This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
1

Collection of Data with Drones in Precision

Agriculture: Analytical Model and LoRa Case Study
Antonio Caruso∗ , Stefano Chessa† , Soledad Escolar‡ Jesús Barba‡ , Juan Carlos López‡
∗ Department of Mathematics and Physics “Ennio De Giorgi”, University of Salento, Collegio Fiorini, Lecce, Italy.
† Department of Computer Science, University of Pisa, Pisa, Italy.
‡ School of Computer Science, University of Castilla-La Mancha, Ciudad Real, Spain.

Abstract—Unmanned Aerial Vehicles (UAVs) are autonomous appeared on the market1 .

devices employed as data collectors in precision agriculture to Typically, they exploit a fixed infrastructure of base stations
support a large number of applications. These UAVs gather
able to provide connectivity to IoT devices with very low
data from on-the-ground wireless sensor networks, especially in
scenarios that lack any kind of fixed communication infrastructurebandwidth requirements over a range much larger than usual
or where the available infrastructure does not fit the applicationWi-Fi, ZigBee or Bluetooth technologies (it can even reach
requirements. Sensors on the ground can store sensing data, and a few kilometers, depending on the shape and conditions of
in scenarios that do not require a real-time observation and the ground). An alternative approach is to replace the fixed
analysis of the data, like in smart farming, a drone can be used
infrastructure of base stations with mobile base stations that
maybe once or two times each day to collect and report the data
to a command-and-control center that use them directly, without are carried by Unmanned Aerial Vehicles (UAV), hereafter
any other infrastructure (cloud or edges). In this paper, we studycalled “drones”. These vehicles have now reached a level of
analytically how close the drone, that uses a LoRa radio, needs reliability at affordable costs that make them appealing for civil
to fly over the sensors to collect data with a given quality of data
applications, and their market is growing fast, with exponential
collection. This can be used to properly spacing the sensor on
growth expectations [9]. By means of drones, a single mobile
the field at deployment time, to select among different type of
drones, and to properly solve some tradeoff related to field size base station may provide connectivity to IoT devices over an
vs autonomy of the drones and the path used by the latter when area that may otherwise require even several fixed base stations
collecting the data. to be covered.
Index Terms—Internet of Things; Precision Agriculture; Opti- In this paper we consider this latter approach, and we study
mal Data Collection; Drones; LoRa. how an architecture comprising drones and sensors in the field
can be configured in order to reach minimum connectivity
constraints. We focus in particular on large row crop farming,
I. INTRODUCTION which are densely and regularly planted areas that can be
managed by machines across the entire field and that are widely
The adoption of Internet of Things (IoT) technologies [1] studied in the literature [10], [11].
is a recent trend in precision agriculture [2], [3]. In fact, due In this context we consider a regular deployment of sensors
to the ability to combine sensing, actuation, processing and in the field (namely in a grid), which maps well over the
Internet connectivity into low-cost devices, it is possible to regular structure of the crops. We also assume that a drone
monitor the crops, the soil and the environment where they travels periodically over the field carrying a base station to
grow with an unprecedented granularity. The consequence is communicate with the sensors in the ground, and that it moves
however the continuous streaming of sensor data that need to along a predefined path to ensure uniform coverage of all the
be collected and analyzed, in order to be transformed into high- sensors in the field. Under these assumptions we provide an
level information that can be interpreted by the agronomists analytic model that can be used to configure the path of the
and that can be used to activate the automatisms that feed and drone. The model can be used to determine the minimum-
assist the crops. On the side of the data analysis requirements, maximum distance that should be kept between each sensor
the recent developments in artificial intelligence and data and the drone, to guarantee a given probability of collecting
analytics are already being explored by researchers to produce the sensor data. The analytic model is general and can be
novel models suitable to interpret such data and relate them adapted to different wireless technologies with different ranges
to key parameters of the crops [4]–[6]. On the side of the and throughputs. As a case study we analyze, by means of
data collection requirement however, the connectivity to the simulation, the use of the model in a LoRa scenario, considering
Internet of IoT devices, which is provided by 4G/5G networks different specifications of commercial drones and field sizes.
in populated areas, cannot be given for granted in rural and The rest of the paper is organized as follows. After reviewing
remote areas, which is the typical case in large-scale, precision the related work in Section II, in Section III we dip into LoRa
agriculture sites. A study on the coverage vs. costs provided to review the set of low-level parameters that an application
by different communication technologies can be found in [7].
Also for this reason, new connectivity services based on 1 https://www.semtech.com/lora/ecosystem/networks,
long-range, low power technologies (such as LoRa [8]) have https://www.a2asmartcity.it/utilty-week-paris/

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
2

designer needs to know as they impact significantly on the In [23] the authors study the problem of maximizing the
communication performance. In Section IV we present a throughput of the down-link channel (so the focus is more
data collection model to compute a regular deployment of related to different application scenarios than data collection),
sensors on the target field that guarantees a given probability by proper scheduling and UAV’s trajectory planning. They
of successful data communication between the drone and the formulated a mixed integer non-convex optimization problem
sensors through point-to-point communications. We study the and solved it efficiently. In [24] the authors propose the use of
minimum-maximum distance for small and large deployments drones and LoRa to collect data from sensors, they change the
and discuss the analytical results and their impact on the default Aloha MAC in LoRaWan with a time-scheduled policy
drones autonomy and field size. In Section V we discuss to reduce the probability of packet collisions. Their simulations
the conditions for deployments admissibility and evaluate show that a drone can collect the data produced by 80 sensors
the feasibility of using LoRa as communication protocol for in a day (5760 bytes) in an area of 1.5km2 without packet
the model described, as well as its performance in terms of collisions. In [25], [26] the authors assume that sensors are
collisions and retransmissions. Finally, in Section VI we draw deployed randomly on a field and the data are preliminary
the main conclusions and suggestions for further research. collected by a set of cluster heads, it is also assumed that
the communication area of different cluster heads do not
overlap. The problem of computing the optimal (minimum
I I . R E L AT E D W O R K
latency) trajectory for a drone through all cluster heads, in
In a typical organization, wireless sensor networks (WSN) order to collect all data from them, has been studied. The
are organized as a multi-hop tree structure rooted in a device drone could move freely in the area, the number and locations
called sink [12]. Each sensing device (node) thus occupies a of the cluster heads, and the amount of data to be transferred,
position in this tree, and acts as a router by forwarding towards from them to the drone, are known ex ante. The problem
the sink all the data produced by the nodes in its subtree. This of minimizing the latency from the start of the drone to its
network organization can be managed by simple protocols, return to the base reduces to two sub-problems: an extension
but it does not scale well because the nodes closer to the of the TSP to obtain the optimal circuit through the cluster
sink in the network become overburdened by the flow of data heads, and the computation of the minimum hovering time
coming from the rest of the network and thus deplete faster for each cluster head so that all data can be transmitted to the
their batteries, thus implicitly producing a disconnection of the drone; since the problem is clearly NP-Complete, two different
network from the sink and thus making the WSN unavailable approximation algorithms are presented. In [27] the authors
[13]. A solution to overcome this problem consists in the use study how to dynamically control the speed of the drone in
of a controlled mobile sink (in practice a drone) that may order to maximize the data collection efficiency, while reducing
travel across the network to collect the data from the nodes the access congestion for the UAV-based base stations.
[14], [15]. A good survey of the literature on these approaches Another relevant aspects, that also concerns the use of drones,
can be found in [16] and [17], [18]. In the context of precision is related to cyber attacks and security. For these aspects we
agriculture [19], [2], the use of aerial observation, to better refer the reader to the literature [28], [29].
understand and monitor large fields from altitude, is widely The problem of data collection using LoRa has been faced
practiced. In recent years, there has been an increasing interest up in several works. In a typical deployment, the LoRa
in employing UAVs instead of ground mobile sinks. In the gateway is a fixed device installed on a radio mast or antenna
following we cover some works that use UAVs as a mobile tower connected to the power line, and the sensors transmit
data sink. data towards the gateway in one hop, thus implementing
In Wang et. al [20] an entire framework called FPPWR a simple tree topology of depth 1 (also known as a start
(Fast Path Planning with Rules) is proposed, it solves the topology), that may cover distances of up to around 15 km
problems of nodes positioning, anchor points searching, path in open areas as reported in [30], and which is sufficient
planning of the UAV, and data collection. They evaluated the for most large deployments. as reported in [30], which is
performance of their proposal by studying the time required to sufficient for most large deployments. However, even these
collect all data, the flight path distance, and the volume of data long distances could not be enough to interconnect end devices-
collected. In [21] the authors consider how to optimize the data to-gateways at rural areas that lack any network infrastructure,
collection time for multiple drones, when data transmission at outdoor areas with obstacles between the end devices and
time should be respected. They propose a predefined route gateways (where the wireless signal strength may be reduced
scheme and formulate it into the problem of route selection resulting into data losses and communication errors) and at
and communication association, that is solved with a log(m)- indoor scenarios. In such scenarios two main alternatives have
approximation ratio using a greedy algorithm, where m is the been studied: mesh networks as opposite to star topologies
number of cluster-heads. They show that the proposed greedy and mobile gateways, typically based on UAVs. In [30] the
heuristic is 154% at most and 145% on average of optimal authors address real experiments for three mobile scenarios,
solution. In [22], it is assumed that the UAV flies horizontally where a LoRa device is installed on a rotary lathe, on a
with a fixed altitude, and their authors study energy-efficient dash board of a car and on a boat, respectively, to measure
UAV communication with a ground terminal, optimizing the the LoRa/LoRaWAN performance (packet delivery ratio and
UAV’s trajectory, the communication throughput and the UAV’s coverage) under different effects and velocities. Their results
energy consumption. conclude that for speeds larger than 40 km/h, the LoRa

