ry.
Ip-
9-
ey
ir-
he
ds
ds,
"n,
nd
id:
JN
h-
d-
1)
an
D-
APPLICATION OF ELASTIC REGISTRATION TO IMAGERY FROM AIRBORNE SCANNERS
Rafael Wiemker*, Karl Rohr**, Lutz Binder**,
Rainer Sprengel"", H. Siegfried Stieh!**
* Universität Hamburg, ll. Institut für Experimentalphysik
** Universität Hamburg, FB Informatik, AB Kognitive Systeme
Mail: KOGS, Vogt-Kölln-Str. 30, 22527 Hamburg, FRG
WWW: http://kogs-www.informatik.uni-hamburg.de/projects/censis/remotesens.html
E-mail: wiemker@informatik.uni-hamburg.de
Commission IV, Working Group 4
KEY WORDS: Registration, Distortion, Multispectral, Multitemporal, Change Detection, Digital Orthoimage, Elastic Regis-
tration, Airborne Scanner Imagery
ABSTRACT
This paper discusses the application of advanced registration methods to airborne scanner imagery. We investigate an elastic
registration approach and other locally adaptive techniques for image-to-map registration which show promising results where
conventional global polynomial transformations in general do not suffice.
For most applications in remote sensing, rectification and geocoding is essential for the analysis of image data and subsequent
fusion with other data. In our case, we want to overlay multitemporal multispectral image data in order to conduct change
detection for monitoring purposes.
Comparison shows that locally adaptive registration techniques based on radial basis functions improve the results significantly
when compared to conventional global polynomial transformations. The choice of the radial basis function seems to be of minor
sensitivity. AKIMA local quintic polynomial registration is slightly less satisfactory although computationally less expensive.
1. INTRODUCTION Conventional global polynomial registration techniques use
approximating schemes, i.e., they establish a global coordi-
nate transformation function which minimizes the sum of de-
viations at the given GCPs. In contrast, interpolating tech-
niques, such as the one used by Ehlers & Fogel (1994) and
those investigated in this paper, map each given source con-
trol point exactly onto its respective target control point, and
the transformation for all remaining points is determined by
a certain interpolation scheme. In comparison to the global
polynomial transformations, the interpolating techniques in-
vestigated here are locally adaptive, although they can be
based on global functions.
Photogrammetry has studied extensively how aerial pho-
tographs can be rectified and converted into orthophotos.
Bilinear and second degree polynomial transformation func-
tions have proven useful for aerial photographs. However,
multispectral remotely sensed data are often recorded by line
scanners which are mounted on airborne platforms. Thus, the
process of image formation is not instantaneous but depends
on the flight path and the attitude of the instrument. More-
over, the mapping itself cannot be described by a lens camera
imaging model. Therefore, the ortho-rectification of scanner
recorded images — especially aerial imagery with high spatial ^ this paper, we investigate the application of a one step elas-
frequency distortions — by global affine or polynomial coordi- ^ tic registration approach, which allows local rectifications, to
nate transforms is not satisfactory. lt is rather necessary to multispectral airborne scanner imagery. In this context, 'elas-
allow for local corrections. tic’ means the use of thin-plate spline radial basis functions
(Bookstein, 1989), which have been successfully applied to
Zhang et al. (1994) have approached the problem of rectify- ^ the registration of medical images. The transformation con-
ing scanner data by trying to reconstruct the flight path and sists of a global affine as well as a pure elastic part, where the
the sensor orientation, which requires continually recorded pure elastic part is a superposition of the radial basis function
comprehensive flight data of the scanner carrying airplane. U(r) = r?Inr. This radial basis function has a sound phys-
If such data is not available, corresponding ground control ical interpretation and is well known from elasticity theory.
points (GCPs) between image and map must be used for ^ The coefficients of the global affine and the pure elastic part
matching. are determined simultaneously by solving a linear system of
equations.
Considering the above mentioned drawbacks of global poly-
nomial registration of scanner data, Ehlers (1994) used a We compare the elastic thin-plate spline registration not only
two step approach: First, a global second degree polynomial to global polynomial registration, but moreover to HARDY
transformation is carried out. Second, a registration based on multiquadric and AKIMA local quintic polynomial registra-
multiquadric interpolation (Hardy, 1971) is performed usinga tion. Multiquadric interpolation (Hardy, 1971) is based on
radial basis function U(r) = r of the distance r between the a global function (just as elastic registration) and uses ra-
current location and all given GCPs. Ehlers & Fogel (1994) dial basis functions of the type U(r) = v6 + r?, which we
mention thin-plate spline radial basis functions and the lack have used with § = 0 and § = 1. In contrast, AKIMA local
of comparison to Hardy's multiquadric basis functions. quintic polynomial registration consists of a number of local
949
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996