Use of archival Landsat imagery to monitor urban spatial growth - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 26

TABLE 2.6 Image resolution impacts upon land use/cover classification.

Area statistics a (ha) Land use/ Kappa index

cover class Original Degraded MSS Original Degraded MSS

TM TM image TM TM image

image image image image

High-density urban use 50 112 49 917 52 086 0.70 0.66 0.60

Low-density urban use 180 153 214 312 256 103 0.72 0.63 0.47

Exposed/cultivated land 34 083 35 164 5767 0.47 0.31 0.36

Cropland or grassland 140 499 136 692 130 494 0.68 0.49 0.61

Forest 619 024 587 587 581 032 0.88 0.79 0.77

Water 20 212 20 351 18 605 0.93 0.82 0.72

Overall Kappa index 0.71 0.65 0.63

NA

Overall accuracy 75.19% 69.63 68.15%

a The total area is approximately 1 044 030 ha.

2002; Schroeder et al ., 2006; Vicente-Serrano, Perez-Cabello
and Lasanta, 2008; Yang and Chen, 2010). The latter becomes
particularly relevant when using satellite data spanning several
decades, such as archival Landsat imagery acquired by different
sensors.
The differences in radiometry among a satellite time series,
caused by such factors as sensor variations, atmospheric prop-
erties, and sensor-target-illumination geometry, must be elim-
inated or minimized so that a common radiometric response
among the data set can be restored. To this end, the relative
radiometric normalization (RRN) method is preferred over the
absolute radiometric correction method as no in situ atmo-
spheric data at the time satellite overpasses are necessary (Hall
et al ., 1991). There are some RRN methods that have been pro-
posed, such as pseudoinvariant features, radiometric control set,
image regression, no change set determined from scattergrams,
histogram matching, among others. A comprehensive review on
these methods was given by Yang and Lo (2000). These methods
are based on different assumptions and hence can perform differ-
ently with respect to the variations in land use/cover distribution,
water-land proportion, topographic relief, similarity between the
reference and the subject scenes, and sample size. Caution must
be taken when using a RRN method being visually and statisti-
cally robust since it can substantially reduce the magnitude of
meaningful spectral differences. Because of the space limit, we
are not able to delve into specific RRN procedures, and readers
can always refer to the work published by Yang and Lo (2000)
for some further discussions.

for urban land use/cover change detection was assessed by Kam
(1995) and Ridd and Liu (1998). The image-to-image comparison
can be conducted by using raw images (e.g., Green, Kempka and
Lackey, 1994), a composite image (e.g., Liu and Lathrop, 2002),
or derived indexes (e.g., Yang and Liu, 2005a). Being gener-
ally accurate, the image-to-image comparison approach provides
detailed information on intra-class land use/cover modifica-
tion or intensification but suffers from the inability to provide
detailed information of how various urban land use/cover cate-
gories change (i.e., inter-class change) (Singh, 1989; Kam, 1995;
Ridd and Liu, 1998; Yang and Liu, 2005a). On the other hand,
the map-to-map comparison can be completed by using two
classified maps, providing a full matrix of change, as demon-
strated in the case study reported in this chapter. However, the
effectiveness of the map-to-map comparison approach is highly
dependent upon the assumptions and techniques used to pro-
duce maps of the same area at different times. Conventional
pattern classifiers are largely built upon parametric statistics.
They generally work well for medium-resolution scenes covering
spectrally homogeneous areas, but not in heterogeneous regions
or when scenes contain severe noises due to the increase of image
spatial resolution.
For years substantial research efforts have been made to
improve the performance of image pattern classification for
working with different types of remote sensor data, and some
strategies have been developed as a result of such efforts. Examples
include: (i) the identification of various hybrid approaches
that combine two or more classifiers, or incorporate pre- and
postclassification image transformation and feature extraction
techniques (e.g., Yang and Liu, 2005b); (ii) the development
of 'soft' classifiers by introducing partial memberships for each
pixel to accommodate the heterogeneous and imprecise nature
of the real world (e.g., Shalan, Arora and Ghosh, 2003); (iii) the
decomposition of each pixel into independent endmembers or
pure materials to conduct image classification at sub-pixel level
(e.g., Verhoeye and Wulf, 2002; Chapter 9); (iv) the incorpora-
tion of the spatial characteristics of the neighboring (contextual)
pixels to develop object-oriented classification (e.g., Walker and
Briggs, 2007; Chapter 7); (v) the fusion of multisensor, multi-
temporal, or multisource data for combining multiple spectral,

2.4.4 Change detectionmethods

After data acquisition and preprocessing, the next step in the
generic urban spatial growth monitoring workflow is to iden-
tify a digital change detection method. Either image-to-image
comparison or map-to-map comparison can be used for this
purpose. General reviews of different algorithms under these
two approaches are given elsewhere (e.g., Singh, 1989; Lu et al .,
2004; Radke et al ., 2005). The effectiveness of these techniques