INTEGRATING OBJECT-BASED CLASSIFICATION WITH ONE-CLASS SUPPORT 
VECTOR MACHINES IN MAPPING A SPECIFIC LAND CLASS FROM HIGH SPATIAL 
RESOLUTION IMAGES 
Qinghua Guo\ Guoming Du ab , Yu Liu a,c , Desheng Liu d 
“School of Engineering, University of California Merced, P.O. Box 2039, Merced, CA, USA 95344 
b School of Geography Sciences and Planning, Sun Yat-sen University, Guangzhou, China, 510275 
institute of Remote Sensing and Geographic Information Systems, Peking University, Beijing, China 100871 
department of Geography, Ohio State University, 154 North Oval Mall, Columbus, OH, USA 43210 
Commission WGs WG IV/9 
KEY WORDS: Object-Based Classification, One-Class Support Vector Machines, High Spatial Resolution Imagery 
ABSTRACT: 
Remote sensing techniques have been commonly used to map land cover and land use types. For many applications, users may only 
be interested in a specific land class in an image such as extracting urban areas from an image, or retrieving dead trees from a forest. 
This could be referred to as a one-class classification problem. In addition, with the increasing availability of high spatial resolution 
imagery, earth objects can be mapped in detail, which enable us to quickly update and monitor the change of a specific class. 
However, conventional pixel-based classification methods have difficulty in dealing with high spatial resolution remote sensing data. 
In this study, we use urban house extraction as an example, and propose to classify houses from high spatial resolution images by 
integrating one-class Support Vector Machines (SVMs) and object-based classifiers. We also compared the performance from the 
proposed method with the one-class SVMs and pixel-based method. The results indicate that the proposed method outperforms the 
pixel based method, and could be a promising way to provide relatively quick and efficient way in extracting a specific land class 
from high spatial resolution images. 
1. INTRODUCTION 
With the increasing availability of remote sensing data over the 
past few decades, remote sensing data have been commonly 
used in a wide variety of urban and environmental applications 
such as: monitoring land use change, mapping suitable habitat, 
and detecting invasive species. Traditionally, all the land types 
in an image were completely mapped via remote sensing 
classification methods. However, for some applications, we 
may only be interested in a specific class without considering 
other land types (Foody et al. 2006). For example, if the 
objective of the project is to extract roads from remote sensing 
data, we may not be interested in classifying forests, and 
agricultural lands. Another reason for mapping a specific class 
of interest is to reduce significant amount of efforts in 
collecting, training, and testing ground truth data; as it is very 
time consuming to collect ground-truth data for all the land 
classes. Therefore, it is needed to develop methods to retrieve 
only one land type from the remote sensing data. The idea is to 
separate a specific class from the rest of the land classes. This 
question could be referred to as a one-class classification 
problem. Other disciplines also have similar issues. For 
example, species collection data from natural museums often 
contain presence-only data (i.e. absence data are often not 
collected, or unreliable to collect such an animals or invasive 
species), scientists are interested in classifying the habitat based 
on presence-only data in order to find suitable or potential 
habitat for species. Another example in a handwritten number 
recognition problem is to classify the handwriting number such 
as “8” when we only have a sample a set of handwritten “8”s. 
One common solution to deal with one-class classification 
problem is based on similarity matching. Numerous machine 
learning and non machine learning approaches can be applied to 
this problem. Among many methods, support vector machines 
(SVMs), originally developed by Vapnik (1995), are considered 
to be a new generation of learning algorithms. SVM have 
several appealing characteristics for modellers, including: they 
are statistically based models rather than loose analogies with 
natural learning systems, and they theoretically guarantee 
performance (Cristianini and Scholkopf, 2002). SVM have 
been applied successfully to a range of remote sensing 
classification applications (Huang et al., 2002). Recently, 
Scholkopf et al. (1999) developed one-class SVM to deal with 
the one-class problem. This method has proved useful in 
document classification, texture segmentation, and image 
retrieval. 
Moreover, with the advance of sensor techniques, high spatial 
resolution remotely sensed images have become commercially 
available and increasingly used in various aspects of 
environmental monitoring and management (Mumby and 
Edwards 2002). Conventional pixel-based classifiers such as 
maximum likelihood classification (MLC) and Iterative Self- 
Organizing Data Analysis Technique (ISODATA), which label 
unknown areas pixel by pixel based on spectral similarity, do 
not perform well with high spatial resolution images ( Xia 
1996). This is because the inherent spectral variability in 
specific ground targets increases as resolution becomes finer 
(Martin and Howarth 1989). Therefore, retrieving one land 
class from high spatial resolution imagery based on pixel-based 
method may result in significant misclassification. In recent 
years, object-based methods have gained much attention as 
alternative methods for classifying high resolution images. An 
“object” is defined here as a group of spectrally similar 
contiguous pixels, and ideally, it should represent a physically