1

Ensuring Reliable Connectivity to

Cellular-connected UAVs with Up-tilted Antennas
and Interference Coordination
Md Moin Uddin Chowdhury, İsmail Güvenç, Walid Saad, and Arupjyoti Bhuyan

Abstract—To integrate unmanned aerial vehicles (UAVs) in the network capacity and efficiency [1]–[4]. While UAVs were
arXiv:2108.05090v2 [eess.SP] 21 Dec 2021

future large-scale deployments, a new wireless communication originally developed for military applications, due to their
paradigm, namely, the cellular-connected UAV has recently fluid mobility, line-of-sight (LOS) transmission, and steadily
attracted interest. However, the line-of-sight dominant air-to-
ground channels along with the antenna pattern of the cellular decreasing production costs, UAVs have been widely used
ground base stations (GBSs) introduce critical interference issues in various new civilian applications, such as packet delivery,
in cellular-connected UAV communications. In particular, the search and rescue, video surveillance, aerial photography,
complex antenna pattern and the ground reflection (GR) from airborne communications, among others [5]–[8].
the down-tilted antennas create both coverage holes and patchy However, most commercial UAVs acting as aerial users are
coverage for the UAVs in the sky, which leads to unreliable
connectivity from the underlying cellular network. To overcome still dependent on the instructions/maneuvers sent to them by
these challenges, in this paper, we propose a new cellular archi- their associated ground pilots through simple direct point-to-
tecture that employs an extra set of co-channel antennas oriented point communications. More specifically, this, in turn, limits
towards the sky to support UAVs on top of the existing down- the UAV use cases to the visual or radio LOS range only.
tilted antennas for ground user equipment (GUE). To model Thus, to take full advantage of large-scale UAV deployment,
the GR stemming from the down-tilted antennas, we propose a
path-loss model, which takes both antenna radiation pattern and beyond visual line of sight (BVLOS) UAV operations are
configuration into account. Next, we formulate an optimization of critical importance where the UAVs can reliably obtain
problem to maximize the minimum signal-to-interference ratio command and control (C&C) communication in the downlink
(SIR) of the UAVs by tuning the up-tilt (UT) angles of the up-tilted (DL) for safe autonomous operations. In light of such require-
antennas. Since this is an NP-hard problem, we propose a genetic ments, existing cellular networks can be a strong candidate
algorithm (GA) based heuristic method to optimize the UT
angles of these antennas. After obtaining the optimal UT angles, for deploying autonomous UAVs in BVLOS scenarios with
we integrate the 3GPP Release-10 specified enhanced inter- their widespread footprints [2], [9]. In fact, field trials from
cell interference coordination (eICIC) to reduce the interference separate industrial entities reported that the existing long-
stemming from the down-tilted antennas. Our simulation results term evolution (LTE) network is capable of meeting some
based on the hexagonal cell layout show that the proposed basic requirements of UAV-ground communications [2], [10].
interference mitigation method can ensure higher minimum SIRs
for the UAVs over baseline methods while creating minimal However, these studies and the Third Generation Partnership
impact on the SIR of GUEs. Project (3GPP) also pointed out several challenges such as
strong inter-cell interference and service of UAVs through
Index Terms—3GPP, advanced aerial mobility (AAM), antenna
radiation, drone corridor, enhanced inter-cell interference coor- antenna side lobes, among others. These challenges come
dination (eICIC), genetic algorithm, ground reflection, hexagonal into play due to the fact that traditional cellular networks
cell layout, interference, unmanned aerial vehicle (UAV), un- are optimized for ground user equipment (GUE) by tilting
manned aircraft system (UAS), UAS traffic management (UTM), the main lobe of the antennas towards the GUEs. Hence,
urban air mobility (UAM). UAVs flying in the sky are only served by the upper antenna
side lobes and experience abrupt signal fluctuations as the
I. I NTRODUCTION UAVs change their locations. Moreover, UAVs also obtain
more frequent LOS channels than GUEs. This results in severe
As the development of the fifth-generation (5G) and beyond interference in the DL from the nearby ground base stations
wireless networks is underway, unmanned aerial vehicles (GBSs) to the UAVs.
(UAVs) are expected to play an instrumental role in improving The down-tilted antennas of the existing GBSs can also
create another source of interference for the UAVs through the
M.M.U. Chowdhury and İ. Güvenç are with the Department of Electrical
and Computer Engineering, North Carolina State University, Raleigh, NC reflected signal from the down-tilted antennas [11]. The main
27606 (e-mail: {mchowdh,iguvenc}@ncsu.edu). lobe of the antenna hits the ground with an incident angle
W. Saad is with the Wireless@VT, Electrical and Computer Engineering and the reflected signal can cause non-trivial interference to
Department, Virginia Tech, VA 24060 (e-mail: walids@vt.edu).
A. Bhuyan is with the Idaho National Laboratory (INL), Idaho Falls, ID the UAVs flying in the sky. The non-trivial impact of ground
83402 (e-mail: arupjyoti.bhuyan@inl.gov). reflection (GR) at millimeter-wave (mmWave) bands is also
This work has been supported by NSF grants CNS-1453678, CNS-1910153, discussed in [12], [13], where authors introduce the concept
CNS-1909372, as well as by Idaho National Laboratory Directed Research
Development (LDRD) Program under DOE Idaho Operations Office Contract of co-channel up-tilted and down-tilted antennas for serving
DEAC07-05ID14517. UAVs and GUEs in the mmWave domain. Their ray-tracing-
 2

TABLE I
L ITERATURE REVIEW.

Ref. Goal Interference mitigation Antenna radia- up-tilted GR Co-channel UAV

technique tion pattern antenna & GUE
[5] Performance analysis of UAVs 7 directional, array 7 7 7
considering 3D antenna radiation
[14] Provide reliable connectivity and Cooperative transmission directional, array 7 7 7
mobility support for UAVs among GBSs
[15] Simultaneous content delivery to MIMO conjugate beam- directional, array 7 3
GUEs and UAVs forming
[16] Mitigate the strong downlink inter- Cooperative beamforming directional, array 7 7 7
ference to UAVs
[17] Intelligent GBS association for Choosing the best GBS by directional, array 7 7 7
UAVs based on network informa- supervised learning
tion
[18] Maximize the coverage probability Optimizing UAV-BS loca- directional, 7 7 3
and fifth-percentile rate in hetnet tions and ICIC parameters single
using exhaustive search
[19] To reduce disconnectivity time, Finding the optimal UAV directional, array 7 7 7
handover rate, and energy con- velocity by RL
sumption of UAV
[20] Serve both GUEs and UAVs si- Finding the ideal tilting directional, array 7 7 3
multaneously in a co-channel sub-6 angle by RL
GHz network
[21] To ensure robust wireless con- NA directional, array 7 7 7
nectivity and mobility support for
UAVs
[22] Maximize aircraft user throughput Bidirectional deep learn- directional, array 3 7 7
by tuning ISD and UT angles ing
[12] Serve both GUEs and UAVs simul- Finding the ideal tilting directional, 3 3 3
taneously in a co-channel mmWave angle of a single GBS by single
network ray-tracing
This Maximize the minimum UAV SIR Tuning the UT angles by directional, array 3 3 3
work GA

based simulations captured the impact of the angular separa- along with existing down-tilted antennas for GUEs. Our key
tions between these two antennas on the coverage performance contributions can be summarized as follows:
of the network. However, the authors did not consider the • We first introduce and study a new cellular concept to
presence of multiple GBSs in their work. The presence of increase the coverage of cellular-connected UAVs. As
separate co-channel up-tilted antenna sets can help network mentioned earlier, we propose to use extra antennas with
providers to ensure a high signal-to-interference ratio (SIR) UT angles installed on top of the existing down-tilted
for cellular-connected UAVs. However, proper adjustment of antennas for the GUEs. To the best of our knowledge,
the up-tilt (UT) angles is of critical importance since these there are only limited studies in the literature for such
extra antennas can create strong LOS interference towards an architecture [12], [13]. The antenna sets use the
the UAV-GBS links of the network [12]. The works in [1], same time and frequency resources as the existing down-
[2] also suggested such dedicated up-tilted cells for serving tilted antennas. However, they focus their main beams
the UAVs; however, to the best of our knowledge, no prior towards the sky to provide a more efficient and reliable
work considers the problem of tuning the up-tilted antennas connectivity to the UAVs.
for obtaining better UAV SIR performance in a multi-GBS • Unlike other previous works, in our proposed architec-
scenario. ture, we also consider the presence of GR stemming from
Note that, in such a two-antenna setup, the down-tilted the down-tilted antennas while considering the antenna
antennas create interference to the UAVs by antenna side lobes radiation pattern of the down-tilted antennas. To represent
and the GR. Moreover, the down-tilt (DT) angles of the down- the impact of antenna directivity, we modify the GR-
tilted antennas can impact the DL performance of the GUEs as based path-loss model introduced in [11] to capture the
they can be tuned to mitigate the inter-GBS interference for impact of the antenna directivity. Depending on the DT
GUEs. Hence, it may not always be possible or convenient angles of the down-tilted antennas, our analysis shows
to tune the DT angles of cellular networks to optimize cov- that the GR can create stronger interference than the
erage for both ground and aerial users. Thus, to mitigate the antenna’s side lobes when the horizontal distance between
interference stemming from the down-tilted antennas on the the UAV and a GBS increases.
UAVs, we can consider existing inter-cell interference coordi- • By considering an interference-limited DL cellular net-
nation (ICIC) techniques already developed for heterogeneous work, we formulate an optimization problem to maximize
networks, namely, the 3GPP Release-10 specified enhanced the minimum SIR of the UAVs by tuning the UT angles
inter-cell interference coordination (eICIC) [18], [23]. of all the up-tilted antennas in the network. Since this
Motivated by all these factors, the main contribution of this is an NP-hard problem, we propose a simple meta-
paper is a novel cellular architecture that leverages additional heuristics-based technique, which tunes the UT angles
sets of antennas focusing towards the sky to support UAVs of the GBSs to ensure high minimum UAV SIR. Our
 3

