urban
areas, and using land use information. The application of these methods may be, in
fact, considered to respond to different analytical scales, with the first two
applicable at small areas (1 km
2 or less) and the last two for larger (or regional)
and medium scales, respectively (Harvey, 2002a).
Because remotely sensed images are scale dependent, population models derived
from such data are also subject to the impact of scale. Lo (2001) used DMSP-OLS
nighttime light data to model the Chinese population and population densities at
three different spatial scales: province, county, and city. Either allometric growth or
linear regression models were found to be promising in estimating population at all
three levels, but the best models were obtained at the city level. Qiu et al. (2003)
carried out a biscale study of the decennium urban population growth from 1990 to
2000 in the north Dallas–Fort Worth metropolis using models developed with
remote sensing and GIS techniques. Both models yielded comparable results with
that obtained from a more complex commercial demographics model at the city as
well as the census tract levels, yet the GIS model remained robust to the scale
change because of its insensitiveness to the spatial scale. The remote sensing model
was attenuated when moved to the census tract from the city level. Liang et al.
(2007) estimated the residential population of Indianapolis at multiple scales of
census units (block, block group, and tract) using remote sensing–derived
SCALE DEPENDENCY OF URBAN PHENOMENA
71