Ad Hoc Networks 78 (2018) 1–12

Contents lists available at ScienceDirect

Ad Hoc Networks
journal homepage: www.elsevier.com/locate/adhoc

Drone Cellular Networks: Enhancing the Quality Of Experience of

video streaming applications
Ludovico Ferranti a,b,∗, Francesca Cuomo a, Stefania Colonnese a, Tommaso Melodia b
a
Department of Information Engineering, Electronics and Telecommunications, Sapienza – University of Rome, via Eudossiana, 18, 00184 Roma, Italy
b
Department of Electrical and Computer Engineering, Northeastern University, 360 Huntington Avenue, Boston, MA 02115, US

a r t i c l e i n f o a b s t r a c t

Article history: This article addresses the problem of delay mitigation for video streaming applications in congested cellu-
Available online 12 May 2018 lar macro-cells by using a mobile micro-cell mounted on an Unmanned Aerial Vehicle (UAV). Small-scale
UAVs are at a mature stage of development and can carry lightweight commercial micro-cells with small
Keywords:
Unmanned Aerial Vehicles form factor. The mobile micro-cell is used to offload users from a congested macro-cell to optimize the
Videostreaming bandwidth usage of video streaming applications. The paper proposes algorithms and comprehensive de-
HetNets sign criteria for user offload selection (selecting what users need to be offloaded to the micro-cell) and
Drones drone positioning (selecting the position of the UAV that minimizes the network delay). The effectiveness
of the proposed criteria is evaluated through extensive performance analysis. We show that the per-
formance increases consistently in terms of bandwidth requests mitigation and average delay reduction
under different system configurations.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction and motivations
According to the Cisco Visual Networking Index (VNI) Global
Mobile Data Traffic Forecast for 2016–2021 [1], video traffic is ex-
pected to become 78% of total traffic by 2021. This growth will be
even more remarkable in mobile networks. In modern cellular net-
works, the deployment of the Radio Access Network (RAN) is de-
signed mainly based on predictions of long-term spatio-temporal
distributions of the traffic load. Consequently, fixed Base Station
(BS) locations do not provide the necessary flexibility, coverage,
or resources in unpredictable scenarios, thus impacting the per-
formance of the network. To alleviate this problem, the heteroge-
neous network (HetNet) paradigm was recently introduced in 4G
networks [2]. The core idea in HetNets is to distribute several Low
Power Nodes (LPNs), such as micro-, pico- and femto-cells, relays
and distributed antenna systems, under the wide umbrella cover-
age of a macro cell to bring resources to dead zones and increase
the network capacity in highly congested areas [3]. The introduc-
tion of HetNets in 4G systems, and in LTE-Advanced specifically,
offers new opportunities in terms of capacity improvement, macro-
cell offloading, energy saving, and better indoor coverage [3,4].
∗
Corresponding author at: Department of Electrical and Computer Engineering,
Northeastern University, 360 Huntington Avenue, Boston, MA 02115, US.
E-mail addresses: ferranti@ece.neu.edu (L. Ferranti), francesca.cuomo
@uniroma1.it (F. Cuomo), stefania.colonnese@uniroma1.it (S. Colonnese),
melodia@ece.neu.edu (T. Melodia).
https://doi.org/10.1016/j.adhoc.2018.05.003
1570-8705/© 2018 Elsevier B.V. All rights reserved.
However, some major technical challenges, e.g., self-organization,
interference, backhauling, still need an adequate solution. Besides,
the deployment of the LPNs can be expensive as it requires extra
power supply and wired network backhauling, or it is often impos-
sible in remote areas. Wireless nodes with wireless backhaul are,
for this reason, a preferable approach in many scenarios.
In parallel, recent work has proposed to deploy drones in cel-
lular networks [5,11,21] by equipping UAVs with a transceiver sta-
tion (drone-BS) to enhance infrastructure-based wireless networks
[10]. Aerial base stations can provide a cellular network with the
flexibility and agility necessary in scenarios where the traffic de-
mand is hard to predict or the capacity and the coverage cannot be
guaranteed. Examples of these cases include coverage of rural ar-
eas, assisting a congested macro-cell, natural disasters, huge public
events such as concerts or sport events, traffic jams [2,6,10]. Fur-
thermore, the size of a drone-cell, namely the coverage area rel-
ative to a drone-BS, can be adjusted by changing the UAV alti-
tude, transmission power, antenna directivity, and other parame-
ters, providing more adaptability to unstable traffic loads and un-
even user distributions. Drone-BS replacements are facilitated by
the inevitable wireless backhaul.
When considering a system using a UAV carrying an LTE micro-
cell to enhance coverage and capacity of an LTE BS (macro-cell)
aiming at maximizing the Quality of Experience (QoE) of end users
served by video streaming applications, a number of system design
issues arise. In fact, in case of fixed overall available bandwidth, al-
location by the LTE BS (macro-cell) only is challenging because of
the adoption of centralized resource allocation procedures. In this
2 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12

paper, we address a number of questions: since a drone BS has a

limited radio footprint, what is the best drone position given the
actual distribution of users? Given the drone position, how many
users should be served by the drone? And according to which cri-
terion should these users be selected? Finally, what is the best
criterion for balancing bandwidth assigned to the drone-BS versus
LTE-BS?
The structure of the paper is as follows. In Section 2, we report
a summary of related work. In Section 3, we describe the proposed
architecture. Then, in Section 4, we discuss the users’ coverage by
single and multiple drones. Then, in Section 5 we consider a single
drone and discuss criteria for drone positioning and user selection.
In Section 6 we describe the simulation scenario and we report
extensive performance results; Section 7 concludes the paper.