proposed method uses the genetic algorithm (GA), a

well-known meta-heuristics algorithm that can generate
suboptimal solutions efficiently in an iterative method
[24].
• Since the UAVs will experience interference from the
extra up-tilted antenna sets along with the antenna side
lobes and GRs of the down-tilted antennas, here, we
consider the 3GPP Release-10 specified eICIC technique
to ensure the reliable coexistence of cellular-connected
UAVs and GUEs. The basic idea is that the down-tilted
antennas will stop transmission during some portions of
the data transmission duration to reduce interference at
the UAVs in DL. We discuss eICIC briefly later in this
paper.
• We conduct and present extensive simulations to study the
minimum SIR performance of our proposed method. We
first obtain suboptimal solutions from the proposed GA-
Fig. 1. 2-tier hexagonal cell structure with 19 cells and ISD = 500 m. In this
based technique and then use eICIC to further increase paper, we focus on the center cell with GBS location [0,0] km.
the SIR. Our results show that it is possible to obtain
high signal-to-interference (SIR) at the UAVs’ end by
optimizing the UT angles along with considering the summarized the key barriers and their potential solutions
eICIC method. By considering different UAV heights and for widespread commercial deployment of flying UAVs in
inter-GBS distances, we also show the effectiveness and beyond 5G wireless systems. Authors in [29], proposed an
superiority of our method over some baseline methods. optimization method for managing the movement, charging,
Our results also revealed some interesting yet important and service coverage actions of a fleet of UAVs used as flying
design guidelines such as the impact of the number of base stations. By considering a network of UAV base stations
antenna elements and the DT angles while considering (BSs), the work in [30] introduced exact HO probability for
the coexistence of UAVs and GUEs. similar UAV velocity and provided lower bound for UAV BSs
The rest of the paper is organized as follows. We provide with different velocities. The authors in [31] extended the
a literature review related to the interference mitigation tech- results of [30] by providing exact analysis of HO rate and
niques for cellular-connected UAV in Section II. In Section III, sojourn time for different UAV velocities and showed that HO
we describe our system model. Section IV discusses the rate is minimum when UAV BSs move with the same velocity.
UT angle maximization problem. We discuss our proposed However, both of these works treated UAVs as BSs. By using
GA-based UT antenna optimization method in Section V. tools from stochastic geometry, the authors in [32] studied the
Simulation results and the pertinent discussions are presented performance of 3D two-hop cellular networks where UAV-BSs
in Section VI. Finally, conclusions are drawn in Section VII. can obtain wireless backhaul from GBSs. In particular [32]
The notation list of this paper is presented in Table II. considered realistic antenna patterns and dedicated up-tilted
antennas for providing better connectivity in the UAV-to-GBS
links.
II. R ELATED W ORKS
Due to the complex antenna pattern and air-to-ground
Research efforts in integrating UAVs into existing cellular path-loss model, the researcher also relied on learning-based
networks with GUEs have recently attracted substantial atten- frameworks for ensuring reliable integration and operation of
tion from both academia and industry. For instance, in [5], the cellular-connected UAVs. For instance, a supervised learning-
authors explored the impact of practical antenna configurations based association scheme for UAVs was proposed in [17] to
on the mobility of cellular-connected UAVs and showed that associate UAVs with the GBS providing the highest directional
increasing the number of antenna elements can increase the antenna SIR. By tuning the DT angles of the GBSs, the
number of handovers (HOs) for vertically-mobile UAVs. The work in [20] used reinforcement learning (RL) to provide
work in [25] discusses the possibility of using UAVs in good connectivity to both UAVs and GUEs. However, they
wireless networks, with the role of flying base stations and did not consider the SIR at the UAV which plays a critical
relay nodes. role in reliable autonomous UAV deployment. In another
In [14], the same authors provided the upper and lower work [33], the authors proposed a deep-learning-based GBS
bounds on the coverage probability of UAVs considering a association algorithm for cellular-connected UAVs which takes
coordinated multi-point technique. The work in [26] presented the knowledge of the cellular environment into account. In
an analytical framework for a coexisting UAV and GUE con- the recent work in [19], authors study the problem of jointly
sidering a beamforming technique. By conducting extensive optimizing the UAV HO rate, disconnectivity time, UAV flight
3GPP compliant simulations, in [27], the authors showed that duration, and UAV energy consumption by tuning the UAV
the existing cellular networks will be able to support a small velocity. In particular this prior work explored a multi-armed
number of UAVs with good mobility support. In [28], authors bandit RL algorithm to solve the problem and showed that the
 4

TABLE II
N OTATION L IST.

Notation Description
hUAV UAV altitude
PGBS Transmit power of the GBSs
A Set of UAV locations
B Set of GBS
Nt Number of vertically placed antennas
φu Up-tilt angle of the up-tilted antennas
φd Down-tilt angle of down-tilted antennas
h(u)
GBS Height of the up-tilted antennas
h(d)
GBS Height of the down-tilted antennas
hd Height difference between up-tilted and down-tilted antennas
θd Elevation angle w.r.t. down-tilted antennas
Ge (θd ) Element gain w.r.t. down-tilted antennas
Gmax
e Maximum gain of each antenna element
G(d) (θd ) Total antenna gain at elevation angle θd w.r.t. down-tilted antennas
G(u) (θu ) Total antenna gain at elevation angle θu w.r.t. up-tilted antennas
Gm Side-lobe level limit
Pj(u) Received power from the up-tilted antennas of GBS j
Pj(d) Received power from the down-tilted antennas of GBS j
λ Wavelength of the carrier frequency
Ĝ(v)
j (θv ) Height-dependent antenna gain of the direct path
e (d) (h)
G Height-dependent antenna gain of the reflected path
j
ψj Angle of reflection of GBS j
R(ψj ) Ground reflection coefficient for the angle of reflection ψj of GBS j
∆φj Phase difference between the reflected and the direct signal paths of GBS j
α(h) UAV height dependent propagation coefficient
(d)
Ĝj (ψj ) Antenna gain of the incident path on the ground
(u)
γj,usf SIR of a UAV connected to up-tilted antennas of GBS j during uncoordinated subframes
(u)
γj,csf SIR of a UAV connected to up-tilted antennas of GBS j during coordinated subframes
(d)
γj,usf SIR of a UAV connected to down-tilted antennas of GBS j during uncoordinated subframes
(d)
γj,csf SIR of a UAV connected to down-tilted antennas of GBS j during coordinated subframes

perfect parameters can significantly improve the performance maximizing the minimum UAV rate by joint beamforming,
of cellular-connected UAVs. In [21], the authors explored an association, and UAV-height control framework for cellular-
RL algorithm to maximize the received signal quality at a connected multi-UAV scenarios. However, none of these an-
cellular-connected UAV while minimizing the number of HOs. alytical and learning-based works [5], [14], [17], [19]–[21],
An extension of the traditional RL algorithms known as multi- [26]–[28], [30], [31], [33], [38] considered the presence of GR
agent RL has been also introduced for efficient UAV control which plays a critical role in air-to-ground communications
in [34]. Note that these learning-based algorithms will either as an important source of interference for UAVs [11], [12].
require advanced data collection, preprocessing, and training, The most closely related work here is [22], in which the
or sample inefficient repetitive interaction with the cellular authors introduced a bidirectional deep learning-based tech-
networks, which makes the deployment of these algorithms nique to maximize the median capacity of an aircraft flying
challenging for real-world network operators. at a height of 12 km. Using system-level simulation, they
In addition to these learning-based methods, non-linear considered optimizing the inter-GBS distance and dedicated
optimization techniques were also used to provide reliable up-tilted antennas to solve network optimization problems. In
connectivity to UAVs. For instance, in [35], the authors contrast to their work, here, we focus on the UAVs flying
proposed a cooperative interference mitigation scheme to under 400 meters of height where the impact of GR is not
mitigate the strong uplink interference from the UAV to a negligible. Moreover, in our considered system, each GBS can
large number of co-channel GBSs serving terrestrial UEs. individually change its UT angle, in contrast to the similar
The helping GBSs sense the UAV’s power, which is sent UT angles that are assumed for all GBSs in [22]. To further
to the main GBS for further interference processing. Similar increase the minimum SIR, we consider the concept of the
authors introduced a cooperative beamforming and transmis- eICIC to mitigate the interference stemming from the down-
sion scheme to mitigate the interference of cellular-connected tilted antennas at the UAV’s end. Since eICIC was already
UAVs in DL [16]. In [36], they proposed a cooperative non- studied extensively in the last decade for increasing efficiency
orthogonal multiple access (NOMA) technique to the uplink and capacity of the heterogeneous networks [18], [23], it will
communication from a UAV to cellular GBSs, under spectrum be practical to deploy it for mitigating the interference from
sharing with the existing GUEs. The work in [37] discusses the down-tilted antennas. Moreover, the UT angle tuning is
how to integrate UAVs for providing wireless communications based on the GA algorithm, which is also well-studied and
in zones where the deployment of canonical base stations was used extensively in optimizations of different aspects of
is not possible. In [38], authors introduced the problem of wireless networks [39]. For convenience, we summarize and
 5

