Home | Sessions | Authors | Session 1.

Automatic Road Extraction from Satellite Image

B.Sowmya Aashik Hameed

Dept. of Electronics & Control Engg., Dept. of Computer Science Engg.,

Sathyabama Institute of Science & Anna University, Chennai
Technology, Deemed University, Chennai Mobile: 9283466927,Email:
[email protected] [email protected]

Morphological trivial opening[1] has been developed for

Abstract
the obtaining the size information of road network, and to
This paper explains the task of extracting road from any extract the road network from preprocessed images by
given satellite image. Road extraction plays a vital role in removing small objects, narrow path and noises based on
many areas like military, map implementation etc. Road the information provided by the granulometry.
is extracted from satellite image based on its spectral For automatic object detection, the criterion used in
signature. Since there are more than 16 million colours trivial opening can be regarded as a threshold, which
available in any given image and it is difficult to analyse should be determined from the image analysis with
the image on all of its colours, the likely colours are respect to granulometry or the pattern spectrum. The
grouped together by image segmentation. For that granulometry techniques can be used to measure the size
purpose fuzzy clustering has been used. From the and shape of objects in image. Opening operation is used
segmented image spectral information of road is to decompose an image through a series of structure
extracted. Spectral information alone is insufficient as elements with a specific shape. The opened images are
soils and roads have the same spectral signature. In our compared with the original image to generate measures
work, concepts of mathematical morphology are applied with respect to different size of structure element but with
to separate road from other objects having same spectral same shape.
signature. The rest of the paper is organized as follows. Section II
describes image segmentation. Fuzzy clustering
Keywords: Image Segmentation, Morphology, Clustering algorithm used is also discussed here. Section III deals
and Trivial Opening with the steps involved in Mathematical Morphology. In
section IV the software design is explained. In Section V
1. Introduction concluding remarks are given.
It is easy to identify roads in satellite images manually.
But automatically extracting road from satellite image is
2. Image Segmentation
a difficult task. Previous works in this area are based on Image segmentation is an essential preliminary step in
statistical and rule-based techniques [8], which depended most automatic pictorial pattern recognition and scene
primarily upon spectral information. As soils and roads analysis problems. Segmentation of an image entails the
share the same spectral signature it is impossible to division or separation of the image into regions of similar
separate them. We present a new method for extracting attribute [2]. The level to which this subdivision is carried
road from satellite images. It involves image depends on the problem being solved. That is
segmentation using fuzzy clustering and feature segmentation should stop when the objects of interest in
extraction using mathematical morphology. The most an application have been isolated. Fuzzy clustering,
basic attribute for segmentation is image luminance which combines fuzzy logic and cluster analysis
amplitude for a monochrome image and color techniques, has experienced a spur of interest in recent
components for a color image. years owing to its important applications in image
Clustering involves the task of dividing data points into segmentation.
homogeneous classes or clusters so that items in the same In non-fuzzy or hard clustering, data is divided into crisp
class are as similar as possible and items in different clusters, where each data point belongs to exactly one
classes are as dissimilar as possible. In fuzzy clustering, cluster. In fuzzy clustering, a single point can have partial
the data points can belong to more than one cluster, and membership in more than one cluster Fuzzy clustering
associated with each of the points are membership grades, belongs to the group of soft computing techniques (which
which indicate the degree to which the data points belong include neural nets, fuzzy systems, and genetic
to the different clusters. When image segmentation is algorithms).
done using fuzzy clustering, the fuzzy clustering In real applications there is very often no sharp boundary
