GEOREFERENCING FROM ALOS PRISM IMAGERY

VIA LONG-STRIP ADJUSTMENT

C. S. Fraser and M. Ravanbakhsh
Cooperative Research Centre for Spatial Information
Department of Geomatics
University of Melbourne VIC 3010, Australia
Ph. +613 8344 4117, Fax +613 9347 2916
[c.fraser, m.ravanbakhsh]@unimelb.edu.au

Abstract
Numerous recent studies have demonstrated that georeferencing to sub-pixel
accuracy is feasible from high-resolution satellite imagery (HRSI). The main
obstacle to achieving high precision in georeferencing remains the need for
provision of good quality ground control points (GCPs), whether the GCPs are
used to remove biases in RPC triangulation or to support physical sensor
orientation models. The provision of GCPs can be very costly and is often not
feasible in remote regions, the very areas where mapping from satellite imagery
shows significant potential. In order to drastically reduce the number of GCPs
required for accurate georeferencing from HRSI, a rigorous generic sensor
orientation model incorporating strip adjustment capability has been adopted.
Under this approach, the metadata for each separate scene is merged to
produce a single, continuous set of orbit and attitude parameters, such that the
entire strip of tens of images can be treated as a single image, even though the
separate scenes are not merged per se into a long single image. The merging
of orbit data results in a considerable reduction in both the number of unknown
parameters and the number of required GCPs in the sensor orientation
adjustment. RPCs are then generated from the adjusted orientation data for
each image forming the strip or block. The adjustment approach discussed has
previously been shown to be applicable to strips of ALOS and QuickBird
imagery. Application to very long strips of ALOS PRSIM imagery is reported in
this paper. The results of experimental testing indicate that 1-pixel level
accuracy can be achieved over strip lengths of more than 50 ALOS images, or
1500km, with as few as four GCPs.

Introduction
High-Resolution Satellite Imagery (HRSI) is increasingly being adopted for
geospatial information generation due in large part to ever increasing spatial
resolution, and increased availability and frequency of coverage of virtually any
location on Earth. These characteristics have made HRSI suitable for
automated mapping applications such as orthoimagery and DSM generation, as
well as for the updating of medium to large scale topographic maps in remote
areas, where HRSI is often the only available source of imagery. A primary step
in the photogrammetric processing of HRSI is precise georeferencing, the
design accuracy requirements of which are generally at the 1-pixel and even
sub-pixel level. There are basically two options for orientation of HRSI in order

1
to produce optimal geopositioning accuracy: rigorous sensor orientation
modelling or orientation via terrain-independent rational functions (RPCs). In
each case, bias-free georeferencing to optimal accuracy requires the provision
of ground control points (GCPs).
For single-scene georeferencing, the ground control requirements can be as
few as one or two GCPs for RPCs, and four to six for a rigorous sensor model,
depending upon formulation of the model and the level of control point
redundancy adopted. The georeferencing of long strips of imagery via single-
image orientation therefore requires many GCPs, the establishment of which
can be very costly in both remote areas with limited access, and in areas not
conducive to GPS surveying, such as dense forests. Although there may be the
prospect of extracting GCPs from base mapping, such an approach – when
there is existing map/GIS data available - most often yields control of less than
sufficient accuracy, especially at larger scales.
RPC strip and block adjustment has been shown to be a practical means to
achieve high-accuracy georeferencing from HRSI, the number of required
GCPs being as few as one or two, though there is a requirement for tie point
measurements within the images (Fraser & Hanley, 2003; Fraser et al., 2006).
Such a technique can also be applied for 2D and 3D strip and block adjustment
of HRSI via rigorous sensor orientation models. Through utilization of a
combination of ground control and measured tie points, the number of GCPs
can be drastically reduced, for example from around 5n for a strip or block of n
images where single-image orientation is employed, down to perhaps 6-10 for
moderate sized strips and blocks.
There is also an alternative approach to orientation adjustment of single strips
of images from the same orbit via a rigorous model, one that requires GCPs but
does not require the measurement of tie/pass points in the imagery. Under this
approach, the metadata data for each separate scene is merged to produce a
single, continuous set of orbit and attitude parameters, such that the entire strip
of images can be treated as a single image, even though the separate scenes
are not merged per se into a long single image. The merging of orbit data
results in a considerable reduction in the number of unknown orientation
parameters, and thus also in the number of required GCPs, which can then be
as few as two at each end of the strip.
This strip adjustment approach, first reported in Rottensteiner et al. (2008;
2009), makes use of a generic sensor orientation model. However, the
parameters of this model are not necessarily conducive to exploitation by
standard photogrammetric workstations. Thus, in order to further
photogrammetrically process imagery forming the adjusted strip, RPCs can be
generated for each intermediate image from the adjusted sensor orientation.
These then facilitate bias-free georeferencing in any scene without reference to
ground control. Mapping accuracy to sub-pixel level can be achieved from this
strip adjustment approach, which was primarily developed to support automated
orthoimage generation from ALOS PRISM and AVNIR-2 imagery by
Geoscience Australia, (Fraser et al., 2008).