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
3

performance deteriorates while for low-speed mobility (lower regions depending on the local regulations. For example, in
than 25 km/h), the communication is still sufficiently reliable. Europe, LoRa operates in the 863 to 870 Mhz band and in
Many other works in the literature have analyzed the United States it operates in the 902 to 928 Mhz frequency
LoRa/LoRaWAN performance on different real and simulated band. The frequency band is divided in channels, their number
scenarios in terms of packet delivery ratio, maximum number depends in turn on the specific local regulations. In Europe, for
of sensors supported by a gateway (scalability), coverage example, there exist a total of 10 channels, three of which must
range, network density and collisions. For example, in [31] the be implemented by the LoRa end-devices. These channels are
authors question the scalability provided by LoRa and study 868.10, 868.30 and 868.50 Mhz, each one with a bandwidth
its behavior by means of a customized simulator LoRaSim of 125 Khz and a variable bitrate between 0.3 and 5 Kbps.
when different network settings are considered. They report LoRa uses a proprietary spread spectrum modulation named
that, for a typical smart city deployment (with a selected Chirp Spread Spectrum (CSS), where each bit of payload is
conservative transmission settings and one only gateway) with represented by multiple chirps. In LoRa terms, the amount
sensors sending 20 bytes packets each 16 minutes, a gateway of spreading code applied to the original data signal is called
can support 120 sensors per 3.8 ha, which is insufficient for the spreading factor (SF). LoRa modulation has a total of
future IoT deployments. These results are, however, very far six spreading factors (SF7 to SF12). The larger the spreading
from the theoretical results provided in [32]–[34]. factor used, the farther the signal will be able to travel and
In most of work above, we noted an emphasis in optimizing still be received without errors by the RF receiver, but also the
the communication from the sensors to the drone, in the lower bit rate and the larger energy consumption.
scenario that the sensors are limited in power. In our scenario, Another parameter is the coding rate (CR), which is the
if the sensors are equipped with an energy-harvesting device error correction coding given as A/B, where A is the data
they are not battery limited, hence it is the drone that becomes block length (A=4) and B is the codeword length; available
the bottleneck from the energy point of view. Moreover, the rates are 4/5, 4/6, 4/7 and 4/8. The data transmission rate (DR)
application of WSNs in precision agriculture requires that the is the amount of data transmitted by unit of time (bps) and it
deployment is planned beforehand with an agronomist that is depends on the SF, bandwidth, and CR as expressed by DR =
the main responsible to set the best position of the sensor in SF× BW 2SF
× CR, and it ranges between 0.3 and 27 Kbps.
the field. As we see in Section IV we assume that sensors are The transmission power ranges between 5 and 23 dBm,
deployed regularly in a grid, since it is a common and quite where higher power increases the energy consumption. Finally,
practical assumption for this scenario. LoRa specifies the packet format, which is composed of three
elements: a preamble (8 symbols by default), an optional
I I I . L O R A S P E C I F I C AT I O N S A N D C O N S T R A I N T S header, a variable-length payload and a CRC (1 byte). The
preamble enables the receiver to synchronize with the incoming
LoRa (Long Range) [8] is one of the enabling technologies
data; the header if exists, includes metadata to inform the
for Low-Power Wide Area Networks (LP-WANs) [35], which
CR, the payload length and payload CRC presence. The
are characterized by a long range communication, a low data
payload includes the data coded at the CR specified. Note
rate and a low energy consumption. LoRa is especially targeted
that varying the spreading factor (SF) changes the range of
for connecting devices in applications that need to send small
communication between two devices; therefore, in rest of the
amounts of data a few times per day over long distances (in the
paper, any change in the communication range between the
order of kilometers), which fits well with very diverse outdoor
drone and a sensor deployed on the field, directly expressed
monitoring scenarios, as for instance, precision agriculture. In
as a distance in km, can be interpreted as a proper choice
a typical precision agriculture setting, a single device may
of SF in the LoRa radio of both devices, that is able to
embeds several different transducers for the monitoring of the
support that minimum range, if the value exists. The Table
crops and of the soil, and it may transmit all the sampled data
below (from https://lora-developers.semtech.com/) provides
into one single LoRa frame, provided it does not exceeds the
some combination of SF, Bit Rate, Range, and Time on air of
maximum allowed payload. Note that the actual payload size
a packet.
depends on the LoRa configuration and may range from a
few tens to a few hundreds of bytes, and may thus contain
several sensed data. The reader can consult the work in [36]
for a comparison between LoRa and other communications
technologies for LP-WANs as are SigFox and NB-IoT. The
next subsections describe briefly the LoRa specifications and
its major constraints. Note that we do not cover the details of
LoRaWan, a network specification that use the physical layer,
in the rest of the paper we refer only to LoRa and do not use
the LoRaWAN layers. B. Limitations on the duty cycle
Theoretically, hundreds of devices (sensors and gateways)
A. The LoRa physical layer could use the same channel for data transmission. In order to
LoRa uses a wide range of license-free sub-gigahertz radio make a equitable sharing of the channel between the radio-
frequency bands and operates in different bands in different based devices, the governments have regulated, among other

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
4

TABLE I
MOST I M P O R TA N T S Y M B O L S U S E D I N T H E PA P E R W I T H T H E I R
UNITS.
r
N number of sensors - v
g distance between sensors meters d
x field width (g multiple) meters k
y field heigth (g multiple) meters
r range of wireless comm. meters
∆ transmission period seconds
A drone autonomy seconds
D path-length of the drone meters
v = D/A drone’s velocity m/s
d in-range distance meters Fig. 1. The picture is centered on the drone that moves vertically with constant
T = d/v drone time of flight over the sensor seconds speed v. A sensor remains into the communication range for time T = d/v.
pk probability of receiving a packet (0,1) The distance 0 < k < r from sensors to drone is the optimal distance that
maximize the packet reception probability.

