An index based road feature extraction from LANDSAT-8 OLI images

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 11, No. 2, April 2021, pp. 1319∼1336
ISSN: 2088-8708, DOI: 10.11591/ijece.v11i2.pp1319-1336 r 1319
An index based road feature extraction from LANDSAT-8
OLI images
Sama Lenin Kumar Reddy1
, C. V. Rao2
, P. Rajesh Kumar3
, R. V. G. Anjaneyulu4
, B. Gopala Krishna5
1,2,4,5
National Remote Sensing Centre (NRSC), ISRO, Hyderabad, India
3
Department of ECE, Andhra University College of Engineering, India
Article Info
Article history:
Received Apr 17, 2020
Revised Jul 16, 2020
Accepted Sep 25, 2020
Keywords:
Band combination of OLI
Road extraction
Saturation
Shock square filter
Top-Hat transform
ABSTRACT
Road feature extraction from the remote sensing images is an arduous task and has a
significant role in various applications of urban planning, updating the maps, traffic
management, etc. In this paper, a new band combination (B652) to form a road in-
dex (RI) from OLI multispectral bands based on the spectral reflectance of asphalt, is
presented for road feature extraction. The B652 is converted to road index by normal-
ization. The morphological operators (Top-hat or Bottom-hat) uses on RI to enhance
the roads. To sharpen the edges and for better discrimination of features, shock square
filter (SSF), is proposed. Then, an iterative adaptive threshold (IAT) based online
search with variational min-max and markov random fields (MRF) model are used on
the SSF image to segment the roads and non-roads. The roads are extracting by using
the rules based on the connected component analysis. IAT and MRF model segmenta-
tion methods prove the proposed index (RI) able to extract road features productively.
The proposed methodology is a combination of saturation based adaptive thresholding
and morphology (SATM), and saturation based MRF (SMRF), applied to OLI images
of several urban cities of India, producing the satisfactory results. The experimental
results with the quantitative analysis presented in the paper.
This is an open access article under the CC BY-SA license.
Corresponding Author:
Sama Lenin Kumar Reddy
National Remote Sensing Centre (NRSC)
Balanagar, Hyderabad, 500 037, India
Email: leninkumar438@gmail.com
1. INTRODUCTION
Road feature extraction from remote sensing (RS) images is a challenging and one of the most in-
tensive research topic. Roads are very crucial for transportation, providing many ways of utility for human
civilization. The research of road extraction has vital significance for surveying, updating the maps, urban
planning, on-line traffic management, geographical information system (GIS), global positioning system (GPS)
based road transport and so forth. In the absence of automatic extraction methods, manual road drawing from
RS images requires great effort in terms of cost and time.
RS images provides information for various objects of the earth, based on the spatial and spectral prop-
erties. In low resolution (LR) images the roads look like curvilinear structure. From LR imagery, road feature
extraction is always a difficult task, mainly in urban places due to the presence of trees, multistory buildings,
fly-overs and their shadows are major obstacles irrespective of spatial resolution and sensors. Automatic road
extraction from RS images is evolving and most of the approaches are limited in providing solutions with low
to medium accuracy. This is due to the factors affecting the imaging conditions like environment (seasonal
Journal homepage: http://ijece.iaescore.com

1320 r ISSN: 2088-8708
changes), spatial resolution, nature of data [1] and road surface conditions.
Detection and extraction of roads from RS images depend on the width of the road features and spatial
resolution of data [2]. Generally, roads constructed by a mix of Asphalt and Gravel or crushed stone, and the
spectral reflectance of roads at various conditions (in-situ measurements) has given analysis [3, 4]. New asphalt
roads have less reflectance due to the dominance of the hydrocarbon absorption. Aged and paved roads have a
high reflectance than new asphalt road due to the erosion of asphalt mix [4].
Using RS images, normalized indices based methods are being used extensively for feature detection
like normalized difference water index (NDWI) for water [5], normalized difference vegetation index (NDVI)
for vegetation and Built-up area detection by normalized difference building index (NDBI) [6], etc. NDBI
index detects both buildings and roads as a built-up area. In the same way, a new band combination is framed
to form an index for extract the road features alone from OLI images.
Recently, [7, 8] have shared their views on road extraction methods. Most of the methods are using
the panchromatic and RGB images. Spectral-spatial classification [9] was implemented to differentiate the
road and non-road based on the General adaptive neighborhood mathematical morphology (GANMM) and
then morphological profiles [10, 11] created for roads using GANMM. In [12], on segmented objects of very
high resolution (VHR) images, object-based frangi’s filter (OFF) and object-based shape filter (OSF) are used
to enhancing road features and generated training samples to model. In [13], demonstrated that roads from
HR images using directional morphological enhancement and segmentation, which depends on the road seed
templates homogeneity in different directions. In [14], roads extracted from VHR images by using texture
analysis based on the structure feature set standard deviation (SFS-SD) followed by dilation as a preprocessing.
Mathematical morphology (MM) used to distinguish the curvilinear features of edges detected by the Canny
operator. Linear features enhanced by linearness filtering, which is a combination of Hessian matrix-based
filter, local standard variance, and geometrical properties [15]. Supervised classification of SPOT panchromatic
images by spatial information and is extracted by morphological openings with various structuring element size
according to object size.
After the preprocessing, image segmentation is one of the crucial steps in object-based image analysis
(OBIA) and widely used in digital image processing for pattern recognition [16]. A detailed review of RS image
classification and feature extraction presented in [17]. The threshold used for binary image segmentation and
is calculated based on the histogram, intra-class and interclass variance [18], or optimal value of an image.
Segmentation methods are region-based, edge-based and hybrid. Region-based techniques are region growing
[19], split-merge [20] which depends on the homogeneity of regions with seed points; and cluster methods (K
means, Fuzzy C means, etc.) depend on the number of classes present in an image. The edge is one of the finest
features of an image, and it represents the transition of one object to another object. Many approaches used in
literature to detect edge of features by edge operators [21, 22] and edge primitives [23]. In recent literature,
researchers proposed new edge extraction methods [24, 25], and these methods are produced better results than
classical edge detectors like Canny, Prewitt, etc., but generate the unconnected edges.
A method that segments the images by using optimization is called a hybrid method. This paper
adopts two methods, namely the iterative adaptive threshold (IAT) and markov random fields (MRF). IAT,
which finds the edges and updates the edges by stochastic gradient descent optimization [26]. MRF is also
a stochastic model used for segmentation that uses the neighborhood information [27] and optimized by the
iterative condition model (ICM).
The main objective of the work is the extraction of road features (Freeways (Highways), and Arterial
(State and District)) from the RS images at the scale of 1:50,000 scale map. OLI images chosen to extract the
roads at a 1:50 K scale because at this scale, road widths are varying from 20 meters (m) and above [2]. OLI
images have a panchromatic band with 15 m and multi-spectral bands with 30 m spatial resolutions. These
images cover a swath of 185 km with a revisit period of 16 days. Its wide swath provides coverage of large
areas on the earth with sufficient resolution to distinguish various features like land, farms, forests, urban areas,
etc. Here, the OLI pan-sharpened image (15 m) used in road extraction. In OLI imagery, a large portion of
road widths is a minimum of 1, 2, and above 3-pixel roads are very few. For this reason, strategies which
can identify wide roads (10-30 pixels) from aerial and HR images [9-15] are not effective. In this paper, the
extraction of roads presented using a band combination from multispectral images of OLI based on the spectral
range of asphalt to enhance the accuracy of road extraction from LR-RS images.
The rest of the paper describes the proposed methodology in section 2. Various data sets used in this
paper presented in section 3 and section 3.1 describes the analysis of the selection of band combinations the
Int J Elec & Comp Eng, Vol. 11, No. 2, April 2021 : 1319 – 1336