Fig. 2. Illustration of the inter-cell interference at a cellular-connected UAV from the GR signal of a downtilted antenna and the LOS signal from the uptilted
antenna of a nearby base station. Though not shown in the figure, the associated GBS in the right can also create interference by the downtilted antennas.
The signal quality at the UAV will be effected by the UT angles of the uptilted antennas since they will impact both the desired and the interference signals.

compare the state of the art in the literature with our work in antenna tilt angle is obtained by introducing a fixed phase
Table I. shift to the signal of each element. We define h(u) (d)
GBS and hGBS ,
respectively, as the height of the up-tilted antennas and down-
III. S YSTEM M ODEL tilted antennas. The two sets of antenna setups are separated
A. Network Model by a height difference hd , i.e., hd = h(u) (d)
GBS − hGBS . We
consider that all of the GBSs and their sets of antennas share
We consider an interference-limited DL transmission sce-
the same time and frequency resources. The UAVs will be
nario from terrestrial GBSs to cellular-connected UAVs where
associated with the antenna set (up-tilted or down-tilted) of
the 19 GBSs are distributed in a two-tier hexagonal grid with
the GBS providing the highest reference signal received power
a fixed inter-site distance (ISD). An illustration of such a
(RSRP) [5], [42].
network is presented in Fig. 1. Here, we do not consider
wraparound [40], [41] and thus, we will only focus on the B. Antenna radiation pattern
performance of the central hexagonal cell to capture the impact
of inter-cell interference from the neighboring cells. However, The Nt antennas are equally spaced where adjacent ele-
our analysis can easily be extended to larger cellular networks ments are separated by half-wavelength distance. The element
with different GBS distributions. Hereinafter, we will use the power gain (in dB) in the vertical plane at elevation angle
terms ‘GBS’ and ‘cell’ interchangeably. To average out the θd with respect to the down-tilted antennas can be specified
impact of UAV distribution, we divide the center cell into by [40]
(  2 )
discrete grid points, and a UAV is placed on each grid point at max θd
a height hUAV . Note that a closer inter-UAV distance or higher Ge (θd ) = Ge − min 12 , Gm , (1)
θ3dB
grid resolution will provide more fine-grained information
on the cellular network characteristics such as interference, where θd ∈ [−90◦ , 90◦ ], θ3dB refers to the 3 dB beam width
GBS association, received signal strength, etc. at the height with a value of 65◦ , Gmax
e = 8 dBi is the maximum gain of
hUAV . Each UAV is assumed to be equipped with a single each antenna element, and Gm is the side-lobe level limit,
omnidirectional antenna. The set of the UAV locations and respectively, with a value 30 dB [43]. Note that θd = 0◦
the GBSs can be expressed as A and B, respectively. refers to the horizon and the θd = 90◦ represents the case
We also assume that all GBSs have equal altitudes hGBS when the main beam is facing upward perpendicular to the
and transmission power PGBS . The GBSs consist of Nt ver- xy-plane [40]. The array factor Adf (θd ) of the ULA with Nt
tically placed cross-polarized directional antennas down-tilted elements while considering a DT angle φd is given by
by angle φd [5], [6]. We consider the GBS antennas to be 1 sin N2t π (sin θd − sin φd )


(d)
omnidirectional in the horizontal plane but they have a variable Af (θd ) = √ . (2)
Nt sin π2 (sin θd − sin φd )


radiation patterns along the vertical dimension with respect to
(d)
the elevation angle between the antennas and the users [15]. Let us denote Gf (θd ) , 10 log10 (Adf (θd ))2 as the array
Different from the traditional cellular network setting, here, power gain in dB scale. Then the overall antenna gain at
we also consider the presence of another set of antennas on elevation angle θd is given by
top of the previous ones, which can provide connectivity to (d)
G(d) (θd ) = Ge (θd ) + Gf (θd ). (3)
the UAVs using UT angle φu . Since the UAVs served by
only down-tilted antennas suffer from poor connectivity and Similarly, the array factor pertinent to the up-tilted antennas
severe interference, up-tilted antennas can be used to provide with UT angle φu and elevation angle θu can be expressed as:
reliable connectivity to the UAVs [1], [12]. Note that the
 6

1 sin N2t π (sin θu − sin φu )

(u)
Af (θu ) =√
. (4)
Nt sin π2 (sin θu − sin φu )
(u)
The array gain Gf (θu ) , 10 log10 (Auf (θu ))2 can then be
derived and, finally, the overall antenna gain due to the UT
angle φu can be expressed as:
(u)
G(u) (θu ) = Ge (θu ) + Gf (θu ). (5)

C. Ground reflection channel model Fig. 3. Analysis of GR depending on the DT angle φd .

The channel between a GBS and a UAV plays a critical role

in the coverage performance at the UAV’s end and we consider depends on Nt . Finally, the height-dependent propagation
a channel model that is characterized by both distance-based coefficient can be expressed as:
path-loss and GR. To characterize the GR, we modify the   
height-dependent path-loss model introduced in [11] which is α0 − h · (α0 − 2) , h < 2 · h(v) ,

GBS
α(h) = h(v)
GBS (8)
a variant of the two-ray path-loss model [44]. Let the length
(v)

2 h ≥ 2 · hGBS ,

of the 3D Cartesian distance from a UAV to a GBS j be lj and
the length of the incident and reflected paths are r1,j and r2,j , where α0 is the maximum possible attenuation coefficient [11].
respectively. For convenience, we discard the subscript from Here, we do not consider any GR due to the antenna side lobes.
hUAV in the following analysis. Finally, the received power From (7), we can see that the antenna gain is dependent on the
from GBS j at a UAV at height h can be specified as: incident angle ψj , whereas in [11], the gain of the reflected
path is assumed to be constant with respect to ψj . In Fig. 2,
Ĝ(v) e (d) (h)ei∆φj
2 α(h)
j (θv,j ) R(ψj )G

λ j we provide a simple illustration of how a UAV can suffer
Pj(v) = PGBS + ,
4π lj r1,j + r2,j from interference from GR and antenna side lobes. Remark
(6) 1: Due to the the DT angle φd , the main lobe of the down-tilted
where v ∈ {u, d}, θv,j is the elevation angle with respect √ to antenna will not reach the ground level before the horizontal
the up-tilted or down-tilted antenna of GBS j, i = −1 is hGBS
distance (in meter) is away by tan(φ from the GBS. Hence,
d)
the imaginary unit of a complex number, λ is the wavelength UAVs closer to this distance from a GBS will not be impacted
of the carrier frequency, Ĝ(v) e (d)
j (θv ) and Gj (h) represent the by the GR stemming from the down-tilted main lobe of that
height-dependent antenna gain of the direct and reflected particular GBS.
path, respectively, R(ψj ) is the GR coefficient for the Next, for a given UAV height and DT angle, we derive the
angle of reflection ψj with respect to the ground plane, distances from a GBS where the impact of the GR is the most
∆φj = (r1,j + r2,j ) − lj is the phase difference between the effective.
reflected and the direct signal paths, and α(h) is the height Theorem 1: For a given hGBS , hUAV , and DT angle φd ,
dependent propagation coefficient for UAV height h. Here, the impact of the GR from a GBS will mostly be seen between
we do not consider GR from the up-tilted antennas since horizontal distances d1 = hGBS +hUAV
and d2 = hGBS +hUAV
tan(φ1 ) tan(φ2 )
their main beams are oriented towards the sky. from that GBS, where
Note that the GR coefficient for cross-polarized antennas
R (ψ )−R (ψ ) φ1 = φd − 0.5 × θhpbw , (9)
can be calculated as R(ψj ) = H j 2 V j [45], which also
depends on the relative ground permittivity r ≈ 15 [11], re- φ2 = φd + 0.5 × θhpbw , (10)
flection coefficients for horizontal linear polarization RH (ψj )
and vertical linear polarization RV (ψj ). Moreover, Ĝ(v) and θhpbw is half power beam width of the main lobe of the
j (θv )
depends on the instantaneous elevation angle between the GBS down-tilted antenna.
and the UAV by (3) and (5), whereas G e (d) (h) can be expressed Proof : Consider a scenario with a single GBS with antenna
j
as: pattern and height are as specified in Section III. Since GR
 only stems from the down-tilted antennas, here, we consider
Ĝ(d) that the GBS is only equipped with down-tilted antenna with
j (ψj ), h < ht


DT angle φd . Let us consider the half-power beam width


 (d)

 Ĝj (ψj )


 , ht ≤ h ≤ 2ht (HPBW) of the main lobe as θhpbw . Note that the HPBW is
G(d)
e (h) = 2 inversely proportional to the number of elements in the antenna
j (d)
 Ĝj (ψj ) h array [46]. Given the DT angle φd , the two angles of the two
− · (Ĝ(d)
j (ψj ) − 1), 2ht ≤ h ≤ 500






 2 2ht,c end points of the HPBW will be as expressed in (9) and (10).
0.5, h ≥ 500