2
One of the obvious questions that arises with long-strip adjustment centres
upon just how long the strip of images can be. This paper addresses this
question, through a practical evaluation of the adjustment technique to single
and double-strip configurations of ALOS PRISM imagery where the strip length
reaches more than 60 scenes, or 1760km, with GCPs being confined to the
ends of the strip. In the following sections, the sensor orientation model
employed for the strip adjustment is first reviewed, along with the modeling of
the spatial misalignment of the CCDs inside the PRISM sensor. This is followed
by a brief description of the process by which successive scenes are merged
into a single strip with one set of orbit and attitude correction parameters. The
process of RPC generation is also briefly reviewed. Finally, an assessment of
the accuracy achieved in the experimental long strip adjustment of a single strip
of 55 images and a double strip comprising 124 images is presented and
concluding remarks are offered.

Sensor Orientation
Generic Orientation Model
The sensor orientation model used for the long-strip adjustment has been fully
described in Weser et al. (2008a;b). Here, a brief description only will be
presented. The physical model for a pushbroom satellite imaging sensor relates
a point PECS = (XECS, YECS, ZECS)T in an earth-centered object coordinate
system to the position of its projection pI = (xI, yI,0)T in an image coordinate
system. The coordinate yI of an observed image point directly corresponds with
the recording time t for the image row through t = t0 + t · yI, where t0 is the time
of the first recorded row and t the time interval between scans. The framelet
coordinate system, in which the observation pI can be expressed as pF = (xF, yF,
zF)T = (xI, 0, 0)T, refers to an individual CCD array. The relationship between an
observed image point pF and the object point PECS is given as

pF = cF +  · RMT · {RPT(t) · ROT · [PECS – S(t)] – CM} – x (1)

Here, cF = (xFC, yFC, f) describes the interior orientation, ie the position of the
projection centre in the framelet coordinate system. The vector δx formally
describes corrections for systematic errors such as velocity aberration and
atmospheric refraction. It can also be expanded to model camera distortion or
other systematic error effects. The shift CM and the rotation matrix RM describe
a rigid motion of the camera with respect to the satellite. They are referred to as
the camera mounting parameters.
The satellite orbit path is modelled by time-dependant functions, ie
S(t)=[X(t),Y(t),Z(t)]T, and the sensor attitudes are described by a concatenation
of a time-constant rotation matrix RO and a matrix RP(t). These are
parameterized by time-dependant functions describing three rotation angles,
roll(t), pitch(t) and yaw(t). The components of the orbit path and the time-
dependant rotation angles are in turn modelled by cubic spline functions. The
rotation matrix RO rotates from the earth-centred coordinate system to one that

3
is nearly parallel to the satellite orbit path and this matrix can be computed from
the satellite position and velocity at the scene centre.
PRISM Sensor
A feature of ALOS PRISM is that depending on the imaging mode, four or six
CCD arrays are used to record a scene, each array producing a separate image
file. These sub-images share their exterior orientation and camera mounting
parameters, and the focal length. However, each has its own framelet
coordinate system, so the coordinates of the principal point can be different for
each (see Weser et al., 2008b). In order to treat the sub-images as a single
image, a correction vector x = (x,y, 0)T is used. The interior orientation
parameters cF = (xFC, yFC, f) are assumed to be identical for all sub-scenes, with
corrections (xi, yi) modelling the relative alignment of the CCD chips then
becoming:

xi = xS + a0i + a1i · xFi + a2i ·xFi2 (2)

