PE&RS June 2015 - page 509

the Training the Rule Set Section. Classification metrics such
as user’s and producer’s accuracy (Congalton, 1991) were
used to quantify the performance of the rule set in the specific
areas, see the Validation of the Rule Set Section for details.
The data sources consisted of (a) a lidar data set from 2011
from which the raw point data were acquired, filtered, and
interpolated using linear squares interpolation into 1 m reso-
lution lidar raster DTMs by TopScan(http://www.TopScan.
de/); and (b) a Color Infrared ( CIR) orthorecitified mosaic with
0.25 m resolution produced from aerial photographs acquired
in 2001. The CIR data set is comprised of bands in the near-
infrared (NIR), green and red light. Based on the lidar dataset
several LSP s have been calculated using ArcGIS® 10.2 and
python/GDAL: Shaded Relief, Slope Angle, Relative Elevation
(REL, i.e., percentage of grid cells lower than a center grid cell
in a given moving window), and Topographic Openness (Yo-
koyama et al., 2002). The main steps of the analysis are based
on the approaches of Anders et al. (2011).
The GEOBIA workflow is presented in Figure 2 and was car-
ried out using eCognition® Developer 8.8. The numbers in the
flow chart refer to processing steps which are explained in the
following section.
Training the Rule Set
The Gargellen-West area was selected as training area because
of the well-developed glacial features and the small size of
the study area. Based on visual interpretation of the lidar LSPs
and CIR imagery of the Gargellen-West area, three training
samples per cirque component were manually digitized (Step
2 in Figure 2; see Plate 2). Only two training samples of cirque
lakes were digitized due to the absence of more representative
features in the area.
The training samples were used to calculate frequency dis-
tribution matrices of LSP values within the enclosed polygon
boundaries (Anders et al. 2011; Anders, 2013, step 3 in Figure
2). The selection of LSPs used for the frequency distribution
matrices was based on expert knowledge, so that unique prop-
erties of the three main components for this particular land-
form are captured. For cirque lake NIR values were used, as NIR
images clearly show water bodies due to high absorption of
the NIR light wavelengths. Slope Angle and Relative Elevation
(measured within a 51 m × 51 m window) were used to create
frequency distribution matrices of cirque divides, because
divides are only found high in the landscape and landform
units/elements are well separated by slope units. Slope Angle
and Topographic Openness (measured within a 251 m × 251 m
window) were used to create frequency distribution matrices
of the cirque floor and headwall components, because both
components can be differentiated by slope angle (relatively
steep slopes at the cirque headwalls, and relatively gentle
slopes on the cirque floors) and Topographic Openness clearly
depicts boundaries between different landforms (Anders,
2013; Anders et al. 2013). The window sizes were manually
selected to provide the required detail and texture for the scale
of the landforms (i.e., Topographic Openness) or to provide the
required regional information (i.e., Relative Elevation).
Subsequently, eCognition Developer (8.8) was used to cre-
ate sets of image objects using the multi-resolution (MR) seg-
mentation algorithm described by Baatz and Sch pe (2000).
The MR segmentation algorithm is a region-growing procedure
where neighboring grid cells and objects are merged (Baatz
and Sch pe, 2000). The merging is rejected if the standard
deviation of the objects before and after merging is higher
than a given threshold. This threshold is set by a “scale”
parameter. The theoretical range of scale parameter values is
from 1 to infinity, where a value of 1 produces objects of one
or few grid cells, and a large value results in the clustering of
all grid cells into a single object. Due to the nature of the MR
segmentation algorithm, the actual relation between the scale
parameter value and object size depends on the spatial resolu-
tion and standard deviation of values in the data set. Multiple
sets of image objects were created with different scale param-
eters (step 4 in Figure 2), which greatly affects the number of
grid cells being clustered thus the size of individual objects.
The sets of objects were created with scale parameter values
of 10, 15, 20, 25, 30, 35, 40, 45, 50, 75, 100, 125, 150, 175,
200, 225, 250, 300, 500, and 999, respectively. This range
was found large enough so that the optimal object size could
be identified for each cirque component. The MR segmenta-
tion algorithm also requires the definition of a “shape” and
Figure 2. A schematic overview of the stratified GEOBIA workflow.
Table 1. Subset Area Characteristics
Area
nr Area name
Mean elevation
[m.a.s.l.]
Area size
[km2]
Altitudinal
zone
1 Gargellen-West 2285 1 +
2 Hochm derer 2421 15 ++
3 Zitterklapfen 1676 29 -
4 Winterstaude 1349 20 --
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING June 2015 509