k drone-sensors distance meters

Since the drone passes occasionally over the sensors, each

parameters, the maximum duty cycle (DC), i.e. the fraction of sensor is programmed to log the sensed data in its memory,
time during which a device using unlicensed bands can occupy to aggregate and compress these data, and to transmit the data
a channel [37]. In Europe, the ETSI EN300.220 standard when the link with the drone is available. We assume that the
defines the maximum limits of DC for each sub-band in the sensors and the drone are only weakly synchronized, hence
spectrum. For example, in the 863 to 870 Mhz frequency band, the sensors have knowledge of the approximate time of the
these limits range between 0.1% and 10%. The duty cycle next flight of the drone, so that they can activate their wireless
limit applies to the total transmission time in the period of interfaces when the drone is expected to come. Once activated,
reference of one hour. For example, with a DC of 1% a LoRa they start transmitting periodically their data (according to the
device (both a sensor and a gateway) can transmit up to 36 duty cycle they can sustain with their radio interface) and stop
seconds per hour in each sub-band for each end-device. transmitting either if the drone acknowledges their transmission
In practice, each application may specify the transmission or when the time frame in which the drone is flying over them
mode by defining the most adequate SF, bandwidth and CR ends. Since the sensors are not synchronized among themselves,
and compute, in advance, the maximum time of transmission collisions in transmission are also possible. If at the end of the
for their messages (airtime) based on the calculations of LoRa transmission time frame a sensor was unable to send its data
modulation [38]. This limitation of the duty cycle may reduce to the drone, it will try to send the data again at the next flight
not only the suitability of LoRa for some IoT applications of the drone, along with the new data sensed meanwhile.
(e.g. the ones that require real-time data transmission at high Usually the sensors have a low duty cycle, within which
frequencies or the transmission of huge amounts of data) but each sensor transmits its data with a period ∆. In order to
it may also reduce the amount of sensors that a gateway can transmit successfully data, the sensor must be in the range of
successfully support. In Section V, we will use also a LoRa the drone flying over it. The drone has an autonomy of flight
simulator to test the feasibility of using LoRa radios for the of A hours; hence, letting D be the distance covered in its
data transmission following the model described in the next flight plan its average speed is v = D A . Hereafter, we assume
section. the drone always moves at constant speed v. Note that v, D
and A are not free variables but are, in fact, constrained. For
example, D depends on the size of the field and on the flight
I V. T H E O R E T I C A L M O D E L A N D A N A LY S I S
plan (in turn also dependent on the wireless coverage of the
Let us consider a field of size x · y, where a set of N drone), the autonomy depends on the size of the battery that
sensors are deployed in a grid, at a distance g from each can be carried by the drone and on its consumption, while
other (we   that x and y are both multiple of g), and the speed depends on the technical specifications of the drone,
 assume
N = g + 1 yg + 1 . The sensors collect micro-climatic but it is also limited by the maximum speed that allows the
x

data and data about the state of the soil and surrounding wireless interface to successfully communicate.
plants. We assume that sensors are wireless-connected, battery- While swiping the field, a radio link of length r does not
powered, and (possibly) energy harvesting [39], so that they guarantee that we can successfully receive a packet from all
do not require deployment of wired infrastructures as they are sensors in that range, since the delay and the probability of
completely autonomous. The data collection is performed by collisions between packets of different sensors must be taken
a drone that acts as data mule and gateway of the network. It into account. In the following, we study the upper bound k
swipes the field according to a predetermined flight plan and (k < r) to the minimum distance between the drone and a
it collects the available data from the sensors that are into its sensor such that the probability of receiving a packet pk can
communication range. We assume that this range is such that be maximized (we also call k the maximum minimum distance
the drone can communicate with sensors placed up to distance of the drone from the sensors). In particular, we analyze the
r from its projection on the ground. A summary of the symbols relationship between k and the other parameters in the model,
and parameters used in this work is presented in Table I by providing proper bounds to the probability pk with respect

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
5

to k. Figure 1 shows the relationship between k and the time

100%
of flight of the drone over a sensor, for a given k < r the
drone flies over such a sensor for a time T = vd which is
the time elapsed from when the drone gets in range with the 80%
sensor, to the time
√ in which it gets out of range. Since the

k as percentage of r
distance2 d is 2 r2 − k 2 we √ have that the time of flight over 60%
r 2 −k2
the sensor is T = vd = 2A D and during
 T this
 period the
sensor transmits up to tx times, where tx ≥ ∆ times, which
T 40%
we lower bound with tx ≥ ∆ − 1. pc
Considering that the sensors that transmit the data to the 1%
drone are not synchronized, it is possible that some communi- 20% 5%
cations collide. This is especially true when the transmission 10%
range is very large (this is the case if LoRa is used for example), 15%
0%
and thus the number of sensors within the transmission range 0.5 1 1.5 2 2.5 3 3.5 4
of the drone is also very large. Furthermore, packet loss is Communication Range r [km]
possible due to environmental conditions; we model packet
loss with the parameter pc that expresses the probability that
Fig. 2. Maximum value of k as a percentage of the communication range r.
a packet transmitted by a sensor is not received by the drone
(see Figure 13 for plots of pc in LoRa). Consequently, the
probability pk can be estimated as: shows the maximum k as a percentage of r when: vmax =
pk = 1 − pTc /∆−1 . (1) 11.11 m/s = 40 km/h, ∆ = 60s, p = 0.98 and pc equals to
1%, 5%, 10%, 15% for different curves. We see that when r
In general, to assess the value of parameter k, we need is small, in particular below 1km, the maximum value of k
to meet the requirement that the probability of successful could be even lower than half the range. The value of k is
transmission is at least3 p0 , i.e. pk ≥ p0 . Recalling that also a decreasing function of pc , on the other hand, when r
T T /∆−1 T /∆−1
tx ≥ ∆ −1, follows that √
1−pc ≥ p0 ⇒ pc ≤ 1−p0 . increases over 2 km, k is over 80% of r since the negative
2
r −k 2
Introducing T = 2A D , under the constraint pk ≥ p0 part inside the square root is a constant, and it is negligible
follows: as r increases. The gap in the lower left part of some curves
√ is where k is not defined: when r is too small, the expression
T 2A r2 − k 2 log(1 − p0 )
= ≥ +1 (2) inside the root is negative. Until this point we do not have
∆ D∆ log pc specified the flight path followed by the drone. In the next
Note that if k is very close to r then the time T (numerator subsections, we improve the details of our model. To explicitly
above) decreases. Hence, as k approaches r, T goes to 0 and introduce the relation between the path length D and k, we
so does pk . On the other hand, when k is small and v is limited, define two simple flight plans, one for very large fields and
the length of the path D increases (it is inversely proportional one for medium/small fields.
to k intuitively, if the drone flights closer to the sensors, the
path length increases) and, based on the autonomy of the drone,
A. A model for large fields
it could be infeasible to cover all its length. So, the goals of
maximizing pk and minimizing the path length D are mutually In large fields we assume that g < r  x and g < r  y,
contrasting, and we need to study an optimal trade-off among which can be obtained if the field is big as compared to the
the two. From inequality (2) follows that: transmission range. This condition in practice can be obtained
also if the field is not so big but the transmission range is small
log(1 − p0 ) 1
p  
2
r −k ≥v·∆2 + (for example when WiFi is used to communicate between the
2 log pc 2 drone and the sensors). The assumed flight plan follows a
and solving for k we obtain: conventional plan suitable to cover a regular distribution of
s the sensors in the field. It has the same starting and end point,
2
0 so that the drone can return to the base after the flight. An
 
