International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
AUTOMATIC CLASSIFICATION OF URBAN ENVIRONMENTS FOR DATABASE REVISION USING LIDAR AND
COLOR AERIAL IMAGERY
N. Haala, V. Walter
Institute of Photogrammetry, University of Stuttgart, Geschwister-Scholl-Str. 24, Stuttgart, D-72110, Germany,
Norbert.Haala@ifp.uni-stuttgart.de, Volker.Walter@ifp.uni-stuttgart.de
KEYWORDS: Classification, Building Extraction, Data Fusion, Laser Scanning.
ABSTRACT
This article describes the combination of Digital Surface Models (DSM) resulting from airborne laser scanning and colour aerial
imagery for landuse classification in urban environments. The laser DSM is used to obtain information on the local height above the
terrain surface for each pixel. This information is used in a following step in order to separate objects higher than the terrain like
buildings and trees from objects which are at terrain level, like streets and grass-covered areas. The basic idea of the proposed
algorithm is to combine this geometric information with additional multispectral information provided from colour aerial imagery.
Both types of information are integrated in a pixel-based classification, where the information on the local height above the terrain is
integrated as an additional channel together with the spectral channels. It is demonstrated that, by this approach, the classification of
urban scenes can be improved considerably compared to a scene labelling, where only one type of data is applied.
1. INTRODUCTION
Spectral information has been widely used as a data source for
thematic mapping applications. Surface material information
can be derived by traditional classification techniques from
multispectral imagery and can for example be used for mapping
of man-made structures and natural features in complex urban
scenes. During the past years, numerous classification
algorithms have been developed for thematic mapping
applications. They can be divided into unsupervised and
supervised approaches. During an unsupervised classification,
pixels are grouped into different spectral (and textural) classes
by clustering algorithms without using prior information. After
clustering, the spectral classes have to be associated with the
land cover classes by an operator. In a supervised classification
two basic steps are carried out. First, in a training stage an
operator digitises training areas that describe typical spectral
and textural characteristics of the dataset. In the following
classification stage, each pixel of the dataset is assigned to a
land cover class. For this classification stage, a lot of different
approaches such as minimum-distance, parallelepiped or
maximum likelihood classification are available (Lillesand and
Kiefer (1994)).
The main problem during thematic mapping of man-made
structures and natural features is the limited accuracy and
reliability of the results, as well as the frequently arising
difficulties to discriminate a sufficient number of object
categories. A common goal during data acquisition in built-up
areas is the detection of objects like streets and buildings.
However, this can be difficult if only spectral information is
used, since, for some areas, roofs and streets are built of very
similar material. This complicates or even prevents the
discrimination of these objects due to their similar reflectance.
The same problem can arise, if trees and grass-covered areas
have to be differentiated.
Traditionally, multispectral data has been the dominating source
of information for landuse classification. Usually, the input data
for the classification are multispectral images. Optionally,
textural patterns, which are derived from the original spectral
data, can be included. Up to now, auxiliary information on the
surface topography has been mainly used for the geometric
correction of the spectral data by the generation of ortho
images. Additionally, surface topography information has been
used for the radiometric correction of the multispectral imagery.
This can be necessary, especially in hilly or mountainous
terrain, since spectral reflectance values are influenced by the
inclination and orientation of the terrain surface. In contrast to
this, in our approach height data is not only applied for
correction of the multispectral data, but also integrated as a
complementary data source of equally important information.
During data acquisition in urban environments, trees and
buildings are object classes of major interest since these objects
are very relevant for many applications like visualisations or
simulations of the propagation of electro-magnetic radiation. In
order to separate these types of objects from their surrounding,
information on the local height above terrain can be very
valuable since trees and buildings are higher than then-
surrounding, whereas other objects of interest like streets and
grass-covered areas are at the terrain level. The required
information can be derived at least approximately from a DSM,
which for our application is acquired by airborne laser scanning.
Based on this DSM, the local height above terrain of the visible
surface is derived. The result consists of all objects rising above
the terrain approximately put on a plane. This surface is
combined with the multispectral bands during a standard
classification in order to improve the extraction of the required
object classes.