Int J Elec & Comp Eng ISSN: 2088-8708 r 1321
experimental results with quantitative values shown in section 4.
2. PROPOSED METHODOLOGY
The flow diagram of the road extraction proposed methodology using OLI images shown in
Figure 1. The thick dotted (Green) outline box represents the saturation based adaptive thresholding and mor-
phology (SATM) method, and thin dotted outline boxes represent the saturation based MRF (SMRF) method.
The methodology consists of six stages, and each step described in the following sub-sections.
The proposed methodology is a combination of SATM and SMRF and describes the selection of
multispectral band combination for road index given in sections 2.1 and 2.2 In section 2.4, describes the image
enhancement by using the SSF filter. The segmentation techniques used in the paper given in 2.5 to segment an
image into road and non-road features and road features alone extracted by using the rules based on the shape
parameters 2.6.
Figure 1. Block diagram of proposed methodology
2.1. Band selection criteria for road index
The OLI data is open source and available from USGS earth explorer. The various spectral bands
present in the OLI sensor listed in Table 1.
An index based road feature extraction from LANDSAT-8 OLI images (S. L. K. Reddy)

1322 r ISSN: 2088-8708
Table 1. LANDSAT-8 (OLI) sensors and band names
Band Name Wavelength (µm) Spatial Resolution (m)
B1 - Coastal aerosol 0.43 - 0.45 30
B2 - Blue 0.45 - 0.51 30
B3 - Green 0.53 - 0.59 30
B4 - Red 0.64 - 0.67 30
B5 - NIR 0.85 - 0.88 30
B6 - SWIR 1 1.57 - 1.65 30
B7 - SWIR 2 2.11 - 2.29 30
B8 - Panchromatic 0.50 - 0.68 15
Compared to the usage of the panchromatic image for feature extraction, the multispectral image
contains more information in several spectral bands which used for advantage. Each band in a multispectral
image corresponds to a band of wavelengths in visible, IR, and near infra-red (NIR) regions. From the avail-
able multi-spectral bands of OLI, we propose a band combination to form a road index (RI) for road feature
extraction.
The reflectance of asphalt is more in NIR, short wave infra-red (SWIR) in [4]. USGS Spectroscopy
Lab provided, the spectral reflectance of various minerals, liquids, vegetation, etc. Asphalt road spectral values
are available from https://speclab.cr.usgs.gov/. Figure 2 shows the spectral variation from 0.35 µm to 2.5 µm.
Spectral reflectance of asphalt feature and other features from the OLI image shown in Figure 3. From this,
we can observe the asphalt reflectance variation is almost matching with the provided spectra as shown in
Figure 2 in the bands NIR and SWIR.
In visible band region (0.40 µm to 0.70 µm) asphalt has low reflectance compared to NIR and SWIR
regions (0.85 µm to 2.5 µm). In the spectral range of 1.8 µm to 2.5 µm, which is above the SWIR wavelength,
the reflectance of the asphalt roads is varying. From this, it concludes that asphalt roads have a favorable
spectral response in NIR and SWIR wavelengths.
The OLI sensor has different bands of wavelengths given in Table 1. According to Spectral Reflectance
of asphalt road from Figures 2 and 3, the suitable bands for road detection in OLI are NIR, SWIR1, and SWIR2
bands.The NIR (0.85 µm - 0.88 µm) and SWIR1 (1.57 µm - 1.65 µm) present in OLI have narrow spectral
bandwidth with good reflectance for road features (linearly increasing). In SWIR2 (2.11 µm - 2.29 µm) from
2.11 µm to 2.2 µm reflectance of the roads is increasing and in the range 2.2 µm to 2.3 µm it is decreasing
result in uncertainty in feature recognition. Additionally, the SWIR band can also distinguish the moisture
content of soil and vegetation. The blue and green bands are having very low reflectance of asphalt road in the
entire wavelength range.
Figure 2. Spectral reflectance vs. wavelength of asphalt and concrete
Int J Elec & Comp Eng, Vol. 11, No. 2, April 2021 : 1319 – 1336

