Geocarto International

ISSN: 1010-6049 (Print) 1752-0762 (Online) Journal homepage: http://www.tandfonline.com/loi/tgei20

A cognitive framework for road detection from

high-resolution satellite images

Naveen Chandra, Jayanta Kumar Ghosh & Ashu Sharma

To cite this article: Naveen Chandra, Jayanta Kumar Ghosh & Ashu Sharma (2018): A cognitive
framework for road detection from high-resolution satellite images, Geocarto International, DOI:
10.1080/10106049.2018.1450451

To link to this article: https://doi.org/10.1080/10106049.2018.1450451

Accepted author version posted online: 14

Mar 2018.
Published online: 29 Mar 2018.

Submit your article to this journal

Article views: 8

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tgei20
Geocarto International, 2018
https://doi.org/10.1080/10106049.2018.1450451

A cognitive framework for road detection from high-resolution

satellite images
Naveen Chandra , Jayanta Kumar Ghosh and Ashu Sharma
Geomatics Engineering Group, Civil Engineering Department, IIT Roorkee, Roorkee, India

ABSTRACT ARTICLE HISTORY

Road network extraction from high-resolution satellite (HRS) imagery is a Received 7 September 2017
complex task. It is an important field of research and is widely used in various Accepted 31 January 2018
cartographic applications such as updating and generating maps. The
KEYWORDS
objective of this research work is to develop a novel framework, emulating Cognitive; knowledge;
human cognition, for detection of roads from HRS images. Roads network support vector machine
from HRS images are detected using support vector machines within the
different stages of cognitive task analysis. In the first stage, basic information
about the cognitive parameters which are required for image interpretation is
collected. In the second stage, the rule-based method is used for knowledge
representation. Lastly, during knowledge elicitation, the developed rules are
used to extract roads from HRS images. The proposed method is validated
using 16 HRS images of developed suburban, developed urban, emerging
suburban and emerging urban region.

1. Introduction
Road detection from the urban area using HRS images in an important field of research in computer
vision (Sirmacek and Unsalan 2011). In past, different types of methods namely; semantic model,
neural network-based method, optimal search-based method, knowledge-based method, road opera-
tor-based model and snake model are being developed for detecting road network from VHR satellite
images (Mirnalinee 2007). In spite of the simple shapes, roads are still difficult to detect when compared
to the buildings from satellite images (Sirmacek and Unsalan 2011). Due to the complex properties of
the roads, automated method for road detection still remains a challenging task (Mirnalinee 2007).
Due to the availability of various commercial HRS imaging sensors (Quick-bird and IKONOS), the
large amount of information is available for several remote sensing applications. It is necessary to
have the correct and updated information of the road network in the geographic information system
(GIS) databases for urban planning, transportation and navigation (Mirnalinee 2007). It has been
observed that extraction of roads from VHR imagery using a manual approach is a laborious and
time-consuming task and automated approaches have the capability to speed-up the process of road
detection, therefore, they are widely used (Mirnalinee 2007). Road network is generally made up of
different colour and there is a variation in the width of the roads. Thus, there is a need for advanced
approaches for detecting roads from HRS images (Sirmacek and Unsalan 2011). Various automated and

CONTACT Naveen Chandra naveenchandra0408@gmail.com

© 2018 Informa UK Limited, trading as Taylor & Francis Group
2  N. CHANDRA ET AL.

semi-automated algorithms have been developed for road network detection (Sirmacek and Unsalan
2011) from HRS images in which an assumption was made that the roads segment were labelled earlier
(Udomhunsakul 2004; Dell’Acqua et al. 2007; Ameri et al. 2008). In Yang and Wang (2007), a method
was developed to detect roads from HRS images in which they initially detected the road primitives
(homogeneous areas and straight lines) which were further integrated together to detect roads. The
limitation of this method was that it was unable to detect road network of the urban area and the
occluded road regions. In Ma et al. (2007) discontinuous road segments were linked together through
perceptual rules to detect roads from panchromatic enhanced thematic mapper (ETM) images. In
Gamba et al. (2007), built-up areas and the road network was detected in synthetic aperture radar
(SAR) images. They introduced a clustering-based method for detecting built-up regions and the
straight-line edges in which longitudinal gaps were assumed as road networks. In Rianto et al. (2000)
a method was proposed for detecting main roads from SPOT satellite images. They used Hough
transformation algorithm for road network detection in which the parallel and straight-line segments
were assumed as roads. However, their method was unable to detect complex and curvilinear roads
of the urban area. Further road-tracking approach which is performed on the basis of profiles taken
perpendicular to road axis can be found in Quam (1978). In Barzohar and Cooper (1996), stochastic
and geometric methods were combined together for road detection. Some research work which uses
mathematical morphology i.e. a method for processing the digital images using geometrical shapes
are images that are based on geometrical shape is given in Mena (2003). In Baumgartner et al. (2002),
a semi-automated method is presented for extraction of the road network and in Zhao et al. (2002), a
similar semi-automated approach is proposed for creating and updating roads in suburban and urban
regions using HRS images. The further nearest neighbour method was implemented in Chandra and
Ghosh (2016) for extracting roads from HRS images.
Few algorithms were developed which were able to detect roads and buildings. In Unsalan and
Boyer (2005) road networks, as well as the buildings, were detected from satellite images using clus-
tering, decomposition of binary images, vegetation indices and graph theory. In Akcay and Aksoy
(2008), unsupervised segmentation was used for detecting roads and built-up regions from HRS
images. A novel method was developed in Montesinos and Alquier (1996) where thin objects were
detected from noisy images. They used active contours for carrying out the perceptual grouping on
the segmented edges. To validate their approach, two types of images were used. Aerial images were
used to detect roads and medical images were used to detect blood vessels. In Sirmacek and Unsalan
(2010), a spatial voting-based model and an edge detection method were proposed for detecting roads
from panchromatic satellite images. A deep review on road network detection from satellite images
can be found in Baumgartner et al. (1997) and Wang et al. (2016). A comparative analysis of methods
for road detection is given in Table 1. Therefore, in this research work, cognition-based multi-stage
method is implemented to extract roads from HRS images using cognitive task analysis (CTA) (Clark
et al. 2008; Chandra and Ghosh 2017a, 2017b). The paper is described as follows: Part II contains the
methodology used for extracting roads; part III contains experimental results and discussion and part
IV presents the conclusion.