log(1 − p ) 1
k ≤ r2 − v · ∆ + (3) example of such a flight plan is shown in Figure 3, it is a
2 log pc 2
back-and-forth or zig-zag path, studied in several papers (see
The maximum of k is obtained from the bound above with [40], [41]).
v = vmax , with vmax be the maximum speed determined both Note that, for the sake of simplicity, in the rest of the section
x
by the drone specs or by the specs of the communication we assume that 2k is integer and odd, so that the flight plan of
interfaces that may not work well at higher speeds. Figure 2 the drone is a straightforward generalization of that shown in
Figure 3. Note also that the flight plan can be easily generalized
2 We assume that the drone flights linearly over short distances so we can x
to the case in which 2k is even. If the drone follows the path in
always approximate d with a linear segment, so we are disregarding turns off Figure 3, the distance traveled by the drone can be computed
the drone.
3 In a single pass of the drone over the sensor, as in the flight plan the drone as follows: The number of movements of the drone along
may pass over the same sensor more than once. the y-axis are: x−2k x
2k + 1 = 2k , The total distance covered

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
6

k 2k general higher. Secondly, the model disregards the fact that

the time of flight above the sensors that are at the turns of the
drone (the corners of the flight plan) will be smaller than for the
other sensors (this means that these sensors will have a smaller
probability of successful transmission). Under the assumption
g that the field is big as compared to the transmission range, the
y
number of sensors in this situation is small. Furthermore, this
can be dealt with by letting the drone slow down its speed
during the turns, so to give more time to these sensors to
communicate. Again, under the assumption that r  x, y the
effect of these slow-downs would be negligible on the model.
In the next model, for small fields, we instead analyze this
effect in more detail.

B. A model for small fields

x
When the field is small, for example, when the communi-
cation range r between the sensors on the ground and the
Fig. 3. An example of a zig-zag or serpentine path followed by the drone. drone is very large and is close to at least one of the two
The sensors are spaced regularly in the field every g meters, the domain is dimensions x or y, i.e. x and y are comparable to k ≤ r,
oriented such that x ≤ y and the path follow a direction along the y. a more accurate model is possible. In this model, the drone
performs a rectangular path, as shown in Figure 4 above, and
we consider the case y = min(x, y) ≈ r, the path length is
by the drone along the y-axis (parallel to the y-axes) is thus
x 2 ∗ (x − 2k) + 4k for a total distance D = 2x. Note that in
2k (y − 4k) + 4k; The distance covered by the drone along the this case we bound the height of the domain by y ≤ 4k to
x-axis (parallel to the x axis) is: 2(x − 2k) = 2x − 4k, hence:
ensure an appropriate coverage also of the central part of the
D = 2k x
(y − 4k) + 4k + 2x − 4k = xy 2k . field. Similar to the previous model we have the problem of
Taking the distance as a constant, we can express the speed
xy the sensors that are at the corner of the field, that can be at
of the drone as v = 2kA , and introducing T = vd , from (2) we
distances greater than k. We solve this problem altering the
obtain:
time velocity of the drone.
√ For what concerns the time of flight of the drone on the
4kA r2 − k 2 log(1 − p0 )
 
sensors, if we exclude the sensors√that are close to the corners
≥∆ +1 ⇒
xy log pc of the field, we have that T = 2 r2 − k 2 /v. For the sensors
xy∆ log(1 − p0 ) that are closer to the corners instead, and especially for those
p  
k r2 − k2 ≥ +1 that are exactly at the four corners, the time of flight is shorter
4A log pc
to compensate to the larger distance to the drone. The situation
0
 
Letting h = xy∆
4A
log(1−p )
log pc + 1 , follows k 4 −k 2 r2 +h2 ≤ of these sensors is shown in the bottom part of Figure 4.
0. Considering the constraint r > k > 0, must be: Considering the sensor at the corner, the distance covered
by√ the drone while in range of this sensor will
√ be d0 =
s √ s √ 2 2 0 2 r 2 −k2 −2k
2 4
r − r − 4h 2 r2 + r4 − 4h2 2 r − k − 2k, and the time of flight T = v =
≤k≤ (4) T − 2k . In order to provide the same coverage of these sensors
2 2 v
as well, we assume that the drone stops in each corner for a
Now, to maximize pk we need to take the value of k that time equal to 2k/v, so to ensure a time of flight of those sensors
2
−2k2 )
maximizes T . The first derivative of T is 4A(r√
xy r 2 −k2
, which is of at least T . To consider these delays into the drone flight
null for k = ± √r2 (under the constraint r > k > 0). Since T autonomy, assuming the drone moves at the slowest possible
grows in the range [0, √r2 ], k = √r2 corresponds to a maximum speed v to cover the entire distance within A (note that this is
of T , and the value of k that maximizes pk is hence: desirable because a lower speed favors a longer time of flight
 s √
 over the sensors and thus a higher probability of receiving the
r r 2+ r 4 − 4h2
 = √r transmissions), we have: A = 2x 2k
v + 4 v and then v =
2x+8k
A .
k = min  √ , (5)
2 2 2 In summary we have:
• D = 2x
2x+8k
So, the optimum value of k is a very simple function of • v =
√A
r, and insensitive to other parameters of the model. There A
• T = 2 r 2 − k 2 2x+8k
0
are some additional considerations. This model determines the T /∆−1
• pc ≤ 1 − p0 and then ∆ T
≥ log(1−p
log pc
)
+1
value of k aiming at maximizing the probability that the drone
From which, introducing T , we obtain:
receives the transmission from each sensor with a single pass.
Actually, for many sensors the drone actually does two passes √
r2 − k2 ∆ log(1 − p0 )
 
(remember that the actual transmission range is r > k), which ≥ +1 (6)
means that the probability of receiving the data will be, in (x + 4k) A log pc

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
7

x
100%

4k y

k
98%
p0 = 98%

Probability of correct reception pk

k∗ = √r
2

96%

94%

pc
k
1%
92% 5%
10%
15%
20%
Path of
r d0 the drone 25%
90%
400 600 800 1000 1200 1400 1600 1800 2000
Values of k [m]

Fig. 5. Large Fields: Probability pk as a function of k plotted for different

values of pc . Others parameters: r = 2km, k∗ = 1414.2km, x = 8484m,
y = 11312m, A = 1 hour, vmax = 40km/h.

Fig. 4. top) Example of a reduced field to be covered by a drone. The A. Analysis of Cases and Application of the Model
rectangle is created with parameters: x = 22k, and y = 4k so N = 36.
bottom) Distance d0 inside the range of a sensor at the corner of the field.
In the following we report some analysis of the two models
described in Section IV-A and IV-B for different scenarios.
We start with large fields, then move to small ones. In all the
analysis the range of communication is considered a constant
log(1−p0 )
 
