Use of archival Landsat imagery to monitor urban spatial growth - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 25

FIGURE 2.5 A generic workflow for urban growth monitoring by remote sensing. Note that defining research questions is
critical for designing specific technical procedures.

medium resolution, the three generations of Landsat sensors
provide multispectral images with spatial resolution varying
from 30 - 80 m, excluding the thermal bands. When a study must
combine the use of data acquired by these sensors, caution should
be taken as the information contents revealed by the images with
changing resolutions may vary, as indicated by several empirical
studies (e.g., Markham and Townshend, 1981; Townshend and
Justice, 1981; Mumby and Edwards, 2002).
In a related project, the author evaluated the impacts of image
resolutions upon land use/cover classification by using three
images of the same day in 1987 for the Atlanta metropolitan
area: a MSS image, a TM image (excluding the thermal band),
and a degraded TM image (Table 2.6). The degraded TM image
comprises bands 2, 3, and 4 of the original TM image but the
radiometric quantification has been downgraded to 7 bits. Thus,
the degraded TM image has similar spectral and radiometric
resolutions to the MSS image but the spatial resolution is iden-
tical to that of the original TM image. All three images were
classified using unsupervised method described in Section 2.3,
excluding the spatial reclassification procedures. The thematic
accuracy for each classified map is evaluated and the area for each
land use/cover category is compared. Results indicate that the
improved spectral and radiometric resolutions are quite helpful
in achieving better classification accuracies. But the improved
spatial resolution does not help much in this regard. This is not
a surprise because improved spatial resolution can lead to an
increase not only in the interclass variability but also in the intr-
aclass variability, which can undermine classification accuracies
if a classic pixel-by-pixel classification method is used without
refinements (Markam and Townshend, 1981; Williams et al .,
1984; Haack, Bryant and Adams, 1987).
When looking at the specific statistics in Table 2.6, the original
TM image consistently gives a better classification accuracy for
each category than that of the MSS, especially for the low-density
urban use. If the land use/cover category area statistics extracted
from the original TM image are used as the reference, the MSS
data tend to greatly overestimate the area of low-density urban

use, a suburban housing category associated with woodland,
lawns, and open space. Therefore, ''low-density urban use''
clusters of pixels produced by unsupervised classification tend to
exhibit mixed reflectance from residential use, forest land, and
exposed/cultivated land. Because of the coarser spatial resolution,
the MSS image mixes the reflectance of these component land
cover types per pixel more than that of the TM image. This
explains why the areas for ''exposed/cultivated land'' and ''forest''
were underestimated from the MSS image relative to the original
TM image.
The observations from our image resolution impact assess-
ment experiment can be used to guide the design and implemen-
tation of digital image processing strategies for ensuring reliable
results being derived from multidate, multiresolution Landsat
data. First, because the spectral and radiometric characteristics of
an image can be significant for digital land use/cover mapping,
any enhancements to spectral and radiometric resolutions should
be pursued. Second, appropriate remedial procedures, such as
image preprocessing, should be used to improve the quality of
Landsat MSS data. Finally, proper image processing procedures
should be identified to ensure improved classification accuracies
being achieved from the Landsat TM data with higher resolutions.

2.4.3 Image preprocessing

As a prerequisite for change detection, geometric rectification
makes possible the production of spatially corrected maps of
land use/cover through time. Ideally, sufficient ground control
data should be collected to rectify at least one good-quality
image within a satellite time series, which can be further used
as the 'master' image to rectify other ''slave'' images. While this
preprocessing procedure has been conducted in virtually every
change detection project, the need for radiometric correction
among a satellite time series has recently been stressed (e.g.,
Cihlar, 2000; Yang and Lo, 2000, 2002; Du, Teillet and Cihlar,