Int J Elec & Comp Eng ISSN: 2088-8708 r 1323
Figure 3. Spectral reflectance vs. bands in OLI
From the available OLI multispectral bands, the bands BLUE, NIR, SWIR1 used to form an RGB
image. These bands proved to be the better combination for road extraction in the section 3.1. The bands B6,
B5 and B2 are stacked, as RGB (B652) image. A fusion of B652 image with the Panchromatic band (B8)
carried out using Brovey transform. This method [28] is chosen as it is faster and does not affect the band
spectral relationship for road features as shown in Figure 4 with very fine spatial details.
(a) (b)
Figure 4. Spectral response of Asphalt from OLI, (a) B652 image (b) Fused image
2.2. Road index: Saturation
HSI color model most widely used in digital color image processing and developed by [29]. The road
index (RI) component derived as saturation (S) formulae from the selected band combination B652, and the
roads represented in uniform and with low values compared to surrounding features due to minimum reflectance
value from three bands is dividing with the summation of all band values. In the RI (or S) image, roads show
high discrimination from the other features. S image has normalized values within the range of [0,1]. Here, S
is used (as termed in the paper) as road index (RI) and defined as the following 1 for B652.
S = RI = 1 −
3 ∗ min(SWIR1, NIR, BLUE)
(SWIR1 + NIR + BLUE)
(1)

1324 r ISSN: 2088-8708
2.3. Morphological transforms
The MM is a set theory approach consisting of union, intersection and complement. Top-hat (TH)
transform used to enhance the objects smaller than the structuring element. It preserves sharp peaks and im-
proves contrast. There exist two types of TH transform. White TH or TH transform, which highlights the
’Bright’ features and defined as subtraction of morphological opening (◦) image from the initial (before open-
ing) image. Black TH or Bottom TH (simply bot-hat (BH)) transform defined as dual as the difference between
morphological closing (•) image and the input image. It highlights the ’Dark’ features that are smaller than the
structuring element.
TH = Sth = S − (S ◦ B) (2)
BH = Sbh = (S • B) − S (3)
Where, Sth represents the TH applied on RI image , B is structuring element ’disk’ with (suitable) radius, and
Sbh represents BH. TH and BH transforms are uses in various image processing steps in pre or post-processing
of the feature extraction process. BH transformation applied to the RI image, which highlights the roads and
suppresses the remaining features.
2.4. Shock square filter
Shock filter is one of the classes of morphological enhancement method and given by [30] as
It = −sgn(4I)|∇I| (4)
Here, I represent any gray-scale image. The Laplacian of I i.e., (4I) is act as edge detector. The 4I
replaced by 4(G ~ I) in [31]. In this notation G is Gaussian and ~ is convolution.
IG = −sgn(4(G ~ Sbh))|∇Sbh| (5)
Here, Sbh is BH of S image as derived in (3).
The shock filter used to sharpen edges but forms maze-like structures in regions [32]. To eliminate
this, we proposed a filter, which is a combination of Shock filter [31, 30] followed by a modified shock filter.
This smooth the regions as well as keep the sharp edges of features.
Here, the motive is to sharpen the edges of the image and to smooth the regions (inside the objective
feature).
The direction of shocks measured by using the hyperbolic tangent function Tsh described in [32] on
the E, which is edge detector. The modified shock filter Ish is
Ish =

1 − S.
sgn(E)
2

|∇IG| (6)
sgn(x) =





1, if x 0,
0, if x = 0,
−1, if x 0.
(7)
where (4I) replaced by Weickert’s coherence shock filter [33]. It is one of the best edge detectors and
multiplied by the Road index (S). In this way, the regions smoothed along with sharp edges in the image.
Tsh(E) =
(1 + tanh(λ((1−sign(E)
2 ) − 0.5)))
2
(8)
where λ designed for tuning the sharpness of curve function; hence, it set as the value 6 (fixed). E is an edge
detector, as described in Weickert’s coherence shock filter [33]. Let I = Sbh, then
E = ( I2
x Ixx + 2 Ix Iy Ixy + I2
y Iyy) (9)
Int J Elec Comp Eng, Vol. 11, No. 2, April 2021 : 1319 – 1336

