tanbul 2004
wameters of
CPs. In the
observation
n.Y. The
r equations
accuracy of
ons are then
In addition,
ameters are
wledge and
yrvent the
put errors.
reduce the
least-square
15 of the 3D
inearization
e values for
s initialized
information
ata can be
'Ssing are:
model;
:h GCP and
ns for each
MS errors;
each point.
he image to
11 elevation
s) to avoid
; than the
P residuals
the GCPs
o obtain an
modelling
P accuracy,
1are (RMS)
adjustment
GCP RMS
e ICP RMS
des feature
of the final
er, the final
ling will be
hus normal
adjustment
3CP error.
rally in the
. which is,
n of image
n of Z-error
the general
. stable and
rating local
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
errors, regardless the GCP accuracy and number. These
statements are mainly supported by the ICP errors, as an
unbiased validation of the modelling. However, these ICP
errors include both the cartographic errors and the extraction
error (due to the image content) of ICP features, and are thus a
good estimation of the restitution accuracy. The internal
accuracy of the modelling is in fact better, in the order of pixel
or sub-pixel.
p GCP | GCPX-Y | ^ GCP/ | GCP RMS ICP RMS
| Method | Accuracy | ICP Residuals Errors
mE | 10m | 30358 ^60 | 53 132 l 4S
i Photos 3-5m .| 20/38 3.9 4.3 4.2 2.6
|pGPs | 02m | 1038 | 02 | 02 | es [08
Table 1. Results of the least-square bundle adjustment of the
3D physical model using the different GCP
collection: with GCP accuracy the number of GCPs
and ICPs, RMS residuals and errors (in metres) on
GCPs and ICPs, respectively.
Presently there is no apparent explanation as to why the good
quality results (6 m and 5.3 m), as related to input accuracy (10
m), were obtained with the GCPs from a 1:50,000 topographic
map. The ICP error is in fact at least twice better than the input
accuracy, with the cartographic coordinate error as the
predominant error. A good potential reason is that the map has
a good homogeneity and a good relative and internal accuracy
(small random error), which are thus reflected in strong bundle
geometry of the QuickBird image. The systematic error of the
map is thus compensated by a translation in the image
modelling. Care must be taken in the extrapolation of these
results with other maps; however, these results have been
confirmed in an unpublished CCRS study with QuickBird
image over Voisey Bay, Newfoundland and Labrador, Canada.
On the other hand, the medium quality results (3.9 m and 4.3
m), as related to input accuracy (3-5 m), of the geometric
modelling computed with the GCPs collected from the
orthophotos are due to differential errors caused by a lack of
homogeneity between the orthophotos. The ICP error is in fact
almost the same than the input accuracy, with the cartographic
coordinate error as the predominant error. These differential
errors (even the systematic error) create local random errors in
the different parts of the image, which cannot be fully
compensated by the modelling.
The high quality results, as related to input accuracy, obtained
with the DGPS system and the GPS meet the 1:5,000 to
1:10,000 mapping accuracy, respectively. The ICP error is in
fact almost the same than the input accuracy: the cartographic
coordinate error being the predominant error (2-3 m) for hand-
held GPS collection method, while the 1-m image pointing
error being the predominant error for the DGPS collection
method. The DGPS results demonstrate that to achieve the best
accuracy (sub-pixel) with QuickBird image, the predominant
error in this GCP collection method has to be reduced by
choosing well-defined GCPs (natural or artificial targets) in
order to reduce the image pointing error to sub-pixel.
4. CONCLUSIONS
Different GCP collection methods were used for geometrically
processing QuickBird image with CCRS 3D multi-sensor
physical model: 10 m accurate topographic map to 0.2 m
accurate DGPS. First, a larger number of GCPs has to be
collected when their accuracy decreases to reduce the error
propagation in the least-squares bundle adjustment. Then,
positioning errors of few metres were achieved regardless the
collection method: 3-4 m with 1:50,000 map and 3-5 m accurate
orthophotos and around 1 m with 2-3 m accurate GPS and 0.20
m accurate DGPS. With the DGPS method, the predominant
error came from the image pointing: natural and artificial
targets should be used to further reduce the errors.
Consequently, depending upon the positioning accuracy
required by the user and their applications, the appropriate GCP
collection method can be chosen to maximize scientific aspects
such as input and processing, output and accuracy as well as
better manage project aspects such as delivery time, efficiency,
and costs.
ACKNOWLEDGEMENTS
The authors thank Mr. Matthew Woods from DigitalGlobe for
the QuickBird image and Mr. Réjean Matte du Ministère des
Ressources naturelles du Québec, Canada for the cartographic
data. They also thank Kim Lochhead and Doug Scott of the
Geodetic Survey Division, NRCan and Gerry Belanger of
Gemini Positioning Systems Ltd. (http://www.gpsl.com/) for
assisting with the DGPS data collection, as well as Ms. Katrin
Hohlbein and Ines Geisler of Technische Universität Dresden,
Germany for processing the data.
REFERENCES
Mikhail, E.M., 1976. Observations and Least Squares, Harper
& Row Publishers, New York, U.S.A.
Robertson, B., 2003. Rigorous Geometric Modeling and
Correction of QuickBird Imagery, Proceedings of the
International Geoscience and Remote Sensing, IGARSS 2003,
Toulouse, France, 21-25 July 2003, (Toulouse, France: CNES)
|. CD-ROM.
839
Savopol, F., A. Leclerc, Th. Toutin, Y. Carbonneau, 1994. La
correction géométrique d'images satellitaires pour la Base
nationale de données topographiques, Geomatica, 48(3), pp.
193-207.
Toutin, Th., 1995. Multi-Source Data Fusion with an Integrated
and Unified Geometric Modelling, ÆARSeL Advances in
Remote Sensing, 4(2), pp. 118-129.
http://www.cers.nrcan.gc.ca/ccrs/rd/ sci pub/bibpdf/1223.pdf
(accessed 9 May 2004).
Toutin, Th., 2004a. Comparison of Stereo-Extracted DTM from
Different High-Resolution Sensors: SPOT-5, EROS, IKONOS
and QuickBird, — /EEE-TGARS, 42(0), (in press)
http://dweb.cers.nrcan.ge.ca/cers/db/biblio/paper_e.cfm?Bibliol
D=13383 (accessed 9 May 2004).
Toutin, Th., 2004b. Review Paper: Geometric processing of remote
sensing images: models, algorithms and methods, /nternational
Journal of Remote Sensing, 25(10), pp.1893-1924,
http://Www.cers.nrean.ge.ca/cers/rd/sci_pub/bibpdf/13288.pdf
(accessed 9 May 2004).