Letting h = ∆ + 1 , follows (16h 2
+ 1)k 2
+ and set to r = 2km, a conservative value for technologies like
A A log pc
8xh k + h x − r ≤ 0. Considering that r > k > 0 and that LoRa. Enlarging the communication range does not alter0 the
2 2 2 2

h > 0, the latter equation has two real solutions (k1 , k2 ) and qualitative analysis that is obtained here. The parameter p that
the admissible values of k are in [k1 , k2 ]. Also in this case, affects the expected bound on the probability pk of receiving
the value of k that maximizes pk is the one that maximizes T , a packet from the sensors to the drone, is always set to 0.98.
considering that the first derivative of T is − √rA(kx+4r
2
)
, by The plot in Figure 5 represents the values of pk as a function
2 −k 2 (4k+x)2
of k obtained with different values of pc . The values of x, y
means of a simple analysis (and assuming x > 4k), we observe
are set to be multiple of 2k, the optimal value of k is equal to
that T is decreasing for k ∈ [0, r]. Hence the maximum of
1414.2km that is 70% of r, and represented in the figure as
probability pk is reached when k = k1 :
k ∗ . The others two solutions computed in (4) are shown at the
p intersection of the curves with the horizontal line with pk =
−4xh2 + (16h2 + 1)r2 − x2 h2 0
k1 = (7) p ∗ 100 = 98%, we call these two values kmin , and kmax in the
16h2 + 1 following. We note from the plots that increasing pc decreases
the values of pk as expected, but only marginally at k ∗ , this
V. S I M U L AT I O N S A N D E VA L U AT I O N effect is more emphasized for kmax that even if suboptimal will
play an important role in practical application as we will see.
We validate our approach and the model by means of Choosing k = k ∗ guarantees that pk is maximized but it also
simulation. Firstly, in Section V-A we evaluate analytically affects the path length of the drone, and thus the required flight
the two models for large and small fields previously described autonomy to cover the field.
in order to determine how the parameters of the models can be Figure 6 shows the minimum flight autonomy A as a
optimized. Secondly, in Section V-B we look for commercial function of k (note that for each value of x in the legend,
drones to validate their suitability as data collectors in our we set y = 2x). The figure shows that the autonomy of the
scenario, and then we face up the usage of LoRa as physical drone increases when k decrease and for k = k ∗ it exceeds 300
layer technology to evaluate by simulation its performance minutes (5 hours), a value that is far greater than the higher
in terms of collisions and successful reception probability for autonomy of most commercial drones (see Table II). Note also
several deployments and LoRa configurations. that, in practical applications, it is important to distinguish

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
8

541 x [km] 2000 x [km]

5 5
500 7 7
9 9
450 11 1900 11
433
k∗
13 13
400 15 20 15
Minumum Drone Autonomy [min]

382
17 17
19 1800 19
350
337

Values of k [m]
306
300
270 1700
253
250 238
50
216 30
200
181 179 1600
168
100
150 170
128
126
121

100 210
85 90 1500
73 70
51
60 130
50
37 36
26
18 26 51 85 128 179 238 306 382
0 1400
1000 1200 1400 1600 1800 2000 50 100 150 200 250 300 350 400
Values of k [m] Minimum Autonomy of the drone [min]

Fig. 6. Analysis of the minimum autonomy of the drone, as a function of k Fig. 7. Analysis of the value of k with different field side x with y = 2x as
and pc . a function of the autonomy of the drone.

between k and r. In fact, the minimum autonomy that could

be obtained by setting k = r would be much lower, but it would 95%
not take into account the real path that must be followed by the
drone in order to guarantee the desired probability of collecting
90%
the data from the sensors. This effect is particular noticeable
for large fields, since the slope of the curve for larger values
k (r %)

of x is steeper. 85% pc
From all the analysis done so far, we see that setting k = k ∗ 1%
has several limitation in practice, specifically, if we fix the 5%
80%
10%
domain size, the minimum autonomy required for the drone is
15%
too high, this can be approached in different ways. A possibility 20%
is the use of multiple drones, splitting the domain in smaller 75%
25%
parts, another possibility is to consider the data collection phase
2 3 4 5 6 7 8 9 10
to be composed of several flight missions that alternate between Domain width x [km]
different part of the domain, still using a single drone. However,
a simple solution is also to increase the value of k over k ∗ as
Fig. 8. Values of the maximum-minimum distance k as a function of the field
much as possible, i.e. taking k = kmax . In Figure 7, we plotted side length x.
the values of kmax as a function of the autonomy of the drones
(in minutes), for different field size, moreover, we can compare
the values in the curves, with the values required if we used effect can be seen if we increase the side of the domain x. The
k = k ∗ represented as dots at k ∗ = 1414.2. From the plot, we plot in Figure 9 shows that the time T that the drone spends
see clearly that using the maximum value for k, reduces the in the range of a single sensor decreases as k increases, but
autonomy required to cover a field of 19km×38km from 382 as we see in the next section, it is still sufficiently large with
minutes to 210 minutes, a gain of more than 54%. respect to the time required by a packet to be transmitted in
Now we study the case of small fields, using Equation 7. In LoRa. The time is also a decreasing function of x, since a
this case we set the autonomy of the drone to A = 40 minutes bigger domain requires an increased velocity of the drone to
and ∆ = 60 sec. We set p0 = 0, 98 and r = 2 km as in the complete the path.
previous case. Figure 8 shows k as a function of x. Specifically, Finally Figure 10 plots pk for different values of k and for
the curves represent k as a percentage of r for values of pc different packet collision probabilities pc . The values of the
ranging from 1% up to 25%. As expected, the plot shows that optimal values for k for different values of pc , obtained using
if we consider x constant, when pk increases, k decreases (so equation 7 are annotated in the plot. As we see from the plot,
increasing the time spent in the range of a sensor); the same pk decreases with increasing pc and with increasing k, here

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
9

1400
LoRa in terms of collisions and successful reception probability
x for several deployments and LoRa configurations.
4km 1) Admissibility Conditions: As described in Section IV,
1200
6km given a field and a drone defined by the tuple hx·y, r, A, ∆i, we
8km
Interconnection Time T [s]

1000 may compute the corresponding optimal deployment defined

10km
12km
as hk, N, v, Di. Such a deployment is admissible, that is, the
800 14km sensors in the field are all covered by a drone, if two conditions
hold: 1) v ≤ vmax , i.e the velocity must be in the range of
600 velocities supported by the drone, and the maximum speed
allowed by the wireless interface to successfully communicate
400
(in LoRA, for example, and according to [34] this value is
40km/h); and 2) the time required to cross the path should
200
be lower then A (the autonomy of the drone, as provided in
the drone specifications), that is, D v ≤ A. Table II reports the
0 250 500 750 1000 1250 1500 1750 specifications of some current commercial drones. Note that the
Values of k
size of the field is, therefore, the key factor that determines the
admissibility condition, since it affects directly the parameters
Fig. 9. The time T in second, available for communication between a drone v and D.
flying over the field and a sensor on the ground as a function of the maximum-
minimum distance k. 2) LoRa Feasibility and Performance for Admissible Deploy-
ments : Let us assume the drone collects data from N sensors
in a field, by using a LoRa communication interface, and that
100%
the sensors have a maximum duty cycle of 1%. Concerning
97% the feasibility of the LoRa communication interface to our
883 1110 1307 1492 1672 p0 = 98%
purposes, we assess two conditions: 1) the time for a packet
95% transmission is lower than the time in which a sensor remains
into the communication range of the drone, i.e., Tpacket ≤ T ;
92%
Value of pk

