PE&RS June 2015 - page 507

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING June 2015 507
Rule Set Transferability for Object-Based Feature
Extraction: An Example for Cirque Mapping
Niels S. Anders, Arie C. Seijmonsbergen, and Willem Bouten
Abstract
Cirques are complex landforms resulting from glacial erosion
and can be used to estimate Equilibrium Line Altitudes and
infer climate history. Automated extraction of cirques may
help research on glacial geomorphology and climate change.
Our objective was to test the transferability of an object-based
rule set for the extraction of glacial cirques, using lidar data
and color-infrared orthophotos. In Vorarlberg (W-Austria),
we selected one training area with well-developed cirque
components to parameterize segmentation and classification
criteria. The rule set was applied to three test areas that are
positioned in three altitudinal zones. Results indicate that the
rule set was successful (81 percent) in the training area and a
higher situated area (71 percent). Accuracy decreased in the
two lower situated test areas (66 percent and 51 percent). We
conclude that rule sets are transferable to areas with a com-
parable geomorphological history. Yet, significant deviation
from the training area requires a different extraction strategy.
Introduction and Background
Cirques are complex landforms resulting from glacial erosion.
Ivy-Ochs et al. (2008), regard a cirque as “a landform erod-
ed by a glacier positioned in isolated niches in mountains.”
Three main cirque components are commonly recognized (see
Plate 1): (a) an upper semi-circular “cirque divide” bounded
by (b) steep surrounding slopes or “cirque headwall” and (c)
a relatively flat lower surface or “cirque floor” bordering the
headwall. In many cirques, three sub-components related
to (de)glaciation are found on the cirque floor: (1) a ‘cirque
threshold’ developed in bedrock, (2) (a) cirque moraine(s), rep-
resenting recessional phases of the former cirque glacier, and
(3) a “cirque lake” often located in the lowest part of the valley
floor in between cirque moraines and cirque threshold. In
certain areas (e.g., western Austria) cirque glaciers were active
during the waxing and waning stages of former glaciations (De
Graaff et al. 2007), their existence linked to former Equilib-
rium Line Altitudes (ELAs). The identification of cirques and
cirque components can help in the reconstruction of former
ELAs and contribute to studies concerning climate change.
Modern remote sensing data sets such as lidar (Light Detec-
tion and Ranging) elevation data, aerial photography and sat-
ellite imagery are powerful sources for the identification and
analysis of (glacial) landforms such as cirques (Schneevoigt
et al. 2008; Smith et al. 2006; Smith and Pain, 2009). Manual
identification and digitization however can be time-consum-
ing, especially when larger study areas are of interest. Geospa-
tial Object-Based Image Analysis (GEOBIA) is a promising tool
to analyze (glacial) landscapes (e.g., Saha et al., 2011). With
GEOBIA, image grid cells are clustered to form objects that can
be classified based on internal grid cell values, shape charac-
teristics, and topological relations. The clustering of grid cells
makes GEOBIA particularly useful when used with high-resolu-
tion data sets. In this context, detailed morphometric charac-
teristics of cirque components can be addressed in rule sets to
(semi-) automatically map the distribution of cirques.
Transformation of conceptual semantic models for cirques
into rule sets was emphasized by Eisank (2013) as a necessity
to develop transparent workflows in order to improve objec-
tivity and transferability of rule sets. This way, it is possible
to further automate the creation of maps containing different
landscape features (Drağuţ and Blaschke, 2006). Research on
cirques that use a GEOBIA approach are scarce and generic
rule sets for detection of cirques do not yet exist. Eisank et al.
(2010) and Ardelean et al. (2011) both used mean curvature
derived from digital elevation data for the segmentation of
cirques, focusing on the upper divides. Altitudinal thresholds
were used in combination with specific context rules as input
for segmentation and classification. Altitudinal boundary con-
ditions however, prevent the transferability of rule sets to other
areas that are positioned higher or lower in the landscape than
the training area. Anders (2013) has developed rule sets that
can successfully extract (glacial) landforms in mountainous
terrain from DEMs by using selected combinations of Land Sur-
face Parameters (LSPs). By using multiple LSPs, classification
rules may be formulated without altitudinal boundary condi-
tions so that they are potentially transferable to other areas.
A possible concept that can be used to improve cirque classi-
fication is by taking into account the cirques degree of deviation
from a “textbook example” or prototype landform (Evans, 2012),
thus addressing a cirque’s potential polygenetic history. This
degree of deviation can result from a difference in length of the
glaciation phase producing well or less developed cirques, or
can be addressed to differences in the activity of post-glacial
processes. Post-glacial landscape modification, for example
fluvial erosion and accumulation may thus also disguise the
boundaries of the main components of a cirque (Plate 1).
Our objective is to develop and test an object-based rule
set that decomposes lidar DEMs into the three main cirque
components: divide, cirque headwall, and cirque floor, and
into the subcomponent cirque lake. The transferability of the
rule sets will be evaluated for areas that are in different states
of development. Our hypothesis is that rule sets to classi-
fy cirques can successfully be transferred to regions with a
similar glacial and post-glacial history, but will perform less
well in areas with a different geomorphological history. Based
on the results found, this paper addresses a discussion on the
transferability of GEOBIA rule sets in the extraction of geomor-
phological features from digital elevation data.
Niels S. Anders is with Soil Physics and Land Management,
Wageningen University, P.O. Box 47, 6700 AA, Wageningen,
The Netherlands ( niels.anders@wur.nl).
Arie C. Seijmonsbergen and Willem Bouten are with the Insti-
tute for Biodiversity and Ecosystem Dynamics, Computational
Geo-Ecology, University of Amsterdam, P.O. Box 94248, 1090
GE, The Netherlands.
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 6, June 2015, pp. 507–514.
0099-1112/15/507–514
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.6.507