You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects
Baltsavias, Emmanuel P.

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
mentioned and briefly discussed only for reasons of
completeness. Although it is widely acknowledged that such
methods do not at all fulfil serious accuracy requirements, they
are still applied by many users in their remote sensing
applications. Emphasis, however, will be put on methods which
allow a precise coregistration or geocoding of images in order
to provide optimum datasets for subsequent fusion-based
Suitable approaches to precisely fuse data from a geometric
point of view are:
• parametric methods, being based on the consideration of
sensor-specific mapping models following photogrammetric
techniques to transform an image to a map or vice versa.
• coregistration techniques being independent of any sensor-
specific imaging parameters. Here, registration can be
performed in image or map geometry based on one and the
same approach.
Pre-operational and operational image coregistration and
geocoding have been the topic of various research activities
related to remote sensing and GIS at the Institute of Digital
Image Processing. Illustrative examples of these activities are
included to show the potential as well as the drawbacks of the
various methods being described.
A major restriction for the application of multisensor and
multitemporal image data for monitoring applications is the
position accuracy when overlaying the different images. In
order to demonstrate the necessity of high location accuracy a
simulation has been performed to analyse displacement effects
for forest monitoring applications.
The simulation was carried out by shifting a forest type
classification, for which three different forest classes have been
derived from geocoded Landsat TM image data by a magnitude
of 0.5 pixel each to east and north (25 m pixel footprint).
According to many reported Landsat TM geocoding results, 0.5
pixels can be assumed to be the optimal realistic geometric
accuracy that can by reached by parametric methods. The
classification error resulting from the simulated geocoding error
can then be estimated by comparing the shifted classification
with the original one. The class changes due to the geocoding
error for the three forest types are listed in Table 1.
In case of no displacement, or of no effects of the displacement,
the change between the classes being compared should be 0%.
In other words, the comparisons between e.g. coniferous and
coniferous should amount to 100%. In the simulated case,
however, only 86.74% of the coniferous pixels still belong to
the "coniferous"-class after the displacement. This means that
the displacement leads to an error of 13.26%. This error splits
up into the coniferous pixels, which are incorrectly displaced
into the class "mixed forest" (11.17%) and the coniferous
pixels, which are incorrectly displaced into the class "deciduous
forest" (2.09%).
Table 1. Classification changes caused by a geocoding error
of 0.5 pixel to east and north.
Changes of coniferous forest
coniferous => coniferous
86.74 %
coniferous => mixed
11.17 %
coniferous => deciduous
2.09 %
Changes of deciduous forest
deciduous => coniferous
deciduous => deciduous
69.07 %
deciduous => mixed
10.75 %
Changes of mixed forest
mixed => coniferous
8.63 %
mixed => deciduous
mixed => mixed
74.38 %
By shifting the forest type classification by an assumed
geocoding error of 1 pixel each to east and north, the values in
Table 2 were obtained.
Table 2. Classification changes caused by a geocoding error
of 1.0 pixel to east and north.
Changes of coniferous forest
coniferous => coniferous
83.55 %
coniferous => mixed
13.96 %
coniferous => deciduous
2.49 %
Changes of deciduous forest
deciduous => coniferous
25.98 %
deciduous => deciduous
60.07 %
deciduous ==» mixed
13.95 %
Changes of mixed forest
mixed coniferous
10.00 %
mixed => deciduous
22.48 %
mixed => mixed
67.52 %
These simulations show that a pixel-accuracy comparison of
signatures or classification results cannot be recommended for
forest monitoring applications, since changes in forest classes
occur in many cases in subpixel dimensions. This is particularly
true for forest types, which are characterised by heterogeneous
spatial distribution. This can be demonstrated by the
correspondence values of the class "mixed forest", for which
correspondence values of only 74.38% (shift of 0.5 pixel to east
and north) and 67.52% (shift of 1.0 pixel to east and north)
could be noticed.
These classification changes are just one indicator for the
necessity of high-accuracy image geocoding. Methods to fulfil
this demand are discussed in the following section.
3.1. Polynomial Rectification
Straightforward polynomial rectification and image warping
tools have been used to register and/or rectify image data for