Then the down-tilted main beam will reach the ground
(7) and the impact of the HPBW will be within the distances
where ht = 2h(d) GBS + 2 and ht,c = 500 m are threshold hGBS
r1 = tan(φ 1)
hGBS
and r2 = tan(φ 2)
from the GBS as depicted in
heights [11], and Ĝ(d)
j (ψj ) is the antenna gain of the incident Fig. 3. By assuming regular reflection from the ground, the two
path on the ground from the down-tilted antennas which rays will reach the UAV height at a distance d1 = hGBS +hUAV
tan(φ1 )
 7

and the reflected path and the GR can provide more than 10 dB
more signal power than the 3GPP model. For hUAV = 100 m,
as shown in Fig. 4(b), the GR shows a similar kind of trend
but after greater UAV-to-GBS horizontal distance as discussed
in Theorem 1.
Finally, we split the reflected signal from the down-tilted
antennas into its two ingredients: the signal from the antenna
side lobes and the reflected signal from the main beam of
the DT antennas. The relevant results for hUAV = 100 m
are shown in Fig. 5(a), from which we conclude that the GR
path-loss model coincides with the side lobes when the UAV is
close to the GBS. However, after a distance of 400 m, the GR
starts to provide high power through the main lobe which even
compensates the antenna’s side-lobe null at 442 m. Overall, the
(a) hUAV = 50 m. GR keeps dominating the signal from the DT angles till about
900 m. We also study the impact of GR for higher DT angles
in Fig. 5(b). For a DT angle of 10◦ , GR starts dominating the
signal power from about 350 m and can act as the dominant
source of interference for a UAV situated at a distance of
1500 meters. From the above discussion, we can conclude
that the down-tilted antennas can create significant interference
towards the far UAVs by GR. However, other than some works,
the impact of GR is not considered in the literature. Apart from
this, the up-tilted antennas can also create strong interference.
However, we can mitigate the interference from the up-tilted
antennas by tuning the UT angles properly [12]. Hence, to
increase the reliability of the cellular-connected UAVs, we
consider the eICIC method to reduce the interference from
the down-tilted antennas.

(b) hUAV = 100 m. E. Overview of eICIC

Fig. 4. Comparison of GR and 3GPP RMa-AV channel model [41] for
different UAV heights considering the antenna radiation pattern and φd = 6◦ .
To mitigate the interference problems caused by the extra
(a) hUAV = 50 m and (b) hUAV = 100 m. set of antennas, we consider eICIC techniques which have
been specified in LTE Release-10 of 3GPP [47]. The time-
domain eICIC technique provides an interference coordination
and d2 = hGBS +hUAV
tan(φ2 ) , respectively from the GBS, which method based on the subframe blanking, known as almost
completes the proof. blank subframe (ABS) that does not send any traffic channels
Theorem 1 provides us the range of distances from a GBS and sends mostly control channels with very low power.
where a UAV will be impacted significantly by GR for a given In our proposed interference mitigation method, the down-
DT angle φd . From Theorem 1, we can observe that for a tilted antennas will not transmit data while allowing the up-
higher φd , locations closer to the GBSs will be impacted by tilted antennas to serve UAVs suffering from high interference
GR and vice versa. during an ABS. Transmissions from the down-tilted antennas
θ
Remark 2: If φd < hpbw 2 , then the impact GR at the UAV are periodically muted during the entire frame duration. The
will start from the distance d1 and will the impact of the up-tilted antennas can send their data during such an ABS
main lobe will last till infinity. However, due to the path-loss, and avoid interference. Note that certain control signals are
the impact will gradually decrease as the horizontal distance still required to be transmitted even in the muted subframes
increases beyond d1 . to avoid radio link failure [48].
The frame structure of the eICIC is shown in Fig. 6. During
D. Numerical example the uncoordinated subframes (USFs), the down-tilted antennas
By considering φd = 6◦ , in Fig. 4(a), we compare the 3GPP transmit data and control signals at full power PGBS while
RMa-AV model [41] and our proposed height dependent GR during the coordinated subframes (CSFs), they remain muted.
model for hUAV = 50 m, hGBS = 30 m, and PGBS = 30 dBm, We define β as the duty cycle of USFs which refers to the
while considering the antenna radiation pattern as discussed ratio of the number of USFs to the total number of subframes
before. The received signal plot with respect to 2D UAV-BS in a frame. Then, (1 − β) will be the duty cycle of the silent
distance shows that the impact of GR comes into play after a subframes or CSFs. Here, we assume full coordination and
certain horizontal distance. The ripple in the received signal is synchronization among the GBSs and hence, the ABS pattern
created due to the phase difference between the direct LoS path of all the down-tilted antennas will be the same. We will show
 8

common transmission bandwidth and full buffer traffic is used

in every GBS [6], [49]. Then, we can calculate the SIR of a
UAV connected to the up-tilted antennas of GBS j considering
flat-fading channels [23] and antenna pattern during USF by
the following expression:
(u)
(u) Pj
γj,usf = P P (v) (d)
. (11)
Pb + Pj
b∈B,b6=j, v∈{u,d}

Similarly, SIR of a UAV connected to the down-tilt antennas of

GBS j considering flat-fading channels during USF as follows:
(d)
(d) Pj
γj,usf = P P (v) (u)
. (12)
Pb + Pj
(a) hUAV = 100 m, φd = 6◦ . b∈B,b6=j, v∈{u,d}

Note that (6) is used to calculate the received power from

a particular antenna set (up-tilted/down-tilted) of a GBS. We
assume flat-fading channels due to the presence of narrowband
OFDM-based communications in existing cellular networks.
After considering the antenna radiations from the both sets
of antennas and some algebraic calculations, the closed-form
expressions of (11) and (12) are expressed by (16) and (17),
respectively, which are presented on the next page. During
the CSFs, the down-tilt antennas are kept off to protect the
UAVs from interference (GR of the beam’s boresight and the
LOS interference from the beam’s side lobes). Note that the
interference to a UAV served by an up-tilted antenna may
be coming also from the down-tilted antenna located at the
same GBS. Thus, the SIR of a UAV connected to the up-tilted
antennas of GBS j during CSF can be expressed as follows:
(b) hUAV = 100 m, φd = 10◦ .
(u)
Fig. 5. Impact of GR and antenna side lobes on the GR-based path-loss (u) Pj
γj,csf = . (13)
model for hUAV = 100 m. (a) φd = 6◦ and (b) φd = 10◦ . P (u)
Pb
b∈B,b6=j

Finally, we can find the capacity of a UAV connected to up-

tilted antennas of GBS j during USFs as follows:
(u) (u)
Cj,usf = log2 (1 + γj,usf ). (14)
On the other hand, if the UAV is associated with down-tilted
antenna of its serving GBS, it will obtain its data in the DL
Fig. 6. Basic principle of time domain eICIC. For the considered scenario, during the USFs. Hence, the rate can be expressed as
the aerial users can be scheduled in the up-tilted antenna subframes that
overlap with the almost blank subframes of the down-tilted antennas. This will (d) (d)

Cj,usf = β log2 (1 + γj,usf ) . (15)
protect aerial users from the sidelobe interference and the ground reflection
interference coming from the down-tilted antennas, as illustrated in Fig. 2.
Note that the rate of the UAVs associated with down-tilted
antennas will be scaled by the parameter β. Lower values of β
in the next subsection that the choice of β will impact the will increase the SIR performance of the UAVs associated with
capacity/rate of the UAVs/GUEs associated with the down- the up-tilted antennas as shown in (13). However, the UAVs
tilted antennas. However, this is out of the scope of this paper associated with the down-tilted antennas and most importantly,
and will be subject of our future work. the GUEs will suffer from low rates for a low β. This trade-off
will be addressed in our future work.
IV. U P - TILT A NGLE O PTIMIZATION FOR M AXIMIZING SIR
A. SIR definitions over different subframes B. Problem definition
As mentioned earlier, we consider an interference-limited Our goal is to tune the UT angles of the up-tilted antennas
DL sub-6 GHz band for the cellular network, where the individually during the USFs to provide reliable SIR at the
presence of thermal noise is omitted. We also assume that UAVs’ end. Without optimizing the UT angles, the SIR
the GBSs and both up-tilted and down-tilted antennas share a performance will worsen due to the additional interference
 9

"

#α(h)
Nt π
Ge (θu,j ) sin2 sin(θu,j )−sin(φu,j )

2 1
10 10

lj
sin2 π sin(θu,j )−sin(φu,j )
(u) 2
γj,usf = "

#α(h)

α(h)
.
Ge (θu,i ) Nt π Ge (θd,k ) sin2 Nt π e (d) (h)ek∆φ
sin2 sin(θu,i )−sin(φu,i )
sin(θd,k )−sin(φd,k )
R(ψk )G
2 1 2
+ N1 1 k
P P
10 10

li 10 10

lk + r1,k +r2,k
π t π
i∈B,i6=j sin2 sin(θu,i )−sin(φu,i ) k∈B sin2 sin(θd,k )−sin(φd,k )
2 2
(16)

α(h)
Ge (θd,j ) sin2 Nt π e (d) (h)ej∆φ
sin(θd,j )−sin(φd,j )
R(ψj )G
1 2 1 j
Nt 10 10

lj + r1,j +r2,j
sin2 π sin(θd,j )−sin(φd,j )
(d)
γj,usf = "