algorithm compares the colours in a relative way and between clusters so that fuzzy clustering is often better
groups them into clusters. suited for the data. Membership degrees between zero
and one are used in fuzzy clustering [3] instead of crisp The first one is the objective function of the HCM, while
assignments of the data to clusters. The resulting data the second is a regularizing term, forcing the values ujk to
partition improves data understanding and reveals its be greatest as possible, in order that points with a high
internal structure. Partition clustering algorithms divide degree of typicality with respect to a cluster may have
up a data set into clusters or classes, where similar data high ujk values, and points not very representative may
objects are assigned to the same cluster whereas have low ujk values in all clusters. In the above equation
dissimilar data objects should belong to different clusters. Ej(xk)=|xk-yj|2 is the square of the Euclidean distance, and
There are many clustering algorithms like, Hard C means, the parameter ρj depends on the distribution of point in
fuzzy C means, possibilistic fuzzy C means, Probabilistic the j-th cluster.
fuzzy C means etc. In our work, we have used
n
Possibilistic fuzzy C means algorithm. This algorithm
gives better reproduction of the original image than the Σ (ujk)mxk
fuzzy C means algorithm [8], as a possibilstic constraint
is introduced. i.e. In this case, the PSNR value of the k=1
reproduced image is higher. vj= , ∀j (2)
2.1 Possibilistic Fuzzy c Means n
In the possibilistic approach to clustering the membership Σ (ujk)m
function or the degree of typicality of a point in a fuzzy
set (or cluster) is assumed to be absolute [5]. In other k=1
words, the degree of typicality does not depend on the
membership values of the same point in other clusters ujk=exp (- Ej(xk)/ ρj ), ∀j,k (3)
contained in the problem domain.
By contrast, many clustering approaches impose a where m ε (1,∞) is a control parameter of fuzziness.
probabilistic constraint, according to which the sum of This theorem provides the conditions needed in order to
the membership values of a point in all the clusters must minimize the cost function. The above two equations can
be equal to one. As a consequence, HCM, FCM[7] and be interpreted as formulas for recalculating the
many other clustering methods assuming the probabilistic membership functions and the cluster centers.
constraint cannot generate membership functions whose A bootstrap clustering algorithm is anyway needed before
values can be interpreted as degrees of typicality. starting PCM in order to obtain an initial distribution of
PFCM algorithm avoids the assumption of the prototypes in the feature space and to estimate some
probabilistic constraint. The PFCM is based on the parameters used in the algorithm. By considering an FCM
relaxation of the probabilistic constraint in order to bootstrap for the PCM, the following definition of ρj can
interpret in a possibilistic sense the membership function be used.
or degree of typicality. n
The Possibilistic C-Means algorithm I (PCM-I) is based
on a modification of the objective function of FCM. In Σ (ujk)mEj(xk)
this case, one must supply the values of some parameters k=1
such as the fuzzifier parameter, and others regulating the
weight of the spread of membership functions.The ρj= K for all j (4)
Possibilistic C-Means algorithm II (PCM-II) is based on n
modification of the cost function of the HCM (instead of
the FCM) in order to avoid, in this way, the determination Σ (ujk)m
of the fuzzifier parameter. The objective function of the k=1
PCM-II contains two terms as shown below.
where, m is the fuzzifier parameter used by the FCM, and
n c c n K is a proportional parameter.
Jm(U,v)=Σ Σ (ujk)Ej(xk)+ Σ ρj Σ(ujk ln ujk-ujk) (1) This definition makes ρj proportional to the mean value of
the intracluster distance, and critically depends on the
k=1 j=1 j=1 k=1 choice of K .
where, Jm(U,v) is the objective function
Ej(xk) is the Euclidean distance between xk and vj. PFCM Algorithm
ρj is the possibilistic constraint.
U= [ujk] is the c x n fuzzy partition matrix. 1. Find U, v (r) using FCM or any other clustering
c is the number of cluster centers. algorithm.
n is the number of data points. 2. Find the possibilistic constraint ρj using the equation
(4).
3. Find U using the equation (3).
4. Find the cluster centers v (r + 1) using the equation
(2).
5. Find the distance matrix D using the equation
Ej(xk) = ||xk- vj||2 (5)
6. Check for convergence
If max | v(r) - v(r+1) | < E Stop else repeat the steps
2-6