Int J Elec Comp Eng ISSN: 2088-8708 r 1325
The edge detection is performed on the shock filtered image by using (5). An-isotropic Shock filter
[32] is represented as
I0
sh = Tsh.D(Ibh) + (1 − Tsh).E(Ibh) (10)
D(Ibh) and E(Ibh) are dilation and erosion of BH image. SSF is able to get sharp edges and enhance
the feature objects. Also, it is useful in removing the connected noise to the feature. Without SSF, the extracted
roads are having connected noise like building and spike ridges. By using SSF, it is producing better discrimi-
nation as shown in Figures 5a and 5b, i.e., roads in the image are enhanced and sharpen edges (highlighted in
red color boxes). To suppress the other elevated features by the SSF, the TH transform is applied on the SSF
image as aforementioned in section 2.3.
I0
T H = I0
sh − (I0
sh ◦ B) (11)
where I0
sh is SSF image and B is structuring element ’disk’ with suitable radius.
(a) (b)
Figure 5. (a) Road Index i.e., Normalization of B652, (b) SSF on Bot-Hat of S (RI) image
2.5. Segmentation methods
An IAT algorithm is an edge-based image thresholding method and edges updated by iterative using
the stochastic gradient descent optimization method based on the variational min-max principle [26] performed
on the I0
T H image to get the segmented binary image. It helps in the extraction of roads accurately. This process
termed as SATM.
A model, MRF [27] is also used for the segmentation of road features on I0
T H image, when there
is no prior information about the model parameters. MRF depends on the neighborhood information for its
probability distribution. The parameters estimated by using the expectation-maximization (EM) algorithm.
The segmenting classes are known, and the label image generated by using the optimal threshold on TH, i.e.,
TH of SSF image. This process termed SMRF. MRF model provides the solution with maximum a posterior
(MAP) [34] estimation by the maximizing label class and minimizing using the posterior energy function. The
parameters (mean and variation) estimated by using the EM algorithm [35].
2.6. Connected component analysis
In the final step, by using the connected component analysis (CCA) [36], the segmented binary image
is labeled. For each label, the shape parameters like extent (Ex), eccentricity (Ec) computed [37]. The extent
defined as a ratio number of pixels in the label to the bounding box of the same label. The label whose Ex value
is ≤ 0.12 and also Ec value is ≥ 0.99 considered as road label for extraction of elongated roads. The road
features in OLI images look like curvilinear, elliptical, etc., for this reason, for each label, the shape features
like shape index (SI) and density (D) computed. The SI with high values at the same time the D has moderate
values of the label taken as road features from the labeled image.
3. DATA SETS
Data sets of OLI covering various urban areas used to extract the roads, which are level-1 precision
and terrain (L1TP) corrected used for road extraction. In this study, four OLI data sets covering three major

1326 r ISSN: 2088-8708
urban areas of Hyderabad, Chennai, Bangalore, and one data set covering the rural areas of Rajasthan used to
evaluate the performance of the proposed methodology. The used scene id along with their date of pass for
each chosen area given in Table 2. All images mentioned in the paper are top of atmosphere (TOA) corrected.
The size of the image represents the subset scene image.
Table 2. Data sets with scene ID’s and date of pass
S.No. Date of Pass sub-scene Scene ID Size of Image
1 2016-04-05 Hyderabad LC81440482016096LGN00 3291 x 3099
2 2017-02-21 Chennai LC81420512017052LGN00 1586 x 1994
3 2017-02-03 Bangalore LC81440512017034LGN00 1978 x 1888
4 2018-05-09 Rajasthan LC81480402018129LGN00 600 x 500
3.1. Analysis on various band combinations
As given in section 2.1 for road feature extraction, the chosen band combination is B652. In this
section, the various band combinations of OLI experimented with the proposed methodology. A few interest
band combinations are B652, B653, B654, B752, and B762 selected for analysis and shown in Figure 6 for
importance in road extraction. The band combination B432 is a natural color image, used in this methodology.
B432 combination was not able to produce even major roads as well. By using the proposed methodology,
B432 is not useful for road detection from OLI.
The band combinations B542 and B654 are false color composite (FCC) images. In the B4 band,
the concrete has high, and asphalt has low reflectance with each other, due to which converted S image has
mixed features (building and roads). The B542 and B654 are helpful in the extraction of major roads only but
not able to recognize other roads. The B654 combination image was shown in Figure 6i and corresponding
results are RI index image, Optimal Threshold from I0
T H image and SATM result shown in Figures 6j, 6k and
6l respectively.
The B752 band combination, not able to produce urban area roads of a few major roads also missing.
Due to B7 has uncertainty in the reflectance of concrete and asphalt. From Figure 3, B7 has a low reflectance
of asphalt and almost equal to band B2. In conversion to S, most of the road values mixed with the barren and
built-up areas, the segmentation methods are not able to separate the road features. The corresponding results
of B752 are shown in Figures 6m, 6n, 6o and 6p.
The band combination B762 is able to detect the major roads, and few in urban area roads using the
proposed method. Because of the asphalt and concrete reflectance are higher in bands B7, B6 and low in B2.
The S of B762 is also useful in road detection, but only a few roads are missing. These missed roads detected
by B652. Also, the B653 band combination able to extract roads, but B652 detects more roads with good
accuracy and smoothness. This can be observed in Figure 6t, 6h and 6d. Based on the experimental results and
analysis, the chosen band combination is B652 for road extraction.
One of the widely used methods for the selection band combination is the optimal index factor (OIF),
which depicts the correlation of bands that have the highest to lowest OIF band combinations. By using this, it
observed that the band combined with both SWIR1 and NIR bands have the highest OIF values. The spectral
correlation plot between NIR and SWIR bands with the corresponding optical image shown in Figure 7 by
using ENVI. From this, the water bodies have the highest correlation as well as vegetation. Also, it observed
that asphalt has a high correlation from the other features, especially in urban areas. From this, both the bands
are most useful to form the RI index.
According to OIF, the B651, B652, and B654 band combinations have the highest OIF values from
all combinations. In the two (B651 and B652) combination, the B2 band is a regular band in all types of
sensors, and B1 is a rarely using sensor in RS. Hence, the chosen band combination is B652. The various band
combinations of original images with the RI index and corresponding results of optimal threshold and SATM
shown in Figure 6.

