The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
281
Figure 3. Road extraction from optical satellite image (IKONOS pan-
sharpened). Green: correct extraction; red: missing extraction; blue:
false alarms. For more details, see (Mayer et al. 2006)
3.3 Internal Evaluation: Effectiveness and processing time
An automatic object extraction system as exemplified in the
previous section may not be expected to deliver absolutely per
fect results and, thus, for meeting predefined application re
quirements, a human operator must inspect the automatically
obtained results. In order to speed up the time- and cost
intensive inspection, the system should provide the operator
with confidence values characterizing the system’s performance
- a so-called internal evaluation. This information can only be
derived from redundancies within the underlying data or the
incorporated object knowledge. In this context, “object knowl
edge” means knowledge that is purely described by the object
model and not by other external data. Therefore, internal evalua
tion is strongly linked to the extraction process. The results of
internal evaluation are particularly important if the extraction
results are combined with other data, e.g., if they are fused with
results from other extraction approaches or if they are used for
the update of GIS data. On the other hand, they are also very
useful for guiding a human operator during post-editing the
results of an automatic extraction. In practice, however, this is
rarely the case.
In order to achieve a highly independent evaluation, we utilize
“knowledge redundancies” in the form of object properties that
have not been used during the extraction (see (Hinz & Wiede
mann, 2004) for details). Usually, these properties relate to
global characteristics of objects which can be hardly used during
bottom-up object extraction processes. An extracted road net
work, for instance, must be in accordance with some typical
global network characteristics: few connected components, no
clusters of junctions outside urban areas, convenient connec
tions between various places depending on the terrain type etc.
Such properties are used in (Hinz & Wiedemann, 2004) to
evaluate the reliability of portions of the network with a fuzzy-
set theoretic approach.
To allow a quick and effective inspection by a human operator,
the evaluated road segments are displayed in an overview win
dow and categorized into three classes: green (to be accepted),
yellow (to be checked), and red (to be rejected). In addition, the
average quality of the evaluated network and the distribution are
displayed (see Figure 4).
Figure 4. Overview of self-diagnosis. Top: Length-weighted histogram
of evaluation scores (left-most bar indicates mean value = green in this
example), thin vertical lines indicate thresholds between categories.
Bottom: Overview window
Figure 5. Detailed inspection of evaluation results of a specific portion
of the road network. Bars indicate results of evaluation for each
criterion involved ([0 ; 1]).
Whenever a particular road section is sought to be inspected in
more detail, it can be selected in a separate cutout to investigate
the evaluation details (see Figure 5). Based on the visualization
and the quality information, the operator may decide how to
handle a particular road section — whether it should be re
tained, deleted, or edited.