2.2 Self Estimation Algorithm Figure 1: Resized Satellite Image

If the number of clusters is manually specified the
segmentation may not be effective. Hence there must be a
system to calculate the robust number of clusters. The
self estimation algorithm used here finds the Euclidean
distance between the different cluster centers.
If the maximum Euclidean distance between the cluster
centers is greater than the specified value, then the
number of cluster centers is increased by one else the
clusters are merged.. An alternate method, in which we
start with a high number of clusters and merge on the
Figure 2: Clustered image with 4 classes
result of the iterations, can also be used.

2.3 Peak Signal to Noise Ratio

Signal-to-noise (SNR) measures are estimates of the
quality of a reconstructed image compared with an
original image. The basic idea is to compute a single
number that reflects the quality of the reconstructed
image [4].
Reconstructed images with higher metrics are judged
better. In fact, traditional SNR measures do not equate
with human subjective perception. Several research Figure 3: Extracted output
groups are working on perceptual measures, but for now with spectral Signature of Road
signal-to-noise measures are used because they are easier
to compute. Also to be noted that higher measures do not Typical PSNR values range between 20 and 40. They are
usually reported to two decimal points (e.g., 25.47). The
always mean better quality.
actual value is not meaningful, but the comparison
The actual metric that is computed in this work is the
peak signal-to-reconstructed image measure, which is between two values for different reconstructed images
gives one measure of quality. The MPEG committee used
called PSNR. Assume a source image f(i,j) is given that
an informal threshold of 0.5 dB PSNR to decide whether
contains M by N pixels and a reconstructed image F(i,j)
where F isreconstructed by decoding the encoded version to incorporate a coding optimization because they
believed that an improvement of that magnitude would be
of f(i,j). Error metrics are computed on the luminance
visible.
signal only so the pixel values f(i,j) range between black
Some definitions of PSNR use 255/RMAE rather than
(0) and white (255).
2552/MAE. Either formulation will work because we are
First the mean absolute error(MAE) of the reconstructed
interested in the relative comparison, not the absolute
image is computed as follows
values.In our assignments we used the definition given
1 MN
above.
MAE= Σ Σ |F(i,j)-f(i,j)| (6) Fig.1 shows the resized satellite image which shows an
end of road at Hawaii. This image is processed using
MN i=1 j=1
possibilistic fuzzy c means. The resultant image is shown
The summation is over all pixels. PSNR in decibels(dB) in Fig.2. The no. of clusters obtained in the resultant
is computed by using image by self estimation algorithm, is 4. The peak signal
PSNR = 20 log 10 (2552 / MAE). to noise ratio of processed image is 41.1364.
Fig.3 shows extracted output without applying
mathematical morphology. Comparing it with original
image we can see that the extracted output shows the road
as well as the soils.
 3. Mathematical Morphology X(i), if X(i) satisfies the criterion T
TO =
Mathematical morphology is an image processing tool φ, otherwise
used for extracting features of interest. Mathematical
morphology is a set theory approach, which is based on where TO is the trivial opening associated with criterion
geometrical shape. It uses set operations such as union, T. It is morphological opening, because it is idempotent,
intersection and complementation. Mathematical antiextensive and increasing. In image processing, this
morphology operations include dilation, erosion, operation uses the criterion T to filter the connected
openings, closings and the derived operations such as Hit components that satisfy the criterion T. Trivial opening
or Miss, thinning, their properties and their use. based on criterion T provides a practical means of object
The objective of the work undertaken is to develop an detection and derived area opening from connected and
algorithm for automated road network detection from trivial opening to find the connected regions in an image
high resolution images using mathematical morphology. with a certain area. Trivial opening does not affect the
Firstly a road is segmented from background. The shape and size of the connected regions that are preserved
mathematical morphology tools used are granulometry because it preserves the entire connected regions. In our
and trivial opening. A granulometry analysis of objects in design we have used the structural element disc of size 1