2. Related work
The investigation of UAVs to enhance mobile connectivity is a
rather new branch of research which has received some attention
in recent years. In [10] a study on the coverage performance of
drone small cells is conducted, deriving the optimal altitude for a
single drone. The authors also find a maximum coverage area using
two drones, deriving the optimal distance between cells.
A 3-D placement algorithm for drone cells with the goal of
maximizing the revenue of the cellular networks is proposed in
[11]. In [12], it is shown that aerial network provisioning can be
used for optimizing mobile networks in overload and outage sce-
narios. In this framework the UAVs, equipped with a cellular tech-
nology, are used to temporarily offload traffic into neighbor cells
in 4G networks. In [13], a study on the optimal altitude of aerial
platforms is presented to provide maximum radio coverage on the
ground. A mathematical model based on the statistical parameters
of the urban environment is provided. In [14], the authors focus on
the scheduling of beaconing periods to optimize the energy con-
sumption, the most critical issue for UAVs. Comprehensive surveys
on communication networks for UAVs and satellites are given in
[15] and [16]. In [17] an overview of UAV-aided wireless commu-
nications is provided, by introducing the basic networking archi-
tecture and main channel characteristics, highlighting challenges
and new opportunities. In [18], a statistical propagation model for
predicting the air to-ground path loss between the aerial platform
and the terrestrial terminal is proposed, based on properties of the
urban environment. In [19], a study on power consumption model
for a micro-cell base stations is proposed, comparing the power
consumption to fill coverage holes by using only macro-cell base
stations and by using both macro-cell and micro-cell base stations,
reduction in power consumption in the second case is shown. In
[20], the authors deal with energy efficiency and mobile target
tracking by minimizing the total energy consumption of drones
by adjusting their altitude. In [21], an analysis of the potential
of using small Unmanned-Aerial-Vehicles as wireless relays for
assisting cellular network performances is presented, as well as
a network analysis using both stochastic geometry and multi-cell
simulations results. Perumal et al. [22] addresses the problem of
providing full connectivity between disconnected ground clusters
in wireless network by appropriately placing a minimum number
of Aerial Platforms as relay nodes. In [23], the authors propose to
offload WiFi users to a LTE system, and investigate different offload
schemes according to the channel quality. In [24], the authors opti-
mize location and movement of UAVs to improve the connectivity
of a wireless network. In [26], a framework to support connectiv-
ity via UAVs as a relay network to guarantee the delivery of data
produced by WSN nodes on the ground to the system users. In
[27], a study on the utilization of low-altitude unmanned aerial
platforms equipped with base stations in future wireless networks
is conducted. The authors demonstrated the 3-D placement of one
Fig. 1. Architectural model: a macro-BS covers N users; NcDR are served by the
drone in the cluster where the drone acts as micro-BS while the others N c are
served by the BS.
drone-cell, depending on the drone’s altitude, location, transmis-
sion power, antenna directivity, type of drone, and the character-
istics of the environment. Finally, in [6] a study on the number of
drones used to cover the metropolitan area of the city of Ghent,
Belgium is provided. The authors study two different classes of
drones and develop an algorithm to continuously provide con-
nectivity. As for video streaming in cellular networks, in [25] and
[30] techniques for providing video streams at assured QoE levels
to mobile users served by a heterogeneous cellular network are
investigated. The network is composed by micro and macro base
stations and the authors define a resource allocation algorithm to
offload a subset of users from the micro to the macro base station.
3. Proposed system architecture
We consider a HetNet, i.e., a communication network that
adopts various types of access nodes. It can be defined as a net-
work with composite interworking between macro base stations,
characterized by a high transmission power (5−40 W) and small
heterogeneous low power nodes transmitting at 0.1−2 W and pro-
viding coverage ranges from tens of meters up to 1−2 km.
The considered network architecture is shown in Fig. 1. The sys-
tem consists of a macro-cell covering a wide area comprising a
certain number of mobile end-terminals. The end-terminal clients
request bandwidth to consume video streaming contents from a
video server with a defined QoE to be respected. Based on the
study presented in [7], our system leverages a UAV to increase the
performance of the HetNet and, ultimately, to enhance the video
streaming QoE of the end users. The drone transports a small base
station, specifically a LTE micro-cell, and acts as a mobile access in-
frastructure node. The mobile micro-cell is initially located at the
BS position and is moved around depending on the distribution of
the user resources demand in the entire macro-cell. We assume
that the videos are streamed to the clients by means of Hypertext
Transfer Protocol (HTTP) based streaming. In HTTP based stream-
ing, a video sequence is encoded into multiple versions each at
different qualities/rates. Then, each sequence is divided into small
video segments of a few seconds (“chunks”), which are requested
by the clients and transmitted by using the HTTP protocol.1
We assume that the clients are served either by the macro- or
by the drone-mounted micro-cell which is consistent with the LTE-
A standard. We refer to M-BS to indicate the macro BS and DR-BS
1
Several algorithms have been developed to choose the appropriate quality (ver-
sion) of video chunks that the server should forward based on the user’s terminal
typology and link condition, in order to improve the QoE, but this topic is out of
the scope of this work.
 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12
Table 1
Drone parameters [6,32].
Parameter Value
Average carrier speed 30 km/h
UAV power usage 13.0 A
UAV power capacity 17.33 Ah
UAV battery voltage 22.2 V
Flight time 30 min
Fly height 50 m
Permanency time in the coverage range 60 s
to indicate the one implemented by using the drone. The drone is
assumed to prefetch the client data, according to network aware
DASH elements [8].
The bandwidth requested by each user is estimated by taking
into account the data rate to download a chunk given its size, the
time period for downloading the chunk (denoted as τ ) and the
spectral efficiency as a function of the channel quality experienced
by the user. Specifically, the data rate needed by the ith user for
λ tronics, is assumed to be 0.1 kW as in [33]. This value considers
downloading the kth chunk without introducing delay is: ri = ατi
where λi is the video chunk size, τ is the fixed chunk duration
(set to τ =2 s), and α ∈ (0, 1) is a factor accounting for overhead
caused by retransmissions incurring at lower protocol layers [31].
The bandwidth Bi (expressed in MHz) required by the user i to
support the application layer rate ri is determined by the spec-
tral efficiency of the selected modulation/coding scheme, which in
turns depends on the channel quality as reported by the user via
the LTE parameter known as Channel Quality Indicator (CQI). As
in [30]:
ri
Bi =
ηc i
where η is the spectral efficiency in [bit/s/Hz] realized with the
CQI level c((x, y), i) that depends also on the user position. The CQI
is an integer number that takes values in the range between 1 and
15, with higher values representing better communication quality
channels (higher SNR values) and is computed (see [30]) as: c =
7
1 + 13 (SNRdB + 6 ) with
Ptx · G/APL
SNR =
Pnoise · BWs
where Pnoise represents the power spectral density of the back-
ground thermal noise, BWs the system bandwidth, Ptx the trans-
mission power, G is the antenna gain, and APL is the path loss. The
path loss between the users and the M-BS or the users and the
DR-BS can be computed based on different state-of-the-art mod-
els, such as the COST 231 Walfish-Ikegami model for urban areas
[9] and the air-to-ground path loss model, reported by Mozaffari
et al. [10].
3.1. BS equipped drone features
Here, we discuss here some feasibility issues related to the
adoption of the drone. The most critical parameter is the bat-
tery duration. A thorough analysis of battery duration in terms
of flight time (with and without payload) and maximum distance
covered is presented in this subsection. For this study, state-of-the-
art Micro Drones md4-10 0 0 model [32] was considered, because of
the high level of efficiency, strength and battery performance. The
most relevant parameters are listed in Table 1.
Focusing on the drone battery, we evaluated the power con-
sumption given in [33], calculated as
 