and 2) Tpacket · npacket ≤ Tairtime (remind that Tairtime is

90% limited to 36 seconds per hour with a DC = 1%). Note that we
pc
1%
assume that each sensor performs data aggregation and, thus
87%
5% transmits one only packet of a certain payload during the flight
85% 10% of the drone, i.e. npacket = 1.
15% Figure 11 presents the transmission times Tpacket for packets
82% 20% with payloads between 10 and 100 bytes (amount of data
25%
that a node transmits in an only packet), different SFs and
80%
800 1000 1200 1400 1600 1800 2000 a CR fixed to 4/5. As expected, these times grow with the
Values of k
payload and SF, since the data size to transmit is larger. This
coincides with the numbers already reported in many papers
Fig. 10. Value of pk as a function of different values of k. The optimal value that analytically compute the LoRa performance [32], [42]:
of k are annotated in the plot for different values of pc
a larger SF implies a longer transmission time and a longer
communication range. In fact, an SF i allows to send 2 times
the different model shows a higher sensibility to the value of more bytes than an SF i + 1 in the same time or, alternatively,
k, and the optimal values of k ranges from 883 = 44% of r allows to reduce the time approximately to the half for a
to 1672 = 83% of r. same payload. A higher CR implies both a larger overhead
To conclude, we observe that in small fields we can use and a higher reliability of communications. Note that the times
lower values of k depending on the values of pc , in order Tpacket shown in the figure, hold the two conditions that ensure
to optimize pk , the reduction in k, increases the path length, the LoRa feasibility: 1) Tpacket is always lower than the times
but since here we are not particularly limited by the drone’s T provided in Figure 9 on the right, even for the smallest times
autonomy, there is much more flexibility in the choice of k. that correspond to the largest fields evaluated, with the largest
values of x, y and k; and 2) since we assume npacket = 1
the second condition also holds. Note, however, that if the
B. Case Studies Simulation sensor tries the continuous transmission of packets within the
We now assess the suitability of the specifications of typical same transmission period ∆ and on the same channel (i.e.,
commercial drones and of LoRa radio interfaces for the data npacket > 1), for instance due to retries as consequence of
collection process described in Section IV. To this purpose, we collisions or packet losses, the maximum number of packets
analytically determine the packet transmission times and the transmitted should not exceed the upper bound defined in
maximum density of the sensors (that is, the maximum number Section III-B. We can conclude here, therefore, that LoRa may
of sensors in a transmission range) that can be served by a be successfully adopted as a communication protocol for the
drone. Finally, we evaluate by simulation the performance of data collection process described in this work.

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
10

TABLE II
S P E C I F I C AT I O N S OF COMMERCIAL DRONES. LEGEND: 1: DIAGONAL WEELBASE OR LXWXH DIMENSIONS; 2: WEIGHT INCLUDES
PAY L O A D . N A : N O T A VA I L A B L E .

Name Manufacturer Type Controller Dimensions1 Battery Max Speed Weight2 Max. Flight Price
(mm) (mAh) (km/h) (Kg) Time (min)

eBee SQ senseFly Fixed Wing 1100 4900 LiPo 3S 110 1.25 55 12000$
Lancaster 5 PrecisionHawk Fixed Wing 1500 NA 79 3.55 45 25,000+$
Firefly6 BirdsEyeView Aerobotics Fixed Wing 1520 NA 64 4.5 50-59 8000$
AgDrone System HoneyComb Fixed Wing 1245 8000 3S LiPo 82 2.15 55 >10000$
Mavic 2 Pro DJI Quadcopter 322x242x84 3950 LiPo 4S 72 1.1 31 1499e
Phantom 4 DJI Quadcopter 370x325x220 6000 LiPo 2S 72 1.58 28 1153e
3D IRIS+ 3D Robotics Quadcopter 550(L)x100(H) 5100 3S 39.6 1.48 20 750$
Aero Ready to Fly Intel Quadcopter 360(L)x222(H) 4000 Li-Po 4S 53.76 2.765 20 1099$
Autel EVO Autel Robotics Quadcopter 338 4300 Li-Po 72 1 30 1049$
Anafi Parrot Quadcopter 175x240x65 2700 2 cells 55 0.32 25 699$
HYBRiX-2.1 Quaternium Quadcopter 1630x913x509 Petrol/Battery 80 5 240 NA
H2 Quad EnergyOR Quadcopter 1200 (diagonal) Fuel Cell/Battery NA 10 120 NA
MATRICE 600 PRO DJI Heptacopter 1668x1518x759 6 45 00 LiPo 6S 65 15.1 35 5699$
Xena OnyxStar Octocopter NA 10000/6S 50 5.6 37 NA

4000 1000
SF SF
3500 7 7
8 8
800
Packet transmission times (ms)

3000 9 9
Max. number of nodes

10 10
2500 11 11
600
12 12
2000

400
1500

1000
200
500

0 0
20 40 60 80 100 20 40 60 80 100
Payload (bytes) Payload (bytes)

Fig. 11. Packet transmission times for different payloads and SFs, taking Fig. 12. Maximum number of sensors that can be covered by a single drone
CR=4/5. for different payloads and SFs, taking CR=4/5.

heard simultaneously by a single drone

Let us denote with Nr the maximum admissible number 3) Collision Probability: To analyze the probability of colli-
of sensors per transmission range r and let us assume, for sion pc we have used LoRaFREE4 [43], a comprehensive LoRa
the sake of simplicity, that the drone uses for data collection simulator written in Python that supports packet error model,
a single channel of 125 kbps. Concerning the maximum the orthogonality of SFs, the fading impact, the duty cycle
density of sensors that the drone can serve, we estimate limitation at both the sensors and the gateway, retransmisions
the maximum amount of packets of a certain payload that and acknowledgments, and energy consumption profiling. For
the drone is able to receive during its activity cycle (<1%) our evaluation purposes, we modify this simulator to have an
under perfect synchronization, i.e. without overlapping among adapted version that simplifies the process of data transmission
packets. Figure 12 shows the maximum number of sensors in order to transmit exactly one packet of data. We have also
that can theoretically communicate with the drone, where each adapted the target scenarios to simulate a square field of area
sensor transmits exactly a data packet with different size data x · y where the sensors are regularly distributed on a square,
using a CR=4/5. As expected, the widest coverage is achieved keeping a distance between each pair of sensors of g = 200
when the payload is minimum and the SF is the lowest. meters. In our experiments we assume that the drone is located
On the other hand, the coverage is minimum, with only 9 at the center of the square (0,0) and it has a radius of r. Table III
sensors, for the maximum payload and the highest SF=12. It shows the number of sensors that drop within the radius r,
is clear, therefore, that an application designer should keep
the payload into the limits that ensure that Nr sensors can be 4 LoRaFREE: https://github.com/kqorany/FREE

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
11

TABLE III 100

SCENARIOS S I M U L AT E D W I T H LORAFREE.
86.58%
x, y[km] 1 3 4 5 6 7 8 9
80
r[km] 0.5 1.5 2 2.5 3 3.5 4 4.5

Retransmissions (%)
N 36 49 81 121 169 225 289 361
Nr 16 29 49 81 113 149 197 253 60

40
Payload
1.0 Payload (bytes) 20 bytes
20-pc 20-prx 40 bytes
Collision (pc ) and succesful rx probabilities

40-pc 40-prx
20 60 bytes
60-pc 60-prx
0.8 80-pc 80-prx 80 bytes
100-pc 100-prx
100 bytes
0
50 100 150 200 250
0.6 Number of nodes

0.4 Fig. 14. Percentage of retransmissions for the deployments in Table III.

0.2 pc =0.1664 following, average values are shown.

