×

You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Title
CMRT09
Author
Stilla, Uwe

In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
FUSION OF OPTICAL AND INSAR FEATURES FOR BUILDING RECOGNITION IN
URBAN AREAS
J. D. Wegner a ’ *, A. Thiele b , U. Soergel a
a Institute of Photogrammetry and Geoinformation, Leibniz University Hannover, Hannover, Germany -
(wegner, soergel)@ipi.uni-hannover.de
b FGAN-FOM Research Institute of Optronics and Pattern Recognition, Ettlingen, Germany -
thiele@fom.fgan.de
Commission III, WG III/4
KEY WORDS: Remote Sensing, Fusion, Feature Extraction, InSAR Data, Optical Data, Urban Area
ABSTRACT:
State-of-the-art space borne SAR sensors are capable of acquiring imagery with a geometric resolution of one meter while airborne
SAR systems provide even finer ground sampling distance. In such data, individual objects in urban areas like bridges and buildings
become visible in detail. However, the side-looking sensor principle leads to occlusion and layover effects that hamper
interpretability. As a consequence, SAR data is often analysed in combination with complementary data from topographic maps or
optical remote sensing images. This work focuses on the combination of features from InSAR data and optical aerial imagery for
building recognition in dense urban areas. It is shown that a combined analysis of InSAR and optical data very much improves
detection results compared to building recognition based on merely a single data source.
1. INTRODUCTION
Due to its independence of daylight and all-weather capability,
synthetic aperture radar (SAR) has become a key remote
sensing technique in the last decades. One main application
scenario arises in crisis situations when the acquisition of a
scene is required immediately for rapid hazard response. Urban
areas play a key-role since the lives of thousands of people may
be in danger in a relatively small area. In SAR data of one meter
geometric resolution collected by modern space borne sensors
such as TerraSAR-X and Cosmo-SkyMed, the geometric extent
of individual objects like bridges, buildings and roads becomes
visible. In airborne data such objects are imaged with even more
detail. However, shadowing and layover effects, typical for
SAR image acquisitions in urban areas, complicate
interpretation. Small buildings are often occluded by higher
ones while façades overlap with trees and cars on the streets. In
addition, the appearance of an individual building in the image
highly depends on the sensor's aspect. Buildings that are not
oriented in azimuth direction with respect to the sensor are
often hard to detect. This drawback can be partly overcome by
using SAR images from multiple aspects (Xu and Jin, 2007).
Building recognition and reconstruction can be further
improved based on interferometric SAR (InSAR) acquisitions
from two orthogonal flight directions (Thiele et al., 2007).
Nevertheless, automatic urban scene analysis based on SAR
data alone is hard to conduct. SAR data interpretation can be
supported with additional information from GIS databases or
high-resolution optical imagery. Optical images have the
advantage of being widely available. In (Soergel et al., 2007)
high-resolution airborne InSAR data is combined with an
optical aerial image in order to three-dimensionally reconstruct
bridges over water. Tupin and Roux (2003) propose an
approach to automatically extract footprints of large flat-roofed
buildings based on line features by means of a SAR amplitude
image and an optical aerial image. Furthermore, homogeneous
regions in an aerial photo, represented in a region adjacency
graph, are used in (Tupin and Roux, 2005) to regularize
elevation data derived from radargrammetric processing of a
SAR image pair by means of Markov Random Fields.
In this paper, an approach for building recognition in dense
urban areas is presented that combines line features from mono
aspect InSAR data with classification results from one optical
aerial image. Building corner lines extracted from InSAR data
are introduced as features into a classification framework that is
based on a segmentation of the optical image. Optical features
and InSAR lines are jointly used in order to evaluate building
hypothesis. The focus is on the fusion approach of building
primitive hypothesis.
2. ANALYSIS OF OPTICAL DATA
Optical images provide high resolution multi-spectral
information of urban scenes. For human interpreters they are by
far more intuitive to understand than SAR data since the
imaging geometry corresponds to the human eye. In aerial
imagery of 0.3 meters resolution, like used in this project,
building roofs become visible in great detail. In addition, façade
details may appear in the image if high buildings situated far
away from the nadir point of the sensor are imaged.
2,1 Appearance of Buildings
The appearance of an individual building mapped by any
imaging sensor is both governed by its own properties (e.g.,
material, geometry) as well as by sensor characteristics (e.g.,
principle, spectral domain, pose), which have to be considered
for recognition. For example, in optical images acquired from a
near nadir perspective, building roofs are the most important
features for automatic detection. Shadows are also good
indicators for buildings (Fig. 1) and distinguish them, for
instance, from road segments or parking lots. In western
Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author.