#α(h)2

α(h)
.
Ge (θu,i ) Nt π Ge (θd,k ) sin2 Nt π e (d) (h)ek∆φ
sin2 sin(θu,i )−sin(φu,i )
sin(θd,k )−sin(φd,k )
R(ψk )G
2 1 2
+ N1 1 k
P P
10 10

li 10 10

lk + r1,k +r2,k
π t π
i∈B, sin2 sin(θu,i )−sin(ψu,i ) k∈B,k6=j sin2 sin(θd,k )−sin(φd,k )
2 2
(17)

from the up-tilted antennas [12]. Note that the UAVs can and their SIRs will also be impacted by the choice of the UT
be associated with either up-tilted antennas or down-tilted angles.
antennas depending on the highest RSRP providing antenna
set [12]. Let us consider the vector of SIRs of all UAVs when V. G ENETIC A LGORITHM -BASED U P - TILT A NGLE
they are associated with the highest RSRP providing antenna O PTIMIZATION
sets as: The GA is a stochastic population-based optimization tech-
γ = [γ1,usf , ..., γ|A|,usf ], nique that mimics the metaphor of natural biological evalua-
tion and is an efficient tool in searching for the global opti-
where | · | represents the cardinality of a set. Then, we can mum [24]. It borrows the idea of “survival of the fittest” in its
formulate the problem of maximizing the minimum UAV SIR search process to select and generate individuals (design solu-
as: tions) that are adapted to the underlying objectives/constraints
max min γ of the problem of interest. Hence, GA is well suited to and has
Φu
(18) been extensively applied to solve complex design optimization
s.t. 0 ≤ Φu ≤ 90◦ .
without being guided by stringent mathematical formulation.
Here, the optimization variable Φu = [φu,1 , ..., φu,|B| ] is It can explore the whole search space simultaneously, and
the vector of the UT angles of the up-tilted antennas in the hence, identify high quality solutions more quickly than an
network. Note that only the interference caused by the up-tilted exhaustive search. The detailed principles of a GA scheme
antennas is dependent on the UT angles. We also keep the UT can be found in [24]. In the following subsections, we outline
angles above the horizon level (greater than 0◦ ) for saving our proposed GA-based UT angle tuning method for obtaining
the GUEs from additional interference. However, changing the optimal solution of (18). We assume that each GBS sends
the UT angles will change the association of the serving only its chosen UT angle and the SIR information of the UAVs
GBS/antenna sets. Overall, the optimization problem in (18) is associated with it to a central server. The server can then run
very difficult to solve efficiently since the objective function is the proposed GA-based algorithm and compute the optimum
highly non-convex with respect to decision variables Φu [1]. UT angles.
The search space of the problem is continuous and grows
exponentially with the number of GBS. Moreover, due to the A. Representation
complex antenna pattern and tilting angles involved, it is not At first, some randomly generated candidate solutions for
possible to obtain the closed-form optimal solutions by taking the optimization problem are encoded in a chromosome-
the derivatives of (16) and (17) even under a free-space path- like strings. The collection of these candidate solutions or
loss model and a similar UT angle for all the GBSs. Assuming chromosomes are referred to as population. In other words,
the tilting angles to be 0◦ for simplification as done in [26] will members of the population are the vectors of possible UT
not represent a realistic cellular network scenario. Using an angles for our formulated optimization problem. Note that each
exhaustive search method is also computationally prohibitive
member of the population must provide a complete solution
since its complexity increases exponentially with number of
to the problem. The size of the population does not change
GBSs or up-tilted antenna sets. To overcome these challenges,
over time usually. To meet the constraint, the UT angles of
in the next section, we introduce our GA-based UT angle
the population are generated within the feasible search space.
optimization method for maximizing the minimum UAV SIR.
Note the SIR gain due to the eICIC is not related to tuning
the UT angles and the gain can be calculated by simply not B. Fitness evaluation
considering the received power from the down-tilted antennas. The objective function of the problem is used to evaluate
The rates of the UAVs who are associated with the down-tilted the fitness of each chromosome. In our case, the randomly
antennas will be reduced by the quantity β as shown in (15) generated UT angles are used as inputs to the simulator for
 10

Algorithm 1 Up-tilt Angle Optimization using GA future population, which will have better fitness value (higher
1: Input: minimum SIR in our case) on average.
2: population: Set of UT angles for all GBSs
3: Fitness function (FF): Minimum SIR of the UAV
4: network parameters, GBS and UAV locations
E. Mutation
5: Method:
6: NewPopulation = empty set The last operator is the mutation, which introduces diversity
7: StopCondition: Number of iterations in population characteristics and prevents premature conver-
8: SELECTION: Roulette wheel selection method
9: Create random Population gence. In this step, certain parts of the newly formed children
10: EVALUATE (Population, FF) (new sets of UT angles with better fitness) are subjected to a
11: while (StopCondition is not met) mutation with a low random probability. In our proposed GA-
12: for i = 1 to Population size do based framework, the mutation takes place with a low mutation
13: Parent1 = SELECTION(NewPopulation, FF) probability. We first generate random numbers between −1
14: Parent2 = SELECTION(NewPopulation, FF)
15: Child = Reproduce(Parent1, Parent2) and 1 for each member of the UT angle population. If the
16: if (small random probability) absolute value of a random number is less than the mutation
17: child = MUTATE(Child) probability, that particular random number is added to that
18: add child to NewPopulation set member (UT angle) of the population.
19: end if
After all of these genetic processes, the members of the
20: end for
21: end while populations with the worst fitness values are replaced by the
22: EVALUATE (NewPopulation, FF) new individuals with better fitness values or higher minimum
23: Args = GetBestSolution (NewPopulation) SIRs. The algorithm continues until good results are obtained
24: Population = Replace (Population, NewPopulation) through iterations in terms of the objective function. The over-
25: Output: Args: Best individuals of the UT angles and the highest
all algorithm is also summarized in Algorithm 1. In essence,
minimum SIR
obtaining high-quality suboptimal solutions from our proposed
method depends on carefully addressing the following issues.
obtaining the minimum SIR of all the discrete UAV locations. • representation of tentative solutions (UT angles) as chro-
The higher the minimum SIR of a solution is, the better the mosomes;
fitness value is associated with it. • initialization of the randomly generated population;
• determination of the fitness function (min SIR);
• selection of genetic operators;
C. Selection
• adjustment of GA parameters (population size, crossover
The selection process determines the pair of candidate and mutation probabilities).
solutions/ UT angles which will act as parents for mating.
Considering the impact of mutation, the work in [51]
After being evaluated by a fitness function, each member
provided the lower bound of the number of iterations required
of the population is assigned a probability to be selected
for obtaining the global optimum for a given population size.
for reproduction. Note that, the worse performing members
In particular, they showed that to obtain the global optimum
should also be given a chance in the evolution process so
with any specified level of confidence, GAs should run for long
that the overall algorithm can maintain a good exploration in
enough. However, later we show that increasing the number
the search space. Here, we consider a simple biased roulette
of iterations or population size will increase the complexity
wheel to select individuals as parents [50]. More explicitly,
and run-time of the proposed algorithm. Hence, we run ex-
each chromosome in the population is assigned a slot in a
tensive simulations for different numbers of population size
roulette wheel, whose size is proportional to its fitness over
and iterations, and check the associated minimum UAV SIRs.
the total sum of fitness in the population. Then, a random
We found that with a the population size of 200, mutation
number between 0 and 1 is generated for each member/ UT
probability of 0.1, and 50 iterations, our algorithm provides
angle set. A chromosome/member is selected as a parent for
high-quality suboptimal solutions.
further genetic operations if the random number is within the
range of its roulette wheel slot.
F. Complexity analysis
D. Crossover As described in the previous subsections, our proposed GA-
The selected parents are then processed by the crossover based UT angle optimization technique randomly generates
operator, which mimics mating in biological populations. It tentative solutions and then produces new better solutions
is considered to be the most significant phase in a GA. from the previous ones iteratively. For a given GBS and UAV
Here, for each pair of parents to be mated, a crossover point distributions, the overall time complexity of the algorithm is
is chosen at random from within the chromosomes. Then O(M 2 I|A||B|), where M represents the number of popula-
offspring/children are created by exchanging the chromosomes tions and I is the iteration number, respectively. Hence, for a
(UT angles) of parents among themselves until the crossover given population size, number of iterations, and number of
point is reached. The crossover operator propagates features GBSs, the complexity of our proposed algorithm increases
of good surviving designs from the current population into the linearly with an increasing number of UAVs.
 11

TABLE III
S IMULATION PARAMETERS .

Parameter Value
P GBS 46 dBm
hUAV 100 m & 200 m
(d)
hGBS 30 m
ISD 500 m & 1000 m
hd 1m
hGUE 1.5 m
λ 0.15 m
α0 3.5 [11]
DT angle (φd ) 6◦

VI. S IMULATION R ESULTS

In this section, we present the simulation results for our
proposed cellular architecture based on a new set of antennas
and eICIC. Unless otherwise stated, the simulation param-
(a) hUAV = 100 m, ISD = 500 m. eters are as listed in Table III. By considering flat fading
channels [23] and hexagonal cells, we report our finding for
two ISDs namely, 500 m and 1000 m while considering the
highest RSRP-based association (HRA). It is worth noting
that in our setup, the HRA association will also provide the
highest SIR among all the available antennas of the network.
For convenience, we refer to our proposed method as ‘optimal
HRA’ hereinafter. To study the performance of our proposed
method we consider also three baseline schemes. These four
scenarios can be summarized as follows.
• optimal HRA: this is our proposed GA-based UT angle
tuning method.
• HRA single: all GBSs pick the same optimal UT angle
which maximizes the minimum SIR. This UT angle is
calculated by exhaustive search method.
• Random: each GBS picks UT angles randomly from the
search space.
(b) hUAV = 200 m, ISD = 500 m. • HRA (no eICIC nor UT antennas): presence of up-tilted
antennas and eICIC is ignored. UAVs associate with the
highest RSRP providing GBS.
As mentioned in Section III, we divide the whole network
into 10 m×10 m grids [23], and a UAV is placed on each
grid point with height hUAV . Such a uniform distribution will
average out the impact of UAV distributions [23]. We only
take the discrete points inside the center hexagonal cell into
consideration.