( m p + mv ) · v
PC = +p
370 · n · r
where mp and mv represent the drone weight and the payload
mass respectively. We consider a drone weight of 2.65 kg and a
payload, carried by the drone, of 1.2 kg. The average speed v is set
3
Table 2
Simulation parameters.
Parameter Value
Average carrier speed (Km/h) 30
UAV power usage/capacity/battery voltage 13.0 A/17.33 Ah/22.2 V
Flight time (min) 30
Fly height (m) 50
Permanency time in the coverage range (s) 60
Cell radius (Km) 2.5 (BS), 0.5 (DR)
Receiver node power/sensitivity (dBm) −97.5/−107.5 dBm
Maximum TX power (dBm) 43 (BS), 36 (DR),
Antenna gain (dB) 13 (BS), 10 dB (DR),
Operating frequency (GHz) 2.1
Bandwidth (MHz) BWBS = 35, BWDR = 5
Number of different videos 5
Number of chunks per video nc = 1200
Video encoding rates [Mbps] V1 = 1.08 V2 =1.33 V3 =1.35
V4 =1.19 V5 =1.17
to 30 km/h (8.3 m/s), and p, the power consumption of the elec-
both the micro-cell transmission power and the power consumed
by electronics, including all sensors. The parameters n and r de-
pict the power transfer efficiency for the motor propeller and the
lift-to-drag ratio, respectively, kept at 0.5 and 3, respectively, to
consider the worst case scenario. The overall power consumption
is 0.30 kW. A high-end lithium ion battery has a specific power
of 0.35 kW/kg [33], therefore a 1 kg battery is sufficient to provide
enough power. The UAV is initially stationed at the macro base sta-
tion and after its maximum flight time, the drone is forced to go
back to the macro cell where its battery will be recharged.
(1)
4. Drone assisted system design: user clustering strategy
In this Section, we introduce the user clustering to obtain a ser-
vice by means of the drone cell. To this aim, and without loss of
generality, we consider as starting point of our analysis a users
clustering by means of the κ -means algorithm, and we consider
the centroid of each cluster as a candidate for drone positioning.
For proper clustering dimensioning, we analyze the potential par-
(2)
tial coverage of the users in the BS cell achievable by means of one
or more drones. It is crucial to evaluate whether the cluster can
be fully covered by the drone micro-cell coverage area (footprint).
Figs. 2 and 3 depict two different scenarios, encompassing N = 150
randomly located users each, with an example cluster subdivision.
These two scenarios have been extracted out of hundreds of sim-
ulations as they represent a widely spread users scenario and a
closely populated one, respectively.
To find the number of cluster κ to be selected to optimize the
network performance, we need to identify how many users fall out
of the drones footprints (the circles in Figs. 2 and 3), for each and
every possible κ . In Figs. 4 and 5, in the upper plot, we show the
percentage of users falling out the κ footprints, as a function of κ ,
and in the lower plot we show the number of clusters which are
fully covered by the drone footprints. In particular, from Fig. 4 we
can notice that for Scenario A, when κ is too low, most of the users
are not actually covered by the κ drone footprints. This means that,
even if κ drones were employed, the centroids are positioned too
far apart from one each other with respect to the drone cell radius
in order to cover all the users. For κ = 3, approximately 30% of the
users are outside of the drone cell reach. Similar results are shown
in Fig. 5 for Scenario B. In a nutshell, it can be argued that the
best number of clusters is that for which the cluster spread best
(3) matches the drone cell footprint area.
As a second step of our analysis, we consider the potential re-
duction of the overall bandwidth required by the users achieved
by introducing the drone assisted service. In a set of experiments,
carried out using the settings in Table 2 and not reported here
4 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12

Fig. 2. Scenario A: widely located users.

Fig. 3. Scenario B: closely located users.

for the sake of compactness, we have studied the bandwidth gains
achievable in the limit, yet unfeasible, case when κ drones were
employed. In this limit case, each user is typically next its serv-
ing BS (located in the cluster centroid) and, as far as the coverage
increases (i.e. for sufficiently large κ ), significant bandwidth reduc-
tions are observed, down to 30−40% of the original bandwidth re-
quest. This reduction is due mainly to users of clusters far apart
from the LTE base station, which adopt robust, low-efficiency mod-
ulation/coding schemes and largely benefit of the drone position-
ing at the cluster centroid.
With this reference bound in mind, we discuss now the band-
width reduction that can be achieved when one drone cycles over
the clusters centroids and serves each of the κ cluster for a cer-
tain fraction of the time. We refer to the two above scenarios, with
κ = 5 and κ = 6 respectively, and assume that all the users which
are not within the drone footprint, either because their cluster is
not served or because, although being in the drone served cluster,
they are further away from the centroid than the drone cell radius,
are served by the LTE macro BS. Then we consider, for each users
cluster, the bandwidth that would be requested either to the macro
 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12 5

Fig. 4. Coverage analysis on Scenario A.