2. Methodology
In this research, CTA is merged with the process of road extraction from HRS images. It is an approach
which is mostly used in cognitive psychology for carrying out difficult tasks. CTA is divided into five
different phases. CTA generally describes the cognitive and psychological processes which an analyst
acquires while performing a complex task. CTA briefly illustrates all the cognitive parameter and the
input that are necessary for obtaining the output of the complex task. CTA is a method that is widely
used in the different decision-making process. Extraction of information using CTA is much easier as
it utilizes the thought process for carrying out a task. This method is mainly oriented towards analysis
and understanding of the task that uses the cognitive processes such as memory, problem-solving and
decision-making. In general, CTA method represents and analyses all the cognitive parameters which
 GEOCARTO INTERNATIONAL  3

Table 1. A comparative analysis of road detection methods (Boyer 2011; Das et al. 2011; Wang et al. 2016).

Number of test images

Methodology used Features used Performance measure (%)
Cooperating snakes 4 Gradient vector flow Precision: 51.88–85.8
Recall: 79.89–98.3
AdaBoost 3 Local binary patterns and Completeness: 64–91
intensity Correctness: 66–90
Snakes 3 Intensity gradient Completeness: 72–84
Correctness: 95–99
SVM 3–6 Intensity, width, edge length Completeness: 75–90
and edge gradient Correctness: 13–55
Grouping and filtering 2–6 Intensity and line direction Completeness: 66–95
Correctness: 80–98
MRF and binarization 5 Value of mean radiance Completeness: 76.9–90.9
Correctness: 63.5–93.4
Mean shift clustering 1–8 Histogram of image Accuracy: 86–90
Region competition 4 Intensity Accuracy: 80
Angular texture signature and 2–4 Edge direction and intensity Accuracy: 70–99
road tracker
Morphology 3–4 Length and line direction Accuracy: 91.76
Fuzzy inference 1–2 Wavelet-based filters Accuracy: 82–93
Artificial neural network 1–2 Edge, intensity and normal- Accuracy: 95
ized distance

are taken as input by various users for carrying out a particular task. A cognitive framework of the
proposed method is given in Figure 1.

2.1. Stage 1: preliminary knowledge collection

In first stage, various procedures which are necessary for the cognitive task analysis are determined.
In this stage, different types of methods are reviewed which can be further used for knowledge elic-
itation. In past, several methods such as observation and unstructured interviews (Clark et al. 2008)
are used for preliminary knowledge collection but in this research document, review has been used.
In this stage, analyst gathers a basic knowledge about the subject or the field in which CTA is to be
integrated. In simple words, the analyst gets familiar with subject knowledge.

2.2. Stage 2: identification of knowledge representation method

Various methods such as concept map, flow chart and semantic network are being used for representing
knowledge (Clark et al. 2008). Here, ‘if-then’ rule-based method is used for representing the knowledge.
Rules are developed on the basis of visual interpretation of an image. ‘If ’ the objects in the image are
interpreted as roads on the basis of the prior knowledge ‘then’ those objects are manually labelled as
roads else non-roads. This method is also known as ‘premises-action’ rule-based approach. Further,
the obtained interpretation results are used for knowledge elicitation.

2.3. Stage 3: identification of knowledge elicitation method

Elicitation is defined as a process of collecting the knowledge or information from several resources.
In this stage, the analyst uses different methods for integrating the information or knowledge identi-
fied in the earlier step. In previous research, different types of methods have been used for knowledge
elicitation, namely prototyping method, observation method and simulation method (Clark et al.
2008). In our research work, classification is used for knowledge elicitation. Support vector machine
(SVM) (Cortes and Vapnik 1995; Maulik and Chakraborty 2017) belonging to classification approach
is considered in this study. Initially, the classifier is trained on the basis of prior knowledge. Further,
4  N. CHANDRA ET AL.

