decade. This site is an ideal test area for land cover change
study by using satellite remotely sensed data.
Korea Peninsula
Hungnam
e
Kwangju
e Pusan
Figure 1. The Study Area in Ansan, Korea.
2.2 Remotely Sensed Data
The remotely sensed data were acquired by TM on Landsat 5,
which covers a region cross Ansan, Incheon, and Seoul, with 7
bands by 2377 lines and 2357 samples i.e. totally 5,602,589
points for each band. The data used in this study were cut from
the original dataset, with 600 lines and 600 samples which
covers the study area shown in Figure 1. Ideally, the remotely
sensed data should be obtained from a sensor system that
acquires data at approximately the same time of day.
Furthermore, whenever possible it is desirable to use remotely
sensed data acquired on anniversary dates. For the purpose of
yearly land cover change detection, four dates of remotely
sensed data were obtained from the years of 1985 to 1993
(Table 1).
Table 1. Landsat TM Data Used in This Study
Time Number of Bands | Data Size of Each Band
May 14, 1985 7 600 x 600 pixels
May 20, 1987 7 600 * 600 pixels
April 26, 1990 7 600 x 600 pixels
May 20, 1993 7 600 x 600 pixels
2.3 Available Conventional Data and Processing
As a pilot study, most useful conventional materials, such as
detailed geographic status, physical environment, land use map,
social and economic development, ecosystem, population and
population growth, GDP and its growth, agricultural
production and production growth for various crops concerned,
meteorological and hydrological data, etc., have not obtained
yet. Only the following conventional data are available for
Ansan area: topographic map, scale 1:250,000 of 1996;
administrative boundary and road network map, scale
1:500,000 of 1995.
30 METHODOLOGY
Numerous algorithms have been developed for change
detection, such as change detection using write function
memory insertion (Price et al., 1992; Jensen et al., 1993);
multi-date composite image change detection (Fung and
LeDrew, 1987; Eastman and Fulk, 1993); image algebra
change detection (Green et al, 1994): univariate image .
differencing (Weismiller et al., 1977; Williams and Stauffer,
1978), image regression (Jenson, 1983; Singh, 1986), image
ratioing (Howarth and Wickware, 1981), vegetation index
differencing (Nelson, 1983), etc.; manual on-screen digitization
of change (Light, 1993; Wang et al., 1992; Lacy, 1992); post-
classification comparison change detection (Rutchey and
Velcheck, 1994); knowledge-based vision systems for detecting
change (Wang, 1993, Gong et al., 1996); multi-date change
detection using a binary change mask applied to date 2; multi-
date change detection using ancillary data source as data 1;
spectral change vector analysis; and so on.
The selection of an appropriate change detection algorithm is
very important (Jensen, 1996). First, it will have a direct
impact on the type of image classification to be performed (if
any). Second, it will dictate whether important “from-to”
information can be extracted from the imagery. Post-
classification comparison change detection was selected to
perform land cover change detection in this study. By using
this method, it requires rectification and classification of each
remotely sensed image. These two maps are then compared on
a pixel-by-pixel basis using a change detection matrix. The
advantage of this method include the detailed from-to
information that can be extracted and the fact that the
classification map for the next base year is already complete
(Jensen, 1996). However, every error in the individual date
classification map will also be presented in the final change
detection map (Rutchey and Velcheck, 1994). Therefore, it is
imperative that the individual classification maps used in the
post-classification change detection method be as accurate as
possible (Augenstein et al., 1991).
To perform change detection by using post-classification
comparison change detection method, classification algorithm
must be developed or selected so as to create accurate
classification maps for various years. Replacing to develop a
new classification method and algorithm, the authors tested the
common and reputable methods, including unsupervised and
supervised classification methods, and selected the most
suitable one for the study area to perform change detection.
The algorithms include: (1) Parallelepiped (PAR), (2)
Minimum Distance (MID), (3) Mahalanobis Distance (MAD),
(4) Maximum Likelihood (MAL), and (S) IsoData
unsupervised classification (ISD).
400 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
anm
H
the
Ih
par
low
eva
doe
to €
32
For
Dis
wh:
It s|
assi
The
Wh
IS |
bet
(Scl