yi = yS + b0i + b1i ·xFi + b2i · xFi2

where i is the index of the CCD array. The parameters xS and yS are the
systematic error corrections for velocity aberration and atmospheric refraction.
The offsets a0 and b0 model the different positions of the origins of the individual
framelet coordinate systems relative to the camera coordinate system. The
linear terms a1i and b1i correct for small rotations and differences in the pixel
sizes of the CCD arrays with respect to each other. The parameter b2i models
any deviation of the CCD array from a straight line, whereas a2i is related to
non-linear variations in pixel size. The required sensor orientation model for
ALOS PRISM is formed through a substitution of Eq. 2 into Eq. 1.
Bundle Adjustment
The aim of bundle adjustment of the strip(s) of imagery is to improve the
parameters of the sensor model formulated in Eqs. 1 and 2 using the framelet
coordinates of image points (both control points and tie points in the case of
multiple strips), the corresponding object coordinates of GCPs, and direct
observations for the orbit path and attitudes derived from the metadata. The
adjustment formulation includes a correction for systematic errors in the direct
observations of orbit path and attitudes. These errors are modeled by
expanding the associated cubic spline functions so that for each orbit
observation pobs at time tobs, a time-constant unknown ∆p is introduced. The
resulting spline function Sp(tobs) is given as

Sp(tobs) = pobs + r + ∆p (3)

where r denotes the random error affecting the orbit observations. Six
systematic error correction parameters per satellite orbit, three offsets (∆X, ∆Y,
∆Z)T for the orbit path points and three offsets (∆roll, ∆pitch, ∆yaw)T for the
rotation angles, have to be determined along with the spline parameters.

4
For stereo imaging configurations or overlapping strips of imagery, bundle
adjustment can directly yield 3D ground point coordinates from the
corresponding observed image point pairs, or possibly triplets for PRISM’s
three-line scanning mode. In the case of a single strip of near nadir scenes, only
a single set of image coordinates can be measured for the corresponding
ground point, i.e. no tie points are present, thus the term bundle adjustment is
perhaps a little misleading since no direct computation of the ground
coordinates of any measured image point is performed. The actual process is
more akin to a spatial resection from GCPs in which exterior orientation
parameters are refined. In such cases 2D georeferencing is then possible via
monolpotting (Willneff et al., 2005) so long as an underlying DEM is available.
3D georeferencing will be possible for tie points between overlapping long strips
of images. Two image strip configurations are considered in this paper, namely
a long single strip and two overlapping strips of nadir imagery.
Strip Adjustment
In the long-strip adjustment, the entire strip is regarded as a single image. The
first step in the process involves an initial merging of scenes along with their
associated orbit and attitude data. A single set of camera mounting and interior
orientation parameters then applies for the orientation of what may now be
considered a single composite image. Also, the six bias correction parameters
for orbit path and attitude relate to the entire strip. The resulting adjusted
orientation and bias-correction parameters for the strip can then be mapped
back to the individual scenes to refine their orientation.

RPC Generation
As mentioned, the measurement of tie points between adjacent images is not
necessary in the long-strip adjustment of a single strip of images, and the strip
adjustment should be viewed as a means of correcting the sensor orbit and
attitude data. The adjusted orientation data can then be delivered to the user in
the form of newly generated RPCs. Georeferencing can then follow as a
separate rather than integral part of the strip adjustment process.
The RPC model, as computed using camera and sensor orientation
parameters, is universally accepted as a valid alternative sensor orientation
model for HRSI (eg Fraser et al., 2006). RPCs facilitate the transformation from
image to object space coordinates in a geographic reference system. Within the
long-strip adjustment scenario, RPCs are generated for each of the images
forming the strip, such that bias-free georeferencing and associated processes
(eg orthoimage generation) can subsequently take place on standard
photogrammetric workstations. While such commercial systems might be able
to directly ingest orbit and attitude data from the sensor metadata files, it is
unlikely that they will accommodate the sensor orientation corrections
generated within the long-strip adjustment (Fraser et al., 2008). The generation
of bias-corrected RPCs offers a means to circumvent this problem.

5
Experimental Evaluation
Description of Datasets
In order to evaluate the strip adjustment process for very long strips,
experimental tests were conducted using two sets of ALOS PRISM nadir
images provided by Geoscience Australia. One set of data formed a 1527km
long, 35km wide strip of 55 images from the same orbit over Eastern Australia,
as shown in Fig. 1. Each scene comprised sub-images from four CCD arrays. A
total of 194 ground points were employed, either as GCPs or check points.
Figure 2a indicates their location within the strip. The majority of the ground
points were GPS-surveyed to 1m or better accuracy in planimetry and height.
Some 19 points were determined using existing orthophotos and DEMs of the
areas where more accurate reference data was not available. These points
were distributed across the 17 northern-most scenes.

