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
Fig 2. Workflow within a common classification process.
A main aim of our project is to create a training set database
once (for the first scene or date). In fact, we use the outlines
of the training sets and apply them to other scenes. The main
advantage is, that we don’t run into the pitfall that the same
class has variations in spectral behaviour because of temporal
and solar differences - this would definitely occur, if we just
transfer collected signatures which belong to the first scene.
Ideally, training data should be collected from multiple
sources. Our training sets consist mostly of ATKIS data,
Topo-Maps and ground truth data. This data should be used,
on one hand for extracting training areas, and on the other for
evaluation of the results. In principle, we could introduce
errors, if we use training set boundaries of different sources
in different scales (spatial/time). To avoid this, we use the
one database with the best accuracy, which fits spatially and
temporally to the image acquisition. If this can’t be carried
out, there are three possibilities to handle this problem:
Fig. 4. Process of signature extraction.
(1) we could use the intersection of both databases by
applying overlay techniques, (2) we could use the union by
utilising join processes, and (3) we could use both databases
separately. We avoid the 3 rd method because it causes a
statistical overweighting in the classification process. After
each GIS process flow, we perform a histogram check for
each new signature. If we get a valid signature, we keep the
file, otherwise it will be excluded from our database (Fig. 4).
Within this workflow, our strategy is not to detect areas
where changes have occurred but those without changes. If
changes have occurred during the process of transferring
training set boundaries to a new scene, we have to modify the
geometry of the signature boundaries by using GIS
capabilities (Fig. 4), such as buffering (dilate, shrink), until
the signature becomes unimodal. We do not make use of
thresholds between peaks of multimodal histograms to
separate the different signatures, because this technique is not
practical for automation. In the worst case, multimodal
signatures have to be eliminated. A problem is to treat
classes, which are unimodal in the first and second time stage
Fig. 3. Overview of a modem workflow within an integrated GIS and Remote Sensing environment.