ORTHORECTIFICATION OF SATELLITE IMAGES USING EXTERNAL DEMS FROM

IFSAR
J. Bryan Mercer, Jeremy Allan, Natalie Glass, Johnathon Rasmussen, Michael Wollersheim

Intermap Technologies Corp., #1000, 736 8th Ave. S.W., Calgary, Canada, T2P 1H4
bmercer@intermap.ca
Commission II, Working Group II/2
KEY WORDS: Mapping, Orthorectification, SAR, Interferometer, DTM/DEM, IKONOS, Quickbird

ABSTRACT:
Images from high-resolution satellites are now in relatively wide use for a variety of applications. For use in most medium and largescale mapping applications, it is desirable that the imagery be orthorectified. While this may be accomplished by utilizing stereo pairs
acquired from these agile-pointing systems, the price and delivery schedule may be significantly impacted by this requirement. On the
other hand, savings of time, resources and cost may be realized through use of an external Digital Elevation Model (DEM) of
appropriate quality to perform the orthorectification on single, non-stereo, images. The quality of the DEM is a key factor in the
ultimate accuracy and robustness of the orthorectification. DEM quality in this sense is equated to vertical accuracy and to sample
spacing (posting). It has been shown that as the off-nadir viewing angle increases, the vertical accuracy required must improve for any
given horizontal accuracy requirement. DEMs created by the STAR-3i IFSAR system have been used to test and demonstrate
orthorectification of IKONOS, QuickBird and EROS-A1 images in addition to aerial photography. Several methodologies have been
used including that of a simple Rational Functions approach. Together with input of GCPs (Ground Control Points) taken from the
STAR-3i Orthorectified Radar Images (ORIs) that accompany the DEMs, mapping scale accuracies suitable for 1:4800 and larger
have been achieved. In this presentation we demonstrate several examples of orthorectification performed with single images from
IKONOS and QuickBird. Quantitative and visual results are presented.
1. INTRODUCTION
In order for imagery to be utilised in GIS and other applications
for which pixels must be correctly geo-located, image
orthorectification is a necessity.
The particular application
determines the factors required specifically the image
resolution and horizontal accuracy. These are often defined in
terms of desired mapping scale. Orthorectification implies that
both sensor system and terrain-related errors are reduced to
some specified level and that the image pixels are resampled to a
particular projection and reference system.
Aerial photography and satellite images exhibit certain similarities
and differences in terms of the parameters and issues to be dealt
with during orthorectification. In both cases, the horizontal pixel
displacement caused by off-nadir viewing increases with viewing
angle . Specifically, =h tan for aerial photography, as a
back-of-the-envelope calculation shows, where = the terrain
displacement and h = the terrain height of the pixel. For typical
high-resolution satellites, increases somewhat faster than tan
as shown by Zhang, et.al. (2000). This parallax displacement
error is usually corrected in a pixel-by-pixel manner during
orthorectification, provided the elevation, h, of the ground pixel
is known. However, the consequence of the viewing angle
dependence is that the displacement correction required depends
more critically on the vertical accuracy of the elevation
information as increases. A second elevation-related factor
impacts the accuracy of the orthorectification: the spatial
frequency of the terrain relief drives the requirement for spatial
sampling density of the terrain elevation. For example, elevations
sampled at 30 meter spacing would be inappropriate for retaining
orthoimage accuracy along ditches or other land features that
change more rapidly.

Traditional orthorectification involves the use of stereo pairs

from which elevation is extracted directly to be used for
production of contours and DEMs in addition to
orthorectification of the images. However the cost of acquiring
and processing stereo may not be justified if the purpose is to
do orthorectification only. In the case of the high-resolution
satellites, there is also an opportunity cost for stereo acquisition:
the time spent in acquiring the stereo partner is time lost in
acquiring additional primary images. Given the small footprint,
(100-200) km2 per image, this is a serious cost factor. Provided
a suitable DEM is available from external sources, it seems
plausible and cost-effective to acquire single images and perform
the orthorectification without invoking stereo requirements.
The purpose of this paper is to demonstrate orthorectification
results achievable using external DTMs derived from airborne
IFSAR specifically the Intermap STAR-3i system. We
provide visual examples and quantitative accuracy results relating
to the orthorectification of IKONOS and QuickBird images.
Aside from the technical merits we also address the cost benefits
of the process.
We provide a brief review of the IFSAR data in section 2,
describe the orthorectification methodologies used in section 3,
and the data sets in section 4. We then present the results of the
accuracy analysis along with visual examples in section 5. A
demonstration of comparative cost is provided in section 6
followed by concluding remarks section 7.
2. THE IFSAR DATA
STAR-3i is an X-band interferometric SAR (IFSAR) carried in a
Learjet (Mercer and Schnick (2000)) and has been operated
commercially by Intermap Technologies since 1997. The core
products generated include a Digital Surface Model (DSM) and
an Orthorectified Radar Image (ORI). A Digital Terrain Model