Prior Knowledge High Resolution

Collection Satellite Image

Creating Training
Knowledge Representation
Data Set

Selection of SVM Support Vector

Knowledge Elicitation
Parameters Machines

Kernel Penalty
Detected Road Segments Ground Truth
Type Parameter
Images

Threshold
value
Verification and Analysis
of Results
Qualitative & Quantitative

Compiling Theoretical and Numerical Results

Figure 1. A cognitive framework for detecting roads from HRS images.

SVM classifies the input data in two classes i.e. roads and others. Support vector machine is a classifi-
cation method which was developed at AT&T BELL Laboratory (Burges 1998). It concentrates on
determining the balanced relation between the amount of accuracy obtained on training data-set and
the capability of the system. It is widely used in different application such as face recognition, object
detection, handwriting and text recognition (Blanz et al. 1996). It is noticed that SVM performs better
than other traditional algorithms such as neural networks (Cortes and Vapnik 1995). SVM basically
divided the classes by a hyperplane surface to determine the boundary between them. If H represent
the line of classification and H1 and H2 denotes the line passing through sample of each classification
which are nearer to H then distance between H1 and H2 is known as separating margin. In feature
space, Rd SVM decision function is given by:
f (x) = sign(𝜔 ⋅ 𝜑(x) + b) (1)
where f (x) represents the discriminate function, b represents the threshold value and ω represents the
weighted vector. The optimal hyperplane is a combination of two parameters: (a) margin maximization;
(b) error minimization is given by:
V
1 �
Ψ(𝜔, 𝜉) = ‖𝜔‖2 + C 𝜉i (2)
2 i=1

( ( ) )
yi 𝜔 ⋅ 𝜙 xi + b ≥ 1 − 𝜉i (3)
 GEOCARTO INTERNATIONAL  5

where 𝜉i ≥ 0 and i = i = 1, 2, … , V ;𝜉i is known as slack variables and constant C is regularization

parameter. The solution to the optimization problem is to maximize quadratic programming using
Lagrange function given as:
v,v v
∑ ( ) ∑
LD (a) ai aj yi yj K xi .x + ai (4)
i,j=1 i=1
v
∑
under the condition 0 ≤ ai ≤ C; i = 1, 2, … , V and ai yi = 0; where a = (a1 , a2 , … , an ) denotes vec-
tor of Lagrange multipliers. Thus, final discriminanti=1functions in feature space X is given by:
v
∑ ( )
f (x) = sign( ai yi K xi .x + b) (5)
i=1
( )
where K xi .x represents a kernel function. There are various kernel function, which satisfy Mercer’s
theorem. Kernel function plays a vital role in the overall performance of the machine. A kernel function
represented by Gaussian function is given as:
( )
K xi .x = exp(−𝛾xi − x 2 ) (6)

In this study, radial bias function is used which performs better than polynomial and linear kernel
function. In SVM, three parameters are used. Firstly, the gamma in kernel function is set to 0.333.
Gamma represents the overall spread of the data points for defining a decision region. If the value of
gamma will be low, the obtained spread/curve of the decision boundaries will also be low. However, if
the value of gamma will be high, the curve obtained will also be high. The support vector is unable to
determine the shape of the data points if the value of gamma is set very low. Secondly, penalty param-
eter is set to 120. Penalty parameter is used to avoid the misclassification of each training sample. It
controls the training error. If the value of penalty parameter is low, the produced decision surface will
be smooth. Thirdly, the threshold for classification probability is defined as 0.05. It has been found
that various types of parameters such as the availability of different objects like buildings, vegetation
and water bodies in the satellite image should be analysed for determining the proper threshold value
(Singh and Garg 2013).
Later, all these parameters are used for detecting roads from satellite images. SVM is a robust
machine learning algorithm and not widely used in remote sensing domain. The key advantage of
SVM is that it produces better results with less number of training samples with minimum chances of
overfitting compared to other machine learning algorithms. SVM delivers an effective out-of-sample
generalization, with respect to the appropriate selection of the defined parameters (kernel function
and penalty parameter) i.e. by selecting a proper generalization grade; SVMs can prove to be robust,
even if the selected training sample has some bias. This is also the key advantage of SVM over neural
networks, which have multiple solutions associated with local minima and for this reason may not
be robust over various samples.

2.4. Stage 4: verification and analysis of the results

The output of CTA may vary due to the selection of method used for elicitation of knowledge. Thus, it
is necessary to quantitatively evaluate the method so that it can be intended for various applications.
In this research work, cognitive task used for validating the output is known as a comparison and the
overall process of validation is defined as cognitive resonance. Thus, output image, which contains
the extracted road network, is compared with the reference data-set for quantitative assessment and
validation. In past, various types of methods (Singh and Garg 2014a, 2014b) have been used for
determining the accuracy of the proposed methods for detecting road networks from satellite images.
6  N. CHANDRA ET AL.