(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
(m) (n) (o) (p)
(q) (r) (s) (t)
Figure 6. (i). OLI data sets of band combinations (a) 652 (e) 653 (i) 654 (m) 752 (q) 762 , (ii). RI of (i), (iii).
Using Optimal Threshold on RI of BH images (i) (iv) SATM
(a) (b)
Figure 7. (a) OLI B652 image (b) Spectral plot of B6 Vs. B5

1328 r ISSN: 2088-8708
4. EXPERIMENTAL RESULTS AND DISCUSSION
The proposed methodology is applied to sub-scene images of a full scenes, as mentioned in
Table 2. All these sub-scene images are multispectral band combination of B652, as proposed in
section 2.1 and based on analysis in section 3.1. The small scenes of sub-scenes Figure 8a (Chennai) and
Figure 9a (Hyderabad) chosen for proposed methodology verification, has the roads and as well as water bod-
ies. The proposed methodology consists of two methods are SATM and SMRF. One method SMRF applied
to Figure 8a, for extraction of roads, but observed that water body edges also extracted. SATM applied to
Figure 9a, also observed that water body edges are present along with roads. A method proposed for the water
body and water canal extraction [38] used to remove the water body edges from the images.
4.1. Removal of water body edges
As mentioned above, by using the proposed methodology, the water body edges are also extracted as
roads. Due to the water body (lakes, reservoirs and canals) boundaries are man-made structures, by concrete,
gravel, and stone. These water body edges removed by the proposed method MNDWI2 of water bodies in the
same images and dilated. That is, MNDWI2 [38] index used to extract the water bodies and water canals from
the same image. From this method, extracted water bodies from the images shown in Figures 8c and 9c (SR is
30 m). After extraction of these features, image interpolated by cubic convolution to the SR of 15 m and dilated
by the structure element ’disk’ with radius 2 to match with edges. Wherever water body edges present (if water
body positions are matches in both images 8b and 8c) are removed by using logical operators. Water body
edges removed image shown in Figure 8d and similar procedure applied for Figure 9. Further results shown in
subsection 4.2 used the same methodology for the removal of water bodies edges from the resultant images of
the proposed methodology.
(a) (b) (c) (d)
Figure 8. (a) OLI B652 image (Chennai) (b) SMRF (c) MNDWI2 [38] (d) Final roads
(a) (b) (c) (d)
Figure 9. (a) OLI B652 image (b) SATM (c) MNDWI2 [38] (d) Final roads
4.2. Results and discussion
The proposed methodology performed on OLI data sets given in Table 2. The sub-scenes of data sets
taken as which covers the urban areas and one sub-scene (Rajasthan) with arid type rural area. The different
steps involved in the proposed methodology resultant image shown in figures.
As mentioned in section 2.1 the proposed new multi-spectral band combination B652 sub-scene im-
ages of full scenes as shown in Table 2 are shown in Figures 10a, 11a and 12a. All the mages mentioned
in the paper are pan-sharpened by the Brovey method with SR of 15 m. These images are converted into
indexed RI image by using the 1 and are shown in Figures 10b, 11b and 12b. These results illustrate the
proposed index (RI) image, and the roads having low values at the same time differentiating from the other fea-
tures. To enhance the road feature, BH is applied and elevates road features as well as suppress other features.

Figures 10c, 11c, and 12c, illustrates the proposed filter (After applying the BH and SSF), sharpen the road
edges as well as smooth the road regions.
SSF images segmented by IAT, Otsu, and MRF, which segment as the road and non-road feature.
From these segmented images, road features alone extracted by using CCA and are shown in Figures 10d, 10e,
and 10f. As described in section 4.1 the water body’s edges also extracted as roads due to the features are like
roads and those water bodies removed from segmented images.
Figure 10g describes the proposed methodology of the combined results of SATM and SMRF. In
combining the results, we can observe that the roads little more than individuals. Also observed, roads which
are freeways and arterial roads extracted with high accuracy. Similarly, road features are extracted from the
remaining images and given in Figures 11, 12 and 13. SATM results are shown in Figures 11d and 12d.
Figures 11e and 12e represents the SMRF results. The final results (Combined images) of proposed methodol-
ogy are shown in Figures 11f and 12f.
(a) (b) (c)
(d) (e) (f)
(g)
Figure 10. (a) OLI B652 (Hyderabad) image, (b) Road index, (c) SSF, (d) SATM, (e) Optimal threshold,
(f) SMRF, (g) Proposed methodology

1330 r ISSN: 2088-8708
In OLI, pan-sharpened images, each pixel represents 15 m on the ground. Generally, road widths are
7 m (two-lane), 15 m (4 lanes), 30 m, 45 m and vary. Highway roads have 4 to 6 (or more) lanes with a varying
width from 15 m to 45 m, which have smooth curves represents one to 3 pixels in the image. District (Arterial)
roads are mostly 2 or 4 lanes with paved or unpaved roads varying width from 8 m to 20 m. Minor (residential)
roads are mostly irregular with less than 7 m width.
By using the proposed methodology results illustrates in the Figures 10, 11 and 12, and the roads width
22 m (1.5 pixels) are extracted consistently. The roads with 7 m (1/2 pixel) width are also extracted based
on contrast and separation of road and background. These images show the proposed methodology extracting
the roads with good accuracy.
Figure 13a represents a sub-scene of the rural area (residential type) roads of Rajasthan with a band
combination of B652. RI and enhanced images are shown in Figures 13b and 13c. In the arid type of conditions,
roads highlighted due to high asphalt reflectance and observed the width of the roads is varied from 7 m to 3
m shown in Figure 13. These types of roads also are extracted efficiently by the proposed method, as shown in
Figure 13d. The road widths less than 7 m also extracted in such arid areas due to obstacles not present beside
roads.
(a) (b) (c)
(d) (e) (f)
Figure 11. (a) OLI B652 (Chennai) image, (b) Road index, (c) SSF, (d) SATM, (e) SMRF,
(f) Proposed methodology
From the results, it observed that more than 15 m width of roads extracted with 100 percent, when
not present multi-storey buildings and trees beside the road. The roads width less than 30 m, which are present
in between the buildings and trees, pixel reflectance changes for roads due to having the high reflectance of
buildings and trees, and not able to identify as the road in such dense urban areas.
The observed gaps in between the roads because of not enough width of roads, not consistent and
urban area roads with less than 30 m. Gaps with less than 3 pixels in between roads connected, by opening

