S are
AS is
, the
indle
m or
ions
high
ingle
nage
aries
nage
nally
) get
TM
> the
d by
ation
ates
nage
e of
the
TM
vhen
d on
jatic
E in
Jeffrey Shan
Tab.4 RMSE at the given 25 points
Blunder threshold | No. of points | Root mean square errors (RMSE, m)
in resection For resection X Y 7
1.8 85 0.878 0.938 0.0667
1.90 262 0.732 0.833 0.0381
2.00 454 0.735 0.7035 0.0594
Tab.5 RMSE at manually measured points
Blunder threshold | No. of points | Root mean square errors (RMSE,m) | No. points for
in resection For resection | X Y Z Statistics
1.8 o0 85 1.138 1.177 0.170 61
1.9 o0 262 1.071 1.158 0.231 60
2.000 454 1.028 1.076 0.196 58
Another evaluation is done by using the manual measurements as check points. After exterior orientation parameters are
obtained through automatic image matching, mono-image intersection is performed for all check points to calculate
their ground coordinates, which are then compared with the ones derived from measurements on the orthoimage. The
RMSE is listed in Tab.5. Since the measurement is done manually on the screen in mono mode, its precision is expected
no better than 1 pixel and several blunders have been detected and deleted. Considering this fact, the RMSE for
planimetric coordinates listed in Tab.5 actually has two components, errors caused by space resection and by
measurement errors on check points, namely
si ei ”
(RMSE) “XY = S resection tS check points (5)
Taking the average of all (RMSE )'xr for X and Y in Tab.5 and counting S poi, as 0.8m (1 pixel), the actual
affection of space resection on planimetric coordinates on the ground is then estimated as oresection = (1.23 — 0.82) =
0.77m (< 1 pixel). It should be noted that this value is also an accuracy estimation for the new updated orthoimage. As
the elevation of ground points is virtually determined via DTM interpolation, its accuracy (RMSE); is solely dependent
on terrain slope and the planimetric accuracy, namely
(RMSE), = A(F, + F,’) (RMSE), (6)
where Fy and Fy are terrain slopes at X and Y directions respectively. For flat area as covered by this test image, the
slope angle is less than 5 degree, therefore the elevation accuracy is estimated better than /.474*tan5 #J1.23 =
0.14m or 0.003% flying height.
5 CONCLUDING REMARKS
An approach for automatic exterior image orientation is proposed based on DTM and orthoimage in current GIS
database. The approach uses automatic feature extraction and matching within image patches in aerial image and
orthoimage. Large amount of feature points (ca.2000) is included in the bundle adjustment, which may reach a precision
of 1/3 pixel for orientation parameters due to the large redundancy. To evaluate the quality of this approach, a novel
mono-image intersection algorithm is developed. Analyses and comparison based on results from manual measurements
indicate that the accuracy of planimetric coordinates of ground points, determined by mono-image intersection, is better
than 1 pixel. This is also the accuracy estimation for the new updated orthoimage. The accuracy of elevation is solely
dependent on DTM interpolation. Further research will be focused on linear features detection and using them along
with point features in GIS to automate the exterior orientation and orthoimage update procedure.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 835