image is performed and the size information of road and 2 as the threshold i.e. criterion T.
network is obtained. The road is extracted preprocessed Since the width of street is less than that of the main road,
images and differentiated other features with similar opening operation with the structure element whose size
properties as roads using Trivial Opening and closing. It smaller than main road but slightly larger than that of
is done based on the size information provided by the street remove small paths. Further it removes the paths
granulometry. connecting road network to the houses. The side effect of
opening is that it also removes the road parts, which are
3.1 Granulometry under shadow caused by houses and trees. It is necessary
to develop tools to reconstruct these road parts.
Granulometry[6] is one of the most useful and versatile
Mathematical morphology based window operation is
tools of morphological image analysis. Granulometry
designed for this purpose.
techniques find wide range of applications in feature
extraction, texture characterization, size estimation image
4. Software Design
segmentation etc. The concept behind granulometry is to
determine the size and shape distribution of objects The steps involved in the road extraction algorithm are as
present within an image. follows.
Granulometry is a field that deals principally with 1. Load the image from which road has to be extracted
determining the size distribution of particles in an image. 2. Segment the image using PFCM
Opening operations with structuring elements of 3. Extract spectral information of road
increasing size are performed on the original image. The 4. Perform Granulometry
difference between the original image and the its opening 5. Remove noises and small path using trivial opening
is computed after each pass when a different structuring 6. Close the small openings caused by trivial opening
element is completed. At the end of the process, these 7. Display the extracted road
differences are normalized and then used to construct a
histogram of particle size distribution. This approach is 4.1 Estimation Algorithm
based on the idea that opening operations of a particular The self estimation algorithm for finding no. of clusters
size have the most effect on regions of the input image in the image automatically is given below
that contain particles of similar size. 1. Start with c = 2
Thus, a measure of the relative number of such particles 2. Find the cluster center matrix v using equation (2)
is obtained by computing the difference between the 3. Compute Euclidean distance between cluster center
input and output images. These measures can be used as values, q
shape and size signature of the original image and can be 4. If min ( q ) > E, increment c and repeat steps 2, 3
plotted as a pattern spectrum.
4.2 Image Segmentation
3.2 Morphological Trivial Opening
The following steps do the image segmentation
Morphological Trivial Opening is defined as one, which 1. Load the image file to be segmented
provides a practical mean of object detection and 2. Get the mode of operation
identification. It does not affect the shape and size of the 3. If the mode is manual get the no. of clusters
objects of interest. Let X be an image, {X(n)|n = 1, 2, 3, 4. If the mode is automatic calculate the no. of clusters
…,N} is a series of connected components in the image, using self estimation algorithm
x(i) is a point in X(i). We define the trivial opening with a 5. Using PFCM segment the image
criterion T, as follows. 6. Determine PSNR
7. Display the segmented image and its PSNR value spectral Ikonos image . This image is clustered using
possibilistic fuzzy c means. The resultant image is shown
5. Conclusion in Fig.5. By self estimation algorithm, the no. of clusters
obtained in the resultant image is 7. The peak signal to
noise ratio of processed image is 38.286. From this
clustered image objects of road like spectral values are
extracted. It is shown in fig.6.
It is seen from fig.6 that shadows, small paths are also
extracted. Some of the road parts are also removed. In
Fig.7 the extracted road is shown. For road extraction
trivial opening been used. The structural elements used
are disk and rectangle. For closing the paths line structure
has been used.
Figure4: Original Image Even after applying morphological methods some part of
road is missing. This is because image resizing which
resulted in poor image resolution. It takes longer time for
the system to process large image. Hence, resizing been
done.
Some part of the image, which has the same spectral
feature of the road, is also obtained. But compared with
the PFCM segmented image of Fig.6 in the image of
Fig.7 the road is quite visible.