Figure 13 shows the collision probability pc (lower curves)
0.0
and probability of successful reception (upper curves) reported
50 100 150 200 250 by the simulator for the deployments in Table III. As shown,
Number of Nodes
pc keeps always under 0.2, it increases with the number of
sensors until a certain point when this probability reduces
Fig. 13. pc (lower curves) and probability of successful reception (higher slightly due to the ratio between the number of collisions
curves) for the deployments in Table III.
and the total number of packets transmitted starts to decrease
(number of collisions does not increase significantly but the
denoted as Nr for different fields. Note, therefore, that in our number of packets transmitted is much higher). The probability
simulation scenario not all sensors are covered by the drone of successful reception drops (below 0.2 in the worst case)
with a radius r, but the optimal paths for large and small when the number of sensors increases because the load of the
fields described in Section IV ensures a complete coverage of network grows, either with more packets and longer packets,
sensors. Note, also, that this simulation scenario represents a and therefore the probability of errors also increases (e.g. due to
small fragment of the flight of the drone and, with its movement, not only collisions but packet losses or bad formed packets that
the drone enters similar fragments (squares). need to be disregarded). Note that for the smallest deployments
In our simulations each sensor transmits one only packet in this probability is close to 1.0.
each transmission period ∆, directly to the drone and in a single Since the probability of successful reception lower than (1 −
hop, using an optimal LoRa radio configuration that selects pc ), this means that, with the current network load managed,
the lowest SF that guarantees the sensor has a higher received the number of collisions is small and it does not represent
signal strength indicator (RSSI) than the receiver sensitivity, the main reason for packet retransmissions, which may be due
a CR=4/5 and a channel of 125 kbps. After transmitting the mainly to other causes as previously mentioned. This effect
packet, each sensor waits a time for an acknowledgment from can be observed in Figure 14, where we show the percentage
the drone; if ACK is received, the sensor waits for the next of retransmissions, which increases with the number of sensors,
transmission period; otherwise, it retransmits the packet. computed as the quotient between the total number of packets
transmitted, i.e., original plus retransmissions, and the total
A collision occurs at the receiver when two or more LoRa
number of original packets. For a coverage of Nr > 100 the
packets overlap at time. We compute the probability of collision
percentage of retransmissions is above 50% and for Nr > 250
pc as the quotient between the number of packets collided
this percentage achieves 80%, which makes inefficient this
and the total number of packets sent. We also compute the
network configuration and suggests the need of adding one or
probability of successful reception, that is computed as the
more gateways to balance the load network.
quotient between the number of successful received packets
and the total number of packets transmitted.
The number of packets successful received is affected both VI. CONCLUSIONS AND FUTURE WORK
by the number of collisions and the unconfirmed traffic (bad We address the problem of data collection from sensors for
formed packets that need to be disregarded, lost packets and precision agriculture by means of drones acting as data mules.
lost ack) that generate a new transmission of the original packet, We consider a scenario in which the sensors are deployed in a
so the probability of successful reception is always lower than regular pattern, and the drone, that acts as a mobile gateway,
(1 − pc ). All simulations were repeated 5 times and, in the follows a pre-determined path to provide wireless coverage

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
12

of the field. This scenario gives rise to a complex trade-off gence and Smart Environments, vol. 11, no. 1, pp. 23–43,
that we solve by introducing an analytic model, according to Jan. 2019.
which it is possible to relate the field parameters (e.g. the [2] S. E. Dìaz, J. C. Pérez, A. C. Mateos, et al., “A novel
field size) with the specs of the communication technology methodology for the monitoring of the agricultural
and of the drone (e.g. radius, velocity, autonomy), in order production process based on wireless sensor networks,”
to determine the parameters of the drone path that achieve a Computers and Electronics in Agriculture, vol. 76, no. 2,
desired target probability of successful data collection. The pp. 252–265, 2011.
output of our model includes the optimal distance between [3] I. Becker-Reshef, C. Justice, M. Sullivan, et al., “Monitor-
nodes, the time in which a sensor is in the range of the drone ing Global Croplands with Coarse Resolution Earth Ob-
for a successful communication, and the velocity that needs servations: The Global Agriculture Monitoring (GLAM)
to keep the drone to complete its flight plan. The analytic Project,” Remote Sensing, vol. 2, no. 6, pp. 1589–1609,
model is general and can be adapted to different sort/long 2010.
range communication technologies and to different drone [4] A. Kocian, D. Massa, S. Cannazzaro, et al., “Dynamic
specifications. In particular, it is built over few parameters that Bayesian network for crop growth prediction in green-
are technology-dependent, namely the maximum drone speed houses,” Computers and Electronics in Agriculture,
and autonomy, the duty cycle of the sensors and the range and vol. 160, p. 105 167, Jan. 2020.
packet loss of the communications. These parameters can then [5] R. Pydipati, T. Burks, and W. Lee, “Identification of cit-
be set and analyzed one for all, as we showed in Section V, rus disease using color texture features and discriminant
for a given technology, to adapt the model. We assess the use analysis,” Computers and Electronics in Agriculture,
of the model in a scenario where the drone employs a LoRa vol. 52, no. 1-2, pp. 49–59, 2006.
communication technology to interact with the sensors in the [6] W. Shen, M. Zhou, F. Yang, et al., “Multi-crop convolu-
field and show how the parameters can be optimized in this tional neural networks for lung nodule malignancy sus-
case. As a final remark, note that our approach is not limited to piciousness classification,” Pattern Recognition, vol. 61,
precision agriculture applications, but may be applied also to pp. 663–673, 2017.
other scenarios in in which sensors are deployed in a grid and [7] M. Lauridsen, H. Nguyen, B. Vejlgaard, et al., “Coverage
the data collection is executed by means of a drone that moves comparison of gprs, nb-iot, lora, and sigfox in a 7800
along a pre-defined path on the field (similar requirements km area,” vol. 2017-June, 2017.
may be found in the monitoring of large polluted areas for [8] LoRa, LoRa, https://www.lora-alliance.org, May 2019.
example). However, if the sensors are deployed with a regular [9] M. Boyle, “The race for drones,” Orbis, vol. 59, Dec.
structure different than a grid, and the drone moves along a 2015.
different path, the general methodology of our work remains [10] A. Ahmadi, L. Nardi, N. Chebrolu, and C. Stachniss,
valid, although the equations would need to be adapted to the “Visual servoing-based navigation for monitoring row-
specific case. crop fields,” in 2020 IEEE International Conference on
As further research we plan to analyze scenario with different Robotics and Automation (ICRA), 2020, pp. 4920–4926.
field shapes, flight plans for the drone, and different deployment [11] B. Zhang, W. Fan, X. Xu, and Y. Liu, “FAPAR and BRDF
of the sensors. We also plan to investigate the feasibility of simulation for row crop using Monte Carlo method,” in
combining the use of the drone with a classical scheme sense- 2014 IEEE Geoscience and Remote Sensing Symposium,
store-forward in the sensors. Other promising future works 2014, pp. 812–815.
concern the study of flight strategies for the drone in the context [12] C. Intanagonwiwat, R. Govindan, and D. Estrin, “Di-
of hybrid data collection architectures that combine drones and rected diffusion: A scalable and robust communication
fixed base stations, and also the study of flight plan strategies paradigm for sensor networks,” in Proceedings of the
that may meet time constraints in the data collection procedure, 6th Annual International Conference on Mobile Comput-
so to enable not only off-line data analysis but also timely ing and Networking, ser. MobiCom ’00, Boston, Mas-
actuation in response to the field conditions. sachusetts, USA: Association for Computing Machinery,
2000, pp. 56–67.
AC K N OW L E D G M E N T [13] S. Basagni, A. Carosi, C. Petrioli, and C. A. Phillips,
This work has been partly funded by the EU’s Horizon
“Coordinated and controlled mobility of multiple sinks
2020 programme under project SHAPES (GA Nº 857159), by
for maximizing the lifetime of wireless sensor networks,”
the Spanish Ministry of Economy and Competitiveness under
Wireless Networks, vol. 17, no. 3, pp. 759–778, 2011.
project PLATINO (TEC2017-86722-C4-4-R) and by the Regional
[14] A. Chakrabarti, A. Sabharwal, and B. Aazhang, “Us-
Government of Castilla-La Mancha under project SymbIoT (SB-
ing Predictable Observer Mobility for Power Efficient
PLY/17/180501/000334).
Design of Sensor Networks,” in Information Process-
ing in Sensor Networks, F. Zhao and L. Guibas, Eds.,
ser. IPSN’03, Berlin, Heidelberg: Springer Berlin Hei-
REFERENCES
delberg, 2003, pp. 129–145.
[1] C. Gomez, S. Chessa, A. Fleury, et al., “Internet of [15] A. A. Somasundara, A. Ramamoorthy, and M. B. Sri-
Things for enabling smart environments: A technology- vastava, “Mobile Element Scheduling for Efficient Data
centric perspective,” JAISE - Journal of Ambient Intelli- Collection in Wireless Sensor Networks with Dynamic