Fig. 5. Coverage analysis on Scenario B.

Fig. 6. Bandwidth analysis of Scenario A. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

BS (M-BS) or to the drone BS (DR-BS), under different cases: (i) the
drone serves each cluster for 1/κ of the time, depicted in Figs. 6
and 7, (ii) the drone serves the cluster for all the time, see Figs. 8
and 9. The first case is analyzed in Figs. 6 and 7; each cluster gen-
erates a bandwidth request towards the drone for 1/κ of the time
(blue bar) and a bandwidth request towards the macro BS (yellow
bar); the sum of the two components is slightly smaller than the
bandwidth that would be requested to the M-BS only (russet bar).
In Figs. 8 and 9 we show the M-BS only bandwidth (russet bar)
and the DR-BS only bandwidth (blue bar) versus the cluster index
when the drone is serving the cluster all the time. It is interesting
to observe that the bandwidth reduction is variable, depending on
6 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12

Fig. 7. Bandwidth analysis on Scenario B. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

Fig. 8. Bandwidth analysis of Scenario A. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

Fig. 9. Bandwidth analysis on Scenario B. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)
 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12 7

the average distance of the cluster users to the macro BS and to The remaining N − NcDR users of the scenario are served by the
the cluster centroid; there are clusters in which the bandwidth is (i )
base station using BWs = BWBS and the bandwidth requests BBS
k
not reduced at all.
with respect to the BS.2
From these observations it clearly stems out that, if one drone
only is considered, simply cycling over the different cluster brings
almost no advantage w.r.t. the M-BS only solution. On the contrary, 6. Simulation results
the best solution would be to stay 100% of the time over one out
of the κ clusters, selected according to a rule that should aim at To validate our approach we carried out numerical analysis
minimizing the bandwidth request. using a simulation tool in Matlab. We considered a BS covering
In the following section, we compare different methods for a hexagonal area spanning for 2.5 km. The BS serves video con-
cluster selection and address the further question of which clus- tents via HTTP streaming to a total of N = 150 users randomly dis-
ter users are to be offloaded from the M-BS to the DR-BS. tributed under its coverage area. More specifically, the users den-
sity is not uniformly distributed, but rather we generate the posi-
tions of the users around 15 randomly distributed central points,
as we suppose that people tend to gather in specific areas in real
5. Drone assisted system design: users offloading strategies
life. Users are then clustered in 4 clusters in accordance to a κ -
means algorithm simply based on their geographical coordinates
As previously mentioned, all users connected to the base sta-
in the plane. We assume this information to be available to the BS
tion are divided in clusters. The proposed approach then follows
Network Administrator. The streamed videos are taken from real
two main steps. The first aims to identify a cluster to be served in
traces in [28]. Five different videos are considered with the average
accordance to a Selection Method SM; the drone selects the cluster
requested rates reported in Table 2, where all the main parameters
presenting:
used to set up the simulator are reported. These parameters have
been selected in order to replicate as much as possible the realistic
• SM1 : the minimum average CQI;
behavior of a drone carrying a micro-cell.
• SM2 : the maximum requested bandwidth to the BS;
We consider a scenario in which the drone flies at an altitude
• SM3 the maximum number of users in the cluster.
higher than the building height. Typical building height in urban
area is about 20 m; thereby, without loss of generality, we consider
Then, once a cluster is identified, users to be offloaded from
the case of a drone flying in an urban area at an altitude 50 m,
the BS are selected as follows. Let us denote as N the set of users
covering a circular area with a diameter of 500 m on the ground,
of cardinality N in the BS radio range. Nc is the number of users
and experiencing a path loss described by a Non Line Of Sight
in the selected cluster c and Nc the set of these users. Besides,
Walfish-Ikegami model. Specifically, the adopted Walfish-Ikegami
we assume that in general only a subset of the cluster users, NcDR ,
model expresses the path loss as: P L = LF S + Lrts + Lmsd where LFS
is offloaded to the DR (to have an example of the users sets see
is the free space loss, Lrts is the rooftop to street diffraction and
Fig. 1). We consider three different criteria to identify the set of
scatter loss, and Lmsd is the multi-screen loss, as defined in [9]. In
cluster users NcDR ∈ Nc that are offloaded to the drone, namely:
the simulations we adopt the settings: building height of 20 m, av-
erage road width equal to 20 m, distance between building 36 m,
• C1 : All the Nc cluster’s users are served by the drone, i.e. NcDR ≡ receiver height 1.5 m, road orientation w.r.t. the radio path 10°.
Nc ;
• C2 : Only the subset NcDR of cluster users having BS-related CQI
below a threshold θ CQI are served by the drone and the others 6.1. Cluster selection method
remain connected to the M-BS;
• C3 : Only the subset NcDR users with bandwidth requests toward Firstly we analyze the effect of the 4 different offloading crite-
the M-BS greater then a threshold θ BW are served by the drone ria as a function on the method that can be applied to select the
and the others stay connected to the M-BS. cluster where the drone moves to serve the users.
The scenario is generated by considering 15 random central
points and generating 10 users around each central point. The
Let us denote as BDR the requested bandwidth to the drone and
users are then clusterized according to their locations using a κ -
BBS the one requested to the M-BS. The bandwidth requests origi-
means algorithm with 4 clusters, and the minimum mean CQI
nated by the set of offloaded users derived in accordance to one of
measured in the cluster is selected for offloading. The thresholds
the Cl , l = 1, 2, 3 criteria are served by the drone. In more detail,
for offloading of the C2 and C3 criteria are set to θCQI = 10 and
for each chunk the DR bandwidth is allocated in accordance to the
algorithm Dynamic Minimum Average Delay (D-MAD), proposed in
θBW = 1800 KHz, respectively. Without loss of generality we as-
sume both for the M-BS and the DR-BS the bandwidth allocation
[29], with NcDR offloaded users jointly allocated on an overall band-
mechanism D-MAD in [29]. This scheme allocates bandwidth dy-
width of BWs = BWDR . The rest of the users in the set N NcDR is
namically chunk by chunk and jointly for all users and it mini-
served by the BS with BWs = BWBS .
mizes the average transmission delays experienced by the users.
The case when all users request the bandwidth to the BS, so no
In order to analyze the performance of the different schemes,
cluster users are served by the drone, is indicated as C0 , i.e., all the
the QoE metric is defined as the average per chunk delay. Here we
N users are served by the M-BS.
consider 20 runs and for each run the delay δk(i ) is averaged over
Based on the criterion applied, we compute for the ith user the
(i ) the overall number of chunks and on different set of users under
bandwidth requests BDR with respect to the drone micro-cell, allo-
k the lth Cl criterion. Thus, the average has been computed as:
cate the overall drone bandwidth BWDR and compute the assigned
bandwidth B˜k(i ) . (S )  δk(i)
The kth chunk delay experienced by the ith user can be evalu- δ (Cl ) = (5)
nc · |S |
ated as [31]: i∈S k

