2.3 Typical Adjustment Results
Over the last several years there have been several
projects which have had sufficient ground control points
to establish a level of confidence in the reliability of the
adjustment processing. It has been found in most cases
that the RMS residual control point errors after the ad
justment are at the magnitude of the pixel resolution.
Also, the RMS intersection errors for the stereo conjugate
points are usually less than the pixel resolution (Gibson,
1987). These figures can vary depending on the resolu
tion of the imagery and on the quality of the ground con
trol coordinates. For example, if the uncertainty in the
control point coordinates derived from maps (typically
1:10000 or 1:20000 scale), is greater than the pixel reso
lution, it has been found that the residual errors from the
adjustment tend to reflect the errors in the control point
coordinates rather than the pixel resolution.
3 Ortho-image Generation
One of the desired end-products from the geometric cor
rection processing has been mosaiced ortho-images. It
has been possible to demonstrate that such products may
be generated with relative ease and without requiring hu
man intervention to stitch overlapping strips of imagery
together.
3.1 Terrain Height Correction
One of the necessary elements for the production of ortho-
images is accurate terrain height data and the ability to
incorporate that data into the image resampling process.
The terrain height information is incorporated into the
resampling algorithm by computing the intersection of
the pixel pointing vector (zj or xl in Figure 1) for each
pixel with the surface of the earth which in this case is
represented by a digital elevation model. The intersection
calculations are iterative, based on estimating a scale fac
tor for the pointing vector to extend it to the intersection
point. The scale factor is initialized as the difference be
tween the flying height and the average terrain height,
normalized for a unit focal length lens. The elements of
the pointing vector are multiplied by the scale factor and
the coordinates of the vector are computed. The height
component of the vector coordinates is compared with
the terrain height value for the same horizontal location.
The scale factor is then adjusted based on the ratio of
the height difference and the original normalized flying
height. Once the algorithm has been initialized in this
fashion for each scan line of imagery, the scale factor may
be passed from pixel to pixel with only minor corrections
needed for each. Figure 2 shows the results of comput
ing two ortho-image products and then merging them to
gether in the common overlap region with a simple linear
weighting algorithm. The details of the imagery sample
are as follows:
• The imagery covers 2 km by 2 km of Cambridge,
Ontario.
• The flying height was 1400 metres which resulted in
a ground resolution of 1 metre.
• The overlap region of the two flight lines is approx
imately 100 metres wide. The extent of each of the
separate images is indicated in Figure 2.
• The terrain height variation in the area covered has
a range of 40 metres. The misregistration between
the images for this amount of height variation would
have been up to 28 metres if the terrain height cor
rection had not been applied.
• The terrain height data was supplied by the On
tario Ministry of Natural Resources as spot eleva
tions which were subsequently interpolated to a reg
ular grid for the use in the resampling processing.
• The two input images were radiometrically balanced
and adjusted to have the same average intensities
before they were combined. This was done by a sim
ple empirical method of deriving a scale factor and
an offset function which varies across each image to
adjust the radiometric histograms to have matching
statistics.
The imagery was generated as part of a project sponsored
by the Ontario Centre for Remote Sensing (OCRS) while
the authors were employed by Rem/Sense Mapping Tech
nologies Inc.,(R/SMT), (R/SMT, 1989). An initial check
of the quality of the corrected imagery was performed by
overlaying the Ontario 1:10000 base map of the area. A
good agreement of features was noted and it was imme
diately obvious where new features existed that were not
present on the map, however a more formal evaluation
which was to follow has not been completed.
3.2 Further Development
Although on the surface, the combined imagery appears
to look seamless, there are several areas where there is
some blurring of features in the region of overlap. This
is partially due to the fact that the terrain height data
was obtained from another source and also that it did
not contain data for ground cover or buildings. It is felt
that the best results in generating ortho-images will be
obtained when they are computed using terrain height
derived from the MEIS imagery itself so that there will
be a greater degree of consistency between the features in
the imagery and in the terrain height data. The current
software effort for MEIS imagery is therefore centered on
the development and refinement of an automated terrain
height extraction package using the forward and aft stereo
channels of MEIS. The software is being developed on a
computer graphics workstation in order to facilitate the
rapid display of both imagery and derived height prod
ucts. One of the key features of the system will be a