and closing operators of morphology used. Gaps less than 10 pixels in between roads connected by the circle
method, that is, at every endpoint placed a circle ( with diameter 10) and thinned by morphological Operation.
Where two endpoints of circles are connected, there itself connect, and remaining are left. The proposed
method with gaps filled results converted into vector form by using ArcGIS 10.2, and these results used for
validation.
(a) (b) (c)
(d) (e) (f)
Figure 12. (a) OLI B652 (Bangalore) image , (b) Road index, (c) SSF, (d) SATM, (e) SMRF,
(f) Proposed methodology
(a) (b) (c) (d)
Figure 13. (a) OLI B652 (Rajasthan)image (b) Road index (c) SSF (d) Proposed methodology
The accuracy of the proposed methodology in terms of length of extracted roads using proposed
one of method SATM and Reference lengths(manually drawn on OLI test case images with QGIS) given in
Table 3. The accuracy of SATM method for sub-scenes given in Table 4. Also, the length of extracted roads
by the proposed methodology (SATM + SMRF) in Table 5 and accuracy presented in in Table 6 for the same.
From these tables, the index based methods SATM and proposed methodology able to extract the road features
with an overall accuracy of 85%.
Table 3. Accuracy in Length (in km) of Roads for SATM
sub-scene Hyderabad Chennai Bangalore
SATM 1117 436 304
Reference 925 411 322
TP 869 390 249
FN 56 21 73
FP 248 46 55

1332 r ISSN: 2088-8708
Table 4. Accuracy of SATM
sub-scene Precision (%) Recall (%) F1-Score
Hyderabad 77.80 93.95 85.11
Chennai 89.45 94.89 92.09
Bangalore 81.91 77.33 79.55
Table 5. Length (in km) of extracted roads by proposed methodology
sub-scene Hyderabad Chennai Bangalore
SATM+SMRF 1209 463 325
Reference 925 411 322
TP 910 399 263
FN 15 12 59
FP 299 64 62
Table 6. Accuracy of proposed methodology
sub-scene Precision (%) Recall (%) F1-Score
Hyderabad 75.27 98.38 85.29
Chennai 86.18 97.08 91.30
Bangalore 80.92 81.68 81.30
4.3. Validation
Reference roads are obtained by digitizing the test images [7], to evaluate the road feature extraction
framework, at a 1:50,000 scale. This reference utilized as ground truth for an evaluation of the extracted roads.
Also, the extracted roads are converted into the vector format using ArcGIS and spikes with a length of 150
meters (10 pixels) removed. The reference layer is a buffer layer with a width of 15 m. The extracted roads fit
within the reference roads network called as true positive (TP). Similarly, false positive (FP) and false negative
(FN) calculated as in [39]. These Three metrics used to assess the accuracy of extracted road networks and
calculated precision (P), recall (R), and F1-score [40]. Overlay of extracted roads of TP on the reference roads
shown in Figure 14. Roads features extracted using RI from OLI images for the urban areas of Hyderabad,
Chennai, and Bangalore with an overall F1-score of 85%.
F1 score = 2 ·
precision · recall
precision + recall
(12)
Figure 14. Overlay of vector layers, reference and proposed method (TP only)

4.3.1. Comparison
Here, a classification algorithm i.e., pixel-based support vector machine (SVM) [41] classification
with RBF kernel and polynomial order 2 is used to extract roads. Accuracy of the proposed methodology is
comparing with the extraction of roads using the SVM on the same (B652) band combination. By using the
SVM method, only major (width) roads get detected, due to limitation of SVM misclassification. The results
of SVM and SATM images are shown in Figure 15. SVM method was able to classify the roads where roads
are clear and remaining gaps are occurred due to the misclassification with building type. The accuracy of road
feature extraction using the SVM and SATM methods given in Table 7.
(a) (b)
(c) (d)
Figure 15. (a) Original (B652), (b) SVM, (c) After Post-processing of (b), (d) SATM
Table 7. Accuracy comparison
Extracted TP FP FN P R F1
SVM 39 39 0 29 98.21 65.25 65.12
SATM 66 61 5 7 96.42 94.64 95.96
5. CONCLUSION
In this paper, we proposed a methodology for road feature extraction based on road index using IAT
and MRF (SATM and SMRF) from RS images. This methodology applied for the OLI images of the Indian
urban areas. From the results, it observed that the extracted road widths are regularly greater than 30 m and
airport runways with accurate. Also, observed that, in results, discontinuity is present in the road networks
due to insufficient road width, presence of trees, and multi-storey buildings. These small gaps in between
roads connected by the morphological operations, circle method that is, placing a circle at every endpoint of
extracted roads and thinned. The remaining large gaps are not able to connect and consider for future work.
The methodology brings advantages of extraction of the freeways and arterial roads by using OLI (LR images)
instead of the HR images. Hence, the proposed methodology with RI is producing satisfactory results from
OLI images. This methodology can extend for other sensors like Sentinel 2.
ACKNOWLEDGMENT
We wish to express our sincere gratitude to Dr. Shantanu Chowdhary, Director, National Remote
Sensing Centre for their encouragement and guidance in bringing out this publication.

