004 
Arce 
Sr". 
06 
tion 
asp 
ject 
ofra 
useo 
and 
LE. 
hank 
ty of 
FOREST TYPES CLASSIFICATION USING ETM+ DATA IN THE NORTH OF IRAN/ 
COMPARISON OF OBJECT-ORIENTED WITH PIXEL-BASED CLASSIFICATION 
TECHNIQUES 
S. Shataee ^ * T. Kellenberger", A. A. Darvishsefat € 
* Forestry faculty, Gorgan University, Beheshti Street, 49165, Gorgan, Iran — shatace@gau.ac. ir 
b : i TAPAS . : ^ . . : - . 
Remote sensing laboratories, Department of Geography, University of Zurich, CH-8057 Zurich, Switzerland- knelle@geo.unizh.ch 
Natural Resources Faculty, Tehran University, Karadj, Iran — adarvish@chamran.ut.ac.ir 
Commission PS, Working Group VII/I 
KEY WORDS: Forestry, Mapping, Segmentation, Comparison, Classification, Method, Object, Pixel 
Different classification methods have been used for classification of satellite data by researchers. Addition to current classification, 
pixel-based methods, developments in segmentation and object oriented techniques offer the suitable analyses to classify satellite 
data. To compare the pixel-based with object-oriented classification approaches to extract forest types, a case study in a small area 
have been accomplished in the northern forests of Iran. The ETM+ data and Some processed bands, which extracted by suitable 
processing analyses, were used due to high spectral resolution. Pre-processing of data was done for geometric correction of images 
and corresponding to ground truth map. The best suitable data sets have been chosen by seperability indexes, In the pixel-based 
classification approach, the maximum likelihood classifier classified images of data set. In the object-oriented approach, images were 
segmented to homogenous area as forest types by suitable parameters in some level. Classification of segments was done trough three 
classification methods of nearest neighbor, membership function and combination of both methods. A sample ground truth map of 
forest type did the accuracy assessment of the results. It was generated trough sampling method by 193 plots of one hectare. The 
accuracy assessment of the results showed that the object-oriented classification approach could improve considerably the results in 
compare to pixel based classification approach (19%), However, increasing of kappa coefficient from 25.5 % in the pixel based 
classification to 44.4 % in the object-oriented approach shows capability of multiresolution Segmentation of data, which provide 
other useful attributes for classification in addition to Spectral information (or overall accuracy from 44 % to 61%). The results of 
study indicate that integration of nearest neighbor with membership function technique can improve the results more than the both 
techniques individually. More researches to Survey on these classification techniques will be necessary in future. 
1. INTRUDUCTION On the other hand, the eCognition software was designed based 
on the object-oriented classification, including new classifiers 
rent fielding ways, is time and techniques by providing the new possibilities for 
consuming and cost-intensive. Using satellite Imagery and its multiresoultion segmentation ‚of the images and the object- 
potentials are new tools in order to managing and mapping Oriented fuzzy-rule classification. With these facilities, it Was 
successfully applied in several different applications (Gorte, 
1998; Baatz and Schape, 1999; Blascke et. al., 2000). The 
object-oriented classification has been used in some studies 
related to forestry. Application of these methods, following with 
à comparison of their results With pixel-based classification was 
examined in some case studies [Willhauck, 2000, De koke et. 
al., 1999]. They applied these methods on different data set and 
could get better results in compare to the pixel-based 
Forest types mapping trough cur 
forest-covered area. Since, different classification methods and 
Satellite imagery have been used. Beside these, introducing a 
suitable data and method of classification are the main 
interesting for researchers. 
During the last decades, researchers have mainly focused on the 
pixel-based classification in different applications. The pixel- 
based classification is à current method because satellite data 
Sets are acquired digitally on the basis of pixel units. In the P get à 
pixel-based classification method, Image Statistics is a base that classification approaches. Schwarz and et. al (2001) applied the 
pixel-based classification and pattern recognition are established Pixel based and the object oriented classification methods on 
on it. Current pixel based classifiers such as maximum SPOT and IKONOS data to recognize forest damage areas by 
likelihood; minimum distance and parallel-epiped were ^ storm. The results showed that extraction of the damaged area 
designed on this base. Since these approaches are based on IKONOS data by the object-oriented method could be done 
exclusively on the digital number of pixels and are used only better in compare lo the pixel-based method. ; 
spectral information, the results look like slat-peppery images. In the mixed forest like which forest exist In the north of Iran, 
In order to overcome these problems and to produce more grouping the similar specious of trees as stands or forest types is, 
homogenous pattern for classification, the researchers have also difficult and Is generally determined by dominant Shegies and 
experienced per-parcel or per-field classification methods percent of their existences IN an area. Therefore, In the satellite 
(Haralick et. al, 1973, Jensen, 1993, alpine et al., 1999). In a data, reflectance of stands is referred to various. trees together 
per-field analysis or ‘pixel in polygon: analysis, pixel With open cover area. At the result, delineating of homogeneous 
information is linked to a spatial database and boundary maps. area as stands or forest types is difficult and normally with 
Field or parcel refers to homogenous patches of land 
error. Then, forest type classification by using of traditional 
(agricultural fields, gardens, urban structures or roads) which pixel based methods often leads to noisy results, which refer to 
already exist and are superimposed on the image. 
existent of heterogeneous pixels in a homogenous area. 
1105