Network Size Estimation for Direct-to-Satellite IoT

Pablo Ilabaca, Samuel Montejo-Sánchez, Juan Fraire, Richard Demo Souza,

Sandra Céspedes

To cite this version:

Pablo Ilabaca, Samuel Montejo-Sánchez, Juan Fraire, Richard Demo Souza, Sandra Céspedes.
Network Size Estimation for Direct-to-Satellite IoT. IEEE Internet of Things Journal, 2022,
�10.1109/JIOT.2022.3224678�. �hal-03935406�

HAL Id: hal-03935406

https://inria.hal.science/hal-03935406v1
Submitted on 11 Jan 2023

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 1

Network Size Estimation for Direct-to-Satellite IoT

Pablo Ilabaca, Samuel Montejo-Sánchez, Senior Member, IEEE, Juan A. Fraire, Senior Member, IEEE,
Richard Demo Souza, Senior Member, IEEE, and Sandra Céspedes, Senior Member, IEEE

Abstract—The worldwide adoption of the Internet of things I. I NTRODUCTION

(IoT) depends on the massive deployment of sensor nodes and
timely data collection. However, installing the required ground The Internet of things (IoT) market is thriving. By 2025,
infrastructure in remote or inaccessible areas can be economically the annual revenue of IoT will reach 11 trillion dollars
unattractive or unfeasible. Cost-effective nanosatellites deployed
in low Earth orbits (LEO) are emerging as an alternative solu-
obtained from more than 30 billion connected devices [1],
tion: on-board IoT gateways provide access to remote IoT devices, [2]. Applications enabled by IoT networks include monitoring
according to direct-to-satellite IoT (DtS-IoT) architectures. One the environment, meteorological or environmental phenomena,
of the main challenges of DtS-IoT is to devise communication animals or plants as well as tracking and positioning ser-
protocols that scale to thousands of highly constrained devices vices. Applications for logistics and location are also available
served by likewise constrained orbiting gateways. In this paper,
we tackle this issue by first estimating the (varying) size of
among many others. However, many applications require the
the device set underneath the (mobile) nanosatellite footprint. IoT devices to be deployed in remote areas with minimal or
Then, we demonstrate applicability of the estimation when no human presence. In cases where connectivity via terrestrial
used to intelligently throttle DtS-IoT access protocols. Since techniques might be unappealing or not possible, satellite tech-
recent works have shown that MAC protocols improve the nology emerges as the appropriate approach. Indeed, satellite
throughput and energy efficiency of a DtS-IoT network when
a network size estimation is available, we present here a novel
IoT has recently been recognized as instrumental to ensure
and computationally-efficient network size estimator in DtS- data collection over remote regions of the planet, enabling a
IoT: our optimistic collision information (OCI) based estimator. truly global IoT solution [3].
We evaluate OCI’s effectiveness with extensive simulations of The ad-hoc nature of space missions driving the deploy-
DtS-IoT scenarios. Results show that when using network size ment of satellite infrastructure typically demands a full re-
estimations, the scalability of a frame slotted Aloha-based DtS-
IoT network is boosted 8-fold, serving up to 4 × 103 devices, engineering of the complete space and ground segments—a
without energy efficiency penalties. We also show the effectiveness factor that has kept most countries out of satellite technology