A. Optimal UT angle analysis

After obtaining the best solutions of UT angles by using (11)
and (12) and our proposed GA-based method, we calculate the
UAV SIRs in USFs for the two ISDs and UAV heights. Then
eICIC is used to get the pertinent UAV SIRs in CSFs. For
ISD = 500 m and hUAV = 100 m and 200 m, the best solutions
obtained from the proposed GA-based algorithm are presented
(c) hUAV = 100 m, ISD = 1000 m. in Fig. 7(a) and Fig. 7(b), respectively. Our results show that
one of the six neighboring GBS chooses a relatively smaller
Fig. 7. Optimal UT angles obtained from the proposed GA algorithm for
ISD = 500 m for (a) hUAV = 100 m, (b) hUAV = 200 m, and (c) for UT angle and provides high received power to the UAVs for
ISD = 1000 m and hUAV = 100 m. hUAV = 100 m. The other GBSs overall maintain higher UT
angles to reduce the interference from the side lobes.
A similar conclusion can also be drawn for hUAV = 200 m,
while one big exception is that the UAVs are supported by s
 12

tier-2 GBS as shown in Fig. 7(b). Due to the compact GBS

locations and higher UAV height, the tier-2 GBSs can provide
better SIR by choosing an angle that covers most of the
discrete UAV locations for hUAV = 200 m. For ISD = 1000 m,
both UAV heights show the similar trend as Fig. 7(a) and
in Fig. 7(c), we report the best solutions of UT angles for
hUAV = 100 m. Overall, the GBSs tend to choose lower
UT angles for larger ISD to reduce inter-cell interference. A
similar case of obtaining lower UT angles for higher ISD was
also reported in [22].
For ISD = 500 m and hUAV = 100 m and 200 m, the
respective UAV SIR cumulative distribution function (CDF)
plots are presented in Fig. 8(a) and Fig. 8(b), respectively.
From both figures, we can conclude that our proposed optimal
HRA scheme provides higher minimum SIR (about −1.36 dB
for hUAV = 100 m and about 10 dB for hUAV = 200 m)
than the other baseline methods. The optimization framework (a) hUAV = 100 m.
considers the minimum UAV SIR inside the center cell and
thus the interfering GBSs choose UT angles which create less
interference towards the UAVs. During the CSFs, turning the
down-tilted antennas off increases the minimum SIR to about
6 dB for hUAV = 100 m and about 12.5 dB for hUAV = 200 m.
One interesting observation is that the overall SIR with eICIC
is higher for hUAV = 100 m. This is because the UAVs suffer
more interference from the down-tilted antennas for lower
UAV heights via GR and antenna side lobes. Moreover, the
path-loss is also lower for hUAV = 100 m than hUAV = 200 m.
Hence, muting the down-tilted antennas provide higher SIR
gain in the CSFs for hUAV = 100 m.
In the HRA single scheme, the GBSs choose the same
optimal angle, which result in less degree of freedom to im-
prove the SIR performance. Hence, it provides comparatively
lower SIR (about −11 dB for hUAV = 100 m and about
−8 dB for hUAV = 200 m) than our proposed method. Even
(b) hUAV = 200 m.
with the ICIC, the overall gain in the minimum SIR is still
significantly lower than without the ICIC minimum SIR of our Fig. 8. UAV SIR CDFs for ISD = 500 m for (a) hUAV = 100 m and (b)
hUAV = 200 m.
proposed scheme. The random scheme chooses the UT angles
for each of the GBSs and thus provides better performance
than HRA single. Thus, it is evident from the discussion that the same rate in USF and CSF. Similar observations are also
it is critical to tune the UT angles of the GBSs individually for obtained for other UAV height and ISD.
the successful integration of the up-tilted antenna sets. Finally,
for the case in which the UAVs are served by only down-tilted
antennas and without the ICIC scheme, the overall SIR is very B. Impact of the down-tilted antenna
low (less than −8 dB) for both of the UAV heights. For larger DT angles can create a significant impact on the overall
cell sizes or ISD = 1000 m and the two UAV heights, we performance of the network since they play a major role
can conclude from Fig. 9(a) and Fig. 9(b) that our method in determining the inter-cell interference. Higher DT angles
outperforms the other baseline schemes significantly in terms decrease the interference towards other nearby GBSs which
of the minimum UAV SIR during the USFs i.e., without ICIC. translates to a better coverage for GUEs. However, for UAVs
Fig. 10 shows the rates (bps/Hz) for the baseline schemes using flying in the sky, the DT angles can create interference by
(14) and (15). From Fig. 10, we can observe that our proposed both side lobes and GR. This motivated us to study the impact
optimal HRA scheme provides a higher minimum rate, 50th- of DT angles of the down-tilted antenna sets and report the
percentile rate, and sum rate than other baseline schemes. The relevant results in Fig. 11.
HRA (no ICIC or UT antennas) scheme is excluded in the In Fig. 11(a), we show the SIR CDFs for hUAV = 100 m
rate comparison due to its very low SIR performance (less and 200 m by calculating the optimal UT angles using an
than −8 dB). Due to the higher SIR obtained with eICIC, optimal HRA scheme for three DT angles namely, 0◦ , 6◦ , and
overall the rates increase significantly in the CSFs. The UAV 12◦ , respectively. From this figure, we can conclude that the
with the minimum SIR in the HRA single scheme is associated 0◦ DT angle overall provides low SIR in both USF and CSF
with the down-tilted antennas and thus, HRA single provides frames due to the higher interference stemming from the main
 13

(a) hUAV = 100 m. (a)

(b) hUAV = 200 m. (b)

Fig. 9. UAV SIR CDFs for D = 1000 m for (a) hUAV = 100 m and (b)
hUAV = 200 m.

beam of the down-tilted antennas. Though the impact of GR is

trivial for φd = 0◦ as discussed in Theorem 1, the focus of the
main beam causes severe interference to the far away UAVs,
which degrades the overall SIR performance. Although higher
DT angles are beneficial for GUEs, our results show that
6◦ provides better SIR performance than its 12◦ counterpart.
This is because, for a 12◦ DT angle, the UAVs faces more
interference by GR from the closest GBS as described in
Theorem 1. For a 6◦ DT angle, UAVs usually suffer less severe
interference in GR from neighbor GBSs due to higher path-
loss since the GR signals have to travel longer to reach the
UAV.
For the CSFs, we obtain high SIR for both 6◦ and 12◦ . (c)
Due to the higher GR interference of 12◦ , this angle provides
Fig. 10. Rate (bps/Hz) analysis for hUAV = 100 m and ISD = 500 m. (a)
the highest SIRs in the CSFs by muting the down-tilted min rate, (b) 50th-percentile rate, and (c) sum rate.
antennas. From Fig. 11(b), we can make similar observations
for hUAV = 200 m. However, in Fig. 11(b), the UAVs achieve
better SIRs than those of lower heights. This is due to the fact become weak when they reach UAVs. Moreover, the interfer-
that the GRs from the GBSs face higher path-loss and thus ence due to the side lobes also weakens due to the increased
 14

(a) hUAV = 100 m. (a)

(b) hUAV = 200 m. (b)

Fig. 11. UAV SIR CDFs for ISD = 500 m for (a) hUAV = 100 m and (b) Fig. 12. (a) Vertical antenna pattern of a GBS considering cross-polarized
hUAV = 200 m. elements, each with 65◦ half power beam width and φd = 6◦ . (b) UAV SIR
CDFs for hUAV = 100 m and ISD = 500 m during the USFs.

distances from the GBSs. Interestingly, 6◦ provides slightly

finding in Fig. 12(b). Since antenna with low Nt provides
better SIRs because this angle provides better antenna gain
lower gain, the SIRs corresponding to Nt = 4 obtains lower
through the side lobes from its other DT angle counterparts at
values. For instance, about 20% of the UAVs suffer from very
hUAV = 200 m.
low SIR (less than −5 dB).
For the other two Nt plots, we can see an interesting trade-
C. Impact of the number of antenna elements off. When Nt = 16 is considered, Fig. 13(b) verifies that it
The number of antenna elements has a direct impact on provides better minimum SIR (greater than 0 dB) than Nt = 8,
the antenna array gain and the beam width of the antenna thanks to its higher antenna gain. However, due to its wider
pattern [46]. Here, we focus on how the number of antenna beam width, with Nt = 8, GBSs can cover a larger area in
elements at the GBS can influence the SIR performance of the sky with higher gains. This translates into the fact that
the UAVs. Note that increasing the element number increases about 70% of the UAVs achieve a higher SIR compared to the
the antenna array gain but reduces the beam width and vice case when GBSs are equipped with 16 antennas each. This
versa [46]. In Fig. 12(a), we plot the antenna gains in dB interesting insight can help the network operators better plan
scale for Nt = 4, 16, and 32 using (3) and φd = 6◦ . As the number of antenna elements they need depending on their
expected, the antenna gain increases by 3 dB for doubling performance requirements.
the antenna elements and at the same time, the main beam
becomes narrower. To study the impact of this phenomenon, D. Impact of the physical separation of the antenna sets
we use the proposed optimal HRA method to calculate the We also study the impact of the antenna separation distance
optimal UT angles in USFs for different Nt and report the hd between the up-tilted and the down-tilted antenna sets. We
 15

(a) ISD 500 m.

