Detection of Rooftop Regions in Rural Areas Using Support Vector Machine
Uploaded by
ijsretDetection of Rooftop Regions in Rural Areas Using Support Vector Machine
Uploaded by
ijsretInternational Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 5, May 2015
Detection of Rooftop Regions in Rural Areas Using
Support Vector Machine
Liya Joseph1, Laya Devadas2
1
(M Tech Scholar, Department of Computer Science, College of Engineering Munnar, Kerala)
(Associate Professor, Department of Computer Science, College of Engineering Munnar, Kerala)
ABSTRACT
2. RELATED WORK
Rooftop detection in rural areas is an important task in
many applications including vegetation identification,
land encroachment detection, route planning to rural areas
etc. This paper proposes a new approach for rooftop
detection using machine learning techniques.
The rural area selected for this study is Munnar in
Kerala. In the first step satellite images of Munnar are
randomly collected from Google Maps. It consists of both
rooftop and non-rooftop images. An initial Support
Vector Machine (SVM) classifier is used to detect rooftop
images. This rooftop image is segmented into different
candidate rooftop regions using k-means clustering
algorithm. Then each candidate region is given to a final
Support Vector Machine classifier, which predicts the
true rooftop candidate.
The performance of this method is evaluated using
Mean Square Error (MSE) and Peak Signal-to-Noise
Ratio (PSNR). The Results show that proposed method
has lower value of MSE and higher value of PSNR
compared to existing methods.
Rooftop detection is a tedious task, however many
rooftop detection methods are exists. Most of the earlier
work on rooftop detection has based on edge detection,
corner detection, and image segmentation.
Detection of building with polygonal shapes rooftops
[1] is an edge and corner based detection technique. It is
based on detecting lines and their intersections using a
graph representation. Then find a polygonal shape in the
graph which corresponds to loop in the graph. It detects
only polygonal shaped rooftops.
An automatic building extraction from remote sensing
images [2] is based on both region growing and
morphological methods. But this approach could not
detect buildings with dark rooftops. Another method
based on both edge detection and Hough transform
algorithm [3].
The idea behind many modern approaches is machine
learning techniques. Machine Learning deals with the
construction and study of systems that can learn from data
rather than follow any explicitly programmed
instructions. In the proposed system it first generates
rooftop candidates using image segmentation method and
detects true rooftops from it using Artificial Neural
Network [4] and SVM [5].
The existing methods focus only on detection of
rooftops in urban area. But the proposed method mainly
focus on detection of rooftops in rural area particularly
Munnar at Idukki district in Kerala. Munnar is a hill
station area with enchanting range of vegetation. The
proposed method is capable of detecting rooftops in urban
areas also.
Keywords Feature extraction, Image segmentation, K
means clustering, Machine learning techniques, Support
Vector Machine
1. INTRODUCTION
In recent years, there has been an increasing demand for
rooftop detection due to its variety of applications. All
these methods are mainly focus on urban area rooftop
detection. But this paper proposes a new approach for
rooftop detection in rural areas. In rural areas, the crimes
like unofficial settlements in government land, land
encroachment are increases day by day. To detect these
crimes as well as for the other applications like vegetation
identification, change detection, route planning to rural
areas based on rooftop density, tourism development can
be detected by this new approach.
Section 2 describes related work, section 3 describes
proposed method, section 4 describes results and
discussions and section 5 concludes the paper.
3. PROPOSED METHOD
Proposed rooftop detection system consists of following
steps:
www.ijsret.org
549
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 5, May 2015
Major axis
Minor axis
Fig. 2 Image showing major and minor axis
Fig. 1 Rooftop detection using machine learning
techniques
3.1 DETECTION OF AN IMAGE AS ROOFTOP OR
NON ROOFTOP
In the first step, satellite images of Munnar are randomly
collected from Google maps. Initial training dataset
consists of both rooftops and non-rooftops images. Then
features of input image are extracted and an initial SVM
classifier is used to distinguish between rooftop images
from the non-rooftop images.
3.2 FEATURE EXTRACTION AND INITIAL SVM
CLASSIFICATION
Features are numerical attributes which characterize the
object to be classified. So the extracted features are those
which hold the properties which can help to distinguish
rooftops and non rooftops in an image [6].
