Use of archival Landsat imagery to monitor urban spatial growth - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 18

the TM and ETM + images because they have been processed to
good radiometric quality.

through which some ''manual'' operations can be implemented
automatically. With the use of this method, four major types
of spectral confusion were identified: low-density urban (mostly
residential) versus forest (clearcuts, sparse forest, and wetlands);
low-density urban (sparse residential) versus cropland or grass-
land; forest (sparse forest and shrubs) versus cropland/grassland;
and high-density urban (large open roof buildings, air fields,
and multilane highways) versus cultivated or exposed land (large
barren landmass, river beach, fallowed land). Spectral confusions
were found to be generally more serious for the MSS data than for
the TM and ETM + data. Spectrally confused clusters were first
identified, and AOI layers were created by on-screen digitizing.
The AOI layers served as masks for splitting confused clusters.
Finally, GIS-based reclassification functionality was employed to
recode the split clusters into correct land classes. This was an
interactive process until an acceptable accuracy was obtained.

2.3.3.2 Unsupervised classification

The 1973, 1979, 1987, 1993, 1999, and 2007 images were classified
by using a two-step unsupervised method. Firstly, the ISODATA
(Iterative Self Organizing Data Analysis) algorithm was used
to identify spectral clusters from image data, excluding the
thermal bands for the TM and ETM + images because of their
coarse spatial resolutions. It was implemented without assigning
predefined signature sets as starting clusters to avoid the impacts
of sampling characteristics (Vanderee and Ehrlich, 1995). Some
important clustering parameters specified include number of
classes, convergence value, and maximum number of iterations.
To determine the optimum class number, we tested 20, 40, 60,
and 80, and found that 60 for the TM images or 80 for the ETM +
image allows the resultant clusters to be better interpreted in
relation to the classification scheme. The convergence value was
specified as 0.990 for all types of data. Finally, for the TM or MSS
data, 60 was used for the maximum number of iterations while
80 for the ETM

2.3.3.4 Thematic accuracy assessment
Due to the limited availability of ground truth data, it is impos-
sible to perform accuracy assessment for each map exhaustively.
The strategy adopted here was to assess the maps produced from
each type of imagery covering the study area. Three maps were
selected for the thematic accuracy assessment: the 1988 map from
MSS data, the 1997/1998 map from a summer 1997 and a winter
1998 TM images, and the 1999 map from ETM

scene.
The resultant clusters were assigned into one of the six land
use/cover classes through visual inspection of the original images
in relation to appropriate ancillary data. To label the clusters, the
original image and the clustered map were displayed side by side
and then spatially linked. The class assignment for specific clusters
was based on an examination of the cluster at the two different
detail levels. At the large-scale level, image color was mainly used
in decision making; at the small-scale level, however, additional
image elements such as association and site were combined to
improve classification quality. Also a note was taken for any
spectral cluster containing more than one land use/cover type.
This happens when spectral contents of more than two different
land use/cover classes are similar. Whenever this occurred, the
cluster was initially labeled as one of the most likely land use/cover
classes. And at a later stage, these clusters were split into smaller
clusters for the correct land use/cover labels using the spatial
reclassification procedures described below.

+

+

data. The first
two were produced by the author for other projects using the
same method described in this chapter. The accuracy assessment
was carried out by using a standard method recommended by
Congalton (1991). Results revealed that the land use/cover maps
based on TM or ETM + images yielded a slightly better accuracy
than that from MSS data (Table 2.2). The maps derived from TM
or ETM data show a slightly higher kappa index for low-density

urban, cropland/grassland, and forest than that from MSS data.
This could be the result of higher spatial, spectral, and radio-
metric resolution of the TM or ETM + data. However, the map
derived from MSS data is compatibly accurate in every respect to
the ones from TM or ETM + data. Overall, all these maps meet
the minimum 85 percent accuracy stipulated by the Anderson
classification scheme (Anderson et al ., 1976), indicating that the
image processing procedures adopted here have been effective in
producing compatible land use/cover data over time, despite the
differences in spatial, spectral, and radiometric resolution of the
three generations of Landsat data used in this project. Other maps
were produced using the same procedures. It is anticipated that
the same level of accuracy should be maintained. Land use/cover
classification maps for six dates are shown in Fig. 2.3. The statis-
tics of each classification map are summarized in Table 2.3 and
the land use/cover changes during different periods are shown in
Table 2.4.

+

2.3.3.3 Spatial reclassification
Spatial reclassification was used to reduce the two types of
misclassification errors on the initial maps produced through the
unsupervised classification. The first type was the boundary error
due to the occurrence of spectral mixing within a pixel, which
is small relative to the areas of correct classification. Within a
class there are some anomalous pixels representing the noise in
the data. A modified 3 3 modal filter with the four corner

neighbors disabled was used to suppress the boundary errors.
The second type of misclassification errors was the spectral
confusion error due to the spectral similarity between different
land use/cover classes, which is inevitable for an image acquired
with a broad-band sensor and tends to be more perceptible in an
urban scene than in a rural one. Defining the spectral confusion
involves the use of spatial and contextual properties through an
image interpretation method that can be incorporated effectively
into a digital classification procedure with on-screen digitizing,
multiple zooming, Area Of Interest (AOI) functionality, and
other relevant spatial analysis tools. In addition, several image
processing programs permit advanced tools for geoprocessing

×

2.3.4 Change detection

Change detection was used to examine urban growth and the
nature of change. To analyze urban growth, the spatial distribu-
tion of each urban land class was extracted from each map in
the time series. The change in urban land was summarized by
using the GIS minimum dominate overlay method, which allows
the smallest amount of high-density or low-density urban use
in the earliest year to show up fully, while only the net addition