3b. Beijing 2008 
Scale behaviour 
ndscape objects. 
mmetry, Remote 
»1. XXXVI, Part 
tic extraction of 
imagery. ISPRS 
58, pp. 83-98. 
ler Straßennetze 
;h, 131 p. 
iges. Biological 
'omputer Vision. 
mic Publishers, 
., 1999. Vehicle 
nee Evaluation. 
:attle, WA. 
events and their 
y RS Journal of 
52-75. 
ammetric image 
Remote Sensing, 
> Kontextobjekte 
sity of Hannover, 
>., and Ender, J. 
ction in Urban 
} hotogrammetry, 
es, Vol. XXXV, 
ke, C.-E., 1999. 
Sensing Images 
ering & Remote 
oceedings of the 
ence, Karlsruhe, 
n for automated 
f imagery and 
Remote Sensing, 
n low resolution 
pp. 693-703. 
OBJECT-ORIENTED BUINDING EXTRACTION 
BY DSM AND VERY HIGH-RESOLUTION ORTHOIMAGES 
N. Jiang a,b ’*, J.X. Zhang a , H.T. Li a , X.G. Lin a,c 
d Chinese Academy of Surveying and Mapping, Beijing 100039, P.R. China 
b Shandong University of Science and Technology, Qingdao 266510, P.R. China 
c Wuhan University, Wuhan 430079, P.R. China -jiangnal23321@163.com 
Commission III, WG III/4 
KEY WORDS: building extraction, high resolution, segmentation, DSM, object-oriented, eCognition 
ABSTRACT: 
High-resolution remote sensing images accelerate the development of information extraction and 3D city reconstruction. This paper 
concentrates on object-oriented methods to solve building extraction problems. With object-oriented methods, not only the spectral 
information but also the shape, contextual and semantic information can be used to extract objects. The object-oriented building 
extraction typically includes several steps: data pre-processing, multi-scale image segmentation, the definition of features used to 
extract buildings, building extraction, post-processing and accuracy evaluation. Among the features of buildings extraction, we 
consider the height of the buildings as the most useful information. In this paper DSM is used to extract high objects as trees and 
buildings; then the work of distinguishing trees and buildings will be applied. With the removal of trees, building information is 
extracted. We achieve this goal in the software package eCognition. The result shows that the method works well. 
1. INTRODUCTION 
1.1 General Instruction 
High-resolution imagery provides an important new data 
source for building extraction and 3D reconstruction. How to 
ucture of buildings has evolved most rapidly in the last years 
(Forstner, 1999). 
Detection and description of buildings from aerial and spatial 
images is a practical application of 3-D object description. 
Building extraction is also the key problem in urban 
information updates construction of digital cities, large-scale 
mapping, and also motivated by the importance of geographic 
information systems (GIS): the need for data acquisition and 
update for GIS. 
A couple of years ago the main input data for the production of 
building extraction and 3D city models were aerial images, 
terrestrial images, map data, and data derived from classical 
surveying (Fuchs et al, 1998), so the main feature used is the 
spectral and textural information in the images. However, with 
the development of data acquisition methods, multi sensory data, 
e.g. SAR, infrared, stereo or laser scan images, is available as 
additional information; the image processing methods have 
developed, so different cues as color, texture, semantics, edges 
and color edges elevation data features can be used to detect 
and reconstruct buildings. Different input data combining with 
different processin 1 g means lead to different models. 
Parametric models and generic models are used to describe the 
buildings; DSM and DEM also appear to solve these extraction 
and reconstruction problems. Both semiautomatic and 
automatic methods are applied to building extraction; the 
semiautomatic way that allows for efficient human interaction 
can meet the high precision but the automatic procedures seem 
to be the only way to satisfy the developing trend in the future. 
* Corresponding author. 
extract topographic objects as buildings, roads, trees and pipes 
in urban areas form images automatically, rapidly and 
accurately, now has become a hot spot of imagery information 
extraction and application. The need for 3D str 
The development tendencies of the building extraction 
according to Jiang(2004) are: From single image to multiple 
images; From gray information to color information; Interaction 
between 2-D and 3-D; From single image information grouping 
to grouping with multiple images; More imaging geometry, 
object knowledge and spatial reasoning are used; Building 
model develops from single to complex building of plane 
patches; Multiple sources information integration, such as with 
LIDAR data or map data. 
1.2 Overview of Related Work 
A brief literature overview concerning building extraction from 
image data is given here. Many of the early building extraction 
systems have used a single intensity image. Nevatia ( Nevatia et 
al, 1997)described a method for detecting rectilinear buildings 
and constructing their 3-D shape descriptions from a single 
aerial image of a general viewpoint. They use the geometric and 
projective constraints to make hypotheses for the presence of 
building roofs from the low-level features and to verify by 
using 3D cues. Shadows, wall vertical and base line are 
important cues in these methods. Then stereo or multi-view 
analysis is focused because of widely available data. In(Roux et 
al, 1994)3D descriptions of buildings are generated from 
matched lines and junctions. 
DSM, DEM and LIDAR data are used in the extraction of 
buildings. Gerke( Gerke et al.) have done a lot of work on 
automatic detection and extraction of trees and buildings from 
aerial CIR orthoimages and normalized digital surface models. 
They use a hierarchical strategy to solve the complex models 
and complex images problem. Orhner and Descombes (Orhner 
441