(DTM) is extracted from the DSM in a process referred to as

TerrainFit (Wang et. al., 2001). Current specifications for these
products are shown in Table 1. A detailed description of these
products may be found in Intermaps Core Products Handbook
at www.intermaptechnologies.com. The range of vertical
accuracy specifications relate to economic trade-offs that are
associated with the flying altitude (6,000 meters flying altitude
permits better accuracy but 6 km swath compared to 9km
altitude and 10 km swath at reduced accuracy) as well as other
GPS and operational considerations.
ORI

DSM

DTM

Units

1.25, 2.5

meters

Posting

meters

Accuracy
(vertical)

0.5, 1.0, 3.0

0.7, 1.0

m (RMSE)

2.0

2.5

m (RMSE)

Resolution

Accuracy
(horizontal)

Table 1. STAR-3i core product specifications. Note that

prior to 2001, ORI resolution was 2.5 meters while after a
major upgrade it was reduced to 1.25 meters.

A database of Core Products (currently >1 million km2) has

been created and is growing rapidly. National programs have
been completed recently and it is the intent to continue with the
creation of similar data sets. The relevance to this paper relates
to the ability to use the orthorectification possibilities from
archival data that enhances the cost impact of the method. In the
discussions that follow, it is the DTM rather than the DSM that
is generally used for orthorectification in order to remove the
effect of objects such as buildings or trees.
3. METHODOLOGIES
There are several approaches that may be taken to perform the
orthorectification of satellite images. In the current work we
focus on a specific Rational Functions approach. It is
particularly suitable for the class of high-resolution satellites
typified by IKONOS , QuickBird, EROS-A1, and Orbview-3
which are agile pointing and characterized by potentially large
off-nadir viewing angles which are relatively uniform across
each image.
3.1 Rational Functions
In the absence of camera model or other support information,
one approach is to use a Rational Functions Model ( e.g. Tao
and Hu (2000), Dowman and Doloff (2000), Fraser, et.al.
(2001)). This involves solving the rational function expressions
in equation (1) for their polynomial coefficients
(Aijk , Bijk ) which relate the normalized line and pixel numbers
(l,p) of GCPs (Ground Control Points) identified in the raw
satellite image to their object space coordinates (x, y, z).
l=

P1 (x, y, z)
Q1 (x, y, z)

p=

P2 ( x, y, z )
Q2 ( x, y, z)