Figure 1. Locations of ALOS PRISM nadir scene strips, a single strip of 55

scenes and a double-strip block of 124 scenes.
The second data set comprised two overlapping ALOS PRISM strips of 1760
km length and 120 km width over Western Australia, also shown in Figure 1.
The block comprised 124 scenes, each with 6 adjacent sub-scenes (separate
CCD arrays), covering an area of 40 km in length and 70 km in width. The side
overlap between the strips amounted to 25% percent. The only available
sources of reference data for determination of 3D coordinates of GCPs over the
whole 215,000 km2 area of the block were orthophotos and DEMs. The ground
points derived from orthophotos were assigned height values by bilinear
interpolation from the DEMs. Of the 100 points obtained, at the locations shown
in Figure 2c, six were used as GCPs, located at the northern and southern end
of the block, and the rest as check points. The positional accuracy of the points
was modest, at between 5 m and 10 m, the height data having been derived
from the 1 arc second SRTM DEM.

6
Data Processing
All image measurement, sensor orientation adjustment, RPC generation and
subsequent georeferencing via monoplotting were carried out using the Barista
software system (Barista, 2009). The workflow begins with import of ALOS
PRISM nadir scenes into Barista, along with the associated metadata, and the
merging of the orbit and attitude observational data for each strip. This is
followed by the collection of ground points, these being GCPs as well as check
points, and the subsequent measurement of image coordinates. Next, strip
adjustment is performed using a small number of GCPs, and checkpoint
discrepancies (residuals) are computed in image space via back projection
utilising the adjusted (and bias corrected) exterior orientation for each image.
Monoplotting is then employed to assess the planimetric accuracy in object
space. Finally, RPCs are generated for each scene from the adjusted orbit
orientation parameters, and the same accuracy evaluation in both image and
object space is again performed to validate the computation of the RPCs. Under
the assumption that the RPCs model the adjusted sensor orientation data to an
accuracy of better than 0.1 pixel, the results of the initial and final accuracy
assessment should correspond for all practical purposes.

Results
A number of adjustment computations were performed using different GCP
configurations, with similar results being obtained for the cases of most interest,
namely those with the smallest number of GCPs deemed necessary. For the
55-image single strip, the case to be considered here is that of four GCPs, two
GPS-surveyed points at the southern end and two points derived from
orthophotography and DEM data at the Northern end, as indicated by the point
locations shown in Figure 2a. The rest of the points are treated as check
points, and since there is overlap between successive scenes, some check
points appear twice. For the 124-scene block, in which tie points were
measured within the overlap area between the strips, the case of six end-point
GCPs is considered.
Table 1 lists the RMS values of checkpoint discrepancies in image space
obtained from both the adjusted orbit data computed in the long-strip
adjustment, and from application of the subsequently generated RPCs. The
results for both the 55-scene strip and the 124-scene block show that 1-pixel
accuracy is achievable. The corresponding values in object space, in UTM
coordinates, are listed in Table 2 and the discrepancy vectors are shown in
Figure 2. Here the checkpoint discrepancies were determined via monoplotting,
using the 1-second SRTM DEM. For the 124-scene block, accuracy at the 1.1
pixel level in achieved with little difference between discrepancies along- and
across-track. In the 55-scene strip, however, although 1-pixel across-track
accuracy is achieved, the RMS error in the along-track direction reaches 1.6
pixels or 3.9m. The cause of this larger error is thought attributable to the
GCPs/checkpoints being of inhomogeneous accuracy, 19 having been
determined from orthoimagery and DEM information.

7
 Check points
∆ GCPs

Distance scale Residual scale

(a) (b)

Figure 2a,b. 55-scene strip: (a) Distribution of GCPs and check points. (b) Check point
discrepancy vectors in planimetry

8
 Check points
∆ GCPs

Distance scale
Residual scale

(c) (d)
Figure 2c,d. 124-scene 2-strip block: (c) Distribution of GCPs and check points. (d)
Check point discrepancy vectors in planimetry.