  
Bki 2 (i )
δi = max −1 · τ, 0 (4) Notice that BWDR depends on the CQI computed as a function of the SNR to the
˜ (i )
Bk
(i )
drone, while BBS depends on the CQI computed as a function of the SNR to the BS.
8 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12

2000 2000
C0: BDR =0 MHz, BBS ≈ 60.59 C0: BDR =0 MHz, BBS ≈ 61.41

C1 BDR =3.9 MHz, BBS ≈ 42.58 C1 BDR =5 MHz, BBS ≈ 33.4

1500 1500
C2 BDR =3.3 MHz, BBS ≈ 43.74 C2 BDR =4.5 MHz, BBS ≈ 34.15
1000 C3 BDR =3.6 MHz, BBS ≈ 43.05 1000 C3 BDR =4.8 MHz, BBS ≈ 33.75

500 500
m

m
0 0

-500 -500

-1000 -1000

-1500 -1500

-2000 -2000

-2500 -2000 -1500 -1000 -500 0 500 1000 1500 2000 2500 -2500 -2000 -1500 -1000 -500 0 500 1000 1500 2000 2500
m m

(a) SM1 : Minimum CQI (b) SM2 : Maximum bandwidth

2000
C0: BDR =0 MHz, BBS ≈ 54.03

C1 BDR =5.9 MHz, BBS ≈ 27.73

1500
C2 BDR =4.8 MHz, BBS ≈ 29.46
1000 C3 BDR =5.6 MHz, BBS ≈ 28.85

500
m

-500

-1000

-1500

-2000

-2500 -2000 -1500 -1000 -500 0 500 1000 1500 2000 2500
m

(c) SM3 : Maximum number of users

Fig. 10. Different scenarios resulting from the three cluster selection methods SMs; comparison of the resulting requested bandwidths to the DR and to the M-BS as for the
three offloading criteria (C1 , C2 , C3 case of no offloading C0 ).

where the average is over the overall number of chunks nc and
over the users’ set S of cardinality |S |. Different sets of users
have been considered, namely, (i) all the cell users in N ; (ii) se-
lected cluster drone offloaded users in NcDR , (iii) selected cluster
BS served users, in Nc NcDR ; iv) out of cluster users in N NcDR .
Furthermore, for each run, the ratio between the overall aver-
age delay achieved in the M-BS+DR configuration and the overall
average delay achieved using only the M-BS has been calculated.
Namely, the Average Delay Reduction Factor (ADRF) is defined as:
(N )
δ (Cl )
ADRF (Cl ) = (N )
, l = 1, 2, 3
δ (C0 )
where N denotes the set of all the users.
Three snapshots of the scenarios of a simulation run for the
three SMs are reported in Fig. 10. We can notice main differences
in the adopted criteria: the CQI-based SM1 criterion basically seeks
for the farthest cluster, with users experiencing bad channel con-
ditions; the BW-based SM2 accounts jointly for the channel qual-
ity and the application layer bandwidth requests; the population
based one, SM3 , is the simplest and fastest criterion, that requires
minimal knowledge about the user characteristics. Criterion SM2
maximally relieves the M-BS from bandwidth requests but requires
maximal knowledge about the user characteristics, including the
actual throughput rate. The SM1 criterion represents a trade-off
among simplicity, prior knowledge and efficacy in reducing the
bandwidth requests to the BS.
6.2. Offloaded users selection criterion
In the following, we consider as selection method the SM1 . In
(6)
Fig. 11 we present the scatter plot of the per-run average delay vs
the corresponding CQI on different set of users and under different
criteria. Specifically, in Fig. 11(a) in green we plot the delay aver-
aged over the nc chunks and on the set I of all the users when
no user is offloaded. Besides, we show the average delay over dif-
ferent users’ subsets, namely the cluster users served by the DR
(blue) and the out-of cluster users (red). In Fig. 11(b)–(d) we show
in green the corresponding average delay when drone offloading
is performed using the C1 − C3 criteria, respectively. Notice that in
 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12 9

0.5
N 0.5 N
(BS) (DR)
0.4 NC NC
(BS) 0.4 N \ NC
N \ NC
Average Delay

Average delay
0.3 0.3

0.2 0.2

0.1 0.1

0 0
5 10 15 5 10 15
CQI CQI

(a) C0 (b) C1

0.5 0.5
N N
(DR)
0.4 NC 0.4 NC
(DR)
(BS)
NC (BS)
Average delay

Average delay NC
0.3 N \ NC 0.3 N \ NC
0.2 0.2

0.1 0.1

0 0
5 10 15 5 10 15
CQI CQI

(c) C2 (d) C3
Fig. 11. Average delay versus CQI per run for different offloading criteria. (For interpretation of the references to color in this figure, the reader is referred to the web version
of this article.)

