1095
USE Of LINE SEGMENTS IN HIGH RESOLUTION SATELLITE IMAGE
REGISTRATION
J.Heikkinen, A.Laiho-Heikkinen, H.Haggren
Helsinki University of Technology, Department of Surveying Sciences,
Group of Photogrammetry and Remote Sensing
FINLAND
- Jussi.Heikkinen@tkk.fi
Anita.Laiho - Heikkinen@tkk.fi
Henrik. Haggren@tkk. fi
Commission VII, WG VII/6
KEY WORDS: Remote Sensing, Adjustment, Registration, Data Fusion, Georeferencing, High Resolution, QuickBird
ABSTRACT:
In recent years the research activity in image registration has grown at the same pace as the high resolution satellite images have
found their way to end users. Much of this research activity has concentrated on methods improving the accuracy of georeferencing
provided as RPC values. In this study the image registration via projective transformation based on straight line segments is
investigated. The transformation parameters are solved based on real data extracted from topographic database and measurements
done on QuickBird image. Same data is processed also with point wise method and accuracy numbers in selected check points are
calculated. The RMS values computed in same check points prove that with the line based method the equivalent accuracy can be
achieved as with the point wise method computed with minimum number of observations. In a smaller sub image the accuracy of
transformation with line segments could be verified to be in size of a pixel.
1. INTRODUCTION
Georeferencing is an essential part of the process when
combining satellite images acquired in different epochs or with
different sensors. If data fusion is needed for image
interpretation, classification, or change detection, precision of
image registration has to fulfil task specific requirements. In
point-wise method georeferencing can easily be accomplished
within precision of one pixel. However, this precision usually
requires manual work in selecting good points and satisfying
proper point distribution. This is tolerable if the number of
images to process is reasonable, but in case the data processing
is regular and rather frequent, some automation for image
registration is needed. Taking care of point distribution is a
fairly straight forward task, but to recognize correspondences
automatically is a demanding task. At the moment, most of the
implementations of automatic image registration do rely on
point-wise observations and area-based matching strategies. In
a simple case they do succeed reasonably well, but in a more
complex case they tend to suffer from some drawbacks. Most
problematic in determination of correspondences is to deal with
differences of varying radiances of an object point in case of
images from different epochs. Also, when using point-wise
methods and area-based matching strategies, discrepancies of
occlusion patterns and casting shadows due to varying sensor
orientations and time of acquisition affect results. By using
feature-based matching approach there are better chances to
detect these pitfalls automatically.
Feature-based matching algorithms exploit the power of
interest-operators in order to extract large number of feature-
point observations on cost of accuracy of an individual
measurement. However, there is an option to extract feature
lines instead of feature-points. The fact that straight lines
project as straight line segments on images speaks up for using
these straight line segments in image registration. While
searching correspondences multiple matching clues connected
with line-features do enhance and encourage to exploit
automation in image registration.
Full strength of line-based methods can be exploited if there are
line segments on images which are substantially long compared
to image dimensions. Unfortunately, this requirement is rarely
fully met with satellite images. This drawback has been avoided
in investigation of Barakat (Barakat et.all.,2004) by measuring
few fairly good control points on images and constructing
Active lines between those measured points. The selected points
were considerable long distance apart from each other
constructing a solid base for line-based projective
transformation. This way it was possible to solve transformation
with fewer points than what would have been required to
achieve the same accuracy in estimation with equations based
only on point observations.
However, line segments to be detected on high resolution
satellite images are long enough to be used alone in solving the
image registration. Earlier, line-based methods have
successfully been used with aerial images for map revision
processes in order to solve sensor orientation and reconstruct of
object features. The stability of line-features have been found
robust and feasible in such tasks (Mikhail&Mulawa,1988),
(Mulawa, 1989),(Heikkinen, 1994), (Mikhail&Weerawong, 1994),
(Habib, 1999).
In this paper it will be shown that image registration can be
accomplished within precision of few pixels using purely line