2.5. Stage 5: formatting results for several applications

Lastly, all the statistical and theoretical findings of the CTA are compiled together. These results are
further produced in the form of a report or a document containing the overall description. This cog-
nitive method can be used in the different application of the remote sensing field.

3. Experimental results and discussion

3.1. Data-set
The overall performance of the proposed methodology is tested using IITM road network extraction
data-set available at http://www.cse.iitm.ac.in/~vplab/satellite.html. The data-set consists of 16 mul-
tispectral images of four different regions, namely developed suburban, developed urban, emerging
suburban and emerging urban (16) (http://www.cse.iitm.ac.in/~vplab/satellite.html) shown in Figures
1(a), 2(a), 3(a) and 4(a), respectively, along with the corresponding ground truth images. The ground
truth image for each image is prepared by the human operator. The categorization of the images into
different regions was done with the help of GIS expert. The images were acquired from Wikimapia
with the spatial resolution of 1-m/pixel. The size of each image is 512*512 pixels. The assessment is
done by comparing the obtained results with the available ground truth data-set for each image. This
data-set is also used by the other for validating the performance of their methodology. In Mirnalinee
et al. (2011), this data-set was used for validating and comparative evaluation of their proposed frame-
work. They used two road features (local linear trajectory and spectral contrast) for designing their
framework. In Tuncer (2007), this data-set is also used for evaluating their fully automatic method
for road network extraction from satellite imagery. In PCI Geomatics: http://www.pcigeomatics.com,
artificial neural network-based approach for road extraction is evaluated IITM road extraction data.
In Mokhtarzade and Zoej (2007), this data-set is again used for evaluating the overall performance
of their integrated method for road network extraction from VHR images. They also compared their
approach with two state-of-the-art technique using this data-set.

3.2. Method for validation of the results

The qualitative and quantitative evaluation is required in order to validate the performance of the
proposed method. First, the qualitative evaluation of the detected roads is done through visual anal-
ysis of the obtained output. Second, the quantitative assessment of the method is performed using
three basic quality measures (Aksoy et al. 2012; Ok 2013; Ghaffarian 2014) given in (7), (8) and (9).

(True positive)
Precision = ( ) (7)
True positive + False positive

(True positive)
Recall = ( ) (8)
True positive + False Negative

(2 * Precision * Recall)
f-score = (9)
(Precision + Recall)
 GEOCARTO INTERNATIONAL  7

Figure 2. Detected roads from developed suburban region: (a) test images (b) results of PCI Geomatics: http://www.pcigeomatics.
com (c) results of Mokhtarzade and Zoej (2007) (d) results of proposed method and (e) corresponding ground truth images.

3.3. Results and discussion

3.3.1. Qualitative analysis
The results of the developed suburban region show that in first input image, the roads are connected
with each other but they are not wide. Thus, it is difficult to extract the road segment with a smaller
width, however; the proposed method is able to extract the entire road network with negligible false
pixels of non-roads. Second input image includes the major road segment in the middle of the image.
8  N. CHANDRA ET AL.

Figure 3. Detected roads from developed urban region: (a) test images (b) results of PCI Geomatics: http://www.pcigeomatics.com
(c) results of Mokhtarzade and Zoej (2007) (d) results of proposed method and (e) corresponding ground truth images.

Various smaller road segments are originating from the main road segment. Road segments present
in the second input image are extracted and some of the false pixels of houses are detected as road
network because of the high density of the residential buildings. Further, it is seen that third input
image includes the interconnected road network of different width. Third input image also includes the
highly dense residential buildings of different shape and size. The road segments of the entire image
are extracted with some other objects. Few pixels of the road segment at the right bottom of the input
image are not detected as shown in the output image. Fourth input image consists of road networks as
 GEOCARTO INTERNATIONAL  9

Figure 4. Detected roads from emerging suburban region: (a) test images (b) results of PCI Geomatics: http://www.pcigeomatics.com
(c) results of Mokhtarzade and Zoej (2007) (d) results of proposed method and (e) corresponding ground truth images.

well as the other object such houses of different shape and size. Based on the visual representation, it
can be interpreted that the road network is made up of concrete material. Thus, it is easy to extract the
complete road network without including a large number of false pixels of non-roads. The qualitative
results of developed suburban region are shown in Figure 2.
On visual examination of the results of the developed urban region, it is seen that first input image
includes the interconnected road network of different width. It also includes the highly dense residential
buildings of different shape and size. The road segments of the entire image are extracted with some
10  N. CHANDRA ET AL.