Fig. 11(c) and (d) referring to case C2 and C3 we further distin- 2.5
guish between DR served cluster users, still in blue, and BS served C1
cluster user, in black. The proposed offloading architecture system- C2
atically improves the performance in terms of average delay for all 2
C3
the C1 − C3 criteria. Still, the criteria require an increasing knowl-
edge of the users characteristics, ranging from users location to the 1.5
ADRF

knowledge of the application layer user bandwidth. We recognize

that C1 does not distinguish between different users in the cluster,
C2 offloads and improves the performance of the low CQI users, 1
C3 accounts for the requested bandwidth. In medium load condi-
tions, which is the case addressed in Fig. 11, averaging the delays 0.5
(N )
δ (Cl ), l = 0, 1, 2, 3 in Fig. 11 over all the runs yields mean de-
lay values of 0.438, 0.154, 0.150, 0.162, respectively. Thereby, the
C2 criterion offers interesting performance. In high load conditions 0
0 2 4 6 8
this performance ranking changes, as discussed in the following.
BWDR(MHz)
In Fig. 12 we finally plot the ADRF defined in Eq. (6). We
recognize that the maximum performance improvement is found
Fig. 12. Average delay reduction factor versus the drone bandwidth.
for offloaded bandwidth approximating the drone bandwidth, as
shown in Fig. 12 and the same behavior is true if we replace in
the x-axis the number of clusters users. This confirms the intu- As final results in Table 3 we show some performance referred
itive conjecture that the cluster formation can aim at achieving to different number of users. Great advantage is always found by
suitably populated clusters exploiting the majority of the drone the drone offloading; still we observe that the offloading criteria
bandwidth. differently perform depending on the overall cell load. We can no-
10 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12
Table 3
(I )
Mean value of the average delay δ (Ci ) for different num-
ber of users.
BS C0 BS+DR C1 BS+DR C2 BS+DR C3 N ing in the software setup and for providing simulations. The work
0.6782 0.23543 0.24525 0.2401 150
0.37586 0.14607 0.14754 0.15695 120
0.14523 0.041321 0.04001 0.042024 100
tice that for low load (low N) both the C2 and C3 attain a per-
formance improvement; this occurs without overloading the drone
bandwidth. On the contrary when the load is high, the C1 criterion
better relieves the BS.
6.3. Key take-aways
After the previous analysis we here provide some answers to
the questions raised in the introduction. Firstly, since a drone BS
has a limited radio footprint (500 m in our study), the best drone
positioning given the actual users distribution is related to the
cluster that presents a dense population but is located not too
close to the BS. If this conditions are not met, the disadvantage
of dividing the bandwidth between the drone and the BS is not
compensated by the advantage of having better CQIs with respect
to the BS. Obviously, in the design of such an architectural model
also the tradeoff of putting a drone far form the M-BS (if we de-
sign that the drone starts its flight from the M-BS position) should
be taken into account in terms of energy cost to reach such a po-
sition. Secondly, given a suitable drone positioning, the advantage
of using the drone is always evident; nonetheless, the different
criteria on how many users should be served by the drone may
change the attained performance depending on the size informa-
tion available on the offloaded users. Thirdly, the tuning of the
drone-BS versus LTE-BS bandwidth assignment should take into
account of the spatial user distribution, both in terms of loca-
tion and densities, since dense cluster of users located far apart
from the LTE-BS definitely benefit of the devoted drone-BS sup-
port; the expected performance improvement is maximum when
the drone-BS supports for the average users requests plus a mar-
gin to account for video rate random fluctuations. This paves the
way to an optimal allocation problem which is left for further
study.
7. Conclusions
In this paper we consider an Heterogeneous Network where
a UAV is used as mobile LTE micro-cell performing user offload-
ing and optimizing the bandwidth usage of a macro-cell during
video streaming. After dimensioning of the UAV cell, we con-
sider a video streaming service offered to all the users, and par-
tition the users into clusters within the drone coverage area.
Then, we analyze the impact of offloading the cluster users gen-
erated bandwidth requests to the drone, by considering differ-
ent methods for selecting the cluster to be served by the drone
and different criteria for selecting the offloaded users within clus-
ter. We show that, even though independent allocation of the
DR and M-BS bandwidths is sub-optimal in principle, offloading
yields significant improvement as far as the drone serves clus-
ters characterized by bad channel quality and selectively offload
users within the served cluster. Future work will be dedicated to:
(i) combine the clustering and the cluster selection mechanisms
with a suitable offloading criteria; (ii) plan the drone flight un-
der a trajectory constrained by the drone battery and fly char-
acteristics but optimized in order to achieve an high offloading
gain.
Acknowledgments
The authors would like to thank Andrea Iocolano for contribut-
of Ferranti and Melodia was supported in part by the US National
Science Foundation under grant CNS-1618727.
References
[1] Cisco Visual Networking Index: Global Mobile Data Traffic Forecast Up-
date, 2016–2021 White Paper. http://www.cisco.com/c/en/us/solutions/
collateral/service- provider/visual- networking- index- vni/mobile- white- paper-
c11-520862.html.
[2] I. Bor-Yaliniz, H. Yanikomeroglu, The new frontier in ran heterogeneity: multi-
-tier drone-cells, IEEE Commun. Mag. 54 (11) (2016) 48–55.
[3] A. Ghosh, N. Mangalvedhe, R. Ratasuk, B. Mondal, M. Cudak, E. Visotsky,
T.A. Thomas, J.G. Andrews, P. Xia, H.S. Jo, et al., Heterogeneous cellular net-
works: From theory to practice, IEEE Commun. Mag. 50 (6) (2012) 54–64.
[4] D. Lopez-Perez, I. Guvenc, G. De la Roche, M. Kountouris, T.Q. Quek, J. Zhang,
Enhanced intercell interference coordination challenges in heterogeneous net-