of the OCI mechanism given realistic detection ratios and development. However, the CubeSat standard provided a com-
demonstrate its low computational cost implementation, making mon mechanical and electrical framework for the development
it a strong candidate for network estimation in DtS-IoT. of nanosatellites [4]. Since its introduction, the cost of these
spacecraft has dropped dramatically, allowing new space actors
Index Terms—Direct-to-Satellite, Internet of Things, Medium
Access Control, Network Size Estimation, Satellite Communica- to access a more “democratic” space [5]. A 1U (1 unit)
tions. CubeSat is characterized by a weight of 1 kilogram and a
cubic shape of 10×10×10 centimeters. CubeSat platforms of
3U, 6U, and 12U are also popular in the industry. These
homogeneous characteristics reduce manufacturing, procure-
Manuscript received July 18, 2022; revised September 26, 2022; accepted
ment, and launching costs but, at the same time, impose
November 11, 2022 limiting conditions such as lower processing, energy, and
P. Ilabaca is with the Department of Computer Science and storage capabilities. Thus, CubeSat-based missions tend to
Software Engineering, Concordia University, Montreal, QC, Canada,
[email protected].
be heavily constrained in terms of resources. The research
S. Montejo-Sánchez is with Programa Institucional de Fomento a la In- presented in the remainder of this work is inspired by CubeSat
vestigación, Desarrollo e Innovación, Universidad Tecnológica Metropolitana, platforms devoted to providing direct-to-satellite IoT (DtS-
Santiago, Chile, [email protected].
J. A. Fraire is with Univ Lyon, Inria, INSA Lyon, CITI, F-69621 Villeur-
IoT) services.
banne, France, and the Argentinian research council (CONICET), Córdoba, In the DtS-IoT architecture, actuators and sensors on the
Argentina. He is also with the Saarland University, Germany. ground communicate directly with the passing-by low Earth
R. D. Souza is with Federal University of Santa Catarina (UFSC), Flo-
rianópolis, Brazil, [email protected]. orbit (LEO) satellite leveraging technologies such as LoRa,
S. Céspedes is with the Department of Computer Science & Software long range-frequency hopping spread spectrum (LR-FHSS),
Engineering, Concordia University, Montreal, QC, Canada. Correspondence: narrowband-IoT (NB-IoT) and the fifth generation (5G) mobile
[email protected]
This work has been supported by the Project ANID Fondecyt Regular No. communication [6]–[8]. Since a ground-based IoT gateway
1201893, the ANID Basal Project FB0008, the ANID FONDECYT Iniciación is not required, this architecture is particularly appealing for
No. 11200659, FONDEQUIP-EQM180180, Project STARS STICAMSUD efficiently servicing remote areas such as oceans, mountains,
21-STIC-12 Folio STIC2020003, and by CNPq Brazil (402378/2021-0,
305021/2021-4). This work started out while P. Ilabaca and S. Céspedes or inaccessible regions, as well as disaster zones [9].
were with the Department of Electrical Engineering and with the NIC Chile The main challenge of DtS-IoT is to ensure good perfor-
Research Labs, Universidad de Chile, Santiago, Chile. mance in the uplink shared channel as the number of nodes in
Copyright © 2022 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be coverage may be significantly higher than in traditional terres-
obtained from the IEEE by sending a request to [email protected]. trial networks. Indeed, link saturation is already achieved with
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 2

