CURVILINEAR NETWORK EXTRACTION FROM REMOTELY SENSED ... - CVL

MVA'94 IAPR Workshop on Machine Vision Applications Dec. 13-15, 1994, Kawasaki

CURVILINEAR NETWORK EXTRACTION FROM REMOTELY
SENSED IMAGES

Mark R. Dobie, Paul H. Lewis and Mark S. Nixon

Department of Electronics a n d C o m p u t e r Science

University of Southampton, England, S0171BJ.

email [email protected]

ABSTRACT still of wide interest, not only in the analysis of
remotely sensed images.
We describe a new approach t o the extraction of
networks of narrow curvilinear features such as MINIMUM COST PATHS
road and river networks from remotely sensed
images. The approach begins with the iden- In a recent paper[4] we describe an approach
tification of points in the image with a high t o the problem of curvilinear feature extrac-
probability of being on the network. In the tion which uses a combination of selective win-
second stage the broad topology of the feature dowing and minimum cost path techniques to
is extracted using a minimum spanning tree and delineate a minimum cost path between spe-
in the final stage a novel cost minimisation ap- cified points on a required feature. The cost
proach is used t o refine the linear sections of function may be based on image properties and
the network so that they follow the underlying path curvature in a manner similar t o that
structure more closely. used by Kass et a1[5] in their active contour
model formulation, often referred t o as a snake.
INTRODUCTION However, our approach finds a global rather
than a local solution t o the path cost minim-
Curvilinear features are one of the most com- isation, constrained only by a search window
mon feature types found in digital images, specified around the feature of interest. It uses
e.g. arteries in medical images, road and river a minimum cost path algorithm based on an
paths in satellite images and characters and extension to D'Esopo's method[6].
other line work in engineering drawings. Even
if curvilinear features are not present in the raw The method is optimal and an important dis-
image, they are often generated by gradient op- tinction may be drawn between optimal search
erators as an early stage in image analysis and algorithms which find the global minimum cost
their subsequent extraction is an essential task path and search algorithms which use heurist-
in many image interpretation procedures. ics t o reduce the search space. For example,
Montanari [7] introduced dynamic program-
Networks of connected curvilinear features ming for optimal curve detection with respect
also occur widely in many images. In this pro- t o a particular figure of merit and Martelli [8]
ject we have developed a new technique for the introduced heuristic search to detect edges and
automatic extraction of explicit (vector) rep- contours. Since these early papers, many au-
resentations of road and river networks from thors have applied heuristic search techniques
remotely sensed images. These are of partic- t o the problem of boundary and line detection.
ular interest to a range of users, typically for Optimal search times have been significantly
subsequent manipulation in geographic inform- less attractive than search times when powerful
ation systems. In many instances, the lack of heuristics are available, and their computation-
reliable software tools has necessitated the la- ally intensive nature has reduced their popular-
bour intensive process of manual digitisation. ity.
Although there have been several approaches to
curvilinear network extraction[l, 2, 31, methods Although our optimal approach overcame
which seek t o minimise human interaction are some of these difficulties, it had the disadvant-

age that it only extracted a single path, spe- using a minimum spanning tree for network ex-
cified by its end points or by end points and traction was first proposed by Fischler et a1[3] ).
intermediate points indicating the path roughly,
in order to accelerate the search. If a network of The minimum spanning tree algorithm treats
roads or rivers was required it was necessary to the points as nodes in a graph and finds the
undertake the extraction as a series of discrete graph which minimises the total cost of the arcs
stages, each arc in the network being handled whilst simultaneously passing through each of
individually. the selected points. In our current implementa-
tion, the costs used to construct the arcs are dir-
In the new approach described below, all sys- ectly related to the distance between the points
tems of narrow curvilinear features in a given considered. Only if distances are the same are
image which satisfy the chosen criteria are ex- the confidence values of the points taken into
tracted as a single network or several separate account. The operator may also specify a max-
(disconnected) networks of vectors representing imum cost between nodes and if it is not pos-
the paths of roads or rivers, without the need sible to link all selected points, without arcs
for an operator to specify individual points on exceeding the maximum cost, a second network
the features explicitly. and if necessary further networks are initiated
until all selected high confidence points have
NETWORK ALGORITHM been spanned.

For the purposes of description it is useful to The MST imposes straight line connections
view the algorithm as three separate stages. In between the selected feature points, which is
the first stage, points are identified which have quite acceptable for closely spaced selected
a high probability of being on the network to points. However, as the points become more
be extracted. In some cases, where the features widely separated it leads to increasingly inac-
are sufficiently distinct, this may be achieved curate representations of the underlying fea-
simply by thresholding, but more typically it tures. To overcome this problem, we have intro-
may be achieved by the application of line, edge duced an optional third stage to our algorithm.
or road detection operators to transform the In order to ensure that the extracted network is
image so that low pixel values indicate high con- close to the underlying curvilinear feature in the
fidence that the point is on the feature and high original image, we have combined the minimum
pixel values indicate low confidence. The al- cost path approach of our earlier paper[4] with
gorithm has been implemented in an integrated, the minimum spanning tree method. Whenever
window-based toolkit for image analysis which the distance between two adjacent points in the
makes the application of the preprocessing pro- network, extracted by the minimum spanning
cedures quick and easy to apply. An example tree, is above a low threshold, a suitable search
of part of the graphical user interface(GU1) is window is automatically placed around the re-
shown in figure 1. gion containing the two points on the original
image (or an appropriate derived image) and a
A set of high confidence points is selected path between the pair of points is found which
for further consideration by thresholding the minimises a cost function appropriate t o the
confidence image interactively. Options also problem domain and specified initially by the
allow elimination of isolated points and broad operator. The minimum cost path(MCP) al-
feature responses may be thinned to a single gorithm used is the extension to the D'Esopo
pixel in width. algorithm described in our earlier paper[4].

