classified correctly (high user’s accuracy) but many (parts) of
the divides are also missed by the classification (low produc-
er’s accuracy). Yet, cirque floor and headwall components are
dependent on the existence of a cirque divide. This means that
whenever the mean relative elevation values of cirque divides
do not match the training signatures, for example when di-
vides have evolved into more rounded or wider shapes than in
the training area, a large portion of the classification fails.
In summary, the results suggest that to a certain extent,
our rule set is transferable to nearby areas that share common
geological and geomorphological history, if crucial classification
thresholds are adapted to local topographic conditions. Yet, the
rule set fails when geological differences, pre-glacial topography,
or post-glacial geomorphological processes significantly changed
or disguised topographic signatures of cirque components.
As a consequence, the rule set cannot be applied to, for
example, the entire European Alps, to automatically map
all cirque complexes. The question that arises is: “what is
the largest area that can be analyzed with a single object-
based rule set?” There is not a straightforward answer to this
question, as it greatly depends on the diversity of the land-
scape, but should be considered when using Object-Based
Image Analysis for landscape classifications.
Also, data sources and scale are important points of
concern when it comes to transferability of rule sets. For
example, the
MR
segmentation algorithm can work with mul-
tiple gridded data sets, with a different spatial resolution, at
the same time. The scale parameter values are linked to the
gridded data set with the smallest cell size, which is in this
paper is 0.25 m for the
CIR
data set. This means that differ-
ent scale parameter values are required if different resolution
data is used, and as a result, parameter values cannot directly
be compared with, or transferred to those from other studies
which have used different resolution data. In addition, the
value of
LSP
s is different when using different cell sizes. For
example, Eisank
et al
. (2010) used curvature derived from a
10 m DTM as a basis for the extraction cirque divides. With
10 m grid cells, curvature is a meaningful parameter since
the cirque divide is only one or few grid cells wide. In other
words, the curvature is measured at the same scale as the
Area 1
Area 2
Area 3
Area 4
Divide
Headwall
Floor
Cirque lake
0
500
m
0
2,000
m
Plate 3. Final classification results showing the distribution of cirque components in the test areas. Object boundaries have been re-
moved to increase the readability of the maps. Classification metrics are described in Table 4.
T
able
4. O
verview
of
the
C
lassification
A
ccuracy
M
etrics
User’s accuracy [%]
Producer’s accuracy [%]
Overall accuracy [%] KHAT
Divide Headwall
Floor
Lake
Divide Headwall
Floor
Lake
Area 1
90
68
69
100
76
83
63
100
81
0.73
Area 2
84
72
60
100
59
92
70
63
71
0.58
Area 3
93
58
49
67
96
91
76
67
66
0.55
Area 4
86
78
29
0
71
69
89
0
51
0.33
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
513
