ul 2004
—Z—
lution.
nbigu-
metric
hanked
Hutton
hanked
for this
> Natu-
Canada
rol for
14(80),
»mmel,
immar-
). LA-
1al and
y ftware
‘orge.
Limit-
r mod-
itute of
Hf Nav-
-528.
mobile
‘algary,
Roche,
issisted
1,
lins, J.,
yringer-
lvanced
tic GPS
ngress,
Remote
iety of
Vienna,
, 2001.
1 Wiley
e atmo-
eophys-
AUTOMATIC RELATIVE REGISTRATION
OF SPOTS IMAGERY FOR COLOR MERGING
Leong Keong KWOH and Xiaojing HUANG
Centre for Remote Imaging, Sensing and Processing, National University of Singapore,
SOCI, Level 2, Lower Kent Ridge Road, Singapore 119260
crshxj@nus.edu.sg
KEY WORDS: Multispectral, Registration, Correlation, Multiresolution, Orthorectification, SPOT, Sensor Model
ABSTRACT:
An automatic process for relative registering and merging the panchromatic and multispectral images of SPOTS is presented. The
automated hierarchical local registration process makes use of feature information to select tie points from both images. The tie
points are then used to refine the sensor model of the multispectral image by method of least squares solution. Both panchromatic
and multispectral images are then orthorectified to a geo-referenced coordinate system using the available one kilometre gridded
Digital Elevation Models (GLOBE or SRTM DEM). The merged high resolution colour imagery is then obtained by multiplying
each spectral bands with a sharpening factor computed from the intensity values of the orthorectified panchromatic image and the
corresponding pixels of the orthorectified multispectral images.
I. INTRODUCTION
The SPOTS satellite carries two high resolution geometric
instruments (HRG) enabling it to map large area of the Earth
(60km x 60km). Each HRG instrument on-board the SPOTS
satellite images the ground with the panchromatic (PAN) band
(2.5 m or 5 m resolution) pointing slightly forward (0.529?) and
the multispectral (XS) bands (10 m resolution) slightly
backward (-0.529?). This results in a slight relief displacement
in opposite directions for the panchromatic and multispectral
images and also a time interval of about 1.5 seconds between
the PAN and XS images over a same point on the ground.
Therefore there is a need to co-register both images together
and to eliminate the relief displacement before they can be
merged into a high resolution colour image.
Christophe Latry and Bernard Rouge (Christophe Latry, 2003)
used physical sensor models plus massive local image
correlation to register the images and eliminate the relief
displacement. This technique, however, may not be effective if
the images have high amount of cloud cover or water bodies.
Kwoh, in an earlier work (Kwoh, 2003), used a method based
on coarse orthorectified PAN and XS images with the publicly
available Ikm gridded GLOBE DEM to eliminate the relief
displacement differences due to the different look direction in
the Pan and XS image; and then using the simple affine
transformation to register the two orthorectified images. The
method had assumed that both the Pan and XS image have no
relative orientation bias differences, which is logical since the
Pan and XS images were taken at a mere 1.5 second apart.
In this paper, we present an improved method where the PAN
and XS images are registered relative to one another by refining
the physical sensor model of one image relative to the other
(instead of the simple affine transformation). The orthorectified
images generated with the same DEM will then be well
registered to one another and will have no relative relief
displacement errors. Any orientation bias differences between
the PAN and XS image will also be removed during the sensor
model refinement process.
The method is implemented in 4 steps as shown in Figure 1—
(1) Automated image matching and tie point selection; (2)
Relative SPOTS sensor model affinement for XS image with
respect to PAN image with tie points; (3) Orthorectifying both
PAN and XS images to a common georeferenced coordinate
system by using one kilometre resolution GLOBE or SRTM
Digital Elevation Models (DEM); (4) obtaining high resolution
colour image by merging the high resolution (2.5m or 5m)
panchromatic image with the lower resolution (10m)
multispectral image.
XS image — 14
(10 m/pixel)
PAN image — 1A
(2.5 m/pixel)
—línage Resample= — Corner Detection ~~ -— age Resample: >
( | PANimage | [ Comers. [ . XSimage -
|. (Muttiresolution) | — tae |. (Mutti-resolution)
—]linage Matching: >
Ii XS Sensor Model
(Tie Points/ | . .—XS Sensor Model — .
GroundPoints| ^ ~~. Refinement .—
PAN Sensor Model ( XSRefined |
(Sensor Model)
Ls Orthorectify SRTM DEM —Drthorectify -—-.—3À
'XS Ortho image)
v, (0 mipixel)
PAN Ortho Image]
(25 mipixel) |
<<Color Merginÿ —
{Merged multispectral Image)
Ë {2.5 mipixel) |
Figure 1. Relative registration and colour merging diagram