oar
ata
ue
gh
Ire
19
to
1d
e,
Edge Following as Graph Searching
In this stage, Edge Following as Graph Searching
(EFGS) algorithms (Wang and Howarth, 1987,
1989; Wang, 1993) are applied to the combined
channel to identify edges in the image. This
involves three major steps. First, an "edge
operator" is used to obtain the magnitude and
direction of an edge. This is based upon
determining the locations where maximum
changes in digital values occur and what the
directions of these changes are. The procedure
can be applied to detect edges as well as light
lines or dark lines, depending upon the
appearance of the lineaments in the image.
Second, the starting points for the edges are
identified. The starting edge point selection
algorithm identifies the most prominent edges
and it is found that a large number of these
correspond to parts of lineaments. Finally, EFGS
is used to trace all the edges on the image. A
graph can be formed from each starting edge
point. Each arc in the searching graph is
associated with a cost. The cost is a function of
edge magnitude and direction, as well as the
tracing direction. The EFGS algorithm may be
summarized as follows:
(1) Accept all starting points as edge elements.
(2) If there are no more starting points, stop.
Otherwise, assign the next starting point as
the current node.
(3) If there is no neighbor in front of the tracing
direction of the current node, go to Step (2).
(4) Compute the cost for the arc connecting the
current node to each of its neighbors in the
tracing direction. Accept the minimum cost
neighbor as an edge element. If no neighbor
is accepted, go to Step (2) Otherwise,
assign this neighbor as the next current
node and the direction of the arc from the
previous current node to this node as its
tracing direction. Go to Step (4).
Hough Transform
The edge image obtained from the EFGS
algorithms contains edge pixels with a value of 1
and background pixels with a value of 0. The
problem in lineament detection is to locate the
presence of groups of collinear or almost
collinear edge pixels. The problem can be solved
to any desired degree of accuracy by testing the
lines formed by all pairs of points in the picture.
751
However, the computation required for n pixels
is approximately proportional to n? , and may be
prohibitive where n is large. Hough (1962)
proposed an interesting and computationally
efficient procedure for detecting lines in
pictures. It has become known as the Hough
Transform. The main advantages of the Hough
Transform are that it is relatively unaffected by
gaps in lines and by noise. In this paper, Hough
Transform is applied to detect straight lines
which represent geologic lineaments on the
multispectral satellite images. The Hough
Transform method described by Duda and Hart
(1972) is used and modified for lineament
detection. The procedure involves use of the
Hough Transform, finding of local maxima,
application of an inverse Hough Transform and
straight line profile analysis (Wang and Howarth,
1990). Geological lineaments are frequently
discontinuous on the original image. However, the
segments lying on the same line can be joined
together if they are close enough to each other.
In other words, certain gaps are allowed on a line
in this straight line detection algorithm. The size
of the gaps allowed for can be assigned by the
user of the LINDA system (Wang, 1993).
RESULTS
To evaluate the performances of the lineament
detection algorithms on multi-band remote
sensing images versus on single band images, a
study area of part of the Canadian Shield near
Sudbury, Ontario has been selected. It includes
part of the exposed Grenville Province. The
dominant rocks in this area are middle
Precambrian metasediments and an anorthosite
suite of intrusive rocks. Structural control is
suggested by the preponderance of elongate lakes
confined to a few orientations. A subscene of a
Landsat 5 Thematic Mapper (TM) image with
seven channels over the study area is obtained.
The procedure for lineament detection, including
preprocessing, edge detection and Hough
transform, has been applied to each of the seven
Landsat TM channels. From a comparison of all
the seven TM bands, it is observed that Band 4
image (near infrared) displays the lineaments
most clearly (shown in Figure 1) and that Band 4
image also results in the best single-channel
lineament detection (shown in Figure 2) among all
seven Landsat TM bands. However, from Figure
2, it can be seen that not all lineaments are
detected.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996