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3075561, IEEE Internet of
Things Journal
13

Deadlines,” in 25th IEEE International Real-Time Sys- [30] J. Petäjäjärvi, K. Mikhaylov, M. Pettissalo, et al., “Per-
tems Symposium, Dec. 2004, pp. 296–305. formance of a low-power wide-area network based on
[16] M. Di Francesco, S. K. Das, and G. Anastasi, “Data LoRa technology: Doppler robustness, scalability, and
Collection in Wireless Sensor Networks with Mobile coverage,” International Journal of Distributed Sensor
Elements: A Survey,” ACM Transactions on Sensor Networks, vol. 13, no. 3, p. 1 550 147 717 699 412, 2017.
Networks, vol. 8, no. 1, pp. 1–31, Aug. 2011. eprint: https://doi.org/10.1177/1550147717699412.
[17] A. W. Khan, A. H. Abdullah, M. H. Anisi, and J. I. [31] M. C. Bor, U. Roedig, T. Voigt, and J. M. Alonso,
Bangash, “A Comprehensive Study of Data Collection “Do LoRa Low-Power Wide-Area Networks Scale?” In
Schemes Using Mobile Sinks in Wireless Sensor Net- Proceedings of the 19th ACM International Conference
works,” Sensors, vol. 14, no. 2, pp. 2510–2548, 2014. on Modeling, Analysis and Simulation of Wireless and
[18] H. Huang, A. V. Savkin, M. Ding, and C. Huang, Mobile Systems, ser. MSWiM ’16, Malta, Malta: ACM,
“Mobile robots in wireless sensor networks: A survey on 2016, pp. 59–67.
tasks,” Computer Networks, vol. 148, pp. 1–19, 2019. [32] F. Adelantado, X. Vilajosana, P. Tuset-Peiro, et al.,
[19] J. Burrell, T. Brooke, and R. Beckwith, “Vineyard Com- “Understanding the Limits of LoRaWAN,” IEEE Com-
puting: Sensor Networks in Agricultural Production,” munications Magazine, vol. 55, no. 9, pp. 34–40, Sep.
IEEE Pervasive Computing, vol. 3, no. 1, pp. 38–45, 2017.
Jan. 2004. [33] M. Capuzzo, D. Magrin, and A. Zanella, “Confirmed
[20] C. Wang, F. Ma, J. Yan, et al., “Efficient Aerial Data traffic in LoRaWAN: Pitfalls and countermeasures,” in
Collection with UAV in Large-Scale Wireless Sensor 17th Annual Mediterranean Ad Hoc Networking Work-
Networks,” International Journal of Distributed Sensor shop (Med-Hoc-Net), 2018, pp. 1–7.
Networks, vol. 11, 2015. [34] H.-C. Lee and K.-H. Ke, “Monitoring of Large-Area IoT
[21] T. Wu, P. Yang, Y. Yan, et al., “ORSCA: Optimal Route Sensors Using a LoRa Wireless Mesh Network System:
Selection and Communication Association for Drones in Design and Evaluation,” IEEE Transactions on Instru-
WSNs,” in Proceedings - 5th International Conference mentation and Measurement, vol. 67, pp. 2177–2187,
on Advanced Cloud and Big Data, CBD, 2017, pp. 420– 2018.
424. [35] S. Farrell, “Low-Power Wide Area Network (LPWAN)
[22] Y. Zeng and R. Zhang, “Energy-Efficient UAV Com- Overview,” RFC Editor, RFC 8376, May 2018, pp. 1–43.
munication with Trajectory Optimization,” IEEE Trans- [36] K. Mekki, E. Bajic, F. Chaxel, and F. Meyer, “A com-
actions on Wireless Communications, vol. 16, no. 6, parative study of LPWAN technologies for large-scale
pp. 3747–3760, 2017. IoT deployment,” ICT Express, vol. 5, no. 1, pp. 1–7,
[23] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and 2019.
communication design for multi-UAV enabled wireless [37] T. T. Network, Duty Cycle, https://www.thethingsne
networks,” IEEE Transactions on Wireless Communica- twork.org/docs/lorawan/duty-cycle.html, Mar. 2021.
tions, vol. 17, no. 3, pp. 2109–2121, 2018. [38] SEMTECH, LoRa Modem Design Guide, https://www.
[24] D. Zorbas and B. O’Flynn, “Collision-Free Sensor Data semtech . com / uploads / documents / LoraDesignGuide _
Collection using LoRaWAN and Drones,” in Global STD.pdf, Jul. 2013.
Information Infrastructure and Networking Symposium [39] A. Caruso, S. Chessa, S. Escolar, et al., “A Dynamic
(GIIS), Oct. 2018, pp. 1–5. Programming Algorithm for High-Level Task Schedul-
[25] C. Luo, L. Wu, W. Chen, et al., “Trajectory Optimization ing in Energy Harvesting IoT,” IEEE Internet of Things
of UAV for Efficient Data Collection from Wireless Journal, vol. 5, no. 3, pp. 2234–2248, Jun. 2018.
Sensor Networks,” in International Conference on Algo- [40] T. M. Cabreira, L. B. Brisolara, and P. R. Ferreira Jr,
rithmic Applications in Management, Springer US, 2019, “Survey on coverage path planning with unmanned aerial
pp. 223–235. vehicles,” Drones, vol. 3, no. 1, p. 4, 2019.
[26] C. Luo, Y. Wang, Y. Hong, et al., “Minimizing data [41] J. I. Vasquez-Gomez, M. Marciano-Melchor, L. Valentin,
collection latency with unmanned aerial vehicle in and J. C. Herrera-Lozada, “Coverage path planning for
wireless sensor networks,” Journal of Combinatorial 2d convex regions,” Journal of Intelligent & Robotic
Optimization, pp. 1–24, 2019. Systems, vol. 97, no. 1, pp. 81–94, 2020.
[27] Q. Pan, X. Wen, Z. Lu, et al., “Dynamic speed control [42] S. Escolar, F. Rincón, X. D. Toro, et al., “The
of unmanned aerial vehicles for data collection under PLATINO Experience: A LoRa-based Network of
internet of things,” Sensors, vol. 18, no. 11, 2018. Energy-Harvesting Devices for Smart Farming,” in De-
[28] S. Sontowski, M. Gupta, S. S. L. Chukkapalli, et al., sign of Circuits and Integrated Systems, (DCIS). Bilbao
“Cyber attacks on smart farming infrastructure,” in 6th (Spain), Nov. 2019, pp. 1–6.
IEEE International Conference on Collaboration and [43] K. Q. Abdelfadeel, D. Zorbas, V. Cionca, and D. Pesch,
Internet Computing (IEEE CIC 2020), 2020. “FREE —fine-grained scheduling for reliable and energy-
[29] M. Gupta, M. Abdelsalam, S. Khorsandroo, and S. Mittal, efficient data collection in lorawan,” IEEE Internet of
“Security and privacy in smart farming: Challenges and Things Journal, vol. 7, no. 1, pp. 669–683, 2020.
opportunities,” IEEE Access, vol. 8, pp. 34 564–34 584,
2020.

2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on May 21,2021 at 07:44:22 UTC from IEEE Xplore. Restrictions apply.