Full text: Mapping without the sun

GU Haiyan a,b ’ *, LI Haitao 3 , ZHANG feng c , HAN Yanshun 3 , YANG Jinghui 3 
a Institute of Photogrammetry and Remote Sensing, Chinese Academy of Surveying and Mapping, Beijing, 100039, 
b Institute of Surveying and Geography Science, Liaoning Technical University, Liao Ning, 123000, China. 
c Civil Engineering Department, Guilin University of Technology, Gui Lin, 541000, China. 
Commission VI, WG VI/4 
KEY WORDS: Object-oriented Classification, Land-cover Classification, Markov Random Field (MRF), Support Vector Machine 
(SVM), QuickBird 
We present a new object-oriented land cover classification method integrating raster analysis and vector analysis, which adopted 
Markov Random Field (MRF) for segmentation and Support Vector Machine (SVM) for classification using High Resolution (HR) 
QuickBird data. It synthesized the advantage of digital image processing, Geographical Information System (GIS) (vector-based 
feature selection) and Data Mining (intelligent SVM classification) to interpret image from pixels to segments and then to thematic 
information. Compared with the pixel-based SVM classification in Imagelnfo 3.0, both of the accuracy of land cover classification 
by the proposed method and the computational performance for classification were improved. Moreover, the land cover 
classification map can update GIS database in a quick and convenient way. 
For years, many efforts have been made to develop automatic 
procedures for integrating raster analysis and vector analysis. 
Numerous researches have been focusing on updating GIS 
databases using remote sensing data and land cover 
classification technique. The High Resolution (HR) imagery has 
the advantage of high user interpretability, rich information 
content, sharpness, high image clarity, and integrity, which 
provides the unique tool for classification. However, it is not 
practical to classify the HR imagery using traditional pixel- 
based classification methods (e.g., Maximum Likelihood, 
ISODATA, K-means). Recently, researchers have explored new 
classification techniques, such as multi-classifier systems that 
integrate the outputs of several underlying classifier algorithms, 
and the use of structural, spatial, context information from 
imagery. However, these pixel-based classification techniques 
improve classification accuracy with some limitations: 1) they 
have considerable difficulties dealing with the abundant 
information of HR data, 2) they produce a characteristic and 
inconsistent salt-and-pepper classification, 3) they are not 
capable of extracting objects of interest, and 4) they can not 
update GIS database expediently (Baatz, M., et al., 2004). Thus, 
the pixel-based classification methods are no longer applicable 
for HR imagery. 
In this situation, commercial software, eCognition, brought 
forward the revolutionary of object-oriented approach, which 
adopts a bottom-up region-merging segmentation technique and 
fuzzy logical classification method.Recently, some researchers 
have been succeeded in classification with new object-oriented 
classification methods adopting other segmentation and 
classification techniques. Furthermore, in object-oriented image 
analysis, multi-source data fusion technique is helpful to land 
cover classification. Cuozzo,G.(2004) succeeded in 
classification with new object-oriented classification methods 
adopting Markov Random Field (MRF) segmentation 
techniques and MLC algorithm. Bruzzone, L.(2004) proposed a 
multi-level hierarchical approach to classify high spatial 
resolution images with Support Vector Machines (SVM), and 
the results and comparisons confirmed the effectiveness of the 
approach. Li, H.T. (2007) succeeded in object-oriented 
classification based on improved Color Structure Code (CSC) 
and SVM. In this paper, an object-oriented classification 
scheme based on MRF and SVM is developed where QuickBird 
imageries are used. 
2.1 MRF Segmentation Method 
Markov Random Fields (MRF) are commonly used 
probabilistic models for image analysis. The basic idea of MRF 
is to model the contextual correlation among image pixels in 
terms of conditional prior probabilities of individual pixels 
given their neighboring pixels. 
The segmentation is obtained as a cartoon image, which is 
basically a labeling of the original image (Kato, Z., et al, 2006). 
The original image is showed as: 
F = {/ = 0',/);l < i < m, 1 < j < n) , (1) 
The segmentation imageX = {x ? ;x s e {1,2,...,AT}} is a random 
field defined in F . X s is the label of point S . K is the numbers 
*E-mail: haiyanl982709@126.com; Phone : 86 10 88217730; Fax: 86 10 68211420; http:// www.casm.ac.cn

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.