(1)
P = Aijk x y z

i j k

Q = Bijk x y z

i j k

Normally the polynomials are terminated at 3rd order, although as

noted below our best results occur for lower order. The raw
image is then re-sampled on a pixel-by-pixel basis where z(x,y)
comes from the DTM. The software implementation used in this
work is from PCIs Geomatica OrthoEngine Rational Functions
Module. The GCPs are taken from the STAR-3i ORI and are
manually matched to mutually visible features such as road
intersections in the raw satellite image. The advantage of this
approach is that it is usually possible to acquire reasonably large
numbers of GCPs that sample the horizontal as well as vertical
terrain surface. Extensive tests have been performed with real
data and independent checkpoints to ascertain the trade-offs
between accuracy, number of coefficients and number of GCPs
as input. Typically, 20-30 GCPs are acquired, for a single
satellite image of dimension 10-18 km using only the linear terms
of the polynomial expressions.
Several IKONOS images, a
QuickBird image and an EROS-A1 image
have been
orthorectified in this manner with results comparable to those
presented below. There have been no apparent issues with
stability to date, perhaps because of the quantity, quality and
distribution of the GCPs.
The IKONOS test results indicated that most of the signal was
in the linear terms, while the higher terms appeared merely to add
noise. This appears understandable with recognition that
systematic errors remaining in the image are probably linear while
the terrain correction is also linear (to first approximation)
because the viewing angle is almost constant across a single
image (for off-nadir viewing). Hence terrain displacement will
scale with height independently of (x,y) within the image. The
fortunate implication of this result is that there is significant
redundancy possible for a reasonable number of GCPs, thus
allowing matching errors to be partly alleviated by the averaging
effect of the redundancy. Attempts are made to distribute the
GCP samples well in x, y, and z, although experience so far
indicates that the method is not highly sensitive to the z
distribution as long as approximate min/max is included. The
latter statement is specific to the use of low order solutions
where Z is de-coupled from x and y.
An alternative, but still rational functions-based approach to
orthorectification, is through the use of RPCs (Rational
Polynomial Coefficients). Although the camera models for
IKONOS and QuickBird are generally not available to the user,
RPCs, which have been derived from the camera models and
orbital information, are now provided by Space Imaging
(Grodecki and Dial (2001) and by DigitalGlobe. GCPs and
DTM are still required to perform the orthorectification. The use
of these RPCs may provide different results than those
computed directly from the data as described above, at least in
areas of significant topography, since the coefficients in the latter
case will be topography dependent, while the RPCs are not. A
comparative test is currently being done using the Geomatica
RPC module.
3.2 Rigorous Model
A rigorous model, based upon deterministic considerations
has been developed by Toutin (2003) and has been implemented
in the PCI Geomatica software package. Because many of the
required parameters are unknown there is still a need to insert
GCPs into the model implementation (6 GCPs required as a
minimum, but more preferred) as well, of course, as a DTM. In
principle, because it is based on a physical model, it should be
able to produce superior results using fewer GCPs than the
numbers described in the previous section. Moreover the claim

(ibid) is that the method should be more robust in that it is less

dependent upon the number and choice of GCPs. We provide a
comparison below, of the results obtained when using the
method described in section 3.1 versus the rigorous method of
Toutin (2003).

4. THE DATA SETS

uncertainties plus inherent errors in both ORIs result in overall

single-point errors of about 2-3 meters RMSE in northing and
easting.
It should be noted that the IKONOS orthorectification described
below used GCPs acquired from the 2.5-meter radar ORI.
Also, while comparative testing was done on the orthorectified
IKONOS mosaics as a whole (summer/winter), the absolute
error statistics relate only to the area overlapping the photo-truth.

4.1 IKONOS: Morrison

A major test site has developed within the Morrison Quad, a 7.5
x 7.5 USGS mapsheet, located west of Denver CO, USA. Data
sets from several sensors have been acquired in the area over the
past few years, including several STAR-3i data sets, several
IKONOS images, lidar acquisitions, GPS control and the
establishment of some photo-control at various scales. The
relevant details related to this paper are described in Table 2.

Morrison

IKONOS

IKONOS

STAR-3i

Acq'n date Feb 1, 2000 Aug 19, 2000 Oct 23, 1998
Product

Geo

Type

Pan-Sharpened MS

Geo

Resolution

1m

Off-Nadir

23.7 deg

Photo-Truth
1998

4.2 QuickBird: Castle Rock

The Castle Rock test area is to the south of Denver, and covers
an area approximately 17 km x 17 km, about 2/3 of which is
mountainous. The relevant details of the data are shown in Table
3 and a DSM of the area is shown in Figure 6. The QuickBird
files were quite massive and arrived in the form of a 3x3 matrix
of sub-tiles (500 MB each as there were 4 channels of pansharpened 0.6 meter data included). Initially the sub-tiles were
individually orthorectified but this of course led to boundary
problems. However it was possible to create a single large
model into which each of the nine sets of GCPs derived from
the sub-tiles contributed. This worked well and eliminated the
internal boundary issues.

GT2

15 CPs

ORI, DTM

Pan ORI

1m

2.5 m

1m

Castle Rock

QuickBird

STAR-3i

Truth

13.2 deg

30-60 deg

NA

Acq'n date

July 9, 2002

2001

2001

Table 2. IKONOS Morrison tests summary of input data.

