tanbul 2004 
ct features 
he imagery 
ents of the 
in Table 2. 
" coordinate 
on utilising 
mponent of 
1e biases in 
ted ground 
expected to 
ing rows of 
PC bundle 
AP sets. As 
checkpoint 
cular GCPs 
ative of the 
ed in RPC 
0 Bo. It can 
[S, l-sigma) 
btained with 
th one GCP. 
y of Mount 
¢ from the 
inimetry are 
rection. The 
| the 2-GCP 
bias of the 
| any affine 
ration. What 
s with shift 
impact upon 
ses in image 
for the three 
ig. | provide 
y have been 
le from time- 
-hand stereo 
distribution, 
error. Fig. la 
) image, Fig. 
in the along- 
f drift terms, 
due of image 
).25 pixels in 
increase in 
have reported 
anlikely to be 
is obtained in 
this view, 
r pattern seen 
on adequately 
By. it is not 
ameter model 
best indicator 
ereo triplet is 
here the RPC 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004 
  
  
  
  
  
  
  
Testfield Area Elevation Image Coverage Number of Notable Features 
Range (elevation angles) GCPs 
IKONOS, 120 km? sea level to Stereo triplet 110 Full scene; mountainous 
Hobart (11 x 11 km) 1280 m (69°, 75°, 69°) terrain 
QuickBird, 300 km? sea level to Stereo pair 81 Full scene, low relief 
Melbourne | (17.5 x 17.5 km) 50 m (approx. 63° each) area 
  
  
  
  
Table 1. Characteristics of the Hobart and Melbourne testfields. 
bundle adjustment with shift parameters employs all GCPs as 
loosely weighted control thus providing a solution that can be 
thought of as being equivalent to a free-network adjustment 
with inner constraints. Note here the RMS geopositioning 
accuracy of just below '4 pixel in the cross-track direction, and 
close to % pixel in both the along-track direction and in height. 
The results for the stereo pair of Geo images alone, without the 
central image, match very closely those listed in Table 2. From 
RPC bundle adjustments carried out with the Hobart IKONOS 
imagery, as well as with other testfield imagery covering areas 
as large as 2000 km? (e.g. Hanley et al., 2002; Fraser & Hanley, 
2003), we have ample evidence that sub-pixel geopositioning 
accuracy is quite achievable from IKONOS Geo stereo 
imagery. This can be expected to be the case irrespective of the 
nature of the terrain being imaged, the size of the scene, or the 
scanning mode of the satellite (Forward or Reverse). High 
quality GCPs and image coordinate measurement are of course 
prerequisites to the attainment of highest accuracy. 
3.3 QuickBird results 
The same computational procedure as carried out in the Hobart 
testfield was followed with the QuickBird Basic stereo pair 
covering Melbourne. Table 3 lists the results obtained. 
Basically, the  geopositioning accuracy achieved with 
QuickBird was the same as for IKONOS, though QuickBird 
produced in this case slightly lower accuracy in planimetry and 
slightly higher accuracy in height, no doubt as a consequence of 
the higher base-to-height ratio exhibited in the QuickBird stereo 
pair. Whether any component of the minor accuracy 
discrepancy between the two stereo pairs resulted from either 
the degree to which the original RPCs described the rigorous 
sensor model, or the resolution and accuracy of the actual 
orientation sensors on the satellite is not known. What is seen 
with QuickBird, however, are stronger indications of residual 
systematic error which is not being modelled by the bias- 
compensated RPCs. 
Shown in Figs. 2a and 2b are plots of the image coordinate 
residuals arising from the RPC bundle adjustment with shift 
parameters (row 2 of Table 3). The along-track alignment of the 
vectors is suggestive of perturbations in scan velocity, with the 
addition of a first-order scale effect. Thus, we would expect 
some of the error signal to be absorbed by the along-track drift 
parameter, A,. The results listed in Table 3 for the RPC bundle 
adjustment with shift and drift parameters, however, show only 
a modest improvement in accuracy in the cross-track direction 
while there is no impact in along-track or height accuracy. Also. 
the full affine model produces no improvement in accuracy. 
Residual error patterns similar to those seen in Fig. 2 have been 
encountered with other QuickBird stereo pairs (eg Noguchi et 
al., 2004). 
  
  
  
  
  
  
RMS of /, s RMS value of ground checkpoint 
E No. of GCPs 2 ; NI >; es olo 
RPC Bundle Number of image discrepancies. Units are metres and pixels 
Adjustment Solution C m nis residuals Latitude Longitude Hei 
heckpoints) (pixels) (along track) (across track) eight 
Spatial Intersection None (110) - 29 1.2 4.0 
Shift: Ay, By 2 (108) 0.24 0.67 0.29 0.70 
Shift: Ag, Bo 1 at 1200m (109) 0.24 0.60 0.29 0.87 
Drift: Ag, Bg, Aj, B, 6 (104) 0.21 0.68 0.26 0.72 
Affine: Ao - B» 9 (101) 0.20 0.59 0.25 0.78 
Shift: Ay, By 110 (sigma=2m) 0.24 0.54 0.26 0.54 
  
  
  
  
  
Table 2. Results of RPC bundle adjustments with bias compensation for the IKONOS Geo stereo triplet covering Hobart. 
  
  
  
  
  
  
  
s RMS of 7, s RMS value of ground checkpoint 
ao. No oC GCRS image discrepancies. Units are metres (and pixels) 
RPC Solution (Number of iduais = rie 
= ; resiauals Latitude Longitude ; 
Checkpoints) (pixels) (along track) (across track) Hessht 
Spatial Intersection None (81) - 1:0 (1.3) 8.8 (12) 9:2 (12) 
Shift: As, Bo 2 (79) 0.24 0.73 (1.0) 0.38 (0.5) 0.43 (0.6) 
Drift: Ag, Bo, Aj, Bj 6 (75) 0.21 0.74 (1.0) 0.31 (0.4) 0.41 (0.6) 
Affine: Ào - B» 9 (72) 0.19 0.74 (1.0) 0.34 (0.5) 0.36 (0.5) 
Shift: Ay, Bo 81 (sigma=2m) 0.24 0.70 (0.9) 0.36 (0.5) 0.37 (0.5) 
  
  
  
  
Table 3. Results of RPC bundle adjustments with bias compensation for the Melbourne QuickBird Basic stereo pair. 
27