bul 2004 
19-226 
tries. In: 
) Signal 
Robust 
es. In: 
o Signal 
daptive, 
ematical 
g, Paris, 
Beijing, 
IMPROVEMENT OF IMAGE CLASSIFICATION WITH THE INTEGRATION OF 
TOPOGRAPHICAL DATA 
Deniz Gerçek 
Geodetic and Geographic Information Technologies, METU 06531 Ankara, Turkey 
denizger@metu.edu.tr 
KEY WORDS: Remote Sensing, Classification, Land cover, Integration, Accuracy. 
ABSTRACT: 
Remotely sensed data are essentially used for land cover and vegetation classification. However, classes of interest are often imperfectly 
separable in the feature space provided by the spectral data. One of the most common attempts to improve image classification is the 
integration of ancillary data into classification. In this study, an approach for integrating topographic data into land cover classification is 
presented. Integration is basically through selection of training set in order to provide additional sensitivity to topographical 
characteristics associated with each land cover class in the study area. Topographic data including elevation, slope and aspect are tested 
for their correlation with land cover classes and correlated topographic data are used as input. Signatures from topographical data are 
assumed to represent the topographical preferences of land cover classes and are extracted with respect to the spatial position of spectral 
signatures from the remotely sensed images. Initial set of topographical signatures is evaluated and refined statistically. A new training 
set covering both spectral and the topographical signatures is created. New training set is used to supervise the standard Maximum 
Likelihood classification where; topographical raster data together with images is used as input for the classification. Two products are 
derived. First product used remotely sensed data only as input and is trained by spectral information. Second product used bands and 
topographical data as input and it is trained with both spectral and topographical information. Comparison between two products 
conveyed that procedure provided an improvement of 10% in overall accuracy for the classification with the integration of topographical 
data over the one that depended on spectral data only. 
INTRODUCTION 
Use of ancillary data has long been acknowledged as a necessity obvious approach, Logical channel method introduced by Strahler 
in remotely sensed image classification especially when et al. (1978), aims to increase the number of attributes or channels 
discriminating between different types of information classes is of information used in the classification. The second is classifier 
difficult due to low spectral seperability. Image classification modification, which involves changing a priori probabilities 
converts image data into thematic information by categorizing according to areal composition of the expected product based on 
spectral data into classes with respect to statistical decision rules image statistics, ancillary data or a known relationship between 
introduced by classifier algorithm. However, information gathered classes and ancillary data (Harris and Ventura, 1995; Mesev, 
by the classification of remotely sensed data, based solely on 1998). 
spectral variability is often insufficient in accuracy (Janssen et al., 
1990; Bruzzone et al., 1997). Accuracy of image classification can 
be improved with the integration of data and/or information other 
than the imagery (Westmoreland and Stow, 1992; Gahegan and 
Flack, 1996). The data and information, also known as “ancillary 
data” in the literature, are often composed of map-based thematic 
data, terrain data and non-spatial data. There have been numerous 
attempts to increase overall accuracy of classification during the 
period regarding the use of automated classification systems. 
Hutchinson (1982) categorized these attempts into three according 
to their proceeding before, during or after classification. 
Logical channel approach is advantageous for being simple and 
time saving compared to others. However without any 
modification or adjustment of conventional sampling routines 
before class statistics generation, method has obvious limitations. 
Logical channel approach covering simple addition of ancillary 
data as input into classification intuitively lacks the ability to 
handle data of different form and ranges. These limitations may 
cause problems in generating class statistics, furthermore; training 
samples selected conventionally based on spectral signatures can 
not sufficiently represent class properties associated with ancillary 
data. 
Integration before classification so called stratification involves Eiumnoh and Shresta (1997) attempted to explore the effect of 
segmentation of the image into smaller scenes before Digital Terrain Model in accuracy of image classification by 
classification takes place in order to provide spectrally similar simply adding it as a component into classification in a logical 
classes to be classified independently. Integration after channel manner, they achieved certain amount of improvement. A 
classification or post-classification sorting is based on the problem study by Richetti (2000) involves the use of slope map to add 
that a single class of objects may be assigned to more than one information to classification for geological purposes. Logical 
classes due to the fact that a particular class can show different channel and stratification methods were applied and compared to 
spectral characteristics. Integration of ancillary data during spectral classification. Results demonstrated increase in accuracy 
classification mainly has two approaches; first and the most for logical channel method. 
53