a few thousand nodes even when state-of-the-art wide-area TABLE I: List of Acronyms
IoT technologies are implemented [6]. Beyond this scalability Acronym Definition
point, packet collisions and energy inefficiencies undermine 5G Fifth Generation mobile communication
BFS Breadth-First Search
any viable DtS-IoT application. In this context, media ac- CR Capture-Recapture
cess control (MAC) protocols for DtS-IoT are of paramount DtS-IoT Direct-to-Satellite IoT
importance in massive deployments [10]. Furthermore, the eNB evolved Node B
FSA Frame Slotted Aloha
network size under the footprint of the nanosatellite is a priori GEO Geosynchronous Equatorial Orbit
unknown, but an accurate network size estimation can be GPS Global Positioning System
exploited by MAC protocols to optimize the shared uplink IoT Internet Of Things
LEO Low Earth Orbit
channel resources [11], [12]. LR-FHSS Long Range-Frequency Hopping Spread Spectrum
Indeed, estimating a node population’s size is a major LTE Long Term Evolution
factor for achieving scalability and energy efficiency on certain MAC Medium Acces control
MEO Medium Earth Orbit
access control protocols [12], [13]. Additionally, knowing the MTC Machine-Type Communication
number of devices to be served by a CubeSat guarantees NB-IoT NarrowBand-IoT
the quality of service for their applications. Depending on OCI Optimistic Collision Information based estimator
P2P Peer-to-Peer
how overwhelmed a CubeSat will be during operation, it is PDF Probability Distribution Function
possible to justify whether to implement more satellites into RFID Radio Frecuency Identification
the constellation or lower service capacity. Nonetheless, to RMSE Root-Mean-Square Error
ROM Read Only Memory
the best of our knowledge, efficient network size estimation S&F Store and Forward
mechanisms operating in constrained IoT gateways deployed SINR signal-to-interference-plus-noise ratio
in continuously-moving LEO CubeSats have not yet been UE User Equipment
WSN Wireless Sensor Network
addressed in the literature. The problem of an unknown and
massive population of nodes trying to communicate with
nanosatellites in applications that satellite technology supports
is recent. Consequently, the importance in the democratization TABLE II: List of Variables
of low-cost global connectivity and the lack of studies in the Variable Definition
literature motivate the proposal of network size estimation Φ Array of the naive estimates ϕ
ϕ naive estimate ϕ = s + 2c
techniques for this particular scenario. K Maximum number of FSA frames considered
In the past, several network size estimation mechanisms M Number of iterations considered for the estimation phase
have been proposed for terrestrial wireless sensor networks N Number of nodes deployed in the cluster
⃗
N Array of the node populations nk , for k = 1, 2, ..., K
(WSN) [14]. These mechanisms assume static base stations ⃗
P Array of size q formed by the polynomial coefficients
and not necessarily energy-constrained devices, making their T Arbitrary time
direct application difficult in the DtS-IoT context. The work X Vandermonde matrix of dimensions j × q + 1 formed by the
in [15] is aimed at dealing with time-evolving topologies, but elements of Φ
c Number of collided slots in a FSA frame
the method is based on persistent node connectivity, a feature k Frame related to a given CubeSat coverage zone 0 < k ≤ K
not present in DtS-IoT. Conversely, in this work we explore m Iteration number, 0 < m ≤ M
the design of network size estimation mechanisms specifically nk Number of nodes present at a given CubeSat coverage zone
k, 0 < n ≤ N
suited for LEO CubeSats providing DtS-IoT services to thou- p Array of polynomial coefficients in decreasing degree order
sands of IoT nodes. Our preliminary results of the applicability q Degree of the polynomial
of the estimator proposed in [14] to DtS-IoT scenarios are s Number of successful slots in a FSA frame
u Number of idle slots in a FSA frame
presented and analyzed in [16]. w Frame length as number of slots
The contribution of this work is twofold. First, we propose y Estimate correction of the contender nodes by polynomial
the optimistic collision information (OCI)based estimator: a evaluation
y(k,m) Average of m estimates of the contender nodes on a frame k
network size estimation mechanism for DtS-IoT networks.
OCI is a novel scheme with reduced computational complexity,
a high adaptability to different DtS-IoT scenarios in terms of
detection ratios, and an increased numerical stability compared
architecture. The proposed estimation method runs in only
to existing estimators. We demonstrate that OCI achieves high
24.3% of the time required by its most relevant competitor.
scalability and a low estimation error. Second, we assess the
practical value of leveraging the estimations from OCI. To The remainder of this paper is organized as follows. Section
this end, we present extensive simulation results of a DtS-IoT II reviews the related work. Section III provides the description
network governed by a frame slotted Aloha (FSA) protocol, of the scenario of study and the design of the proposed
where device transmissions are throttled by OCI estimations. network size estimator. Section IV analyzes the performance
We show that when OCI is employed, the network achieves a of the OCI mechanism and presents the results of exploiting
throughput close or equal to the theoretical maximum in DtS- size estimations in a DtS-IoT scenario. Section V delves into
IoT regions with more than 4 × 103 devices; furthermore, we practical considerations while Section VII concludes the paper.
provide theoretical and empirical complexity analyses of the Table I offers the definition of the different acronyms and Table
network size estimation mechanisms, using a typical CubeSat II provides a list of variables used in this paper.
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 3