In the proposed method, 7 features are considered for
classification of rooftop and non-rooftop images. The
features are listed below:
1. Area: Area is the number of pixels used to represent
a particular segment. This feature helps to distinguish
rooftop and non-rooftop segments. This feature filters
out objects which are too small or too large.
2. Ratio of Minor to Major Axis Lengths: This is
basically the width to length ratio of the segmented
regions. Figure 2 shows the minor and major axes of
a building rooftop. The minor to major axis ratio for
buildings are comparable, whereas this is drastically
different for elongated non-rooftop objects such as
roads, rivers etc.
3. Variance in intensity: This is the variance of the
pixel intensities within a segmented region. A rooftop
would tend to be fairly homogeneous in appearance,
and as such the corresponding region would also have
a lower variance of intensity when compared to nonrooftop region [7].
4. Visible Vegetation Index: Visible Vegetation Index
(VVI) is the presence of vegetation. Normally, VVI is
calculated for multispectral images. But here for RGB
images VVI can be calculated as follows:
R R0
VVI 1
R R0
G G 0
1
G G 0
B B 0 w
1
B B 0
(1)
Here three variables R, G, B are used to denote the
red, green, and blue intensities in the image, whereas
0 , 0 , 0 are values of red, green, blue used to
reference green colour. Where w is used to adjust
the sensitivity of the scale which is known as weight
component.
5. Homogeneity: Homogeneity measures the closeness
of distribution of pixels in an image. Values of
homogeneity ranges from 0 to 1. For a perfectly
homogeneous image the value of homogeneity is 1.
Since, rooftop regions are more homogeneous, so
they tend to have a high value of homogeneity.
6. Solidity: Solidity can be calculated as the ratio of the
total area of a region to the area of the convex hull of
the region. Because most rooftops are rectangular in
shape, rooftop-related regions in an image are likely
to have higher values of solidity.
www.ijsret.org
550
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 5, May 2015
7. Contrast: Contrast can be defined as the measure of
the intensity difference between a pixel and its
neighbour pixel over the whole image. Usually
rooftop regions in the image are homogeneous, so
they tend to have lower contrast.
For each image, the above 7 features are extracted which
is the initial data set. This data set and a label matrix are
supplied to the initial SVM classifier for training. During
training phase, a prediction rule is created. Based on this
rule, the new data set is classified either it a rooftop
image or non-rooftop image.
3.3 DETECTION OF ROOFTOP REGIONS IN THE
ROOFTOP IMAGE
The previous step predicts given input image is rooftop or
non rooftop image. If it is a non rooftop image further
processing is not possible. Otherwise rooftop image is
segmented into four candidate regions using k-means
clustering algorithm. And each candidate region is given
to final SVM classifier which predicts true rooftop
candidate.
3.3.1 Image Segmentation
Image segmentation is the process of dividing an image
into various segments based on its colour properties.
Image segmentation can be done by using k-means
clustering algorithm. The choice of k as 4 gives the best
result for these images. The k-means algorithm can be
explained here.
3.3.1.1 K-Means Algorithm
The k-means algorithm is used to segment the image into
different clusters, where each clusters center is
represented by the mean value of the objects in that
cluster [8].
Input
K denotes the number of clusters.
D is the data set containing n objects Output
A set of K clusters.
Method
1. Randomly choose K objects from D as the initial
cluster centers. Set it as current cluster centroids.
2. For each object in D, compute Euclidian distance
with each cluster centroid. And find the cluster
center to which the data point having minimum
distance, assign that data point into that particular
cluster. Finally get k clusters.
3. Find mean value of each cluster.
4. Compare mean cluster centroids and current
cluster centroids. If they are equal stop the
process. Otherwise go to step 2
5. (re)assign each object to the cluster
6. Stop
Step-wise description of algorithm
1. Initialization
Initialize K is the number of clusters and D is the
matrix form of the image.
