ul 2004 
——— 
»razione 
olletting 
Itimetry 
sensing 
Journal 
images 
napping 
ivale of 
Claudio 
CLASSIFICATION OF MULTI-SPECTRAL, MULTI-TEMPORAL AND MULTI-SENSOR 
IMAGES USING PRINCIPAL COMPONENTS ANALYSIS AND ARTIFICIAL NEURAL 
NETWORKS: BEYKOZ CASE 
M. Cetin *, T. Kavzoglu »*, N. Musaoglu ° 
“ Dept. of Geodetic and Photogrammetric Engineering, Gebze Institute of Technology, 41400 Gebze-Kocaeli, Turkey - 
(mcetin, kavzoglu)@gyte.edu.tr 
° Istanbul Technical University, Civil Engineering Faculty, 34469 Ayazaga, Istanbul, Turkey - 
nmusaoglu@ins.itu.edu.tr 
KEY WORDS: Land Cover, Classification, Artificial Neural Networks, Principal Components, Maximum Likelihood, PCA 
ABSTRACT: 
The thematic maps derived from remotely-sensed images are invaluable sources of information for various investigations since they 
provide spatial and temporal information about the nature of Earth surface materials and objects. The robustness of classification 
techniques used to produce these thematic maps can be crucial especially for complex classification problems. This study aims to 
determine the level of contributions of multi-temporal and multi-sensor data together with their principal components for 
Maximum Likelihood and Artificial Neural Network classifiers. The performance of a multi-layer perceptron that learns the 
characteristics of the data using backpropagation algorithm is compared to that of Maximum Likelihood classifier in identifying 
major land cover classes present in the study area, Beykoz district of Istanbul, Turkey. The image data available for the study are 
from Landsat ETM+ and Terra ASTER images. Image band combinations are inputted to the neural network for training and the 
success of the classification is tested using test data sets. Results show that the neural network approach is an attractive and 
effective way of extracting land cover information using multi-spectral, multi-temporal and multi-sensor satellite images. It is also 
observed that the level of contribution of principal components to the results is much less than the contribution of multi-temporal 
data in terms of the classification accuracy. 
1. INTRODUCTION 
Satellite images and extracted thematic maps provide top-level 
information for the inventory, monitoring and management of 
natural resources. Given the diversity and heterogeneity of the 
natural and human-altered landscape, it is obvious that the 
time-honoured and laborious method of ground inventory is 
inappropriate for mapping land use and land cover over large 
areas (Civco, 1993). Therefore, the use of remotely sensed 
images is essential for regional or global scale studies. With 
the launch of recent satellites, much image data are available 
from different sources regarding the same area. Naturally, each data 
type represents different characteristics of the area. Using all these 
data may improve the accuracy of classification significantly but 
introduces redundancy and requires more training data. Therefore, 
new data should be added only if they contribute to an improved 
classification. In the literature, a single date satellite image has 
been usually used for classification problems. However, it is 
well known that it is necessary to consider the reflectance 
characteristics of surface objects varying with the scasonal 
conditions to improve the accuracy of the classification. 
One of the most significant recent developments in the field of 
land cover classification using remotely-sensed data has been 
the introduction of Artificial Neural Network (ANN) models. 
They can be thought of as forms of models imitating the 
complicated brain processing in a very simple way. This method 
has been recently used in a wide range of classification and pattern 
recognition problems ranging from signal recognition to image 
  
= (a : 
Corresponding author 
compression. In the remote sensing arena, they have been recently 
applied to many applications, but the most popular application of 
the method in remote sensing is the classification of land cover 
information (Paola and Schowengerdt, 1995; Gopal and Woodcock, 
1996; Sunar Erbek et al. 2004). The use of ANN method has 
become popular due to the unique advantages of the method over 
conventional statistical methods. Perhaps the most important 
characteristic of ANN is that there is no underlying assumption 
about the distribution of data. Furthermore, it is possible to employ 
data from different sources to improve the accuracy of the 
classification. 
For this study the performances of classification methods, the 
Maximum Likelihood and ANN classifiers, were tested for the 
inclusion of multi-temporal, multi-sensor and principal 
component images in classification processes. In order to meet 
the objective of this research Landsat ETM+ and Terra ASTER 
images and their first three principal components accounting 
for the highest variances were employed in land cover 
classification and the results for the two classifiers were 
compared. 
2. STUDY AREA AND DATA 
The study area, Beykoz, located along the north-east side of the 
Bosphorous on the Asian side of Istanbul, is one of the least 
dense regions mainly due to its 60% forest coverage. The 
elevation of Beykoz ranges from 0 to 400 metres. Most