Heliport Detection Using Artificial Neural Networks
Emre Başeski
Abstract
Automatic image exploitation is a critical technology for
quick content analysis of high-resolution remote sensing
images. The presence of a heliport on an image usually
implies an important facility, such as military facilities.
Therefore, detection of heliports can reveal critical infor-
mation about the content of an image. In this article, two
learning-based algorithms are presented that make use of
artificial neural networks to detect H-shaped, light-colored
heliports. The first algorithm is based on shape analysis of
the heliport candidate segments using classical artificial
neural networks. The second algorithm uses deep-learning
techniques. While deep learning can solve difficult prob-
lems successfully, classical-learning approaches can be
tuned easily to obtain fast and reasonable results. There-
fore, although the main objective of this article is heliport
detection, it also compares a deep-learning based approach
with a classical learning-based approach and discusses
advantages and disadvantages of both techniques.
Introduction
The increased resolution and the amount of commercially
available remote sensing data are a new challenge for image
analysis. With the increase in resolution, even a small portion
of an image can contain details that are difficult for a human
operator to analyze. In addition, the increasing number of im-
aging sensors produce a massive amount of imagery data that
is impossible to analyze without intelligent image-processing
algorithms. Therefore, automatic image-exploitation algorithms
are very important for analyzing the content of this huge data.
In this work, the problem of fully automated heliport detec-
tion is discussed and two different approaches are presented.
The presence of a heliport in an image often points to an im-
portant facility, such as government buil
ities. The heliports that are studied in thi
light-colored structures with different siz
work on heliport detection was done by Ba
ş
eski (2018), with
an algorithm analyzing the shape of binary segments.
Automatic detection of artificial structures in remote sens-
ing images is an important area of research. Classical image-
processing techniques and learning-based approaches are
used to solve different problems. Extensive information on
different approaches to artificial-object detection is given by
Cheng and Han (2016) and Blaschke (2010). Classical tech-
niques (e.g., Mueller, Segl and Kaufmann 2004) often use a
combination of region- and edge-based techniques to identify
large and potentially artificial areas, whereas more recent
studies tend to use learning-based approaches (e.g., Inglada
2007; Han
et al.
2015; G.-S. Xia
et al.
2017; Zhu
et al.
2017).
Recent developments in deep learning have led to significant
breakthroughs in the field of image processing. Success in image
classification has been carried to object detection through net-
works that can also locate objects in the image. Networks such
as Fast
R-CNN
(Girshick 2015),
YOLO
(Redmon
et al.
2016), and
SSD
(Liu
et al.
2016) have achieved tremendous performance
improvements over conventional methods in object detection.
The performance problem of these networks on relatively small
objects is solved by Faster
R-CNN
(Ren
et al.
2017).
There seems to have been a significant increase in the use of
deep-learning techniques in recent years by the remote sensing
community. For instance, F. Xia and Li (2018) conducted a
study on airport detection in remote sensing data via the
SSD
algorithm.
SSD
is a network that has difficulty detecting small
objects due to the nature of its anchor-selection mechanism.
The contribution of S. Chen, Zhan, and Zhang (2018) improved
the classical
SSD
algorithm for detecting relatively small objects
in satellite images. Ying
et al.
(2018) present a survey on image
classification for remote sensing. Deep-learning architectures
such as U-Net (Ronneberger, Fischer and Brox 2015) and Dee-
pUNet (Li
et al.
2018) have also been shown to produce quite
successful results in the land use classification problem.
In this work, two different algorithms for heliport detection
are presented. The first algorithm is based on classical artifi-
cial neural networks. The second is an algorithm developed
with deep-learning techniques. The technique based on clas-
sical neural networks is mainly based on the analysis of the
shape of the heliport candidate segments. The deep-learning
technique is based on a Faster
R-CNN
(Ren
et al.
2017) archi-
tecture. Although deep learning seems to solve fairly difficult
problems successfully, classical-learning approaches can be
tuned easily to obtain fast and reasonable results. Therefore,
a side objective of this article is to compare a deep learning-
based approach with a classical learning-based approach and
discuss the advantages and disadvantages of both techniques.
Also, analysis of factors such as shape size, deformation of the
shape, and effect of contrast on heliport detection is discussed.
Algorithm
In this section, two different algorithms that find heliports in
ages are discussed. The first algorithm is
ficial neural networks. The second algo-
arning techniques. The classical neural
network is trained with a data set used for character recogni-
tion. It performs shape analysis on potential heliport candi-
date segments. In the deep learning-based algorithm, a convo-
lutional artificial neural network trained with a large number
of heliport images is used to find the location of heliports.
Automated detection of heliports is a challenging problem
due to variable shape size, changing background, pale edges,
and surrounding shapes. Also, the color of runways fades
over time, and a heliport may have different color shades.
Figure 1 presents some sample heliport images to highlight
the difficulties of the problem.
The evaluation of the presented methods was performed
on 32 images containing 66 heliports. The images were col-
lected from Google Maps and selected from different regions
of the world to include different features (different back-
grounds, image content, vegetation, etc.). By using the same
Emre Ba
ş
eski is with the Image and Video Processing Group,
Havelsan A.S., Ankara, Turkey; and Middle East Technical Uni-
versity Technopolis, Ankara, Turkey (
).
Photogrammetric Engineering & Remote Sensing
Vol. 86, No. 9, September 2020, pp. 541–546.
0099-1112/20/541–546
© 2020 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.86.9.541
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2020
541
