assessed (Table 3). It was expected that the quality of any
given map tended to be lower for a wolf researcher than for a
general user. This is because of the different users’ thematic
similarity weights. As a wolf researcher finds less similarity
between the land cover classes, any mismatch between clas-
sification and reference labels has a more negative impact on
classification accuracy.
T
able
3. W
eighted
O
verall
A
ccuracy
of
C
lassification
of
S
cale
85, 64,
and
71
User
Geometric-only
method
Geometric-thematic
method
Wolf researcher
0.358 (scale 85)
0.390 (scale 64)
General user
0.582 (scale 85)
0.601 (scale 71)
In this paper, the focus is not on the absolute magnitude
of the accuracy of the classifications but the differences that
arise through the use of different segmentations. For the wolf
researcher, weighted overall accuracy reached 0.358 and
0.390 when classifying the segmentations indicated by the
geometric-only method (scale parameter = 85) and the geo-
metric-thematic method (scale parameter = 64), respectively.
For the general user, weighted overall accuracy reached 0.582
and 0.601 when classifying the segmentations indicated by the
geometric-only method (scale parameter = 85) and the geo-
metric-thematic method (scale parameter = 71), respectively.
That is, classification quality increased, especially for the wolf
researcher, with the integration of the user’s specific sensitivi-
ty to misclassification errors into the segmentation assessment.
Traditionally, the comparison of the thematic accuracy of
two maps is based on statistical tests or confidence intervals
calculated around the accuracy measures derived from a test-
ing sample (Foody,2009). For example, if the confidence inter-
vals calculated for the two maps overlap, the relative quality
of one map over the other cannot be considered significant.
In this paper, however, the entire study area (except the area
used for training) was used for accuracy assessment, and thus
the weighted overall accuracy values presented were calcu-
lated effectively from the population. This means that the dif-
ferences found between the maps are not subject to statistical
uncertainty. Therefore, these results confirm the hypothesis
that segmentations indicated as optimal by the geometric-
thematic method lead to higher classification quality than
the optimal segmentation selected using the geometric-only
approach. The increase of classification quality was due to
the inclusion of the users’ view in the assessment of image
segmentation quality.
Conclusions
The results show that by including user-specific thematic
information in an image segmentation quality assessment it
is possible to optimize the segmentation to aid the produc-
tion of a map suitable for that user. The use of metric M
j
to
assess image segmentation quality allowed the segmentations
derived with the scale parameter set at a value suitable for the
Figure 7. Segmentation results: (a) quality assessment of the segmentations based on the geometric-only method and the geometric-the-
matic method for a wolf researcher and a general user (horizontal and vertical dotted lines indicate the estimated optimal results for the
methods), (b) optimal segmentation indicated by M
g
(scale 85, 221 objects), (c) optimal segmentation indicated by M
j
for a wolf research-
er (scale 64, 392 objects), and (d) optimal segmentation indicated by M
j
for a general user (scale 71, 323 objects). The area shown in
(b), (c), and (d) refers to the study area shown in Plate 1. Grey objects in (c) and (d) highlight the objects that are different to those of (b),
while white objects are identical.
458
June 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