other objects. Few pixels of the road segment at the right bottom of the input image are not detected
as shown in the output image. Second input image includes the residential buildings, vegetation and
connected road network. The width of the entire road network is same. All the pixels of the road
network are extracted and some false pixels of non-roads are included in the output image. In third
input image, the width of the road network is similar and they are clearly separated. It also includes
the few residential houses, which are of the same size and shape. It is observed that the each segment
in the input image is clearly extracted. Third input image includes the two major road segments, which
cross each other. The several smaller road segments originate from one of the major road segment. The
width of the one major road segment is large compared to the other. The output of the fourth image
represents that almost all the pixels of the road network is detected and only a few false pixels of the
other object is extracted. Each of the road segments is extracted by the proposed methodology. The
qualitative results of developed urban region are shown in Figure 3.
Further, the first input image of emerging suburban region consists of only few road segments. The
smaller road segments originate from the major road segment, which is in the middle of the image.
The smaller road segment is almost similar in size however the major road segment is wider is size.
The image performs well for the proposed methodology because the entire road network is extracted
properly and only a few false pixels of non-roads are included in the output image. The second image
includes a major road segment, which lies in the middle of the test image. Four smaller road segments
originate from the major road segment. The output image represents the detected road network in
which each segment is extracted properly and some false pixels of non-roads are detected. Third input
image contains an interconnected network of few road segments. It also includes few buildings, which
are located on the either side of the road network. It is seen that two road segment which intersects
each other is detected clearly by the proposed methodology, however, the other remaining segments
are not extracted completely. In fourth input image, each of the road segments is connected to the
other but only one road segment at the bottom is not connected to either of the road segment. Various
types of houses are located on the either sides of the road network. The proposed methodology is
able to extract the entire road network as seen in the output image. It is noted that the smaller road
segment present in the input image are also extracted clearly which improves the performance of the
proposed methodology. The qualitative results of emerging suburban region are shown in Figure 4.
A first input image of emerging urban region includes a complex pattern of the road network. The
road segments are of different shapes i.e. linear, circular and rectangular, however, the entire road
network is connected to each other. It is a difficult task to extract the roads of different shapes even
though the proposed methodology is able to extract the maximum number of road segments. In second
input image, the different shapes of roads are observed. There exists circular, linear and curved shape
of road segment which is difficult to extract. The proposed methodology is able to extract a maximum
number of pixels of different road segments.
The third image includes dense residential region along with various other objects. It consists of an
interconnected pattern of the road network which a number of intersection points. The width of all
of the road segments is similar except the two major road segments, which intersect each other at the
middle of the test image. Maximum numbers of true pixels of the road are extracted along with some
false pixels by the proposed methodology. Finally, the fourth image consists of different types of objects
such as building, roads, trees. The shape and size of objects differ from each other. The two major road
segments intersect each other. The output image includes each of the extracted road segment and only
a few patches of other objects. The qualitative results of emerging urban region are shown in Figure 5.

3.3.2. Quantitative analysis

In order to determine the potential of the proposed method, the results of the detected road network
from the satellite images different regions are compared with the previous studies. First, the results of
the developed suburban area are compared with the results of methods proposed in PCI Geomatics:
http://www.pcigeomatics.com and Mokhtarzade and Zoej (2007). For the first input image, the f-score
produced in PCI Geomatics: http://www.pcigeomatics.com is 64.72%, and Mokhtarzade and Zoej
 GEOCARTO INTERNATIONAL  11

Figure 5. Detected roads from emerging urban region: (a) test images (b) results of PCI Geomatics: http://www.pcigeomatics.com
(c) results of Aksoy et al. (2012) (d) results of proposed method and (e) corresponding ground truth images.

(2007) is 60.42% and, respectively, while the f-score produced by the implemented knowledge elic-
itation method is 71.65%. For the second image f-score produced in PCI Geomatics: http://www.
pcigeomatics.com is 49.0%, Mokhtarzade and Zoej (2007) is 53.56%, respectively, however, the f-score
produced by SVM is 59.25%. For the third input image, the f-score produced in PCI Geomatics: http://
www.pcigeomatics.com is 62.32%, Mokhtarzade and Zoej (2007) is 64.12%, respectively, whereas the
f-score produced by implemented knowledge elicitation method is 71.59%. For the fourth image, the
f-score produced in PCI Geomatics: http://www.pcigeomatics.com is 59.76%, Mokhtarzade and Zoej
12  N. CHANDRA ET AL.

90
80
70
60
F-Score 50
40
30
20
10
0
1 2 3 4
Test Images
Results of [37] Results of [38] Results of proposed methodology

Figure 6. Comparative results of the f-score of the images of the developed suburban region.

80
70
60
50
F-Score

40
30
20
10
0
1 2 3 4
Test Images
Results of [37] Results of [38] Results of proposed methodology

Figure 7. Comparative results of the f-score of the images of the developed urban region.

