In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
University
iRathaus-
[park
/olks-
»arten
Parliament
jBurg-
Igarten
DO
■ meters
Figure 1. Study area located in the city centre of Vienna,
Austria. Diversity in tree species and structure are shown in
detailed photographs (B1-B3)
The full-waveform airborne LiDAR data were obtained in the
framework of the city-wide laser scanning campaign in the
winter season 2006/2007 under leaf-off conditions. Data
acquisition was performed with a Riegl LMS-Q560 system,
which uses near-infrared (1500 nm) laser pulses with a pulse
width of 4 ns. Further settings are a pulse repetition frequency
of 200 kHz, scan angle range of ±22.5° and a beam divergence
of 0.5 mrad. Full-waveform recording can be done with 1 ns
temporal resolution. Decomposition of the waveforms was
performed by using the Riegl software RiANALYZE. The
number of echoes is not limited by the full-waveform recording
sensor system, and therefore, the number of detected echoes per
laser shot can be higher as with traditional discrete echo
recording, particularly in high vegetation. The average echo
density of the LiDAR dataset covering the study area is 50
echoes per m 2 with about 6.8% first echoes, 1% intermediate
echoes (e.g. 2 nd , 3 rd echo), 6.8% last and 85.4% single echoes
(i.e. only one reflection per shot). A Digital Terrain Model
(DTM) with 0.5 m resolution was produced using the software
SCOP++ (SCOP++, 2010). For evaluation and comparison a
reference dataset including single tree positions of alley trees as
well as the official land cover map of Vienna are available:
Mehrzweckkarte (MZK), Flachenmehrzweckkarte 1 (FMZK).
3. METHODS
The very high point density of the LiDAR dataset, i.e. about
50-10 6 echoes per km 2 , does not allow a pure point cloud based
vegetation detection procedure for large areas (400 km 2 for
entire Vienna) in an operational manner. Therefore, a combined
object-based image and point cloud approach is introduced,
taking advantage of both, fast raster processing and detailed
("interpolation-free") point cloud based information extraction
including full-waveform laser point attributes. Furthermore, the
high point density of the point cloud allows the derivation of
high-resolution derivatives (e.g. DTM with 0.5 m cell size),
providing a sufficient planimetrie accuracy of the final results.
1 http://www.wien.gv.at/stadtentwicklung/stadtvenriessung/geod
aten/fmzk/produkt.html (last accessed 31.5.2010)
The developed workflow is implemented in the GRASS GIS
environment (GRASS Development Team, 2010) and
comprises the steps shown in Fig. 2.
Full-waveform Point Cloud
X,Y,Z, amplitude, echo width, echo n
total no. echoes
|~t g™ I
Building Polygons I—I
Data Preparation
?.g. calculation of DSM, nDSM
Segmentation
raster-based
Segment Feature Calculation
geometric and topological features
.^mstcr^vcctor^n^omt^cloud^bascd^
Existing Tree Positions
Figure 2. Workflow of urban vegetation detection using raster
and full-waveform point cloud input data
3.1 Data preparation
The data preparation of the input point cloud ASCII file
includes a classification of the laser points into first, last, single
and intermediate echoes. Each laser point has the attributes
echo number (EN) and total number (TN) of echoes of the
corresponding laser shot, with e.g. single echoes: EN=TN=1;
first echoes: EN=1 && TN>1; intermediate echoes: EN>1 &&
EN!=TN. Using the maximum height per 0.5 m cell a Digital
Surface Model (DSM) is derived and finally normalized with
the input DTM, in order to get a normalized DSM (nDSM).
Empty cells get a normalized height of 0 m. In urban areas
generally complete and high quality cadastral data layers are
available, which can be used as mask for the subsequent steps.
In our study the buildings can be extracted from the official
land cover map and could be used optionally to exclude these
areas from further investigation. In our study the building
polygons are used for evaluation and thus are not considered as
input.
3.2 Segmentation
An object-based raster analysis using an edge-based
segmentation procedure is applied to the high-resolution nDSM.
The procedure has already been tested in densely forested areas
where no buildings are present (Hofle et al., 2008) and is
transferred to the densely built-up urban area of Vienna in this
study. The segmentation aims at delineating convex objects in
the nDSM by finding concave edges in between the objects.
Additionally, constraints on normalized height (nDSM>1.0m)
and the occurrence of multiple reflections are included.
Multiple reflections are parametrized by an echo ratio (ER)
defined as (Eq. 1):
ER [ /0] — (/ifirst ^intermediate) / (^last E ^single) 100 0)
where n is the respective number of echoes per cell. If no echo
is within a cell, ER is set to zero and if no last and single echoes
are found within a cell, the echo ratio is set to 100, exhibiting a
high vertical extension and transparency of the object. High ER
values are assumed to indicate vegetation, which tends to have
a high number of first and intermediate echoes compared to
other elevated objects in urban areas (e.g. buildings; Fig. 3b).