JES
\thens,
ua.gr
napping
ators or
>tectors,
erties of
ines the
ntation,
'eologic
1 image
llowing
thwell,
n were
Finally,
igated.
C).
Lg. the
, Meer
ul tool
are its
ven in
OUGH
d Hart,
scenes,
:ologic
on and
ipplied
n-Cox
d to a
| main
ch was
ament
; with
uch as
Sobel
1995;
'ell as
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
The semi-automatic and automatic lineament
extraction methods, such as edge following, graph
searching (Wang and Howarth, 1990) and edge
linking operators (standard and modified Hough
Transform) (Cross, 1988; Karnieli et.al., 1996; Fitton
and Cox, 1998), novel edge tracing algorithms (STA,
START and ALERT algorithm) (Koike et.al., 1998).
The design of a knowledge-based system, which
could take the measurable lineament information
(length, aspect) from a DEM into account (Morris,
1991).
n2
LU
1.4 Motivation and aim
From the thorough examination of the literature it is inferred
that computer-assisted methods for the detection of structural
(tectonic) lineaments (namely faults and joints), were
exclusively based on edge enhancement or spatial filtering
techniques (directional and / or gradient filters). These methods
produced edge maps requiring further processing (thresholding
and thinning) for the linear segments to appear with one-pixel
thickness. Optimal edge detectors (e.g. the Canny algorithm)
have already been successfully applied on natural scenes with
quite satisfactory results (binary images with one-pixel
thickness, efficient length and pixel connectivity), and this
makes their application in geologic lineament mapping more
tempting and worth investigating. Furthermore, length is stated
as a crucial statistical parameter for lineament interpretation
and classification and optimal edge detection techniques can
produce segments with sufficient length.
The implementation of the selected edge detectors on a satellite
image of a hydrothermal volcanic field has already been
investigated by the authors (Mavrantza and Argialas, 2003),
and the results were quite promising. Since only early vision
operators such as SOBEL, Laplace and Prewitt have been
applied to date for lineament extraction, further application on a
DEM is promising. A furthermore comparison of edge
detection outputs to those of HOUGH Transform with
evaluation metrics was also required in order to investigate the
applicability of these methods for lineament extraction.
2. METHODOLOGY
2.1 Study area and data used
For the implementation and quantitative evaluation of the
applied edge detection algorithms and the HOUGH transform
for lineament mapping, a geothermal terrain was selected, e.g.,
the Island of Nisyros, which is located in the volcanic back-arc
of the Dodekanesse Complex, Aegean Sea, Greece. Nisyros is a
Quaternary strato-volcano, which is characterized by a well
developed caldera. The major part of this caldera is filled with
dacitic and rhyodacitic domes, while andesitic and pillow-lavas
are also present. The tectonic regime of this area is defined by
two major trends in the NW and in the NE direction
(PENED99, 2000).
The data used in the present work were: (a) Band 5 of a
LANDSAT 5-TM image acquired on August 10, 1991, (b) a
scanned topographic map with a scale of 1:50.000, and (c) the
Digital Elevation Model of the same area with 5m spacing.
791
2.1 Image pre-processing
In the pre-processing stage, a LANDSAT 5-TM satellite image
and the DEM of the same area were geometrically corregistered
with the scanned topographic map of the same area (with a
scale of 1:50.000) and geodetically transformed into the
Transverse Mercator Projection and the Hellenic Geodetic
Datum (HGRS87). Band 5 of the LANDSAT 5-TM image was
selected for the implementation of the edge detection
algorithms, because of its usefulness in lithological and
structural mapping (Woldai, 1995). This band was initially
contrast-stretched using a linear transform so as to achieve a
visually better image for input into the edge detection
algorithms.
For the implementation of the Pratt evaluation metric, an
ancillary ground truth (reference) file was required as input.
This ground truth file contained all the visually interpreted
lineaments from the satellite image (and verified on the
ground), represented with their X, Y coordinates and the total
number of the actual edge points (in an ASCII format file).
2.3 Optimal edge detection algorithms: Implementation
For each algorithm, the combinations of input parameter sets
were selected based on trial-and-error experiments and assessed
(a) using mostly the evaluation measures of Pratt and Rosenfeld
(Abdou and Pratt, 1979; Kitchen and Rosenfeld, 1981) (which
will be explained in section 2.5.), and (b) by evaluating the
optical correspondence to the ground map (Figure 1). Due to
the restricted paper length, only the best results of each
category are presented.
Figure 1: Left: Initial band TM-5 of the Nisyros caldera.
Right: Ground map illustrating faults (yellow),
morphostructural segments delineating the caldera crater
(light green), coastline (white), and road network (red) (For
color, see CD).
For the Canny algorithm, the parameter set with the highest
score of the Pratt evaluation metric (0.4680) (Table 1) was for
o=1.25, Tlow=0.40 and Thigh=0.70.
For the Rothwell algorithm, the following combination of input
parameters was the optimal: o=2.00, T low=6.00 and a=0.90
(Figure 2). This parameter set produced a Pratt metric of 0.4508
(Table 1):