works, IEEE Wirel. Commun. 18 (3) (2011) 22–30. Jun.
[5] V. Sharma, K. Srinivasan, H. Chao, K. Hua, W. Cheng, Intelligent deployment of
UAVs in 5g heterogeneous communication environment for improved coverage,
J. Netw. Comput. Appl. 85 (2017) 94–105.
[6] M. Deruyck, J. Wyckmans, L. Martens, W. Joseph, Emergency ad-hoc networks
by using drone mounted base stations for a disaster scenario, in: Proceedings
of the Twelfth IEEE International Conference on Wireless and Mobile Comput-
ing, Networking and Communications (WiMob).
[7] F-cell Technology from Nokia Bell Labs Revolutionizes Small
Cell Deployment by Cutting Wires, Costs and Time, http://www.
nokia.com/en_int/news/releases/2016/10/03/f- cell- technology- from-
nokia- bell- labs- revolutionizes- small- cell- deployment- by-cutting-
wires- costs- and- time.
[8] G. Cofano, L. De Cicco, T. Zinner, A. Nguyen-Ngoc, P. Tran-Gia, S. Mascolo,
Design and experimental evaluation of network-assisted strategies for HTTP
adaptive streaming, in: Proceedings of the Seventh International Conference
on Multimedia Systems, MMSys’16.
[9] Correia, M. Luis, A view of the COST 231-Bertoni-Ikegami model, in: Proceed-
ings of the Third IEEE European Conference on Antennas and Propagation, Eu-
CAP 2009, 2009.
[10] M. Mozaffari, W. Saad, M. Bennis, M. Debbah, Drone small cells in the clouds:
design, deployment and performance analysis, in: Proceedings of the 2015 IEEE
Global Communications Conference (GLOBECOM), 2015.
[11] Bor-Yaliniz, R. Irem, A. El-Keyi, H. Yanikomeroglu, Efficient 3-D placement of
an aerial base station in next generation cellular networks, in: Proceedings of
the 2016 IEEE International Conference on Communications (ICC), 2016.
[12] S. Rohde, C. Wietfeld, Interference aware positioning of aerial relays for cell
overload and outage compensation, in: Proceedings of the 2012 IEEE Vehicular
Technology Conference (VTC Fall), 2012, pp. 1–5.
[13] A.-H. Akram, S. Kandeepan, S. Lardner, Optimal LAP altitude for maximum cov-
erage, IEEE Wirel. Commun. Lett. 3.6 (2014) 569–572.
[14] K. Sara, E. Sabir, T. Taleb, M. Aziz, A green strategic activity scheduling for UAV
networks: a sub-modular game perspective, IEEE Commun. Mag. 54 (5) (2016)
58–64.
[15] A. Torsten, K.A. Hummel, A.P. Schoellig, E. Yanmaz, M. Asadpour, C. Bettstetter,
P. Grippa, H. Hellwagner, S. Sand, S. Zhang, Application-driven design of aerial
communication networks, IEEE Trans. Aerosp. Electron. Syst. 50 (4) (2014)
2715–2735.
[16] S. Kandeepan, K. Gomez, L. Reynaud, Aerial-terrestrial communications: terres-
trial cooperation and energy-efficiently transmissions to aerial base stations,
IEEE Trans. Aerosp. Electron. Syst. 50 (4) (2014) 2715–2735.
[17] Z. Yong, R. Zhang, T.J. Lim, Wireless communications with unmanned aerial
vehicles: opportunities and challenges, IEEE Commun. Mag. 54 (5) (2016)
36–42.
[18] Al-Hourani, Akram, S. Kandeepan, A. Jamalipour, Modeling air-to-ground path
loss for low altitude platforms in urban environments, in: Proceedings of the
2014 IEEE Global Communications Conference, 2014.
[19] D. Margot, E. Tanghe, W. Joseph, L. Martens, Modelling the energy efficiency
of microcell base stations”, in: Proceedings of the First International Confer-
ence on Smart Grids, Green Communications and IT Energy-Aware Technolo-
gies (ENERGY-2011), IARIA, 2011.
[20] D. Zorbas, T. Razafindralambo, F. Guerriero, Energy efficient mobile target
tracking using flying drones, Proc. Comput. Sci. 19 (2013) 80–87. Elsevier.
[21] G. Weisi, C. Devine, S. Wang, Performance analysis of micro unmanned air-
borne communication relays for cellular networks, in: Proceedings of the Ninth
IEEE International Symposium on Communication Systems, Networks and Dig-
ital Signal Processing (CSNDSP), 2014.
[22] S. Perumal, J.S. Baras, C.J. Graff, D.G. Yee, Aerial platform placement algorithms
to satisfy connectivity, capacity and survivability constraints in wireless ad-hoc
networks, in: Proceedings of the 2008 IEEE Military Communications Confer-
ence, MILCOM, IEEE, 2008.
[23] C. Qimei, G. Yu, A. Maaref, G.Y. Li, A. Huang, Rethinking mobile data offloading
for LTE in unlicensed spectrum, IEEE Trans. Wirel. Commun. 15 (7) (2016). July.
 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12
[24] Z. Han, A.L. Swindlehurst, K.J.R. Liu, Optimization of MANET connectivity via
smart deployment/movement of unmanned air vehicles, IEEE Trans. Veh. Tech-
nol. 58 (7) (2009) 3533–3546.
[25] E.P. de Freitas, T. Heimfarth, I.F. Netto, C.E. Lino, C.E. Pereira, A.M. Ferreira,
F.R. Wagner, T. Larsson, UAV relay network to support WSN connectivity, in:
Proceedings of the 2010 IEEE International Congress on Ultra Modern Telecom-
munications and Control Systems and Workshops (ICUMT), 2010.
[26] B.Y. Irem, H. Yanikomeroglu, The new frontier in RAN heterogeneity: multi-tier
drone-cells, IEEE Commun. Mag. 54 (11) (2016). November.
[27] I. Rubin, S. Colonnese, F. Cuomo, F. Calanca, T. Melodia, Mobile HTTP-based
streaming using flexible LTE base station control, in: Proceedings of the Six-
teenth IEEE International Symposium on World of Wireless, Mobile and Mul-
timedia Networks (WoWMoM), 2015.
[28] P. Seeling, M. Reisslein, Video transport evaluation with h.264 video traces,
IEEE Commun. Surv. Tutor. 14 (2012) 1142165. Fourth.
11
[29] S. Colonnese, F. Cuomo, T. Melodia, I. Rubin, A cross-layer bandwidth alloca-
tion scheme for HTTP-based video streaming in LTE cellular networks, IEEE
Commun. Lett. 21 (2) (2017) 386–389.
[30] S. Colonnese, V. Salvatore, L. Chiaraviglio, F. Cuomo, Dynamic and cooperative
mobile video streaming across heterogeneous cellular networks, in: Proceed-
ings of the Seventeenth IEEE International Symposium on A World of Wireless,
Mobile and Multimedia Networks on World of Wireless, Mobile and Multime-
dia Networks (WoWMoM).
[31] S. Colonnese, S. Russo, F. Cuomo, T. Melodia, I. Rubin, Timely delivery versus
bandwidth allocation for DASH-based video streaming over LTE, IEEE Commun.
Lett. 20 (3) (2016) 586–589.
[32] https://www.microdrones.com/en/mdaircraft/md4-10 0 0/.
[33] R. D’Andrea, Guest editorial can drones deliver? IEEE Trans. Autom. Sci. Eng.
11 (3) (2014) 647–648.
12 L. Ferranti et al. / Ad Hoc Networks 78 (2018) 1–12