9
 Table 1. RMS values of checkpoint discrepancies ∆x and ∆y in image
space coordinates, in pixels.

#
Strip/Block Orientation model Check RMSx RMSy ∆xmax ∆ymax
points

55-scene Generic model;

280 1.1 1.7 2.4 4.5
strip, Eastern adjusted orbit
Australia RPCs 280 1.2 1.4 3.4 3.6
124-scene, Generic model;
133 1.1 1.1 2.4 4.2
2-strip block, adjusted orbit
Western
Australia RPCs 133 1.1 1.1 2.8 3.5

Table 2. RMS values of checkpoint discrepancies ∆E and ∆N in object space

coordinates obtained via monoplotting with adjusted orbit observations or generated
RPCs, in metres.

# Check
Strip/Block Orientation model RMSE RMSN
points

55-scene Generic model;

194 2.5 3.9
strip, Eastern adjusted orbit
Australia RPCs 194 2.8 3.1
124-scene, Generic model;
100 2.7 2.7
2-strip block, adjusted orbit
Western
Australia RPCs 100 2.7 2.9

Concluding Remarks
The long-strip sensor orientation adjustment process overviewed in this paper
has previously been shown to facilitate sub-pixel accuracy in georeferencing
from ALOS PRISM imagery over strip lengths of 20 or so scenes (Fraser et al.,
2008; Rottensteiner et al., 2009). The results reported here illustrate that the
long-strip adjustment approach is also applicable to single- and two-strip
configurations with lengths exceeding 50 scenes (>1500km), the outcome being
that 1-pixel georeferencing accuracy is obtained with as few as 4 to 6 GCPs. In
the context of the automated orthoimage production program at Geoscience
Australia, which utilises the Barista photogrammetric software system, this has
led to a very significant reduction in GCP requirements and thus a substantial
boost in productivity.

Acknowledgements
The authors express their gratitude to GeoScience Australia for making the strip
of ALOS PRISM imagery available.

10
References
Barista, 2009, Barista information webpage, http://www.baristasoftware.com.au
[Accessed: 14 Jan. 2010].
Fraser, C.S. and Hanley, H., 2003, Bias compensation in rational functions for
IKONOS satellite imagery. Photogrammetric Engineering & Remote Sensing,
69(1): 53-57.
Fraser, C. S. and Ravanbakhsh, M., 2009, Georeferencing from Geoeye-
1Imagery: Early indications of metric performance. International Archives of
Photogrammetry, Remote Sensing & Spatial Information Sciences, Hannover,
Germany, Vol. 38, Part I-4-7/W5.
Fraser, C. S., Weser, T. and Rottensteiner, F., 2008, Image merging to support
georeferencing and orthoimage generation from ALOS imagery. 29th Asian
Conference on Remote Sensing, ACRS, Colombo, 8 pages (on CDROM).
Fraser, C. S., Dial, G. and Grodecki, J., 2006, Sensor orientation via RPCs.
ISPRS Jnl Photogrammetry and Remote Sensing 60(3):182-194.
Rottensteiner, F., Weser, T. and Fraser, C.S., 2008, Georeferencing and
orthoimage generation from long strips of ALOS imagery. Proceedings of 2nd
ALOS PI Symposium, ESA/JAXA, Rhodes, Greece, 3-7 Nov., 8 pages.
Rottensteiner, F., Weser, T., Lewis, A. and Fraser, C.S., 2009, A Strip
Adjustment Approach for Precise Georeferencing of ALOS Imagery. IEEE
Transactions on Geoscience and Remote Sensing, 47 (12; Part I): 4083-4091.
Weser, T., Rottensteiner, F., Willneff, J., Poon, J. and Fraser, C. S., 2008a,
Development and testing of a generic sensor model for high-resolution satellite
imagery. Photogrammetric Record, 21(123): 255-274.
Weser, T., Rottensteiner, F., Willneff, J. and Fraser, C. S., 2008b, An improved
pushbroom scanner model for precise georeferencing of ALOS PRISM imagery.
International Archives of Photogrammetry, Remote Sensing & Spatial
Information Sciences, Beijing, XXXVII (B1-2): 723-730.
Willneff, J., Poon, J. and Fraser, C.S., 2005, Monoplotting Applied to High-
Resolution Satellite Imagery. Journal of Spatial Science, 50(2):1-11.

11