297
A TEST OF 2D BUILDING CHANGE DETECTION METHODS:
COMPARISON, EVALUATION AND PERSPECTIVES
Nicolas Champion 3 , Leena Matikainen b , Franz Rottensteiner c , Xinlian Liang b , Juha Hyyppa b
a IGN/MATIS, Saint-Mandé, France - nicolas.champion@ign.fr
b FGI, Dept, of Remote Sensing and Photogrammetry, Masala, Finland -
(leena.matikainen, xinlian.liang, juha.hyyppa)@fgi.fi
c Cooperative Research Centre for Spatial Information, Dept, of Geomatics,University of Melbourne, Australia -
franzr@unimelb.edu.au
Commission VI, WG VI/4
KEY WORDS: Change Detection, 2D Vector Databases, Algorithms Comparison, Quality Assessment
ABSTRACT:
In the past few years, 2D topographic databases have been completed in most industrialised countries. Most efforts in National
Mapping and Cadastral Agencies (NMCAs) are now devoted to the update of such databases. Because it is generally carried out
manually, by visual inspection of orthophotos, the updating process is time-consuming and expensive. The development of semi
automatic systems is thus of high interest for NMCAs. The obvious lack of expertise in the domain has driven EuroSDR to set up a
test comparing different change detection approaches. In this paper, we limit the scope of the project to the imagery context. After
describing input data, we shortly introduce the approaches of the working groups that have already submitted results. Preliminary
results are assessed and a discussion enables to bring out first conclusions and directions.
1. INTRODUCTION
2D topographic databases have been completed in most
National Mapping and Cadastral Agencies (NMCAs) during the
last decade. The main issue now concerns the map revision.
This procedure is known to be very tedious, time-consuming
and expensive. There is also a growing need to automate it. The
development of semi-automatic tools that are able to detect the
changes in a database from recent data (typically imagery or
LIDAR) and to present them to a human operator for
verification is therefore highly desirable. Only a few solutions
have been proposed by academic research, even fewer by
private companies. Many questions that have arisen remain
unanswered, e.g. those regarding the most efficient
methodology, the type of primary data to use (LIDAR / imagery)
or the most appropriate spatial resolution to choose. These
considerations have driven EuroSDR (European Spatial Data
Research - http://www.eurosdr.net) to set up a change detection
project. This project is in line with previous EuroSDR projects,
e.g. the project on road updating (Mayer et al., 2006) and the
one on change detection (Steinnocher and Kressler, 2006). The
aim of this new project is to evaluate the feasibility of semi-
automatically detecting changes in a 2D building vector
database from imagery or LIDAR. Three specific topics are
investigated in detail: firstly, the impact of the type of data and
methodology on the performance of the change detection;
secondly, the impact of the spatial resolution of input data;
finally, the impact of the complexity of the scene, especially
with respect to topography and land use. The methodology of
the project consists in a test comparing five different algorithms
that are representative of the current state-of-the- art in the field
of change detection. It is the main goal of this EuroSDR project
to gather enough experience to identify key problems in change
detection and to give promising directions for building an
optimal operational system in the future.
In this paper, preliminary results achieved for three different
algorithms, (Matikainen et al., 2007), (Rottensteiner, 2007) and
(Champion, 2007), are presented. In Section 2, the datasets and
the comparison method are described. In Section 3, the three
approaches of the working groups that have already submitted
results are shortly introduced. Results are given and evaluated
in Section 4. We finally present a summary and conclusions.
2. INPUT DATA AND TEST SET-UP
2.1 Datasets Description
Two test areas are used in this study. The first test area is
situated in Marseille (France). It has an area of about 0.4 km 2
and contains about 1300 buildings. The area corresponds to a
very dense urban settlement and features a complicated urban
configuration (lower buildings connected to higher buildings).
The test area is hilly (with height differences of 150 m) and
vegetated, especially along streets. The second test area is
situated in Toulouse (France). It has an area of about 1 km 2 and
contains about 200 buildings. It features a suburban area and is
composed of detached buildings that are very different to each
other with respect to the size, height, shape and roofing material.
The terrain is also undulating (with height differences of 100 m)
and vegetated. In this study, colour infrared (CIR) aerial images
with a Ground Sample Distance (GSD) of 20 cm and multiple
overlap (a forward and a side lap of minimum 60%) are used
for Marseille. Pléiades tri-stereoscopic Satellite CIR images are
used for Toulouse, with a GSD equal to 50 cm. In both cases, a
Digital Surface Model (DSM) was computed using a stereo
matching algorithm based on the 2D minimization of a cost that
takes into account discontinuities and radiometric similarities
(Pierrot-Deseilligny and Paparoditis, 2006). The GSD of the
DSM is equal to the GSD of the aerial images. CIR orthophotos
were also computed from input DSM and images. Reference