II. BACKGROUND are adaptations of the existing random tour and gossip-based
In this section, we first briefly discuss orbit characteristics aggregation methods to fit controller/agent architectures and
for deploying space technology, and then provide a review of to be compatible with random changes in the topology. The
previous direct and indirect network size estimation mecha- random tour method is based on the transfer of a token,
nisms proposed in the literature. registering each new visited node to make the estimation. The
gossip-based aggregation method is based on communication
between adjacent nodes. The network is divided into clusters,
A. Space Technology and Orbits Characteristics with a controller in each cluster repeating a process in which it
Satellites can be deployed in different orbits with diverse averages a value with the agent nodes, sends the updated value
characteristics that condition the exploitation of the hosted to all agents, and eventually obtains a converged value equal
communication payload. The three main types of satellite to N1 , where N is the number of nodes. The two algorithms
orbits are the following: proposed in [15] require communication among nodes, which
• Geosynchronous Equatorial Orbit (GEO): This orbit has a is most likely not available in the DtS-IoT architecture.
rotation period equal to that of the Earth. A GEO satellite In [20], the authors developed a network size estimation
appears motionless to an observer on the ground. GEO mechanism focused on Peer-to-Peer (P2P) networks. This
satellites have an altitude of 35,786 km, which gives mechanism ”marks” neighboring nodes that are discovered
them a coverage of approximately 30% of the Earth’s when notified by other nodes of the network in a Capture-
surface. This orbit offers services such as weather data Recapture (CR) logic. This approach is based on a Breadth-
and broadcast TV [17]. However, the long distance to First Search (BFS) rooted in the node that wants to discover
the Earth surface results in high latency, on the order the size of the network. The sum of new neighbors discovered
of 120 ms for the best case scenario [10]. Furthermore, at each iteration of the BFS is fitted to a logarithmic function
high transmission power and large antenna gains are whose coefficients are empirically set. Then, this function
required [17]. Also, the process of deploying a GEO is solved using Lambert’s W function [21] to estimate the
satellite is the most expensive among the three orbits. network size. An error of at most 10% is reached, allowing the
• Medium Earth Orbit (MEO): With an altitude ranging BFS to iterate only up to the third level. Later, the algorithm
between 2,000 km and 3,586 km, MEO satellites expe- of [20] is modified to obtain a different logarithmic function,
rience a lower latency than GEO satellites and provide achieving a better performance in terms of accuracy versus
navigation services such as the Global Positioning System the number of operations [22]. The improvement reaches an
(GPS). The launching cost of MEO satellites is midway estimation error of at most 5% while visiting less than 10%
between the cost of the systems located in the other two of the network. Once again, network exploration is based on
orbits. direct communication among peers, making the approaches
• Low Earth Orbit (LEO): Its altitude ranges between 300 non-feasible for DtS-IoT scenarios.
km and 2,000 km, offering the lowest latency, on the In [23], a survey of the CR approach for population es-
order of tens of milliseconds [10]. Due to the smaller timation in computer networks is offered. In general, CR
distance to Earth, LEO satellites require lower transmis- solutions show low precision and the estimation process may
sion power and launching costs. Although smaller and be unsuitable for cases when the network size changes are fast,
cheaper, LEO satellites move at ∼7 km/s, which may which is expected from the DtS-IoT scenario. A more reliable
result in passes of <10 min over a given spot on the and faster estimate can be obtained by storing the identity of
surface. Thus, LEO constellations are needed to provide the elements (sensor nodes) caught in the different samples,
persistent coverage. LEO satellites will play a key role but paying the price of higher storage and computational
in providing ubiquitous and low-cost coverage for IoT requirements to handle the sampling history, which could
devices in the near future [8]. be difficult for a CubeSat in terms of memory. In [24], the
One of the main issues in estimating the size of a network as corner stone of the CR methods, the Lincoln-Petersen Index,
seen from a satellite passing by is that the ground track (the is employed to estimate the size of a Wireless Sensor Network
projection of the satellite trajectory over the Earth surface) (WSN). In [25], this method is used to estimate the number
does not necessarily repeat. In other words, the satellite might of wireless access points in a closed study area, as well as the
observe a given set of nodes in one pass, and a subset of Jackknife method. However, both of these methods work well
those (plus possible others) in subsequent passes. While such on a closed population, i.e. no element is entering or leaving
a condition challenges any network size estimation techniques, the population during the sampling and estimate operations,
orbits that periodically cover the same ground track are often which differs from the dynamic characteristic of DtS-IoT
used in practical space applications (i.e., multiple observations networks. In [23], CR methods for open populations are also
of the same area). These orbits are known as Earth-repeat described, such as the Jolly-Seber model. This model measures
orbits [18], [19] and are described in detail in Section V. several quantities such as the number of individuals caught
in each sample, the number of individuals already marked
or caught for the first time in each sample, the number of
B. Direct Network Size Estimation Approaches marked individuals released after a sample, among others.
In [15], the authors present two algorithms for estimating the These quantities are then used to estimate the number of
size of controller/agent and ad hoc networks. These algorithms individuals in the population up to the last sample, and other
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 4

