In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXV1I1. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
239 
URBAN BUILDING DETECTION FROM OPTICAL AND INSAR FEATURES 
EXPLOITING CONTEXT 
J. D. Wegner“'*, A. O. Ok b , A. Thiele c , F. Rottensteiner a , U. Soergel a 
ll Institute of Photogrammetry and Geoinformation (IPI). Leibniz Universität Hannover, Hannover. Germany- 
(wegner, soergel)@ipi.uni-hannover.de 
b Dept, of Geodetic and Geographic Information Technologies, Middle East Technical University, Ankara. Turkey - 
oozgun@metu.edu.tr 
c Fraunhofer Institute of Optronics. System Technologies and Image Exploitation (IOSB), Ettlingen. Germany - 
an tj e. th i el e@.i osb. fraunh o fer. de 
Commission III, WG 111/4 
KEY WORDS: Conditional Random Fields, Remote Sensing. Fusion. InSAR Data. Optical Stereo Data, Urban Area 
ABSTRACT: 
We investigate the potential of combined features of aerial images and high-resolution interferometric SAR (InSAR) data for 
building detection in urban areas. It is shown that completeness and correctness may be increased if we integrate both InSAR 
double-bounce lines and 3D lines of stereo data in addition to building hints of a single optical orthophoto. In order to exploit 
context information, which is crucial for object detection in urban areas, we use a Conditional Random Field approach. It proves to 
be a valuable method for context-based building detection with multi-sensor features. 
1. INTRODUCTION 
Building detection in urban areas based on merely a single 
aerial photo is often hard to conduct (Mueller and Zaum, 2005). 
Features of additional data sources may be introduced to 
improve detection completeness and correctness. In addition to 
features derived from an orthophoto we use building hints of 
high-resolution InSAR data and an optical stereo image pair. 
Several works have already dealt with the integration of features 
derived from high-resolution optical and SAR (or InSAR) data 
with the goal of building detection. Xiao et al. (1998) detect and 
reconstruct building blocks combining high-resolution optical 
and InSAR data. They classify both data sets separately within a 
multi-layer neural network followed by morphological 
operations. Finally, rectangles are fit to building hypothesis and 
heights are derived. Hepner et al. (1998) jointly use hyper- 
spectral imagery and InSAR data acquired by airborne sensors 
to detect and three-dimensionally reconstruct large buildings in 
urban areas. Tupin and Roux (2003) propose an approach to 
extract footprints of large flat-roofed industrial buildings based 
on line features. In (Tupin and Roux, 2005) the same authors 
represent homogeneous regions of an aerial photo with a region 
adjacency graph. This graph is then used within a Markov 
Random Field framework to regularize building heights 
determined by means of radargrammetry. A discontinuity 
constraint based on the image gradient along segment 
boundaries is introduced into the prior term in order to preserve 
sudden height jumps. Poulain et al. (2009) combine high- 
resolution optical and SAR data with vector data in order to 
detect changes. Since no learning step is conducted all 
classification is performed based on prior knowledge. They 
generate features from previously extracted primitives and set 
up a score for each building site using Dempster-Shafer 
evidential theory. Sportouche et al. (2009) detect and three- 
dimensionally reconstruct large industrial buildings semi- 
automatical ly. They combine features of high-resolution optical 
satellite imagery (Quickbird) with high-resolution SAR data 
(TerraSAR-X). Building hypothesis of the optical data are 
validated or rejected based on a classification of the SAR image 
making use of roof textures, bright lines, and shadows. Building 
heights are derived simultaneously exploiting the different 
optical and SAR sensor geometries. We recently proposed a 
segment-based approach for building detection (Wegner et al., 
2009). Segments of an orthophoto are classified in combination 
with InSAR double-bounce lines. 
In this paper, we use a Conditional Random Field (CRF) 
framework, which is a probabilistic contextual classification 
framework originally introduced by Lafferty et al. (2001) for 
labelling 1D sequential data and later on extended to images by 
Kumar and Hebert (2003). CRFs have already been successfully 
applied to various computer vision tasks (e.g., Rabinovich et al., 
2007; Korc and Forstner, 2008). Nonetheless, CRFs have only 
rarely been applied to remote sensing data (Zhong and Wang, 
2007). Furthermore, to the authors knowledge only one 
publication exploits CRFs for the analysis of SAR data (He et 
al., 2008). 
Our focus is on the suitability of CRFs for combining multi 
sensor remote sensing data using context with the aim of single 
building detection. Although much more sophisticated features 
could potentially be derived from stereo and InSAR data we use 
rather simple ones in order to transparently assess the entire 
framework. More sophisticated features may then be introduced 
in future work. 
We now first give an overview of the entire processing chain. 
Then, features we utilize are explained, the basic theory of 
CRFs is described, and finally building detection results with 
different feature sets as input are compared. 
Corresponding author