In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
13
AUTOMATIC EXTRACTION OF URBAN OBJECTS FROM MULTI-SOURCE AERIAL
DATA
Adriano Mancini, Emanuele Frontoni and Primo Zingaretti
Dipartimento di Ingegneria Informatica Gestionale e dell’Automazione
Università Politecnica delle Marche
Ancona, ITALY
{mancini.frontoni,zinga}@diiga.univpm.it
KEY WORDS: LiDAR, buildings, road extraction, automated classification, city models
ABSTRACT:
Today, one of the main applications of multi-source aerial data is the city modelling. The capability to automatically detect objects of
interest starting from LiDAR and multi-spectral data is a complex and an open problem. The information obtained can be also used for
city planning, change detection, road graph update, land cover/use. In this paper we present an automatic approach to object extraction
in urban area; the proposed approach is based on different sequential stages. The first stage basically solves a multi-class supervised
pixel based classification problem (building, grass, land and tree) using a boosting algorithm; after classification, the next step provides
to extract and filter land areas from classified data; the last step extracts roundabouts by the Hough transform and linear roads by a novel
approach, which is robust to noise (sparse pixels); the final representation of extracted roads is a graph where each node represents a
cross between two or more roads. Results on a real dataset of Mannheim area (Germany) using both LiDAR (first - last pulses) and
multi-spectral high resolution data (Red - Green - Blue - Near Infrared) are presented.
1 INTRODUCTION
T ODAY the availability of high spatial resolution LiDAR and
multi-spectral data collected by aerial vehicles (manned or
unmanned) traces new ways for the possible applications. City
modeling, object extraction (e.g., buildings, roads, bridges, ...),
urban growth analysis, land use/cover, developing 3D models,
are the main studied applications. Usually the analysis of data is
made by a human operator; traditional photo-interpretati on is a
slow and expensive process that requires specialized experts; ac
curacies similar to those of man-made maps can now be reached
by automatic object extraction and classification approaches, but
with considerably less wasted time and money, thus allowing high
update rates.
The ability to automatically classify data starting from a set of
heterogeneous features is fundamental to design an automatic ap
proach. One of the first method used to classify LiDAR data was
the height threshold to a normalized DSM (nDSM) (Weidner and
Forstner, 1995); using this method it is possible to extract objects
as buildings, but its has a lot of well-known drawbacks: high-
density canopy can be classified as building and it is not possible
to distinguish low height objects as lands or roads. Multi-spectral
data allow to extend the set of classified objects producing higher
accuracy. Many machine learning approaches were adopted to
solve the problem of object extraction from multi-source data;
Bayesian maximum likelihood method (Walter, 2004), Dempster-
Shafer (Lu et al., 2006), boosting using AdaBoost (Frontoni et ah,
2008).
Common objects as buildings or roads are the main interesting
features that can be extracted from the classified data; road ex
traction is a classical problem of remote sensing, but not com
pletely solved. A really interesting overview (updated to 2003)
can be found here (Mena, 2003). Using only multi-spectral data
(Bacher and Mayer, 2005), road extraction is an extremely diffi
cult task especially in urban area also using high-resolution im
agery as IKONOS or SPOT. Problems as occlusion (due to the
presence of trees), noise inducted by vehicles or object shadows,
influence the quality of road extraction; moreover, spectral sepa
rability of road respects to other objects (e.g. bituminous roofs)
is not always guaranteed. Snakes/active contours are classical
methodological tools; different version of standard snake (Kass et
ah, 1987) were developed to solve the problem of road extraction
especially in not urban area (Marikhu et ah, 2006). Moreover this
approach requires a wide set of good seed points, which are often
user defined. The fusion of LiDAR and multi-spectral data is a
powerful tool for road extraction; LiDAR helps to distinguish be
tween high objects as buildings or canopies, while multi-spectral
data allow to distinguish between land/road and grass or other
low profile objects (Clode et ah, 2005). SAR imagery can be
also useful for road extraction with results comparable with Li
DAR (Guo et ah, 2007). However the goodness of LiDAR and
multi-spectral data fusion approaches allows to obtain interesting
results in building / road extraction.
In this paper, a classification approach, using boosting classifier
to fuse LiDAR and multi-spectral data, is presented. The Ada
Boost technique with CART classifier as weak learner, classifies
data distinguishing among four classes: building, grass, land and
tree; the ReliefF (Liu and Motoda, 2008) feature selection algo
rithm allows to consider only meaningful features to minimize the
misclassification. The result of classification stage is then used to
extract buildings, roads and roundabouts; the approach here pro
posed extracts and clusters a set of linear roads using a pyramidal
representation to reduce time and memory usage. The procedure
is totally automatic and requires only a minimum interaction with
user; a user-defined training set is necessary to train the classifier
and control the learning accuracy; the training set often can be di
rectly accessible by a web-GIS or a photo-interpretation process
over a very small portion of global area; we use a training set that
covers less than 0.5% of total area.
The paper is organized as follows. Section 2 introduces the method
ology for classification and object extraction; Section 3 explains
the data set used for experiments, the adopted classifer and the
classification results on a four class problem. Section 4 presents
the method and obtained results in road extraction; in Section 5
conclusions and future work are outlined.
