| and 
ensing 
water 
data 
remote 
rocess 
‚As a 
ic and 
! been 
ematic 
errors 
ysis of 
tation, , 
ication 
tation. 
imount 
ensing 
canner 
olution 
HRV), 
Xpress 
Ratio 
etation 
te the 
(12) 
and 4, 
sary to 
) steps. 
_andsat 
ting 3 
and c) 
erform 
(13) 
(14) 
(15) 
band 1, 
o make 
cc.img, 
ined on 
May 20 
  
  
4.4 Classification System 
Certain classification schemes have been developed that can 
readily incorporate land-use and/or land-cover data obtained by 
interpreting remotely sensed data (e.g., U.S. Geological Survey 
Land Use/Land Cover Classification System, U.S. Fish & 
Wildlife Service Wetland Classification System, NOAA 
CoastWatch Land Cover Classification System, Asian Land 
Cover Classification System, etc.) (Jensen, 1996; Tateishi et al, 
1995). The U.S. Geological Survey Land Use/Land Cover 
Classification System was chosen and referred to form the 
classification system for this study. By considering on the four 
levels of the U.S. Geological Survey Land Use/Land Cover 
Classification System (Jensen, 1996) and the type of remotely 
sensed data typically used to provide the information, the 
classification system was created as in Table 2. 
Table 2. Land Cover Classification System in Ansan, Korea 
  
Class Description 
Unclassified 
Forest land 
Rangeland 
Agricultural land 
Wetland 
Barren land 
Urban and/or built-up land 
Saltpan 
Tidal zone 
Water (sea and inland water body) 
  
  
  
  
  
  
  
  
  
  
ojos|-31[o NH DIN = 
  
  
  
  
4.5 Dataset and Algorithm Selection 
The Landsat TM data of May 20, 1993 was used as the test 
dataset to select the optimal dataset and most suitable 
algorithm for land cover classification. Two datasets were 
prepared based on the Landsat TM data of May 20, 1993, one 
is subr9305.img which includes seven TM bands, and the other 
is r*ndvi93.img which includes seven TM bands and NDVI. 
Two training datasets, 9305-1.roi and 9305-2.roi, were 
collected randomly from the false color composite image 
r93fcc.img. One of the two datasets was used to perform 
classification while the other was used to evaluate the accuracy 
of classification maps. 
The training data 9305-1.roi and the classification algorithms 
PAR, MID, MAD, and MAL were used to process the datasets 
subr9305.img and r+ndvi93.img, through which various 
classification maps were produced respectively: r93par.cla, 
rv93par.cla; r93mid.cla, rv93mid.cla; r93mad.cla, rv93mad.cla; 
r93mal.cla, rv93mal.cla; r93cut.cla, rv93cut.cla, where 
r93par.cla and rv93par.cla, for example, are classification 
maps produced from the datasets subr9305.img and 
r+ndvi93.img respectively by using PAR classification 
algorithm and training data 9305-1.roi. Also, the ISODATA 
classification algorithm was used to process the datasets 
subr9305.img and r+ndvi93.img, which produced classification 
maps r93isd.cla and rv93isd.cla. 
The error matrices of the classification maps were analyzed by 
using the training data 9305-2.roi. The overall accuracy and 
K,, statistics were computed for each classification maps. 
The results were listed in Table 3, which showed the 
Maximum Likelihood is obviously the best method while the 
two datasets were not too much different. Therefore, the 
Maximum Likelihood algorithm and the two datasets were 
used to perform classification for the Landsat TM data of all 
four dates in the study area. 
4.6 Classification Maps 
Training datasets 8505-1.roi and 8505-2.roi; 8705-1.roi and 
8705-2.roi; and 9004-1.roi and 9004-2.roi were collected from 
false color composite images r8Sfcc.img, r87fcc.img and 
r90fcc.img respectively. The Maximum Likelihood algorithm 
and the datasets 8505-1.roi, 8705-1.roi and 9004-1.roi were 
used to produce classification maps: r85mal.cla and 
rv85mal.cla; r87mal.cla and rv87mal.cla; and r90mal.cla and 
rv90mal.cla, while the datasets 8505-2.roi, 8705-2.roi and 
9004-2.roi were used to evaluate the accuracy of classification 
maps. Table 4 presents the overall accuracy and K, | statistics 
computed for all classification maps derived from Landsat TM 
data of May 14 of 1985, May 20 of 1987, April 26 of 1990 and 
May 20 of 1993. 
Table 3. Accuracy Comparison of Classification Maps 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Method Overall accuracy | Kei 
All TM 7 band data without NDVI 
MAD 89.86% 87.26% 
MAL 92.88% 90.91% 
MID 81.71% 77.32% 
PAR 85.20% 80.83% 
ISD 64.33% 57.33% 
All TM 7 band data and NDVI 
MAD 90.15% 87.60% 
MAL 93.28% 91.40% 
MID 85.04% 81.17% 
PAR 84.94% 80.52% 
ISD 78.79% 72.72% 
  
  
  
  
Table 4. Accuracy of Classification Maps Derived from 
Landsat TM data by MAL Method 
  
  
  
  
  
  
  
  
  
  
  
  
  
Date Overall accuracy K. 
All TM 7 band data without NDVI 
May 20, 1993 92.88% 90.91% 
April 26, 1990 86.03% 83.02% 
May 20, 1987 92.20% 90.45% 
May 14, 1985 91.39% 89.42% 
All TM 7 band data and NDVI 
May 20, 1993 93.28% 91.40% 
April 26, 1990 85.77% 82.72% 
May 20, 1987 92.70% 91.06% 
May 14, 1985 92.06% 90.24% 
  
  
  
4.7 Land Cover Change Detection 
Based on the accuracy assessment of classification maps, the 
classification maps rv93mal.cla, r90mal.cla, rv87mal.cla and 
rv85mal.cla with higher accuracy were used to detect the 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 403