The Morrison quad (approximately 10 km x 14 km) includes the

Rocky Mountain range in the west, and moderate to steep relief
in the east with urban tracts and other forms of development. A
colour-coded STAR-3i DSM of the area along with the
corresponding IKONOS winter scene is shown in Figures 1 and
2. Topographic relief extends beyond the 700 meters shown on
the colour legend. The DSM is shown to illustrate surface
roughness, although the DTM was used in the orthorectification.
The GCPs (28) extracted from the radar ORI are shown in red,
distributed across the IKONOS scene.
The IKONOS winter scene (Feb 13, 2000) came as two image
sets that covered the whole quad between them. The summer
scene (September 13, 2000), shown in Figure 3, covers only the
western portion of the quad. Another partially overlapping
image acquired August 19, 2000, covered the eastern portion.
Both were orthorectified and mosaicked together to form a
composite full-quad summer scene.
The truth for purposes of this test was an orthorectified airphoto that was created using standard photogrammetry. The
AT control was from GPS and the underlying DEM was from
stereo extraction. The horizontal accuracy of this photo-truth
is 1 meter (CE(90)). The coverage unfortunately only overlapped
about a third of the Morrison tile in the southeast (see Figure 1),
and from this 15 well-distributed independent CPs (check
points) were selected for accuracy validation of the overlapping
orthorectified IKONOS images.
Moreover the air-photo was used to assess the apparent
accuracy of the pre-upgrade (2.5 meter) radar ORI insofar as
uncertainties in matching common features allows.
This
assessment showed that typical manual feature-matching

Product

Basic

GT1

25 Check Pts

Type

Pan-Sharpened MS

ORI, DTM

GPS

Resolution

0.7m

1.25 m

Acc~ 0.5mRMS

Off-Nadir

12 deg

30-60 deg

Table 3. QuickBird Castle Rock tests summary of input

data.

Because there were about 25 GCPs per sub-tile initially acquired

(radar/QuickBird match points), all 206 GCPs were used in the
model determination. Subsequently the GCP numbers were
reduced by factors of 2 for accuracy comparison.
Independent ground points (25), used as checkpoints (CPs)
were supplied by DigitalGlobe and included GPS coordinates
and field descriptions as well as photographs of the features they
referenced. These points are shown, overlaid on the DSM in
Figure 6.
5. RESULTS
5.1 IKONOS (Morrison)
An overview of the topography and the appearance of the
Morrison quad is provided in Figures 1 and 2, showing a STAR3i DSM and the orthorectified IKONOS winter mosaic
respectively. The location of the 28 GCPs used in this particular
example is shown as well as the 15 checkpoints that were
acquired from the independent orthorectified aerial image (the
photo-truth). In Figure 3 we show an enlargement of part of the
orthorectified summer IKONOS scene showing the

Figure 1. Morrison DSM with dotted outline showing the subarea where photo-truth is available.

Figure 2. Morrison IKONOS ORI. There are 28 GCPs (from

radar ORI) overlaid in red, and 15 CPs (Check Points) in light
blue. The dark blue box shows the area enlarged in Figure 3.

centrelines of roads overlaid from the winter scene, illustrating

the small relative error found between the summer and winter
image mosaics overall. Total relief in this sub-area is about 200
meters. The same DTM and GCPs were used in each
orthorectification. As noted in Table 1, the imaging parameters
were quite different. Random inspection of the images when
enlarged to about 1:1,000, as illustrated in Figure 5, indicates
differences of 1-3 meters typically which is consistent with the
absolute accuracies reported in Table 4.

IKONOS

The absolute error statistics are reported in Table 4 and are with
respect to the CPs extracted from the photo-truth. In this
example, 28 GCPs were used in the Rational Function solution.
The CE(90) from both sets is less than 4 meters which is the
specification set by Space Imaging for their Precision Product.
This is the NMAS standard for 1:4,800 scale mapping. It is
interesting to note that the summer image exhibits lower error
than the winter scene. This is probably due to better observable
detail in the summer scene, permitting better feature matching for
both GCPs and CPs.

(west)

It was possible to extract large numbers of GCPs from the radar

ORI and for test purposes over 200 GCPs were acquired. The
Rational Functions coefficients were re-computed and the GCP
and CP residuals were determined and the statistics calculated.
The CE(90) error with respect to the CPs is plotted in Figure 4
as a function of the number of GCPs in the solution. It is
notable that after a spike at the beginning, which is likely a
statistical vagary, the results show CE(90) ~ 3.4 meters
consistently, dropping to about 3.3 meters at 200 GCPs. The
CE(90) results from using the Toutin Rigorous model are also
plotted in Figure 4. Over most of the range they perform
similarly, although for large and small GCP numbers the
accuracy is slightly worse than for the Rational Functions
method.