Fig. 13. UAV SIR CDFs for ISD = 500 m and ISD = 1000 m while
considering hUAV = 100 m. Solid lines represent the SIR with up-tilted
antennas and dashed lines represent the case without up-tilted antennas during
the USFs. Both of these lines overlap with each other.

consider hUAV = 100 m and ISD= 500 m and 1000 m and

show the resulting UAV SIRs for the optimal UT angles in
Fig. 13. For both ISDs, we can conclude that the overall impact
of hd is very trivial for the optimal UT angles during USFs.
The related SIRs are slightly better for hd = 2 m. This is
due to the fact that with higher hd , the main lobes of the two
sets of antennas are more separated from each other and thus
creates less interference.
Another interesting finding is that the impact of hd is more
visible for ISD= 1000 m. This is because the GBSs tend to
pick lower UT angles for covering the cell-edge UAVs for
larger ISDs, and hence, the higher hd helps to keep the main
beams of the up-tilted and down-tilted antennas further away. (b) ISD 1000 m.
This results in lower interference and thus higher SIRs for the Fig. 14. GUE SIR CDFs with height 1.5 m for (a) ISD = 500 m and (b)
UAVs. Whereas for lower ISDs, the GBSs pick higher values ISD = 1000 m during USFs.
of UT angles which are already separated from the main beams
of the down-tilted antennas, and thus the overall impact of hd
is trivial here. are presented with solid lines and scenarios without the up-
tilted antennas are represented by the dashed lines. It is evident
from the plots of both Fig. 14(a) and Fig. 14(b) that the impact
E. Impact on the GUE SIR of up-tilted antennas on the GUE SIRs is trivial and the lines
Thus far, we have focused on scenarios in which the UAVs representing these two scenarios overlap each other. This is
as the only users in the network. After proper tuning of the because the main lobes of the up-tilted antennas are focused
UT angles, the presence of the extra set of up-tilted antennas towards the sky and hence, the only impact they can create
along with the eICIC method can provide a high and reliable is through the side lobes. However, these side lobes of the
SIR for the UAVs flying in the sky. However, the extra set of up-tilted antennas can create little to no impact on the GUEs
antennas can also introduce interference to the existing GUEs. who are associated with GBS providing very high antenna
Hence, in this subsection, we study the impact of our proposed gains. Note that the overall trends will still be the same for
UT angle tuning scheme on the GUEs. 3GPP-based path-loss models [40] for GUEs.
Here, we consider the three DT angles as done before along Note that the SIRs of the GUEs increase with increasing
with the two ISDs and UAV heights to check the impact DT angle since higher DT angles reduce inter-cell interfer-
thoroughly and report the results in Fig. 14. We use the GR- ence. Moreover, larger cell areas or ISDs provide better SIR
based path-loss model with a height of 1.5 m to represent performance due to the reduced interference on the cell-
the GUE cases. We only report the USF results for visual centered GUEs. Other than the plot for ISD= 1000 m and
convenience and the CSF cases show the same trends and hUAV = 200 m, all other plots show that GUE performance is
hence, are omitted here. The cases with the up-tilted antennas invariant of the optimal UT angles of the up-tilted antennas.
 16

For ISD= 1000 m and hUAV = 200 m, the cell-edge users will be included in our future work to make our framework
suffer from less interference since GBSs tend to focus more more efficient.
upwards with higher hUAV .
R EFERENCES
VII. C ONCLUDING R EMARKS AND D ISCUSSION [1] Geraci, G., Garcia-Rodriguez, A., Galati Giordano, L., Lopez-Perez, D.,
and Bjornson, E., “Understanding UAV cellular communications: From
In this paper, we have proposed a novel cellular architecture existing networks to massive MIMO,” IEEE Access, vol. 6, pp. 67853–
by considering an extra set of antennas that are up-tilted to 67865, 2018.
[2] Lin, X., Yajnanarayana, V., Muruganathan, S. D., Gao, S., Asplund, H.,
provide good and reliable connectivity to the UAVs. These Maattanen, H., Bergstrom, M., Euler, S., and Wang, Y. . E., “The sky is
antennas coexist with the traditional down-tilted antennas not the limit: Lte for unmanned aerial vehicles,” IEEE Commun. Mag.,
and use the same time and frequency resources. The down- vol. 56, pp. 204–210, Apr. 2018.
[3] Wu, Q., Zeng, Y., and Zhang, R., “Joint trajectory and communication
tilted antennas can create interference to the UAVs by the design for multi-UAV enabled wireless networks,” IEEE Trans. Wireless
antenna side lobes and GR, and we have proposed a modified Commun., vol. 17, pp. 2109–2121, Mar. 2018.
path-loss model to capture the impact of the GR on the [4] Zeng, Y. and Zhang, R., “Energy-efficient UAV communication with
trajectory optimization,” IEEE Trans. Wireless Commun, vol. 16, no. 6,
UAVs. To ensure high SIR and reliable connectivity, we have pp. 3747–3760, 2017.
formulated an optimization problem with an aim to maximize [5] Amer, R., Saad, W., Galkin, B., and Marchetti, N., “Performance
the minimum UAV SIR by tuning the UT angle of each analysis of mobile cellular-connected drones under practical antenna
configurations,” in Proc. IEEE ICC, (Dublin, Ireland), June 2020.
GBS. Since the problem is NP-hard, we have proposed a GA- [6] Chowdhury, M., Maeng, S., Bulut, E., and Guvenc, I., “3-D Trajectory
based UT angle optimization method to obtain high-quality Optimization in UAV-Assisted Cellular Networks Considering Antenna
suboptimal solutions efficiently. Apart from this, we have also Radiation Pattern and Backhaul Constraint,” IEEE Trans. Aerosp. Elec-
tron. Syst., vol. 56, no. 5, pp. 3735–3750, 2020.
considered the 3GPP specified eICIC to reduce the interference [7] Chowdhury, M. M. U., Erden, F., and Guvenc, I., “RSS-based Q-
caused by the down-tilted antennas. We have run extensive Learning for indoor UAV navigation,” in Proc. IEEE Conf. on Military
simulations to study our proposed method for various cellular Commun. (MILCOM), (Norfolk, VA), Nov. 2019.
[8] Rahmati, A., He, X., Guvenc, I., and Dai, H., “Dynamic mobility-
network deployment configurations such as ISD, UAV height, aware interference avoidance for aerial base stations in cognitive radio
DT angle, number of antenna elements, etc. Our results have networks,” in Proc. IEEE Int. Conf. Computer Commun. (INFOCOM),
shown that overall our proposed method can provide high (Paris, France), Mar. 2019.
[9] Geraci, G., Garcia-Rodriguez, A., Mahdi Azari, M., Lozano, A., Mez-
minimum SIR for the UAVs. Our results have also revealed zavilla, M., Chatzinotas, S., Chen, Y., Rangan, S., and Di Renzo, M.,
some interesting design guidelines such as the impact of the “What Will the Future of UAV Cellular Communications Be? A Flight
number of antenna elements and the DT angles on the UAV from 5G to 6G,” arXiv e-prints, p. arXiv:2105.04842, May 2021.
[10] Lin, X., Wiren, R., Euler, S., Sadam, A., Maattanen, H., Muruganathan,
SIR performance, and most importantly, our method has shown S., Gao, S., Wang, Y. . E., Kauppi, J., Zou, Z., and Yajnanarayana, V.,
little to no impact on the SIRs of the existing GUEs in “Mobile network-connected drones: Field trials, simulations, and design
the network. Thus, the proposed technique can be a strong insights,” IEEE Veh. Technol. Mag., vol. 14, pp. 115–125, Sept. 2019.
[11] Goddemeier, N., Daniel, K., and Wietfeld, C., “Role-Based Connectivity
candidate for deploying large-scale urban aerial systems in Management with Realistic Air-to-Ground Channels for Cooperative
the near future while maintaining the reliable and efficient UAVs,” IEEE J. Sel. Areas Commun., vol. 30, no. 5, pp. 951–963, 2012.
coexistence of UAVs and GUEs. [12] Singh, S., Sunkara, S. L., Guvenc, I., Bhuyan, A., Dai, H., and Sichitiu,
M. L., “Spectrum Reuse among Aerial and Ground Users in mmWave
Our proposed framework can be extended in several ways. Cellular Networks in Urban Settings,” in Proc. IEEE Consumer Com-
First of all, the duty cycle parameter β can be taken into mun. Netw. Conf. (CCNC), pp. 1–6, 2020.
account in the optimization framework to maximize the min- [13] Bhuyan, A., Guvenc, I., Dai, H., Sichitiu, M., Singh, S., Rahmati, A.,
and Maeng, S., “Secure 5g network for a nationwide drone corridor,” in
imum rate (instead of SIR) of both GUE and UAV since Proc. IEEE Aerosp. Conf., (Big sky, MT), pp. 1–10, 2021.
those who are associated with down-tilted antennas suffer from [14] Amer, R., Saad, W., and Marchetti, N., “Mobility in the sky: Perfor-
the reduced rate in our proposed framework. Moreover, the mance and mobility analysis for cellular-connected UAVs,” IEEE Trans.
Commun., 2020.
updated version of eICIC known as further enhanced ICIC [15] Amer, R., Saad, W., and Marchetti, N., “Towards a connected sky:
(FeICIC) can be considered in which traffic data is transmitted Performance of beamforming with down-tilted antennas for ground and
during ABS with relatively low power. Another interesting uav user co-existence,” IEEE Commun. Lett., pp. 1–1, 2019.
[16] Mei, W. and Zhang, R., “Cooperative downlink interference transmis-
study will be providing better connectivity and reliable mo- sion and cancellation for cellular-connected uav: A divide-and-conquer
bility (i.e., reducing ping-pong and handover failures) to the approach,” IEEE Trans. Commun., vol. 68, no. 2, pp. 1297–1311, 2020.
UAVs whose trajectories are known before. It is worth noting [17] Galkin, B., Amer, R., Fonseca, E., and DaSilva, L. A., “Intelligent Base
Station Association for UAV Cellular Users: A Supervised Learning
that, our proposed method will not be able to support UAVs in Approach,” arXiv e-prints, p. arXiv:2003.01287, Mar. 2020.
the regions where cellular infrastructures are not available i.e., [18] Kumbhar, A., Binol, H., Guvenc, I., and Akkaya, K., “Interference
over deserts or oceans. We may need to rely on high-altitude coordination for aerial and terrestrial nodes in three-tier lte-advanced
hetnet,” in Proc. IEEE Radio and Wireless Symposium (RWS), pp. 1–4,
aerials platforms or low earth orbital satellites for providing 2019.
reliable connectivity to UAVs in these extreme cases. [19] Azari, M. M., Arani, A. H., and Rosas, F., “Mobile cellular-connected
Another limitation of our proposed framework is that the uavs: Reinforcement learning for sky limits,” in Proc. IEEE Globecom
Workshops (GC Wkshps, pp. 1–6, 2020.
extra set of antennas will increase the overall energy consump- [20] Chowdhury, M., Saad, W., and Guvenc, I., “Mobility management for
tion of the network. Moreover, the DT angles of the down- cellular-connected uavs: A learning-based approach,” in Proc. IEEE ICC
tilted antennas can impact the SIR performance of the UAVs. Workshops, (Dublin, Ireland), pp. 1–6, 2020.
[21] Chen, Y., Lin, X., Khan, T., and Mozaffari, M., “Efficient drone mobility
Hence, joint optimization of UT angles, transmit power of the support using reinforcement learning,” in Proc. IEEE Wireless Commun.
up-tilted antennas, eICIC/FeICIC parameters, and DT angles Netw. Conf. (WCNC), pp. 1–6, 2020.
 17

