ERROR MODELLING ON REGISTRATION OF HIGH-RESOLUTION SATELLITE
IMAGES AND VECTOR DATA
: à "ja: a ^ t
Pu-Huai Chen? *, Szu-Chi Hsu?^, Ge-Wen Lee"
* Dept. of Surveying and Mapping Eng., Chung-Cheng Institute of Technology, Ta-Hsi, Taoyuan, 335 Taiwan, R.O.C.
phchen@ccit.edu.tw
® Dept. of Civil Eng., Chung-Cheng Institute of Technology, Ta-Hsi, Taoyuan, 335 Taiwan, R.O.C.
Commission I, WG 1/4
KEY WORDS: Geometric, Integration, Matching, Raster, Registration, Structure, Understanding, Vector
ABSTRACT:
Traditionally, image-and-map registration is carried out using low-level image processing techniques. One of inevitable problems
resulted from a low-level image processing technique is the need to decide what the ultimately desired object is. An alternative
way to register images and maps is to use a ‘top-down’ or high-level image understanding approach, for instance, a geometric-
structure-matching (GSM) technique. The algorithm of the proposed GSM technique is validated using a Quickbird image and the
corresponding cadastral map. The boundary lines and polygons of cadastral parcels are used as the elements of geometric structure
in the studied case. An automatic technique has been developed to match image features and the corresponding vector data. In
addition, prior knowledge about the error model in the procedures of image-and-map matching has not been fully understood,
therefore, this paper also concentrates on the error model required to implement the algorithm and to achieve a high level of
automation. The error model is vital to give a threshold for optimising the results of the proposed GSM technique. Preliminary
results show that errors of the order of 5m from the procedures of image-and-map registration are possible, and that error is
comparable with the predicted one. It is possible to eliminate the requirements of manual intervention for registering images and
maps, provided that accurate vector data are available. Potential applications of the proposed algorithm include providing ground
control for automatic photogrammetry and updating data of spatial information systems.
1. INTRODUCTION
High-resolution images taken by advanced sensors with ground
sampling distance (GSD) on the order of less than 1m, such as
Quickbird and Ikonos data, keep flowing in, and users of
various fields demand reasonable solutions from
photogrammetry and remote sensing community to cope with
the needs of map revision and extraction of information
promptly. Automation is always the main consideration for
solving the above-mentioned requests. Many efforts have been
made to understand and to extract information from images,
and this kind of photogrammetric approaches can be called as
forward solutions or ‘bottom-up’ approach. Unfortunately, the
current methods for automatic extraction of information from
satellite images are still far from practical. In general, the
reason why visual brains of human beings are able to draw a
map by using complex images is rather poorly understood, if
not entirely unknown. This explains why the development of
algorithms for automatic extraction of spatial information is
progressing slowly (Sowmya and Trinder, 2001).
Qn the contrary, currently available data, such as vector data,
maps, and digital elevation models (DEMs), representing basic
knowledge about areas of interest, have been proved useful for
providing information of ground control for map revision and
photogrammetric, or radargrammetric, tasks (Morgado and
Dowman, 1997; Chen and Dowman, 2000; Habib and Kelley,
2001). Comparing with traditional photogrammetric approach,
* Corresponding author.
this kind of operations might be called as reverse solutions
or ‘top-down’ approach. Automatic image-and-map
registration is still one of unsolved problems in pursuit of
fully automatic photogrammetry, near-real-time map
revision and smart spatial information systems (Dowman,
1998; Heipke et. al., 2000). Traditionally, image-and-map
registration is carried out using low-level image processing,
or ‘bottom-up’, techniques. One of inevitable problems
resulted from low-level image processing techniques is the
need to decide what the ultimately desired object is. Since
that the decision is made by a human operator after
segmentation of features, obviously, the need of human
interventions in the traditional processing procedures
results in relatively low level of automation. An alternative
way to register images and maps is to use a ‘top-down’ or
‘high-level image understanding approach, provided that
prior knowledge about image-and-map registration is
available and applicable in automatic procedures (Shapiro
and Stockman, 2001; Baltsavias, 2004).
There is no intention in the paper to give a precise
definition of knowledge, however, prior knowledge is
referred to as any geo-spatial data or models available, such
as roads, boundary lines and land parcels, which give
geometric structure of areas of interest. Hence, the paper is
aimed at using geometric structure defined by vector data,
given by existing 2-D maps or spatial information systems,
for image-and-map registration with a higher level of
Internationa
automation.
of particula
focused on t
and land p:
about the €
matching ha
to give rea
procedures «
In Section 2
the algorith
also to pro
the propose
technique is
image and
relevant au
done using
be compar
Automatic t
image featu
Section 4.
Section 5.
2.1 Basis
The basic r
well-define
image and
systems pre
areas of 1
projection.
following |
most of the
hand, the |
such as wa
images tak
Filet.
(©Quickbirc
Figure 1. (
standard 1
cadastral n
Taiwan. T!
Since that
resolution
approxima
straightfor
lines/poly;