(2007) is 53.21%, respectively. The f-score produced by the proposed method is 80.21%. The visual
representations of the output of different existing methods are shown in Figure 2, however, the overall
f-score of all the methods are shown in Figure 6.
Second, the output images of the developed urban area are also compared with the existing meth-
ods. For the first input image, the f-score produced in PCI Geomatics: http://www.pcigeomatics.com
is 51.68%, Mokhtarzade and Zoej (2007) is 45.2%, however, f-score produced by classification-based
knowledge elicitation method is computed as 67.01%. The f-score produced by the second input
image in PCI Geomatics: http://www.pcigeomatics.com is 27.72%, Mokhtarzade and Zoej (2007) is
26.39% and 40.67% by the proposed method. Further, the f-score produced by the third image in PCI
Geomatics: http://www.pcigeomatics.com is 33.27%, Mokhtarzade and Zoej (2007) is 32.55%. The
f-score computed by SVM is 35.17%. Lastly, f-score of the fourth input image of the developed urban
region is produced as 47.95% in PCI Geomatics: http://www.pcigeomatics.com, 61.0% in Mokhtarzade
and Zoej (2007) and 61.24% by the proposed methodology. The output images of the developed urban
region of existing methods are given in Figure 3 and the comparative results of f-score are shown in
Figure 7.
Third, the output images of emerging suburban area are also compared with the results of methods
proposed in PCI Geomatics: http://www.pcigeomatics.com and Mokhtarzade and Zoej (2007). The
f-score produced in PCI Geomatics: http://www.pcigeomatics.com is 20.99%, Mokhtarzade and Zoej
(2007) is 38.94% for the first input image while the f-score produced by the proposed method is 83.59%.
 GEOCARTO INTERNATIONAL  13

90
80
70
60
F-Score 50
40
30
20
10
0
1 2 3 4
Test Images
Results of [37] Results of [38] Results of proposed methodology

Figure 8. Comparative results of the f-score of the images of emerging suburban region.

80
70
60
50
F-Score

40
30
20
10
0
1 2 3 4
Test Images
Results of [37] Results of [38] Results of proposed methodology

Figure 9. Comparative results of the f-score of the images of emerging urban region.

For the second input image, the f-score produced in PCI Geomatics: http://www.pcigeomatics.com
and Mokhtarzade and Zoej (2007) is 50.33% and 35.01%, respectively, however, the f-score produced
by SVM is 62.10%. The third image produced the f-score of 40.17% in PCI Geomatics: http://www.
pcigeomatics.com, 15.34% in Mokhtarzade and Zoej (2007), whereas the f-score produced by proposed
methodology is computed as 72.21%. Finally, f-score of the fourth input image in PCI Geomatics:
http://www.pcigeomatics.com is 34.18%, in Mokhtarzade and Zoej (2007) is 14.63% while the f-score
of the proposed method is 69.15%. Figure 4 provides the output of proposed method along with the
results of the existing methods, whereas in Figure 8 the comparative results of the f-score are given.
The f-score computed for first input image of emerging urban region in PCI Geomatics: http://www.
pcigeomatics.com is 67.83%, Mokhtarzade and Zoej (2007) is 65.82% while the f-score obtained by
SVM is 73.44%. For second input image, the f-score produced in PCI Geomatics: http://www.pcige-
omatics.com is 74.91%, Mokhtarzade and Zoej (2007) is 57.77%, however, the f-score produced by
the proposed method is 76.19%. Third input image produced the f-score produced in PCI Geomatics:
http://www.pcigeomatics.com is 20.6%, Mokhtarzade and Zoej (2007) is 59.93% while, the f-score
produced by SVM is 65.05%. Lastly, the f-score of the fourth image is computed as 73.72% which is
greater than the results of the previous methods PCI Geomatics: http://www.pcigeomatics.com and
Mokhtarzade and Zoej (2007). The results of the detected road network of emerging urban region are
given in Figure 5 and the comparative results of the f-score are shown in Figure 9. The quantitative
results of the proposed methodology are given in Table 2.
14  N. CHANDRA ET AL.

Table 2. Quantitative results of the detected roads: DS (developed suburban), DU (developed urban), ES (emerging suburban), EU
(emerging urban), P (precision), R (recall) and F (f-score).

DS DU ES EU
Test image P (%) R (%) F (%) P (%) R (%) F (%) P (%) R (%) F (%) P (%) R (%) F (%)
1 62.82 83.37 71.65 66.02 68.03 67.01 80.49 86.93 83.59 70.80 76.29 73.44
2 55.28 63.83 59.25 41.04 40.31 40.67 82.69 49.72 62.10 72.60 80.15 76.19
3 69.56 73.73 71.59 29.53 43.49 35.17 72.94 71.49 72.21 67.41 62.85 65.05
4 74.98 86.22 80.21 77.65 61.24 68.47 68.85 69.46 69.15 72.93 74.52 73.72

Figure 10. Comparative results (a) results of the test image of developed suburban region (b) results of the PSVM proposed in
Mirnalinee et al. (2011).

