Validation of the Rule Set
All classifications were evaluated systematically to determine
the performance of the rule set in the different areas. In each
area 200 random points were generated which were manually
classified and labeled as Divide, Headwall, Floor, Lake, or “not
part of a cirque.” Where required, additional points were added
so that all classes are sufficiently covered (i.e., at least 15 points).
The manual data set served as reference to validate the
automated classification results. Here, point labels were
compared with underlying classified polygons and as such
a confusion matrix and derived user’s accuracy, producer’s
accuracy, average accuracy, and KHAT statistic (Congalton,
1991) have been calculated. The classification scores served
as measure to evaluate the performance of the classification
rule set in each test area.
Results and Discussion
The Cirque Rule Set
The three main cirque components, and cirque lake are
segmented based on the parameters presented in Table 2, and
classified based on the criteria presented in Table 3. Cirque
lakes are relatively small and divides are relatively narrow
and are therefore segmented using relatively small scale
parameters (respectively, 100 and 225). Elements of a cirque
headwall and floor are generally larger and are therefore best
outlined by larger objects, thus segmented using a larger scale
parameter value (400).
The components are classified based on 13 individual rules,
where cirque headwalls and floors are separated into core
objects, which have a distinct morphological characterization,
and surrounding objects. The surrounding objects may share
characteristics of the core component, but show signs of dis-
turbance or are in a transition towards another landform type
or cirque component. When these neighboring objects border
the core cirque component and share the topographic signa-
ture to a satisfying degree, they are also classified as such.
The rules have been summarized based on geomorphologi-
cally meaningful criteria. More specifically, cirque lakes are
classified with low
NIR
values, due to the high absorption of
near-infrared light by water. A second criterion is a low mean
slope angle. Divides are commonly the highest landforms
in the landscape, which was the motivation to formulate
rules with relative elevation (measured over 51 × 51 meters
and 251 × 251 meters). In addition, divides are normally not
directly located next to cirque lakes which was formulated as
a second rule, in order to prevent misclassifications of local
maxima elsewhere in the area. Cirque floors can be character-
ized with relatively low slope angles, and are usually found
within a certain distance from a cirque divide. Cirque head-
walls are also found near divides, but have generally steeper
slope angles than cirque floors.
Figure 3 presents the topographic signatures of the cirque
components, based on the digitized training samples. The
aforementioned descriptions can be recognized, and thresh-
old values for the classification rules can be extracted. For
example, divides are the only features with a mean relative el-
evation (over 251 × 251 meters) of more than 70 percent. Also,
cirque headwalls and floors can be differentiated based on
threshold slope angle value of 25 to 30 degrees in the training
area. Cirque lakes generally have a mean slope angle of less
than 10 degrees (including adjacent banks), and
NIR
values of
less than 120 units. Figure 3 also shows that often the same
criteria can be used to distinguish the different components,
but that absolute threshold values are different between the
areas. In area 2 and 3, cirque floors have generally steeper
slope angles compared to floors in area 1 and 4.
Performance of the Rule Set
Plate 3 presents the classified objects in the four areas. Table
4 presents the accuracy metrics of the four areas. The rule set
was trained based on Area 1, which also shows the highest av-
erage accuracy of 81 percent; 90 percent of the divides are cor-
rectly picked up, and 76 percent of the total (manually labeled)
divides are classified, which are represented by the user’s and
producer’s accuracies, respectively. Headwall and floor receive
lower accuracy scores; they are more often confused with one
another. While slope angle is the major criterion to differenti-
ate both components, Figure 3 suggests that the topographic
signatures of cirque headwalls and floors partly overlap and
explains the confusion in the classification. All cirque lakes
present in the area have been correctly classified.
Area 2 is located in a comparable altitudinal zone as Area
1, and so a comparable length of glacier occupation can be
T
able
3. O
verview
of
the
C
lassification
R
ules
Rule nr
Component
Classifier
Data range
Value
Unit
1
Cirque lake
Mean NIR
0-255
< 100
DN
2
Mean slope
0-90
< 10
°
3
Divide
Distance to Cirque lake
0-inf
> 400
Pixels
4
Mean REL251
0-100
> 60
%
5
Mean REL51
0-100
> 60
%
6
Floor A (core)
Mean Slope
0-90
< 21
°
7
Distance to Divide
0-inf
< 900
Pixels
8
Floor B (surroundings)
Border to Floor A
YES/NO
YES
-
9
Mean slope
0-90
< 30
°
10
Headwall A (core)
Mean slope
0-90
> 29
°
11
Distance to Divide
0-inf
< 700
Pixels
12
Headwall B (surroundings)
Mean slope
0-90
> 29
°
13
Enclosed by Floor + Headwall
YES/NO
YES
-
T
able
2. O
verview
of
the
O
ptimal
S
egmentation
P
arameters
Component Raster for segmentation Scale parameter
∈
[1-inf]
Cirque lake NIR
100
Divide
Slope & REL51
225
Floor
Slope & TO251
400
Headwall
Slope & TO251
400
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
511