parameters. However, this model would require a meticulous for the RFID tag population and it actually outperforms the
measurement and control of the number of nodes present in ART method in both the achieved estimation error and the
an area in the pass of a nanosatellite. Characteristics, such as estimation time [31]. This approach bases its operation on
the non-homogeneity of the conditions of each node, the need the iterative adaptation of the FSA frame length to the node
to store the node identities, and the fact that an orbit can take population to be estimated. Once the frame size is adapted, the
several cycles to repeat itself, make this method difficult (if at estimation is calculated by numerically solving a minimum
all possible) to implement in the scenario studied. square error equation. Given the superior performance of
sMMSE over ART and the possibility to adapt it to the DtS-
IoT scenario studied, this scheme is considered one of the
C. Indirect Network Size Estimation Approaches
benchmarks of our proposal.
Other mechanisms approach the estimation problem focus- The other benchmark selected for its performance and
ing on the observed channel load or the number of collisions simplicity is the collision set size estimator proposed by
(and other events) occurring during contention periods [14], Zanella for the FSA protocol in RFID systems [14]. A collision
[26]–[28]. To address the channel load estimation problem in set is the group of nodes contending for a time slot in a
a Long Term Evolution (LTE) network, the authors in [26]– given contention window of FSA. A collision occurs if two
[28] provide a Probability Distribution Function (PDF) of the or more nodes in the set transmit concurrently. The receiver
result of the access requests to ease the management at the can perceive these collisions, keeping track of the number c
evolved Node B (eNB) when the simultaneous activation of of collisions that have occurred, as well as the number of
a large number of Machine-Type Communications (MTC) – successful transmissions s and idle slots i. These three num-
also called User Equipments (UEs) – occurs. bers together compose a time slot observation. Then, Zanella
In [26], the authors derive an explicit expression to compute applies a maximum likelihood estimator to these values, which
the PDF of the number of successful transmissions. Later, the returns the number n of transmitting nodes that maximizes the
authors in [27] compute the PDF of the number of successful conditional probability of having that observation at the end
preamble transmissions using combinatorial analysis. Both of the frame. The number of transmissions in each time slot
contributions show numerical stability issues around 200 de- is considered to be an independent Poisson random variable.
vices, the point at which they stop generating valid results. In Table III presents a summary with a qualitative assessment
[28], the authors calculate the joint PDF of the successful and of the main estimation methods surveyed here and their appli-
collided access requests within a random access opportunity, cability to DtS-IoT scenarios. In addition, Section III provides
increasing the effectiveness up to 300 devices. However, the a detailed description of the selected benchmark schemes and
method is applied to a scenario that differs from the DtS- the adaptations required for their application in the DtS-IoT
IoT scenario since it requires terrestrial infrastructure (i.e., the scenario.
eNB) capable of carrying the computations of PDF and the
estimations. III. N ETWORK S IZE E STIMATION M ECHANISM FOR
In [29], the authors present an estimation technique to D T S-I OT
detect the number of active users in irregular repetition slotted
In this section, we describe the proposed network estimation
ALOHA (IRSA) access protocols based on successive inter-
mechanism employed in a DtS-IoT network, considering the
ference cancellation (SIC) at the receiver’s side. Each device
capabilities and limitations of an IoT gateway deployed on a
transmits multiple replicas of a packet in different slots, and
CubeSat. We first detail the study scenario and assumptions
the interference contribution of already identified devices is
and then describe OCI, a simple yet accurate network estima-
canceled, permitting the identification of new devices. But,
tion mechanism.
this technique requires high processing and memory capacity
at the receiver, which prevents its applicability in CubeSats
[10]. A. System Model and Assumptions
The authors in [30] present the average run-based tag (ART) The network scenario under study follows a DtS IoT archi-
estimation scheme for FSA wireless networks in radio fre- tecture, with a CubeSat nanosatellite acting as the IoT gateway.
quency identification (RFID). Collisions due to simultaneous The CubeSat, whose coverage footprint passes over a cluster
tag responses may occur in the same way as in WSNs. ART of terrestrial IoT nodes, travels around the earth using a LEO
is based on the average run length of 1′ s in the bit string orbit. The orbit in which the satellite is deployed has a height
received at the reader (i.e., receiver), considering 1 as the bit and speed of approximately 500 km and 7.5 km/s, respectively
value of one or multiple transmissions received in a slot and 0 [32]. The satellite completes a turn around the world in 90
as the value of no transmission. These authors also provide a minutes.
set of optimal parameters that save some calculations for wide For network size estimation purposes, the communication
ranges of node populations, but complex and costly numerical between the CubeSat and IoT nodes follows an FSA logic.
solutions are still needed at the end of each frame during the The FSA communication windows are illustrated in Fig. 1;
estimation stage. This makes ART inadequate for a CubeSat each window is divided for downlink and uplink transmissions.
implementation. In the downlink, a beacon serves the satellite to announce the
Another scheme that uses the scalable minimum mean beginning of the frame, which corresponds to an uplink period
square error (sMMSE) is proposed as an estimation method divided into w time slots intended for contention. The beacon
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 5

