International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 Intert 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
transportation in the region relies on a two-lane highway that 3. METHODS 32 ( 
parallels with the coastline. Almost all other roads are quite 
narrow, which causes traffic congestion. The construction of 3.1 Principal Components Analysis Class 
the two bridges in 1973 and 1989, one of which is located in : : : image 
lower left part of the study area (Figure 1), led to better Adjacent bands in a multispectral remotely-sensed image are sensi 
highway transportation network, but it brought the rapid generally correlated. Presence of the correlation among the image 
urbanization problem around the recently constructed bands of the multispectral image implies that there is numb 
transportation areas within Beykoz district. redundancy in the data. In other words, some information is incre 
being repeated. It is the repetition of the information between becau 
In this study, multisensor data including Landsat ETM+ and the bands that is reflected in their inter-correlations (Mather, chara: 
Terra ASTER imagery, acquired in May 2001 and October 1999). Principal components analysis (often called PCA or the c 
2002 respectively, were used for the delineation of eight main Karhumen-Loeve analysis) has proved to be of value in the provi 
land-cover classes. These classes are namely coniferous forest, analysis of multispectral remotely-sensed data. The classi 
deciduous forest, urban, inland water, grassland, bare soil, road transformation of the raw remote sensing data using PCA can resear 
and sea. Due to the short time difference between the ~~ result in new component images that may be more Maxi 
acquisition dates it is assumed that there was no dramatic interpretable than the original data (Ashutosh, 2002). appro 
change on the types of ground cover classes. The area under Additionally, this technique reduces contributions of noise and mathe 
the analysis (Figure 1) was about a 430 km? region covering error. PCA can be used to reduce the information included in Inforn 
Beykoz district. Borders of Beykoz district are also depicted on ‘he raw data into two or three bands without losing significant photo 
Figure 1. information (Monger, 2002). The principal components cover 
analysis can be used for effective classification of land use, prope 
colour representation or visual interpretation with multi-band work 
data and change detection with multi-temporal data (Sunar, on the 
1998). gener: 
categc 
PCA is used in this study to improve the quality of the 
classification. It is applied to the satellite images to obtain 3.2.1 
uncorrelated — (i.e. statistically independent) principal 
components lying on orthogonal axes that the original data are The | 
reprojected. The results of PCA for both images including statist 
eigenvalues and variances of each component are given in in ea 
Tables 1 and 2. Whilst the first three components for the belon: 
Landsat ETM- image represent 9995 of the image data, those locati 
for the Terra ASTER image account for 97% of the image data. the m 
First three components of PCA analyses for both images were 
used to form three-layer images, which are later used in These 
classification processes. memb 
the m 
: for de 
Component | Eigenvalue | Variance (%) | Total (%) | The s 
1 2445.38 71.30 71.39 dimen 
2 884.21 25.81 97.20 orient: 
3 72.65 2.12 09.32 be m 
| 4 11.98 () 35 99.67 trainin 
aus 5 8.88 0.26 99.93 shapes 
6 237 0.07 100.00 dimen 
Figure 1. The location of study area, Beykoz district, Istanbul centre 
Table 1. PCA for Landsat ETM+ image ds; 
The Landsat ETM+ image was rectified to the UTM projection It SH 
using several 1:25000 scale topographic maps. In the geometric ; Mu . à SIVES € 
correction process, a total of 22 Ground Control Points (GCP) Component | Eigenvalue | Variance (%) Total (%) | the 
were used, which resulted in a Root Mean Square Error 1 16094.89 82.07 82.07 classif 
(RMSE) of less than 0.5 pixels. The Terra ASTER image of 2 2118.43 10.80 92.87 data A 
the study area was later rectified to the geometrically corrected 3 883.73 4.51 97,37 | estima 
Landsat ETM-- image using 29 GCPs giving a RMSE of 4 191.99 0.98 98.35 the cla 
around 0.5 pixels. The two corrected images were resampled at 5 109.97 0.56 98.91 value J 
a spatial resolution of 30m using the nearest neighbour 6 85.03 0.43 99.35 pro 
algorithm in order to prevent the formation of new pixel 7 51.58 0.26 99.61 yaar 
values. Subimages of 714 by 668 pixels covering the study area 8 47 79 0.24 99.85 322 
were extracted and used in subsequent analyses. 9 28.81 0.15 100.00 d 
The bz 
Table 2. PCA for Terra ASTER image proces: 
. brain. 
: 932