Geography Reference
In-Depth Information
records external orientation parameters for each image.
It uses an onboard GPS to acquire position and a series
of infrared sensors to acquire orientation. This data is
crucial to the production of georeferenced mosaics.
We now consider a third approach to image geo-
referencing: image co-registration. This approach offers
another potential alternative to traditional field based
georeferencing methods and can be defined as the process
by which multiple images are aligned into a single coor-
dinate system. For most areas of the world, government
agencies or private sector companies have accumulated
vast databases of low to medium resolution imagery
either from high altitude airborne or satellite platforms.
This imagery is often available to the public and is usu-
ally georeferenced with reasonable accuracy. Therefore,
it becomes possible to use pattern matching approaches
in order to identify features in both an un-georeferenced
hyperspatial image (called the sensed image) and a pre-
existing georeferenced image (called the reference image).
Once the sensed image has been localised within the ref-
erence image, the georeferencing data from the reference
image can be used to calculate the position of the sensed
image. Carbonneau et al. (2010) present an automated
approach which applies this image co-registration strategy
in order to georeference large databases of hyperspatial
imagery. This work uses normalised cross correlation
algorithms to match the pixel intensities within a search
window or the entire sensed image to a specific loca-
tion in the reference image. Normalised cross correlation
has now become an established matching algorithm (e.g.
(Wang et al., 2007) which is implemented in commercial
software such as MATLAB). This algorithm computes a
correlation coefficient
γ
from
1 to 1 between a smaller
template image, t, and a larger image f according to
equation (8.3) (Carbonneau et al., 2010):
−
xy
f
(
x
,
y
)
f
uv
t
(
x
t
−
−
u
,
y
−
v
)
−
γ
(
u
,
v
)
=
xy
f
(
x
,
y
)
t
2
(8.3)
Where x and y are the pixel coordinates in the reference
image f, u and v are the pixel coordinates in the sensed
image t, is the mean of the sensed image and is the mean
of the reference image.
The method described by Carbonneau et al. (2010) rel-
breakies on the metric or sub-metric resolution satellite
or airborne data and has produced encouraging results.
Figure 8.8 shows an example of georeferenced sensed
images (with a white frame added) overlain on base
image for the Ste-Marguerite river. Hyperspatial imagery
with a resolution of 3 cm was automatically georeferenced
to airborne imagery with a resolution of 10 cm. The initial
impression is that the white framed sensed images over-
lay extremely well with the underlying reference image.
Closer inspection at the border of the sensed images
reveals that this overlay is not perfect but nonetheless
the agreement is very close. Carbonneau et al. (2010) cite
an RMS error of 1.7 m. Multiple sources of error affect
these results and weighing the relative importance of
these errors is not straightforward. One factor that was
observed as important is the time elapsed between the
acquisition of the sensed and reference images along
with any potential changes in the scenery between the
f
uv
2
xy
t
(
x
−
−
−
−
u
,
y
v
)
(a)
(b)
Figure 8.8
Example of automated georeferencing. a) Three hyperspatial images, unreferenced acquired from a helicopter platform at
a spatial resolution of 3 cm. b) Hyperspatial image with 10 cm resolution with white frames showing the resulting position as
calculated from an automated georeferencing procedure.
Search WWH ::
Custom Search