1334 r ISSN: 2088-8708
REFERENCES
[1] J. Wang and Q. Zhang, “Applicability of a gradient profile algorithm for road network extraction-
sensor, resolution and background considerations,” Canadian Journal of Remote Sensing, vol. 26, no. 5,
pp. 428-439, 2000.
[2] S.Benjamin and L. Gaydos, “Spatial resolution requirements for automated cartographic road extraction,”
Photogrammetric Engineering and Remote Sensing, vol. 56, pp. 93-100, 1990.
[3] P. P. Singh and R. D. Garg, “Study of spectral reflectance characteristics of asphalt road surface using
geomatics techniques,” Int. Conf. on Advan. Comp., Commu. and Inform., pp. 516-520, 2013.
[4] M.Herald and D. A. Roberts, “Spectral characteristics of asphalt road aging and deterioration: implica-
tions for remote-sensing applications,” Applied optics, vol. 44, no. 20, pp. 4327–4334, 2005.
[5] H. Xu, “Modification of normalised difference water index (ndwi) to enhance open water features in
remotely sensed imagery,” International Journal of Remote Sensing, vol. 27, no. 14, pp. 3025-3033, 2006.
[6] S. S. Bhatti and N. K. Tripathi, “Built-up area extraction using landsat 8 oli imagery,” GIScience and
Remote Sensing, vol. 51, no. 4, pp. 445-467, 2014.
[7] S. Das, T. T. Mirnalinee, and K. Varghese, “Use of salient features for the design of a multistage frame-
work to extract roads from high-resolution multispectral satellite images,” IEEE Transactions on Geo-
science and Remote Sensing, vol. 49, no. 10, pp. 3906-3931, 2011.
[8] W. Wang, N. Yang, Y. Zhang, F. Wang, T. Cao, and P. Eklund, “A review of road extraction from remote
sensing images,” Journal of Traffic and Transportation Engineering (English Edition), vol. 3, no. 3, 2016.
[9] W. Shi, Z. Miao, and J. Debayle, “An integrated method for urban main-road centerline extraction from
optical remotely sensed imagery,” IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 6,
pp. 3359-3372, 2014.
[10] P. Gamba, F. Dell’Acqua, and G. Lisini, “Improving urban road extraction in high-resolution images
exploiting directional filtering, perceptual grouping, and simple topological concepts,” IEEE Geoscience
and Remote Sensing Letters, vol. 3, no. 3, pp. 387-391, 2006.
[11] M. Fauvel, et al., “Spectral and spatial classification of hyperspectral data using svms and morphological
profiles,” IEEE International Geoscience and Remote Sensing Symposium, pp. 4834-4837, 2007.
[12] Z. Miao, et al., “An object-based method for road network extraction in vhr satellite images,” IEEE J. of
Selected Topics in Applied Earth Observations and Remote Sensing, vol. 8, no. 10, pp. 4853-4862, 2015.
[13] D. Chaudhuri, N. K. Kushwaha, and A. Samal, “Semi-automated road detection from high resolution
satellite images by directional morphological enhancement and segmentation techniques,” IEEE J. of
Selected Topics in Applied Earth Observations and Remote Sensing, vol. 5, no. 5, pp. 1538-1544, 2012.
[14] M. O. Sghaier and R. Lepage, “Road extraction from very high resolution remote sensing optical images
based on texture analysis and beamlet transform,” IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, vol. 9, no. 5, pp. 1946-1958, 2016.
[15] Z. Miao, W. Shi, A. Samat, G. Lisini, and P. Gamba, “Information fusion for urban road extraction from
vhr optical satellite images,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote
Sensing, vol. 9, no. 5, pp. 1817-1829, 2016.
[16] T. Blaschke, et al, “Geographic object based image analysis-towards a new paradigm,” ISPRS Journal of
Photogrammetry and Remote Sensing, vol. 87, pp. 180-191, 2014.
[17] S. Dhingra and D. Kumar, “A review of remotely sensed satellite image classification,” International
Journal of Electrical and Computer Engineering, vol. 9, no. 3, pp. 1720-1731, 2019.
[18] N. Otsu, “A threshold selection method from gray-level histograms,” IEEE Transactions on Systems, Man,
and Cybernetics, vol. 9, no. 1, pp. 62-66, 1979.
[19] R. Adams and L. Bischof, “Seeded region growing,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 16, no. 6, pp. 641-647, 1994.
[20] S. L. Horowitz and T. Pavlidis, “Picture segmentation by a tree traversal algorithm,” J. ACM, vol. 23,
no. 2, p. 368-388, 1976.
[21] R. K. Reddy, et al., “Comparative analysis of common edge detection algorithms using pre-processing
technique,” International Journal of Electrical and Computer Engineering (IJECE), vol. 7, no. 5, pp. 2574-
2580, 2017.
[22] M. Basu, “Gaussian-based edge-detection methods-a survey,” IEEE Transactions on Systems, Man, and
Cybernetics, Part C (Applications and Reviews), vol. 32, no. 3, pp. 252-260, 2002.
[23] R. Taniguchi and E. Kawaguchi, “Road network extraction from landsat tm image,” Third International