Ludovico Ferranti received his Master of Science in Engineering in Artificial Intelligence and Robotics (110/110) at Sapienza – University of Rome
in 2015. He is a Ph.D. candidate at Northeastern University of Boston, M.A. at Graduate School of Engineering and Sapienza – University of Rome
at DIET (Dipartimento di Ingegneria Elettronica e Telecomunicazioni). His research interests include Unmanned Aerial Vehicles Networks, Software
Defined Robotics Networks and Wireless Sensor Networks.

Francesca Cuomo received her “Laurea” degree in Electrical and Electronic Engineering in 1993, magna cum laude, from the University of Rome
Sapienza. She earned the Ph.D. degree in Information and Communications Engineering in 1998 (Univ. Rome Sapienza). From 2005 she is Associate
Professor at the University of Rome Sapienza and teaches courses in Telecommunication and Network Infrastructures. She has advised numerous
master students in computer engineering, and has been the advisor of 9 Ph.D. students in Networking. Her current research interests focus on:
Vehicular networks and Sensor networks, Low Power Wide Area Networks and IoT, Cognitive Radio Networks, Multimedia Networking, Energy
saving in the Internet and in the wireless system. She has participated in several National and European projects on wireless network systems
such as the RAMON, VICOM, INSYEME, IST WHYLESS, IST EPERSPACE, IST CRUISE, H2020 symbIoTe. She has authored over 120 peer-reviewed
papers published in prominent international journals and conferences. She is IEEE Senior Member. She has been in the editorial board of Computer
Networks (Elsevier) and now is member of the editorial board of the Ad-Hoc Networks (Elsevier) and IEEE Transactions on Mobile Computing. She
has been the TPC co-chair of several editions of the ACM PE-WASUN workshop, TPC Co-Chair of ICCCN 2016, TPC Symposium Chair of IEEE WiMob
2017, Guest Editor of the Ad-Hoc Networks Special issue on Modeling and Performance Evaluation of Wireless Ad-Hoc Networks, 2015, Leader Guest
Editor of the special issue on Green Communications and Networking for IoT, Sensors MPDI, 2018.

Stefania Colonnese is currently Assistant Professor at the Department of Information Engineering, Electronics and Telecommunications (DIET) of
Sapienza University of Rome, Italy. She received her Ph.D. in Electronics Engineering from the University of Roma Tre. She has participated in
the MPEG-4 standardization activity within the ISO MPEG-4 Core Experiment on Automatic Video Segmentation. She is co-author of more than
a hundred journal and conference papers, two book chapters and several ISO MPEG-4 Contributing Documents. Her research interests range from
statistical signal processing, image deconvolution and restoration, biomedical signal processing, to video encoding, processing and networking. She
has served in several conferences as Technical Program Co-chair (IEEE/Eurasip 5th European Workshop on Visual Information EUVIP 2014), Publicity
Chair (Eurasip 26th European Signal Processing Conference – EUSIPCO 2018, IEEE Workshop on Biometric Measurements and Systems for Security
and Medical Applications – IEEE BioMS 2014), Technical Program Committee member (Globecom 2018 MWN, WF-5G’18, IEEE COMNETSAT 2018,
ICA3PP, ISIVC to cite a few). She has served as Associate Editor of the Hindawi International Journal of Digital Multimedia Broadcasting (IJDMB),
devoted to the topics of Multimedia Broadcasting, Standardization, and Quality of Experience. She is IEEE Senior Member.

Tommaso Melodia is Associate Professor with the Department of Electrical and Computer Engineering at Northeastern University, where he directs
the Wireless Networks and Embedded Systems Laboratory. He received the Ph.D. in Electrical and Computer Engineering from the Georgia Institute
of Technology in 2007. He received his M.S. in Telecommunications Engineering and Doctorate from the University of Rome “La Sapienza”, Rome,
Italy, in 2001 and 2005. He is an Associate Editor for IEEE Transactions on Wireless Communications, IEEE Transactions on Mobile Computing, IEEE
Transactions on Multimedia, and Computer Networks. His research is currently supported by several grants from the National Science Foundation,
the Air Force Research Laboratory, the Office of Naval Research, and local and national industrial partners. He is an IEEE Fellow and received a
National Science Foundation CAREER award, and coauthored a paper that was recognized as the “Fast Breaking Paper in the field of Computer
Science” by Thomson ISI Essential Science Indicators and a paper that received the “Elsevier Top Cited Paper Award”. His research interests are
in modeling, optimization, and experimental evaluation of wireless networks, with applications to Internet of Medical Things, multimedia sensor
networks, underwater networks, cognitive and cooperative networks.