2. Assigning Data points into k clusters
The Euclidian distance is calculated for each
centroid from a data point and the data point
having minimum distance from the centroid of a
cluster is assigned to that particular cluster
3. Centroid Updation
Find mean value of each cluster, which is the new
cluster centroid. And then compare new cluster
centroids and previous cluster centroid. If they
are equal, this will be final clusters. Otherwise
repeat process until convergence. Some
convergence conditions are given below:
Stopping when reaching a given or defined
number of iterations.
Stopping when there is no exchange of data
points between the clusters.
Stopping when a threshold value is achieved.
3.3.2 Feature Extraction and Final SVM Classification
In section 3.2, we discussed about feature extraction. The
same procedures are used here. But the features used here
are Area, Ratio of Minor to Major Axis Lengths,
Variance in intensity, and, Visible Vegetation Index. For
each rooftop candidates these four features are extracted,
which is the final data set. The data set and label matrix is
also supplied to the final SVM classifier for training
purpose. During training phase, a prediction rule is
created. Based on this rule, it finds true rooftop candidate.
4. RESULTS AND DISCUSSION
DATA: For detecting rooftop regions in rural areas,
images gathered from Google Maps were used. Google
Maps images are easy to obtain and cost effective. The
rural area selected for this study is Munnar in Kerala.
Munnar data set consists of 50 images, 25 images are
taken as training images and remaining 25 as test images.
Initially, an input image is given to initial SVM
classifier. The classifier predicts if it is a rooftop image or
non rooftop image. If this image is rooftop image, it will
divide into four regions using k-means algorithm. Each
candidate regions are shown below as Cluster1, Cluster2,
Cluster3, Cluster 4 images.
www.ijsret.org
551
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 5, May 2015
Fig. 3 Input image of Munnar
Fig. 6 Cluster 3 image
Fig. 4 Cluster 1 image
Fig. 7 Cluster 4 image
Then final SVM classifier predicts true rooftop candidate
image, which contains most of the rooftop pixels as
shown below.
Fig. 5 Cluster 2 image
Fig. 8 True rooftop image
The evaluation metric used here is mean square error and
peak signal-to-noise ratio.
www.ijsret.org
552
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 5, May 2015
Evaluation
Metrics
MSE
Existing Method
Our Method
0.3473
0.1171
PSNR
52.72
57.4458
5. CONCLUSION
This paper presented a new approach for detecting
rooftop regions in rural areas using Machine Learning
techniques like k- means and SVM. This novel approach
initially detected rooftop images from the set of images.
Most of the existing methods detect only same color
rooftop regions. But this new approach detects rooftops of
any color.
Applications of detecting rooftop regions in rural
areas are vegetation identification, land encroachment
detection, route planning to rural areas etc.
In future, this can be done as an online application to
detect rooftops.
REFERENCES
[1] M. S. Nosrati and P. Saeedi, A novel approach for
polygonal rooftop detection in satellite/aerial
imageries, in Proceedings of the 16th IEEE
International Conference on Image Processing (ICIP
09), pp. 17091712, November 2009.
[2] K. Ren, H. Sun, Q. Jia, and J. Shi, Building
recognition from aerial images combining
segmentation and shadow, in Proceedings of the
IEEE International Conference on Intelligent
[3] H. V. Guducu, Building Detection from Satellite
Images Using Shadow and Color Information, 2008
[4] B. Joshi, H. Baluyan, A. A. Hinai and W. L. Woon
Automatic Rooftop Detection using a Two-stage
Classication,2013
[5] B. Joshi, H. Baluyan, A. A. Hinai and W. L. Woon
Rooftop Detection for planning of solar pv
deployment: A case study in Abu dhabi, in Data
Analytics for Renewable Energy Integration,pp.137149,Springer,2014
[6] Feature
Selection,
Wikipedia,
2013,
http://en.wikipedia.org/ wiki/Feature selection
[7] B. Joshi, H. Baluyan, A. A. Hinai and W. L. Woon
Novel approach for rooftop detection using support
vector machineRooftop Detection, in ISRN Machine
Vision Volume 2013, Article ID 819768, 11 pages
[8] Ms.Chinki Chandhok, Mrs.Soni Chaturvedi, Dr.A.A
Khurshid An Approach to Image Segmentation
using K-means Clustering Algorithm,2012
www.ijsret.org
553