Mean

Std Dev'n

RMSE

CE90

(meters)

DX

Winter

GCPs 0.0

0.0

3.3

1.9

3.3 1.9

5.5

CPs -1.7

0.8

1.0

1.0

2.0 1.2

3.5

GCPs 0.0

0.0

2.1

1.6

2.1 1.6

4.0

0.3

-0.9

0.8

0.9

0.8 1.3

2.3

GCPs 0.0

0.0

2.8

2.7

2.8 2.7

5.8

Summer
(east)
Summer

CPs

CPs

DY

DX

DY

DX

DY

Table 4. IKONOS ORI difference statistics for Morrison.

Results from both winter and summer scenes are shown. DX, DY
are the differences (IKONOS - GCP, CP). 28 GCPs; 15 CPs

5.2 QuickBird
An overview of the Castle Rock topography is presented in
Figure 6. The QuickBird ORI is displayed in Figure 7 as an
overview. The associated inset is an enlargement of the subscene shown in the small box on the overview. The quality of the
ORI sub-area shown here is characteristic of the remainder of
the scene. No local distortions are evident, indicating that even at
this or larger scales the orthorectification remains robust.
Quantitative results are shown in Table 5. Overall, the results
(CE(90) < 4.0 meters) are similar to those obtained for
IKONOS. Thus the method appears appropriate for mapping at
scales of 1:4,800 or better. DigitalGlobe orthorectified the same
image using their camera model and made it available for
comparison. Although we dont show examples here, the results
were consistent with the accuracies quoted in Table 5. Apart
from a systematic easterly offset (~1.5 meters) of unknown
origin, the centrelines of roads and other features

Figure 3. This is an enlargement (about 500 meters x 500 meters)

from the IKONOS summer ORI (see blue box in Figure 1)
shows road centrelines overlaid that were transferred (without
adjustment) from the winter ORI. The box indicates the location
of further enlargement (Figure 5). Also shown is a transferred
ridgeline segment.

Figure 5. Further enlargement, showing winter and summer road

centrelines overlaid. The 2-meter difference is typical of the
differences observed.

Morrison Ikonos
4.60
4.40

Toutin's Rigorous Model

Rational Functions Model

CE90 (m)

4.20
4.00
3.80
3.60
3.40
3.20
3.00
0

25

50

75

100

125

150

175

200

225

Number of GCPS

Figure 4. CE(90) as a function of the number of GCPs used for

the (a) the Rational Functions Model (low order coefficients) and
(b) the Rigorous Model of Toutin

QuickBird
(meters)
GCPs

206

CPs

14

Mean

Std Dev'n

RMSE

DX

DY

DX

DY

DX

DY

0.0
-1.2

0.0
0.6

2.1
1.2

1.9
1.4

2.10
1.63

1.87
1.50

CE90

4.26
3.37

Table 5. QuickBird ORI error statistics for Castle Rock.

were generally within 2 meters of each other in the two resulting
ORIs.

6. COMPARATIVE COSTS
According to the information published on the web site of Space
Imaging, for example, the difference between the prices of pansharpened MS orthorectified images (Precision Product) and

Figure 6. Castle Rock DSM (approximately 16.5 km x16.5 km).

Topographic relief is about 500 meters. CPs are shown overlaid.

their non-ortho images (Geo Product) is in the range 55-150

US$/km2 depending on whether it is a North American or
International scene. On the other hand, the cost of IFSAR
DSMs, DTMs and ORIs in the Intermap archival database is in
the range of 6-20 US$/km2 depending on location. Experience
(in a research, not a production environment) suggests that the
cost of doing the IKONOS orthorectification should be less than

$10/km2 under similar conditions. These considerations indicate

that where archival IFSAR exists there are substantial savings to
be made by the user by following this approach rather than the
stereo approach. Even when new IFSAR data must be
acquired, provided it is for large enough areas, the economics
still appear very favourable for this approach. A smaller
economic benefit is to be recognized for the DigitalGlobe
product only because its current high-end ortho product is
specified at 1:12,000 (USA) and 1:25,000 scale (International).
In this case the major benefit is product accuracy improvement.

was provided. There appears to be strong economic incentive to

use this approach where possible.