Conference on Image Processing and its Applications, pp. 222-226, 1989.
[24] B. Liu, Z. Zhang, X. Liu, and W. Yu, “Edge extraction for polarimetric sar images using degenerate filter
with weighted maximum likelihood estimation,” IEEE Geoscience and Remote Sensing Letters, vol. 11,
no. 12, pp. 2140-2144, 2014.
[25] F. Baselice, G. Ferraioli, and D. Reale, “Edge detection using real and imaginary decomposition of sar
data,” IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 7, pp. 3833-3842, 2014.
[26] B. N. Saha and N. Ray, “Image thresholding by variational minimax optimization,” Pattern Recognition,
vol. 42, no. 5, pp. 843-856, 2009.
[27] S. Krishnamachari and R. Chellappa, “Multiresolution gauss-markov random field models for texture
segmentation,” IEEE Transactions on Image Processing, vol. 6, no. 2, pp. 251-267, 1997.
[28] N. Zhang and Q. Wu, “Effects of brovey transform and wavelet transform on the information capacity of
spot-5 imagery,” Proc. SPIE, 2007.
[29] A. Munsell, ”A Grammar of Color: An Introduction to the Munsell Color System,” The Strathmore Paper
Company, 1921.
[30] S. Osher and L. I. Rudin, ”Feature-oriented image enhancement using shock filters,” SIAM Journal on
Numerical Analysis, vol. 27, no. 4, pp. 919-940, 1990.
[31] L. Alvarez and L. Mazorra, “Signal and image restoration using shock filters and anisotropic diffusion,”
SIAM Journal on Numerical Analysis, vol. 31, pp. 590-605, 04 1994.
[32] Y. Zang, et al., “Joint enhancing filtering for road network ex-traction,” IEEE Transactions on Geoscience
and Remote Sensing, vol. 55, no. 3, pp. 1511-1525, 2017.
[33] J. Weickert, “Coherence-enhancing shock filters,” Pattern Recognition, Springer, pp. 1-8, 2003.
[34] R. C. Dubes, et al., “Mrf model-based algorithms for image segmentation,” Proce., 10th Int. Conference
on Pattern Recognition, vol. 1, 1990, pp. 808-814, 1990.
[35] S. L. K. Reddy, et al., “Automatic road feature extraction using mrf from landsat-8 oli images,” IEEE
Recent Advances in Geoscience and Remote Sensing : Technologies, Standards and Applications (TEN-
GARSS), pp. 15-20, 2019.
[36] C. Sujatha and D. Selvathi, “Connected component-based technique for automatic extraction of road
centerline in high resolution satellite images,” EURASIP Journal on Image and Video Processing, vol.
2015, no. 1, 2015.
[37] M. Yang, K. Kpalma, and J. Ronsin, ”A Survey of Shape Feature Extraction Techniques,” 2008.
[38] S. L. K. Reddy, et al., “A novel method for water and water canal extraction from landsat-8 oli imagery,”
ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sci-
ences, vol. XLII-5, pp. 323-328, 2018.
[39] C. Heipke, H. Mayer, C. Wiedemann, and O. Jamet, “Evaluation of automatic road extraction,” Interna-
tional Archives of Photogrammetry and Remote Sensing,pp. 47-56, 1997.
[40] D. R. Martin, C. C. Fowlkes, and J. Malik, “Learning to detect natural image boundaries using local
brightness, color, and texture cues,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 26, no. 5, pp. 530-549, 2004.
[41] D. Tuia, F. Pacifici, M. Kanevski, and W. J. Emery, “Classification of very high spatial resolution imagery
using mathematical morphology and support vector machines,” IEEE Transactions on Geoscience and
Remote Sensing, vol. 47, no. 11, pp. 3866-3879, 2009.
BIOGRAPHIES OF AUTHORS
Sama Lenin Kumar Reddy received M.Tech. degree in Digital Communications from Kakatiya
University (KU), India, in 2013, and working towards Ph.D. in Andhra University, Visakhapatnam,
India. He is working as Senior Research Fellow (SRF) at National Remote Sensing Centre (NRSC),
Indian Space Research Organization(ISRO), Hyderabad, India. His current areas of research interest
are image processing, pattern recognition, visual perception and computational intelligence.

1336 r ISSN: 2088-8708
C. V. Rao is working as Scientist-SG, Technology and Innovation officer, NRSC, ISRO, Hyderabad,
India. He received Ph.D. in image processing from JNTU Hyderabad in 2010 and also completed
Three Ph.D. thesis under his guidance (currently guiding Three students) and several M.Tech stu-
dents. He has published about 100 papers in international, national journals including conferences.
His current areas of research interest are digital image processing, visual perception, computational
intelligence and pattern recognition.
P.Rajesh Kumar received Ph.D. from Andhra University, Visakhapatnam, 2006. Currently, working
as professor at Andhra University in the department of Electronics and communication Engineering.
He has more than 20 years of teaching and 10 years of research experience. His current areas of
research interest are image processing, signal processing and antenna theory.
R.V.G.Anjaneyulu received M.Tech degree in Electronics and communication Engineering from
Osmania University, Hyderabad, India, in 2000. Presently working as Head of the special products
division, NRSC, ISRO, Hyderabad, India.
B.Gopala Krsihna received M.Tech degree from IIT Kharagpur. He is former Deputy Director of
Data Processing, Products, Archival and Web Applications Area (DPPA WAA) at NRSC, Hyder-
abad, India. He has more than 170 publications to his credit in National and International journals.
He has four software copyrights. His research interests include digital photogrammetry and mapping,
geometrical data processing for remotely sensed data, planetary data processing, image mosaicking,
stereo image analysis, and pattern matching.

An index based road feature extraction from LANDSAT-8 OLI images

More Related Content

What's hot (19)

Similar to An index based road feature extraction from LANDSAT-8 OLI images (20)

More from IJECEIAES (20)

Recently uploaded (20)

An index based road feature extraction from LANDSAT-8 OLI images