[22] Chen, Y., Lin, X., Khan, T., Afshang, M., and Mozaffari, M., “5g air- [46] Venugopal, K., Valenti, M. C., and Heath, R. W., “Device-to-Device
to-ground network design and optimization: A deep learning approach,” Millimeter Wave Communications: Interference, Coverage, Rate, and
in Proc. IEEE Veh. Technol. Conf. (VTC2021-Spring), pp. 1–6, 2021. Finite Topologies,” IEEE Trans. Wireless Commun., vol. 15, no. 9,
[23] Guvenc, I., “Capacity and fairness analysis of heterogeneous networks pp. 6175–6188, 2016.
with range expansion and interference coordination,” IEEE Commun. [47] 3GPP,, “Enhanced ICIC for non-CA based Deployments of Heteroge-
Lett., vol. 15, no. 10, pp. 1084–1087, 2011. neous Networks for LTE,” TSG RP-100383, 3rd Generation Partnership
[24] Goldberg, D. E., Genetic Algorithms in Search, Optimization and Project (3GPP), Mar. 2010.
Machine Learning. USA: Addison-Wesley Longman Publishing Co., [48] Merwaday, A. and Guvenc, I., “Handover Count Based Velocity Esti-
Inc., 1st ed., 1989. mation and Mobility State Detection in Dense HetNets,” IEEE Trans.
[25] Mozaffari, M., Saad, W., Bennis, M., Nam, Y.-H., and Debbah, M., Wireless Commun., vol. 15, pp. 4673–4688, July 2016.
“A tutorial on uavs for wireless networks: Applications, challenges, and [49] Chowdhury, M. M. U., Bulut, E., and Guvenc, I., “Trajectory op-
open problems,” IEEE Commun. Surveys Tuts., vol. 21, no. 3, pp. 2334– timization in UAV-assisted cellular networks under mission duration
2360, 2019. constraint,” in Proc. IEEE Radio and Wireless Symposium (RWS),
[26] Amer, R., Saad, W., and Marchetti, N., “Toward a connected sky: (Orlando, FL), pp. 1–4, Jan 2019.
Performance of beamforming with down-tilted antennas for ground and [50] Elhachmi, J. and Guennoun, Z., “Cognitive radio spectrum allocation us-
uav user co-existence,” IEEE Commun. Lett., vol. 23, pp. 1840–1844, ing genetic algorithm,” EURASIP Journal on Wireless Communications
Oct 2019. and Networking, vol. 2016, no. 1, pp. 1–11, 2016.
[27] Euler, S., Maattanen, H., Lin, X., Zou, Z., Bergström, M., and Sedin, [51] Greenhalgh, D. and Marshall, S., “Convergence criteria for genetic
J., “Mobility support for cellular connected unmanned aerial vehicles: algorithms,” SIAM Journal on Computing, vol. 30, no. 1, pp. 269–282,
Performance and analysis,” in Proc. IEEE Wireless Commun. Netw. Conf. 2000.
(WCNC), pp. 1–6, Apr. 2019.
[28] Mozaffari, M., Lin, X., and Hayes, S., “Towards 6g with connected sky:
Uavs and beyond,” ArXiv, vol. abs/2103.01143, 2021.
[29] Chiaraviglio, L., D’Andreagiovanni, F., Liu, W., Gutierrez, J. A., Blefari-
Melazzi, N., Choo, K.-K. R., and Alouini, M.-S., “Multi-area throughput
and energy optimization of uav-aided cellular networks powered by solar
panels and grid,” IEEE Trans. Mobile Comput., vol. 20, no. 7, pp. 2427–
2444, 2021.
[30] Banagar, M., Chetlur, V. V., and Dhillon, H. S., “Handover probability
in drone cellular networks,” IEEE Wireless Commun. Lett., vol. 9, no. 7,
pp. 933–937, 2020.
[31] Salehi, M. and Hossain, E., “Handover Rate and Sojourn Time Anal-
ysis in Mobile Drone-Assisted Cellular Networks,” arXiv e-prints,
p. arXiv:2006.05019, June 2020.
[32] Banagar, M. and Dhillon, H., “3D Two-Hop Cellular Networks with
Wireless Backhauled UAVs: Modeling and Fundamentals,” arXiv e-
prints, p. arXiv:2105.07055, May 2021.
[33] Galkin, B., Fonseca, E., Amer, R., Dasilva, L., and Dusparic, I., “Reqiba:
Regression and deep q-learning for intelligent uav cellular user to base
station association,” ArXiv, vol. abs/2010.01126, 2020.
[34] Chen, D., Qi, Q., Zhuang, Z., Wang, J., Liao, J., and Han, Z., “Mean
field deep reinforcement learning for fair and efficient uav control,” IEEE
Internet Things J., pp. 1–1, 2020.
[35] Mei, W. and Zhang, R., “Uplink cooperative interference cancellation
for cellular-connected uav: A quantize-and-forward approach,” IEEE
Wireless Commun. Lett., vol. 9, no. 9, pp. 1567–1571, 2020.
[36] Mei, W. and Zhang, R., “Uplink cooperative noma for cellular-connected
uav,” IEEE J. Sel. Topics Signal Process., vol. 13, no. 3, pp. 644–656,
2019.
[37] Yaacoub, E. and Alouini, M.-S., “A key 6g challenge and opportu-
nity—connecting the base of the pyramid: A survey on rural connectiv-
ity,” Proceedings of the IEEE, vol. 108, no. 4, pp. 533–582, 2020.
[38] Hou, J., Deng, Y., and Shikh-Bahaei, M., “Joint Beamforming, User
Association, and Height Control for Cellular-Enabled UAV Communi-
cations,” IEEE Trans. Commun., pp. 1–1, 2021.
[39] Johnson, J. and Rahmat-Samii, Y., “Genetic algorithm optimization
of wireless communication networks,” in Proc. IEEE Antennas and
Propagation Society International Symposium, vol. 4, pp. 1964–1967
vol.4, 1995.
[40] 3GPP,, “Study on channel model for frequencies from 0.5 to 100
GHz,” Technical report (TR) 38.901, 3rd Generation Partnership Project
(3GPP), 2017.
[41] 3GPP, Technical Specification (TS) 36.777, 2018.
[42] Lin, X., Ganti, R. K., Fleming, P. J., and Andrews, J. G., “Towards
understanding the fundamentals of mobility in cellular networks,” IEEE
Trans. Wireless Commun., vol. 12, pp. 1686–1698, Apr. 2013.
[43] Rebato, M., Park, J., Popovski, P., De Carvalho, E., and Zorzi, M.,
“Stochastic geometric coverage analysis in mmwave cellular networks
with realistic channel and antenna radiation models,” IEEE Trans.
Commun., vol. 67, pp. 3736–3752, May 2019.
[44] Goldsmith, A., Wireless Communications. New York, NY, USA:
Cambridge University Press, 2005.
[45] Najibi, N. and Jin, S., “Physical reflectivity and polarization character-
istics for snow and ice-covered surfaces interacting with GPS signals,”
Remote Sensing, vol. 5, no. 8, pp. 4006–4030, 2013.