thematic quality of under-segmented objects mixing these two
classes will stay relatively high for a wolf research but not for
a general user. Generally, the wolf researcher found less simi-
larity between the land cover classes as compared to a general
user (mean of all weights –
w
is ~0.47 and ~0.53, respectively).
Segmentation and Assessment
The two
LISS-III
images were used to derive a series of segmen-
tations employing the multi-resolution algorithm implement-
ed in GeoDMA software (Körting
et al.
, 2013), version 0.2.1,
which is based on the popular algorithm of Baatz and Schäpe
(2000). This algorithm is a region-based algorithm that uses
the spectral and shape characteristics of objects in the deriva-
tion of the segmentations. To balance the relative importance
of each of these variables, the algorithm uses two criteria,
color and shape; both range within the interval [0 - 100] and
color = 100-shape. In addition, the shape variable depends
on two other parameters: compactness and smoothness,
also ranging within the interval [0 - 100] and compactness =
100-smoothness. However, the most influential parameter in
the analysis is scale, which is a threshold of heterogeneity
allowed within objects as regard color and shape. The bigger
the scale parameter value, the larger the spectral/shape het-
erogeneity allowed within objects, typically resulting in the
production of fewer and larger objects.
Since the scale parameter was the greatest influent on the
algorithm performance, only the scale parameter was changed
to produce a set of segmentations. All the remaining param-
eters were kept constant at 50. Segmentations were produced
with the scale parameter set from 50 to 100 with steps of 10.
Using values at the extremities of this scale range produced
over-segmented results (objects were noticeably smaller than
CLC2006’s land cover polygons) and under-segmented results
(objects were noticeably larger than CLC2006’s land cover
polygons), and hence optimal results should be achieved
through the use of intermediate scale values.
From the series of six segmentations derived, the scale
value for which the metric M
j
was closest to zero for each
user, was estimated by linear interpolation. The estimated
optimal segmentation for each user was then generated using
the relevant scale parameter and used for land cover clas-
sification. Similarly, the scale whose M
g
was estimated to be
closest to zero was also generated to represent the segmenta-
tion derived following the standard geometric-only approach
to the assessment of segmentation quality. In total, therefore,
three segmentations were selected for classification and so
three land cover maps produced.
Classification and Assessment
The three segmentations selected were classified to produce
land cover maps. This procedure was performed within GeoD-
MA software (Körting
et al.
, 2013). The classifier used was a
decision tree based on the algorithm C4.5 (Quinlan, 1993).
The classifier was trained with spectral data and thematic
information associated with the randomly located square
patches (Plate 1). The spectral data used was the mean and
standard deviation of image digital numbers of each spectral
band contained by the polygons of the
COS2007
for each land
cover class. The mean value provides a measure of central
tendency whereas standard deviation provides a measure of
variability (texture).
The classifier was applied to the
LISS-III
data using the three
segmentations derived. Since the same training data was used
through, the differences in the accuracy of the resulting clas-
sifications arises from the segmentations used.
Classified segmentations were overlaid with the CLC2006
data not used in the assessment of image segmentation quality
to build a confusion matrix. Several measures of thematic
classification accuracy can be calculated from a confusion ma-
trix (Liu
et al.
, 2007). Standard measures used as kappa, and
overall accuracy were not used because they weight all types
of error equally. In this paper, a weighted version of overall
accuracy was used. Specifically, the measure was calculated
as
∑
v
ij
p
ij
/
∑
v
ij
p
ij
, where v
ij
is the agreement/disagreement
weight between the j
th
class of the reference and the i
th
class of
the classification, p
ij
is the area of the j
th
class of the reference
classified as the i
th
class of the classification, and i = j in the
numerator. Weights v
ij
represent either agreement or disagree-
ment between the classes and were calculated for each user
based on their respective thematic similarity matrix (Tables 1
and 2). Specifically, overall accuracy considers the diagonal
entries of the confusion matrix as contributions for accuracy.
Therefore, diagonal weights represent agreement between the
classes and were calculated as v
ij
= w
ij
(i = j), where w
ij
is the
thematic similarity weight of either Table 1 or Table 2. On the
contrary, the off-diagonal entries of the confusion matrix are
considered as contributions for error. Therefore, off-diagonal
weights represent disagreement between the classes and were
calculated as v
ij
= 1–w
ij
(i
≠
j).
Results and Discussion
The geometric-only method indicated the segmentation
derived with the scale parameter set at 85 as optimal (M
g
~0)
whereas the geometric-thematic method indicated scales 64
and 71 as optimal (M
j
~0) for the wolf researcher and general
user respectively (Figure 7). This means that according to M
j
more, and smaller objects, should be produced ideally for
both users than as indicated by M
g
.
As expected, the geometric-thematic method appears to
favor over-segmentation as compared to the geometric-only
method due to the inclusion of the
TSI
; the
TSI
is sensitive to
mixing of land cover classes within objects, and thus its value
decreases when under-segmentation occurs. Therefore, over-
segmentation is preferred to prevent different land cover classes
from being mixed in the same object, especially if the classes
involved are thematically different from the user’s perspective.
The geometric-thematic method indicated different
optimal results for the two users. A higher level of segmen-
tation was indicated for the wolf researcher. The reason for
this situation is that this user finds less similarity between
the land cover classes in the study area than a general user
(~0.47 in contrast with ~0.53). In other words, the user needs
particular land cover classes to be mapped very accurately
because of their relevance to the study of the wolf popula-
tions in northern Portugal (e.g., forest). Therefore, the user
is relatively intolerant to segmentation errors involving the
most relevant classes (e.g., forest) and less so for others. The
“over-segmented” result delivered by scale 64 gives the user
further reassurance that classes weighted as of little similar-
ity are not mixed within “under-segmented” objects as the
objects tend to be smaller. This is also true for a general user.
However, thematic similarity between the classes for this user
is in general higher (~0.53), and therefore an intermediate
scale (71) was selected for this user, between that for the wolf
researcher, and that indicated by M
g
.
Critically, the standard, geometric-only approach provided
only a limited assessment of the segmentation quality and
did not recognize the user dependent variation in the weight
of segmentation errors. Thus, the geometric-only method
indicated scale 85 as being optimal for whatever the user in
question. That is to say, the geometric-only approach repre-
sents users to whom segmentation errors all have the same
importance regardless of the classes involved.
The three segmentations derived with scale 85, 64, and 71
were classified. The weighted overall accuracy of each was
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
457