Figure 5: Clustered Image

Figure 8: Original Image

Figure 6: Extracted output Fig.8 shows the resized satellite image of Rome, Italy
with 1meter resolution. It is a natural colour multispectral
with spectral Signature of Road Ikonos image

Figure 7: Extracted Road Network Figure 9: Clustered Image

In this paper we have presented a new method for road
extraction. It has been tested on high resolution satellite and
aerial photos. The result shows that mathematical
morphology provides an effective tool for automated road
network detection.
Fig.4 shows the resized satellite image Beijing, China
with 1 meter resolution, which is natural colour multi-
 [1] Allan G. Hanbury and Jean Serra, “Morphological
Operators on the Unit Circle”, IEEE Transactions on
Image Processing, VOL. 10, NO. 12, pp.1842-
1850,DEC 2001.
[2] Schupp, S. Elmoataz, A. Fadili, J. Herlin, P. Bloyet
D. “Image segmentation via multiple active contour
models and fuzzy clustering with biomedical
applications” in 15th International Conference on
Pattern Recognition, 2000. Proceedings, pp. 622-625,
Vol.1,2000.
Figure 6: Extracted output
[3] Songcan Chen, Daoqiang Zhang, “Robust image
with spectral Signature of Road segmentation using FCM with spatial constraints
based on new kernel-induced distance measure” in
IEEE Transactions on Systems, Man and
Cybernetics, Vol.34,No.4,pp: 1907-1916,Aug 2004.
[4] Zhou Wang and Dapeng Zhang, “Restoration of
Impulse Noise Corrupted Images Using Long-range
Correlation” in IEEE Signal Processing Letters, vol.
5, no. 1, pp. 4-7, Jan. 1998.
[5] R. Krishnapuram, J.M. Keller. The possibilistic c
means: insights & recommendations. IEEE Trans.
Fuzzy Systems 4: 385-393, 1996. [6] Scott T. Acton,
Figure 7: Extracted Road Network and Dipti Prasad Mukherjee,Scale Space
Classification Using Area Morphology, IEEE
Transactions on Image Processing, Vol. 9, No. 4,
pp..623-635, April 2000.
[6] Ouzounis, G.K. and Wilkinson, M.H.F., “Countering
over segmentation in partitioning-based
connectivities”, IEEE International Conference on
Image Processing, 2005. ICIP 2005. Volume 3, 11-
14 pp. 744-847,Sept. 2005.
[7] B.Sowmya and Sourav Bhattacharya “Colour Image
Figure 12: Extracted Road Network Segmentation Using Fuzzy Clustering Techniques”
. This image clustered using possibilistic fuzzy c means is in the international conference IEEE Indicon 2005
shown in Fig.9. The no. of clusters obtained in the organised by IEEE Chennai Chapter at IIT, Chennai
resultant image are 4. The peak signal to noise ratio of on 11-13 December 2005.
processed image is 39.8518. [8] Gruen, A. and H. Li, “Road Extraction from aerial
From the clustered image objects of road like spectral and satellite images by dynamic programming,”
values are extracted is shown in fig.10. As with the ISPRS Journal of Photogrammetry and Remote
Beijing image the road network seen in fig.11 after Sensing, pp. 11-21, Vol. 50, No. 4, 1995.
processing with morphological methods is better than that
of fig.10. Since the input picture quality is better the road
is seen well. It is better than that of image of Beijing. Fig.
12 shows the extracted road from image of Fig.1
The proposed method been tested with many satellite images
and it is proven to be effective. The algorithm used works
well with high resolution elongated road images. The
reduction of noise using the PFCM algorithm shows its
effectiveness. The program has also been tested for its
consistency and its reliability. This method can be used in
several industrial applications such as for auto
recognition and for military purposes. It is important to
provide military and other groups with accurate, up-to-
date maps of the road networks in any region of the
world.

6. References