Further, the results of the proposed method are also compared qualitatively and quantitatively
with the results produced in Mirnalinee et al. (2011). Firstly, on visual examination of the result of the
developed suburban region (shown in Figure 10), it is interpreted that output image obtained using
SVM includes all the segments of the roads (along with some false pixels) which are also included in
the results of the probabilistic support vector machines (PSVM). Both the algorithms are able to detect
the major segments of the road network. The method proposed in Mirnalinee et al. (2011) achieved
the higher accuracy particularly for the images of developed region where the roads are very wide.
Secondly, on the basis of quantitative comparison, it is observed that the accuracy obtained for the
three images of developed suburban and one image of developed urban and emerging urban region is
higher than the results of the proposed method. This is due to the fact that in Mirnalinee et al. (2011),
author initially used the PSVM and dominant singular measure (DSM) which were later integrated
within neural network which produced the better results, whereas in our study only SVM is used for
detecting road network. However, it is found that this integration couldn’t produce the satisfactory
output due to the variation in the shape and the material of the road.
The qualitative and quantitative results of the proposed method have been given in detail for each
test image of different region in order to represent the performance of the proposed cognitive method
with the previous research. On the basis quantitative analysis, it is observed that the results of the
proposed method for each test image were better than the previous results which proved the viability
of the method.
On the basis of comparative analysis, it is found that the accuracy of the proposed method is higher
than the existing methods for each image of all regions which prove the novelty of the proposed
 GEOCARTO INTERNATIONAL  15

method. Therefore, on the basis of quantitative and qualitative findings, it is stated that this cognitive
method for extraction of road network from HRS images delivers the encouraging results.

4. Conclusion
In this research, a cognitive and automated method for road network extraction has been implemented
by integrating it with a cognitive task analysis. CTA is a major contribution in psychology and is a
current topic of research which can be used in different domains. The briefly described and illustrated
results of the CTA are directly related to the overall performance of the person performing the task.
CTA is capable of delivering the relevant information which is obtained from the cognitive process
which is acquired by human beings. The results of the experiment on satellite images demonstrate that
this is an effective cognitive method for road network extraction from the suburban area. There are
some limitations such as boundaries of the road network are not clearly separated. Thus, for a future
scope, this method needs to be improved further so that a clear separation between the boundaries of
the road segment can be obtained. Lastly, on the basis of results, it can be concluded that this cognitive
method is able to extract the junctions and all the segments of the road network.

Disclosure statement
No potential conflict of interest was reported by the authors.

ORCID
Naveen Chandra http://orcid.org/0000-0002-0957-097X

References
Akcay HG, Aksoy S. 2008. Automatic detection of geospatial objects using multiple hierarchical segmentations. IEEE
Trans Geosci Remote Sens. 46(7):2097–2111.
Aksoy S, Yalniz IZ, Tasdemir K. 2012. Automatic detection and segmentation of orchards using very high resolution
imagery. IEEE Trans Geosci Remote Sens. 50(8):3117–3131.
Ameri F, Mobarakki A, Zoej MV. 2008. Semi-automatic extraction of different shaped road centerlines from ms and
pan-sharpend ikonos images. In: XXIth Congress of the International Society for Photogrammetry and Remote
Sensing. p. 621–626.
Barzohar M, Cooper DB. 1996. Automatic finding of main roads in aerial images by using geometric-stochastic models
and estimation. IEEE Trans Pattern Anal Mach Intell. 18(7):707–721.
Baumgartner A, Steger C, Mayer H, Eckstein W. 1997. Multi-resolution, semantic objects, and context for road extraction.
In: Förstner W, Plumer L, editors. Semantic modeling for the acquisition of topographic information from images
and maps. Basel: Birkhauser Basel; p. 140–156.
Baumgartner A, Hinz S, Wiedemann C. 2002. Efficient methods and interfaces for road tracking. Int Arch Photogram
Remote Sens Spatial Inf Sci. 34(3/B): 28–31.
Blanz V, Scholkopf B, Bulthoff HCBVV, Burges C, Vapnik V, Vetter T. 1996. Comparison of view-based object recognition
algorithms using realistic 3D models. In: International Conference on Artificial Neural Networks; Springer Berlin
Heidelberg. p. 251–256.
Boyer KL. 2011. Review on building and road detection. In: Multispectral satellite image understanding. Springer
London; p. 139–144.
Burges CJ. 1998. A tutorial on support vector machines for pattern recognition. Data Min Knowl Disc. 2(2):121–167.
Chandra N, Ghosh JK. 2016. A cognitive perspective on road network extraction from high resolution satellite images.
2nd International Conference on Next Generation Computing Technologies (NGCT). p. 772–776.
Chandra N, Ghosh JK. 2017a. A cognitive viewpoint on building detection from remotely sensed multispectral images.
IETE J Res. 1–11.
Chandra N, Ghosh JK. 2017b. A cognitive method for building detection from high-resolution satellite images. Curr
Sci. 112(5):1038–1044.
Clark RE, Feldon D, van Merrienboer JJ, Yates K, Early S. 2008. Cognitive task analysis. Handbook Res Educ Commun
Technol. 3:577–593.
Cortes C, Vapnik V. 1995. Support-vector networks. Mach Learn. 20(3): 273–297.
16  N. CHANDRA ET AL.

