313
EXTRACTION OF BUILDINGS AND TREES USING IMAGES AND LIDAR DATA
N. Demir*, D. Poli, E. Baltsavias
ETH Zurich, Institute of Geodesy and Photogrammetry, Wolfgang Pauli Str. 15, CH-8093 Zurich,
Switzerland-(demir, poli, manos)@geod.baug.ethz.ch
Commission IV, WG IV/3
KEY WORDS: DTMs/DSMs, Lidar Data Processing, Multispectral Classification, Image Matching, Information Fusion, Object
Extraction, Buildings, Trees
ABSTRACT:
The automatic detection and 3D modeling of objects at airports is an important issue for the EU FP6 project PEGASE. PEGASE is
a feasibility study of a new 24/7 navigation system, which could either replace or complement existing systems and would allow a
three-dimensional truly autonomous aircraft landing and take-off primarily for airplanes and secondary for helicopters. In this work,
we focus on the extraction of man-made structures, especially buildings, by combining information from aerial images and Lidar
data. We applied four different methods on a dataset located at Zurich Airport, Switzerland. The first method is based on DSM/DTM
comparison in combination with NDVI analysis; the second one is a supervised multispectral classification refined with height
information from Lidar data. The third approach uses voids in Lidar DTM and NDVI classification, while the last method is based
on the analysis of the vertical density of the raw Lidar DSM data. The accuracy of the building extraction process was evaluated by
comparing the results with reference data and computing the percentage of data correctly extracted and the percentage of missing
reference data. The results are reported and commented.
1. INTRODUCTION
Pegase is a EU FP6 project aiming at studying the feasibility of
a new 24/7 navigation system, which could either replace or
complement existing systems and would allow a three-
dimensional truly autonomous aircraft landing and take-off
primarily for airplanes and secondary for helicopters (Pegase
Web, 2008). Within the project, the acquisition of a reliable
geospatial reference database of the airports, and in particular
the automatic extraction of buildings and obstacles at airports,
has a critical role for aviation safety. Often, 3D information of
airports is not available, is not accurate enough, not complete,
or not updated. Thus, methods are needed for generation of
accurate and complete 3D geodata with high degree of
automation. Buildings and trees are considered as obstacles, so
they should be correctly detected. There are several methods
applied for this purpose, based on image and/or airborne Lidar
data. In our approach, objects are detected and extracted in
aerial images and Lidar data through a combination of
multispectral image classification, DSMs and DTMs
comparisons and density analysis of the raw Lidar point cloud.
This paper will give a brief overview of the related work on this
subject. Then, after the description of the test area at Zurich
Airport, Switzerland, the strategy and algorithms will be
presented and the results will be reported, compared and
commented.
2. LITERATURE REVIEW
Aerial images and Lidar data are common sources for object
extraction. Regarding image-based analysis, multispectral
classification uses spectral properties of objects and tries to
extract them by using supervised or unsupervised methods.
Currently, there are commercial systems (e.g. ENVI, ERDAS,
e-Cognition, IDRISI, etc.), which use expert classification
algorithms such as rule-based, object-oriented and fuzzy
approaches. In digital photogrammetry, features of objects are
extracted using 3D information from image matching or
DSM/DTM data, spectral, textural and other information
sources. Henricsson and Baltsavias (1997) combine image
matching and spectral properties for the extraction process.
Rottensteiner (2001) uses a semi-automated approach based on
feature-based matching techniques. In Weidner and Foerstner
(1995), above-ground objects are estimated using a normalised
DSM (nDSM) by subtracting a DTM, estimated by a
morphological filtering of the DSM, from the original DSM. In
general, the major difficulty using aerial images is the
complexity and variability of objects and their form, especially
in suburban and densely populated urban regions.
Regarding Lidar, building and tree extraction is basically a
filtering problem in the DSM (raw or interpolated) data. Some
algorithms use raw data (Sohn and Dowman, 2002; Roggero,
2001; Axelsson, 2001; Vosselman and Maas, 2001; Sithole,
2001; Pfeifer et al., 1998), while some others use interpolated
data (Elmqvist et al., 2001; Brovelli et al., 2002; Wack and
Wimmer, 2002). The use of raw or interpolated data can
influence the performance of the filtering, but also its speed
being slower for raw data. The algorithms differ also in the
number of points they use at a time. In addition, every filter
makes an assumption about the structure of bare-earth points in
a local neighbourhood. This assumption forms the concept of
the filter. Often, segmentation is performed to find the clusters
which delineate objects and not facets of objects. The mostly
used segmentation methods are based on region-based methods,
like in Brovelli et al. (2002), Crosilla et al. (2005), or use
Hough transform (Tarsha-Kurdi et al., 2007). In Elmqvist et al.
(2001), the terrain surface is estimated by using active contours.
The performance of the algorithm is good for large objects and
all types of vegetation but small objects can not be always
extracted (Sithole and Vosselman, 2003). Sohn and Dowman
Corresponding author.