AUTOMATED 3D MAPPING OF TREES AND BUILDINGS AND IT’S APPLICATION
TO RISK ASSESSMENT OF DOMESTIC SUBSIDENCE IN THE LONDON AREA
Jan-Peter Muller *, Jung-Rack Kim, Jonathan Kelvin
Dept. of Geomatic Engineering, University College London, Gower Street, London, WCIE 6BT UK
jpmuller@ge.ucl.ac.uk *
TS ICWG IVIV
KEY WORDS: IKONOS, LIDAR, Vision, DEM/DTM,Urban, Vegetation, GIS, Geology, Subsidence risk assessment
ABSTRACT:
Very high resolution images (e.g. IKONOS) and airborne lidar have been used for the automated 3D mapping of individual tree and
building locations in a large test area in East London of some 8 x 8km extent with many tens of thousands of buildings and trees.
Initial results of the building and tree detection algorithm for small area assessments were given in Kim & Muller (2002). In this
work we report on the extension of the algorithm to the full area and the refinement of the algorithm to extract tree height. Also
shown will be the building detection’s quantitative assessment using the OS® MasterMap® (Parish et al., 2003, submitted) and the
random sample assessment of tree locations using higher resolution digitised aerial photography from different commercial suppliers.
Overall the detection efficiency is greater than 75% even though the buildings have a huge range in floorplan, height, age and type.
Tree detection efficiency is based on a visual assessment of the degree of overstorey crown overlap but has similarly high values.
1. INTRODUCTION
1.1 Aims
The primary goal of the research work reported here was to
develop practical techniques for the automated production of
dense landscape object models focussing on buildings and trees
in urban area.
With the increasing demands for information on artificial and
natural landscape objects in many application fields (e.g. risk
insurance, mobile telecommunication, city planning, geological
research etc..), newly delivered commercial high-resolution
satellite imagery and LiDAR (Light Detection and Ranging) data
are stimulating the development of automated GIS construction.
This research aims at the retrieval of 3D shape or 2D boundaries
of buildings, which are larger than 10 square metres and have
some regularity, and individual tree crowns with an acceptable
degree of accuracy in very dense urban environments from
high-resolution images and range data. Data processing
algorithms utilising both range and image data or between
image clues are here described to address technical problems
associated with the available data sources such as, insufficient
data resolution to resolve detailed object structure, a very large
search area and irregularity of targets (tree and building)..
1.2 Previous research work
Building detection has been one of the major research topics of.
the photogrammetric community over the last 20 years. A
sample of previous work is provided here relevant to the work
at UCL and other centres. Kim and Muller (1999) developed a
graph- based building reconstruction algorithm using 2D edge
lines extracted from aerial photographs. Roux and McKeown
(1994) used multiple aerial photos to construct 3D roof models
of buildings. Perceptual grouping and shadow information was
used for 3D building reconstruction by Lin and Nevatia (1998).
The AMOBE project at ETH Zurich (Henricsson et al, 1996) is
one of the first examples of the use of colour information for
building extraction. Brenner and Halla(1999) constructed 3D
building models from Lidar data and multi-spectral information.
Marr and Vosselman (1999) suggested algorithms to extract
building structures from invariant moments derived from Lidar.
826
Recently research has begun to examine the application of high
resolution satellite images such as IKONOS for building
extraction (see, for example, Fraser et al., 2001).
Individual tree crown detection is a very recent topic in image
understanding and remote sensing data analysis. Template
matching involving the correlation between a pre-defined model
and an image patch is one method proposed for automated tree
detection (Pollock 1998). Zang (2001) showed the first
promising results using texture analysis in high resolution
optical urban images. Gong et al. (2002) used a semi-automated
interactive 3D model-based tree interpreter from multi-ocular
high-resolution aerial images. Straub (2003) used LiDAR and a
top-down low level operator to extract tree crown.
1.3 Data description
Space Imaging’s IKONOS, which is the first commercial high-
resolution satellite imager of the Earth, has unprecedented
clarity and detail (nominal IfoVx1m). IKONOS products use a
general photogrammetric model, based on Rational Polynomial
Coefficients (Grodecki 2001) Several relevant articles regarding
IKONOS photogrammetric modelling accuracy have recently
been published. A comprehensive review of IKONOS image
radiometric and photogrammetric quality has been performed
by Grodecki and Dial(2001) and Baltsavias et al. (2001)
respectively. In particular, Grodecki and Dial(2001) showed
that, in the case of GCP controlled stereo images (Precision
stereo), the accuracies were within 1 metre horizontally and 2
metres vertically. According to this result, the photogrammetric
quality of any IKONOS precision data set should be acceptable
for urban area mapping, where landscape objects of a few
metres’ scale are present.
The test data-set used in our study consisted of an IKONOS Pro
geocoded single view data set over East London (11 by 11km),
which was pan-sharpened to one metre resolution and contains
R-G-NIR bands, using an unidentified technique by the satellite
supplier. An initial assessment was performed of the planimetric
positioning accuracy through an inter-comparison with
kinematic (k-GPS). This showed that the planimetric accuracy
appeared to be better than it's technical specification.Lidar data
supplied by Infoterra Limited came from an Optech 1020 ALS
(Airborne Laser scanner) (http://www.optech.on.ca/) which was
Int«
use
den
It v
dat
Hei
suc
21
A
info
info
G/R
poir
com
whi
appi
et a
Bar
proc
app
Our
CPL
den:
The
cons
avoi
Cha
heig
whe
Fror
the