In the second stage of the algorithm, the This procedure aims to ensure that the ex-
aim is to extract a network representation of tracted network follows closely the underlying
the broad topology of the underlying feature. curvilinear features in the image and produces
The assumption, a t this stage, is that there a curvilinear rather than a linearly connected
is a sufficiently high density of selected points network. Tools for editing, saving and reloading
in the confidence image, covering the feature, networks are also available in the toolkit.
to obtain the correct topology for the network
by finding the minimum spanning tree(MST) Figure 2 shows part of a Landsat image of
through the selected point set. (The idea of a river system and in figure 3 the extracted
network is shown superimposed in black. No

'1 Lookup Tnhle
R Q ~ W P8101~

. S l ~ l g hhound

.,Cilu* Lvc* l<, rr.u4.y-rr u

Thin non-ze,

Threshold

L o - i---
lnrk spaansd:

nova Isdated Pis+*k ,Thrn Lines

m l g h t l n ~ :155

Figure 1: T h e GUI in action

manual selection of individual feature points all the way t o this ideal.
was required and the resulting network may be If, using distance alone as the cost, the next
saved as a vector representation for use inde-
pendently of the original image. In this partic- point t o be added t o the MST would be a dis-
ular case, preprocessing involved thresholding, tance, d, from the node t o which it would be
noise reduction and thinning before the second connected, calculate the costs of the MCPs t o
and third stages of the algorithm. It can be
seen that not only has the broad river path been all points within some distance, constant * d,
extracted, but also the main tributary together
with other river features which, using the earlier of the node and choose the one with the lowest
approach, would have necessitated several user cost. If the constant is one, the algorithm is
controlled extractions. as described earlier. As the constant becomes
large we will move nearer t o the position in
FURTHER REFINEMENT which the point with the overall lowest cost
M C P is linked in.
In a proposed development, the MST approach
and the M C P approach will be integrated more CONCLUSIONS
closrly. Rather than simply using the M C P t o
refine the path between nodes of the MST after Although extracted networks are not always
it has been constructed, it would be better if the perfect, they provide a useful starting poirit for
cost associated with linking a point to a node the operator and give a significant reduction in
in the MST as it is built, could be the cost the need for manual digitisation. A valuable
of the MCP between the point and the node. feature is the ease of use of the GUI allowing
However, this would be computationally very rapid experimentation t o find the high confid-
expensive as MCPs would have t o be calculated ence set. In the future, the development of more
between all combinations of selected high con- reliable procedures for the accumulation of high
fidence points. The following modification t o confidence feature points and the proposed im-
the MST algorithm takes us either part way or provements to the MST algorithm, described
above, should lead to more robust automatic
extraction of network topologies.

Figure 2: Raw Satellite Image Figure 3: Extracted River Network

ACKNOWLEDGEMENTS [4] M. R. Dobie and P. H. Lewis, "Extracting
curvilinear features from remotely sensed
The authors are grateful for support from a images using minimum cost path tech-
SERC research grant number GR/G19244. niques," in First IEEE International Con-
ference on Image Processing, 1994. t o be
References published.

(11 Z. Aviad and P. D. Carnine, "Road find- [5] M. Kass, A. Witkin, and D. Terzopoulos,

ing for road-network extraction," in Conf. "Snakes: Active contour models," Interna-
Computer Vision and Pattern Recognition, tional Journal of Computer Vision, pp. 321-
(Ann Arbor, Michigan), pp. 814-819, IEEE 331, 1988.
Computer Society, IEEE 88CH2605-4, June
5-9 1988. [6] U. Pape, "Implementation and efficiency
of Moore - algorithms for the shortest
[2] J. Wang, P. M. Treitz, and P. J. Howarth, route problems," Mathematical Pmgmm-
"Road network detection from spot imagery ming, vol. 7, pp. 212-222, 1974.
for updating geographical information sys-
tems in the rural-urban fringe," Interna- [7] U. Montanari, "On the optimal detection
tional Journal of Geographical Information of curves in noisy pictures," Communica-
Systems, vol. 6, no. 2, pp. 141-157, 1992. tions of the Association for Computing Ma-
chinery, vol. 14, pp. 335-345, May 1971.
[3] M. A. Fischler, J. M. Tenenbaum, and H. C.
[8] A. Martelli, "An application of heuristic
Wolf, "Detection of roads and linear struc- search methods to edge and contour detec-
tures in low resolution aerial imagery using tion," Communications of the Association
a multi-source knowledge integration tech- for Computing Machinery, vol. 19, pp. 73-
nique," Computer Graphics and Image Pro- 83, February 1976.
cessing, vol. 15, pp. 201-223, Mar. 1981.