ACKNOWLEDGEMENTS
We wish to acknowledge with gratitude, Space Imaging for
providing both sets of IKONOS Morrison images and Digital
Globe for providing the QuickBird Castle Rock images and
checkpoints. In particular we thank Gene Dial and Jacek
Grodecki for information and encouragement.

REFERENCES
Di, K., Ma, R., and R.X. Li, 2003, Rational Functions and
Potential
for
Rigorous
Sensor
Model
Recovery.
Photogrammetric Engineering & Remote Sensing, Vol. 69, No.
1, January 2003, pp. 33-41.
Dowman, I., and J.T. Dolloff, 2000. An Evaluation of Rational
Functions for Photogrammetric Restitution, International
Archives of Photogrammetry and Remote Sensing,
33(B3/1):252-266.
Fraser, C.S., H.B Hanley, and T. Yamakawa, 2001. Sub-metre
Geopositioning with IKONOS Geo Imagery, Proc. Joint ISPRS
Workshop High Resolution Mapping from Space 2001, 1921 September, Hannover, Germany, (on CD ROM).

Figure 7. QuickBird orthorectified image of Castle Rock area.

The red box in the overview image is enlarged in the inset to
about 250 meters extent.

7. CONCLUSIONS
Orthorectification of IKONOS and QuickBird images at offnadir viewing angles to 23 degrees, has been demonstrated using
a Rational Functions method. CE(90) values better than 4 meters
were obtained when compared to independent check points over
full image scenes. The method utilized external IFSAR-derived
DTMs and associated ORIs as prolific sources of GCPs. The
ORIs (2.5 meter pixels) and DTMs ( 1-2 meter vertical accuracy,
extracted from 5 meter-posted DSMs) were from the STAR-3i
airborne IFSAR system prior to its recent upgrade. Both sets of
images sampled highly diverse topography. Two IKONOS
images were compared over the same area (the Morrison, CO
quad, a 7.5 tile) but at different times and with different viewing
angles, while the larger QuickBird image was of the Castle Rock,
CO area. Relative comparison of features (usually road
centrelines) showed consistency with the absolute checkpoint
measurements. An assessment of the absolute accuracy as a
function of GCP numbers indicated reasonable performance at
about 20 GCPs improving somewhat out to 200 GCPs. A
comparison between the Rational Functions method and the
Toutin model showed similar results as a function of GCP
number. We believe this is because the dominant terms in both
models are linear. Lastly, a brief review of the comparative
economics of this approach versus a stereo extraction approach

Mercer, J. B., and S. Schnick, (1999). Comparison of DEMs

from STAR-3i Interferometric SAR and Scanning Laser:
Proceedings of the Joint Workshop of ISPRS III/5 and III/2, La
Jolla, November, 1999
Grodecki, J., and G. Dial, 2001, IKONOS geometric accuracy,
Proceedings of Joint International Workshop on High
Resolution mapping from Space, 19-21 September 2001,
Hanover, Germany, pp. 77-86 (CD ROM).
Tao, C.V., and Y. Hu, 2001. 3D reconstruction algorithms with
the rational function model and their applications for IKONOS
stereo imagery, Proc. Joint ISPRS Workshop High Resolution
Mapping from Space 2001, 19-21 September, Hannover,
Germany, 12 p. (on CD ROM).
Toutin, T., 2003, Error Tracking in IKONOS Geometric
Processing Using a 3D Parametric Model, Photogrammetric
Engineering & Remote Sensing, Vol. 69, No. 1, January 2003,
pp. 43-51.
Wang, Y., B. Mercer, V. C. Tao, J. Sharma, and S. Crawford,
2001, Automatic Generation of Bald Earth Digital Elevation
Models From Digital Surface Models Created Using Airborne
IFSAR, ASPRS 2001 Proceedings, (on CD ROM).
Zhang, Y., V. Tao, and J.B. Mercer, 2000, Assessment of the
Influences of Satellite Viewing Angle, Earth Curvature, Relief
Height, Geographical Location, and DEM Characteristics on
Planimetric Displacement of High-Resolution Satellite Imagery,
ASPRS
2000
Proceedings,
(on
CD
ROM).