Inter] 
——— 
COMPARISON OF PIXEL-BASED AND OBJECT-ORIENTED CLASSIFICATION uM 
APPROACHES USING LANDSAT-7 ETM SPECTRAL BANDS to fc 
Like 
objec 
M. Oruc* *, A. M. Marangoz *, G. Buyuksalih * analy 
ii eCog 
* ZKU, Engineering Faculty, 67100 Zonguldak, Turkey - (oruc, marangoz, buyuksalih)@jeodezi.karaelmas.edu.tr 
PS ThS 11 
( 
KEY WORDS: Remote Sensing, Land Cover, Classification, Landsat, Multispectral 
ABSTRACT: 
In this study, land cover types in Zonguldak test area were analysed on the basis of the classification results acquired using the pixel 
based and object-oriented image analysis approaches. Landsat-7 ETM with 6 spectral bands was used to carry out the image 
classification and ground truth data were collected from the available maps, aerial photographs, personal knowledge and 
communication with the local people. In pixel-based image analysis, firstly unsupervised classification based ISODATA algorithm 
was realised to provide priori knowledge on the possible candidate spectral classes exist in the experimental area. Then supervised 
classification was performed using the three different approaches of minimum-distance, paralellepiped and maximum-likelihood. On 
the other hand, object-oriented image analysis was evaluated through the eCognition software. During the implementation, several 
different sets of parameters were tested for image segmentation and nearest neigbour was used as the classifier. Outcome from the 7 
classification works show that the object-oriented approach gave more accurale results (including higher producer's and user's 
accuracy for most of the land cover classes) than those achieved by pixel-based classification algorithms. 
pixel-based and object-based approaches with the detailed 
1. INTRODUCTION explanation of the obtained results. : 
2. CLASSICAL VERSUS OBJECT-ORIENTED 
Classification based on pixel-based approaches to image CLASSIFICATION TECHNIQUES 
analysis is limited at present. Typically, they have considerable 
difficulties dealing with the rich information content of high- The overall objective of classical image classification 
resolution data e.g. Ikonos images, they produce inconsistent procedures is to automatically categorize all pixels in an image 
classification results and they are far beyond the expectations in into land cover classes or themes. Normally, multispectral data These 
extracting the object of interest. This situation brings ^ are used to perform the classification and, indeed, the spectral eCogr 
meaningful operator intervention to the implementation. Due to pattern present within the data for each pixel is used as the vas 
mentioned nature of classical methods, new and object-oriented numerical basis for categorization. That is, different feature and 
image analysis of eCognition software can be used. Such types manifest different combinations of DNs based on thei | 
algorithm requires one or more image segmentations which inherent spectral reflectance and emittance properties. In this Li 
should aliso be supported by the additional information like light, a spectral “pattern” is not at all geometric in character SEN 
contextual or textual to make the segments more appropriate for Rather, the term pattern refers to the set of radiant reality 
improve classifications. measurements obtained in the various wavelength bands for S 
: each pixel. Spectral pattern recognition refers to the family of Seale 
Object-oriented approach takes the form, textures and spectral classification procedures that utilizes this pixel-by-pixe du 
information into account. Its classification phase starts with the spectral information as the basis for automated land cov maxim 
crucial initial step of grouping neighboring pixels into classification. In this study, three different classical plus on ee 
meaningful areas, which can be handled in the later step of  objecvt-oriented classification techniques have been employe N er 
classification. Such segmentation and topology generation must in sequence. For the classical methods e.g. minimum-distan ig 
be set according to the resolution and the scale of the expected parallelepiped and maximum likelihood, detailed information iR 
objects. By this method, not single pixels are classified but can be found at Lillesand and Kiefer, 1994. In the following s uc 
homogenous image objects are extracted during a previous paragraphs, main background of eCognition 3.0 software bast mos 
segmentation step. This segmentation can be done in multiple on object-oriented image analysis is given. m 
resolutions, thus allowing to differentiate several levels of ex 
object categories. Segmentation is the first and important phase in the eCognition Sem 
software and its aim is to create meaningful objects. This mei . e d 
In this study, Landsat-7 ETM image Zonguldak testfield of that the shape of each object in question should ideally X RR 
Turkey has been realized by PCI Geomatica V9.1.2 and represented by an according image object. This shape combined s m 
eCognition 3.0 software packages. Classical and the object- with further derivative colour and texture properties can be use C ek 
based classification techniques have been implemented. Several to initially classify the image by classifying the generated which 
tests have been carried out to match with the successful image objects. Thereby the classes are organised within a clas descri 
segmentation, then the classification by entering different hierarchy. Each class can have a sub- or super-class and ths do b 
parameters to the used software. Authors, finally compare the inherit its properties from one or more super-classes Or to 15 each 
  
  
  
* Corresponding author. 
1118