In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
A TEST OF AUTOMATIC BUILDING CHANGE DETECTION APPROACHES
Nicolas Champion 3 , Franz Rottensteiner b , Leena Matikainen c , Xinlian Liang c , Juha Hyyppä c and Brian P. 01sen d
d 1GN, MATIS, Saint-Mandé, France-Nicolas.Champion@ign.fr
b Institute of Photogrammetry and Geolnformation,
Leibniz Universität Hannover, Germany - Rottensteiner@ipi.uni-hannover.de
c FGI, Dept, of Remote Sensing and Photogrammetry, Masala, Finland -
{Leena.Matikainen, Xinlian.Liang, Juha.Hyyppa}@fgi.fi
d National Survey and Cadastre (KMS), Copenhagen, Denmark - bpo@kms.dk
Commission III, WG III/4
KEY WORDS: Change Detection, Building, 2D Vector Databases, Algorithms Comparison, Quality Assessment
ABSTRACT:
The update of databases - in particular 2D building databases - has become a topical issue, especially in the developed countries
where such databases have been completed during the last decade. The main issue here concerns the long and costly change
detection step, which might be automated by using recently acquired sensor data. The current deficits in automation and the lack of
expertise in the domain have driven the EuroSDR to launch a test comparing different change detection approaches, representative of
the current state-of-the-art. The main goal of this paper is to present the test bed of this comparison and the results that have been
obtained for three different contexts (aerial imagery, satellite imagery, and LIDAR). In addition, we give the overall findings that
emerged from our experiences and some promising directions to follow for building an optimal operative system in the future.
1. INTRODUCTION
The production of 2D topographic databases has been
completed in many industrialised countries. Presently, most
efforts in the National Mapping and Cadastral Agencies
(NMCAs) are devoted to the update of such databases. As the
update process is generally carried out manually by visual
inspection of orthophotos, it is time-consuming and expensive.
As a consequence, its automation is of high practical interest for
the NMCAs. The update procedure can be split into two steps:
change detection, in which the outdated database is compared
to recently collected sensor data in order to detect changes, and
vectorization, i.e. the digitization of the correct geometry of the
changed objects. Given the state-of-the-art in automatic object
detection (Mayer, 2008), only the automation of the change
detection step seems to be possible at this time. The key idea is
to focus the operator’s attention on the areas that may have
changed. Work is saved because the operator needs not inspect
areas classified as unchanged by the automatic procedure.
The current deficits in automation and the lack of expertise
within the NMCAs have driven the EuroSDR (European Spatial
Data Research - http://www.eurosdr.net) to lauch a project
about change detection. It also aims at evaluating the feasibility
of semi-automatically detecting changes in a 2D building vector
database from optical imagery or LIDAR. Three subtopics are
investigated in detail, firstly the impact of methodology;
secondly, the impact of the type and spatial resolution of input
data; lastly, the impact of the complexity of the scene in terms
of interfering objects such as roads. The methodology consists
in comparing four different algorithms representative for the
current state-of-the-art in the field of change detection. First
results, achieved for the cases where only aerial and satellite
images are used, were presented in (Champion et al., 2008). The
results obtained there showed the limitations of change
detection methods, especially in relation to the quality of input
data. The main goal of this paper is to present the final results of
the project, including a LIDAR dataset, and to give a detailed
evaluation of the outcomes delivered by the approaches
compared here.
After describing the datasets and the evaluation procedure
(Section 2), the methods compared in the test are concisely
introduced (Section 3). In Section 4, a thorough evaluation is
carried out, including an analysis of the performance of change
detection with respect to the update status of the buildings and
the building size. The weak and strong points are then identified
both for the datasets and the methodologies, and they used to
give overall findings and recommendations for building an
optimal operative system for change detection in the future.
2. INPUT DATA AND TEST SET-UP
Three test areas are used for the comparison: Marseille (France),
Toulouse (France), and Lyngby (Denmark). The area covered
by the test sites is 0.9 x 0.4 km 2 in Marseille, 1.1 x 1.1 km 2 in
Toulouse, and 2.0 x 2.0 km 2 in Lingby. The test areas differ
considerably regarding topography, land use, urban
configuration and roofing material. The terrain is hilly in
Marseille and Toulouse and relatively flat in Lyngby. Marseille
features a densely built-up area consisting of small buildings of
variable height, all connected to each other and mostly covered
with red tile. Toulouse and Lyngby feature a suburban area,
mostly composed of detached buildings and characterised by a
large variety of roofing materials such as slate, gravel, or
concrete. Colour Infrared (CIR) orthophotos and Digital
Surface Model (DSMs) are available for all test areas. In
Marseille and Toulouse an image matching algorithm (Pierrot-
Deseilligny and Paparoditis, 2006) was used to derive the DSM
from input images. In Marseille, these images are multiple aerial
images having a forward and side overlap of 60%. The Ground