Das S, Mirnalinee TT, Varghese K. 2011. Use of salient features for the design of a multistage framework to extract roads
from high-resolution multispectral satellite images. IEEE Trans Geosci Remote Sens. 49(10):3906–3931.
Dell’Acqua F, Gamba P, Lisini G. 2007. A semi-automatic high resolution SAR data interpretation procedure. Proceedings
of Photogrammetric Image Analysis.
Gamba P, Dell’Acqua F, Lisini G, Trianni G. 2007. Improved VHR urban area mapping exploiting object boundaries.
IEEE Trans Geosci Remote Sens. 45(8):2676–2682.
Ghaffarian S. 2014. Automatic building detection based on supervised classification using high resolution Google Earth
images. J Photogr Remote Sens. XL-3:101–106.
Ma H, Qin Q, Du S, Wang L, Jin C. 2007. Road extraction from ETM panchromatic image based on dual-edge following.
In: 2007 IEEE International Geoscience and Remote Sensing Symposium; IEEE. p. 460–463.
Maulik U, Chakraborty D. 2017. Remote sensing image classification: a survey of support-vector-machine-based
advanced techniques. IEEE Geosci Remote Sens Mag. 5(1):33–52.
Mena JB. 2003. State of the art on automatic road extraction for GIS update: a novel classification. Pattern Recogn Lett.
24(16):3037–3058.
Mirnalinee TT. 2007. An approach to cartographic object Extraction from high resolution satellite imagery. In:
Proceedings of International Conference on Computational Intelligence and Multimedia Applications. ISBN 0-
7695-3050-8/07.
Mirnalinee TT, Das S, Varghese K. 2011. An integrated multistage framework for automatic road extraction from high
resolution satellite imagery. J Indian Soc Remote Sens. 39(1):1–25.
Mokhtarzade M, Zoej MJV. 2007 Feb. Road detection from high-resolution satellite images using artificial neural
networks. Int J Appl Earth Obs Geoinf. 9(1):32–40.
Montesinos P, Alquier L. 1996. Perceptual organization of thin networks with active contour functions applied to
medical and aerial images. In: Proceedings of the 13th International Conference on Pattern Recognition; IEEE. vol.
1, p. 647–651.
Ok AO. 2013. Automated detection of buildings from single VHR multispectral images using shadow information and
graph cuts. ISPRS J Photogram Remote Sens. 86:21–40.
Quam LH. 1978. Road tracking and anomaly detection in aerial imagery. Sri international menlo park ca artificial
intelligence center.
Rianto Y, Kondo S, Kim T. 2000. Detection of roads from satellite images using optimal search. Int J Pattern Recognit
Artif Intell. 14(08):1009–1023.
Singh PP, Garg RD. 2013. Automatic road extraction from high resolution satellite image using adaptive global
thresholding and morphological operations. J Indian Soc Remote Sens. 41(3):631–640.
Singh PP, Garg RD. 2014a. A two-stage framework for road extraction from high-resolution satellite images by using
prominent features of impervious surfaces. Int J Remote Sens. 35(24):8074–8107.
Singh PP, Garg RD. 2014b. Road detection from remote sensing images using impervious surface characteristics: review
and implication. ISPRS – Inter Arch Photogram Remote Sens Spatial Inf Sci. 40(8):955.
Sirmacek B, Unsalan C. 2010. Road network extraction using edge detection and spatial voting. In: 2010 20th International
Conference on Pattern Recognition (ICPR); IEEE. p. 3113–3116.
Sirmacek B, Unsalan C. 2011. Road detection from remotely sensed images using color features. In: 2011 5th International
Conference on Recent Advances in Space Technologies (RAST); IEEE. p. 112–115.
Tuncer O. 2007. Fully automatic road network extraction from satellite images. In: 2007 3rd International Conference
on Recent Advances in Space Technologies. p. 708–714.
Udomhunsakul S. 2004. Semi-automatic road detection from satellite imagery. In: 2004 International Conference on
Image Processing, 2004 ICIP’04; IEEE. vol. 3; p. 1723–1726.
Unsalan C, Boyer KL. 2005. A system to detect houses and residential street networks in multispectral satellite images.
Comput Vis Image Underst. 98(3):423–461.
Wang W, Yang N, Zhang Y, Wang F, Cao T, Eklund P. 2016. A review of road extraction from remote sensing images. J
Traffic Transp Eng (Eng Ed). 3(3):271–282.
Yang J, Wang RS. 2007. Classified road detection from satellite images based on perceptual organization. Int J Remote
Sens. 28(20):4653–4669.
Zhao H, Kumagai J, Nakagawa M, Shibasaki R. 2002. Semi-automatic road extraction from high-resolution satellite
image. Int Arch Photogram Remote Sens Spatial Inf Sci. 34(3/A):406–411.