TABLE III: State of the Art

Reference Strategy Suitability for Dts-IoT
[15] Random tour and gossip-based aggregation algorithms rely on Unsuitable, requires interaction between nodes which is not
the interaction between adjacent nodes of the network. assured in the DtS-IoT architecture .
[20] & [22] BFS approach based on CR logic for P2P networks. Nodes Unsuitable, nodes require P2P communication with other net-
discover the network size by interacting with neighboring nodes work’s nodes.
[23]–[25] Survey of CR approaches for population estimation in closed Unsuitable for closed and open populations due to the rapid
and open computer networks [23], and CR approaches for closed network size changes, the node identity storage requirement, and
wireless networks [24], [25]. meticulous control of all nodes present in a region.
[26]–[28] Estimation mechanism according to two PDFs of the number of Unsuitable. The calculations to compute the PDF and the esti-
successful and collided preambles during contention periods of mation are unfeasible considering CubeSat’s limitation. Further-
LTE networks. Estimation effectiveness for up to 300 devices. more, the effectiveness is desired for bigger node population.
[29] Scheme based on the irregular repeat slotted ALOHA (IRSA) Unsuitable for CubeSats since SIC techniques require high
protocol. User devices transmit multiple replicas of their packets processing and memory capacity at the receiver. Furthermore,
in random slots. Their interference contribution is canceled by the high expected channel congestion impairs the operation of
applying (SIC) techniques at the receiver, allowing the identifi- this technique if a frame of adequate size is not used, which
cation of new transmitter devices. risks being excessively large.
[30] Estimation of the tag population size of RFID networks based Unsuitable for a CubeSat serving multiple regions due to the
on the average run size of successful and collided slots in an high processing cost involved in performing complex numerical
FSA frame. calculations at the beginning and end of each frame.
[31] Dynamically and iteratively adapting the FSA frame length to Not directly applicable. This scheme requires modifications to
the RFID network size according to the channel load. Network bear the cost of adapting the frame for each coverage region.
size is estimated using a minimum mean square error function, Frame lengths risk being excessively large (e.g. 4096 slots for
whose result is obtained through numerical solutions. 1500 nodes with the proposed estimator).
[14] Collision set size estimator based on a Poisson simplification of Suitable. The channel observation is feasible in a DtS-IoT
the maximum likelihood estimator of contender nodes given the scenario, and the complexity of the calculations is viable on a
observation of the channel load in RFID systems. CubeSat-type nanosatellite.

is an invite to the IoT nodes against which the only possible announcing the beginning of new frames and decoding the
response is a transmission attempt, i.e., nodes do not remain received node transmissions at each time slot. This work aims
idle. Consequently, during the estimation process, every node to estimate the number of contender nodes nk , using the
competes in each available frame for which it receives a information acquired during the contention of each frame k.
beacon. We assume a cluster is stable for the duration of a The estimated number of contender nodes is noted yk . We
frame, as depicted in Fig. 1. consider each frame k to be related to a specific region of the
Time synchronization between nodes and the satellite is globe, and thus nk corresponds to the number of nodes present
assumed to properly operate the FSA logic. Although the in that region. For this, we assume the application of Earth-
synchronization aspect is out of the scope of this work, recent repeat orbits, further discussed in Section V. Multiple passes
studies show it is feasible to maintain synchronization in over the same region may be needed to estimate nk . In that
DtS-IoT under certain constellation designs [33]. During the case, we consider each pass an iteration of the process, and the
estimation process, the information sent by the IoT nodes estimate of frame/region k at iteration m is denoted y(k,m) .
may be successfully decoded by the CubeSat or not. The M denotes the total number of iterations for the estimation
failures in the decoding process can be due to a) collisions, phase. This means that after M passes, the CubeSat advances
i.e. simultaneous reception of signals from different nodes, or to the operational phase, using the estimates to throttle the
to b) insufficient received signal power due to path loss and communication with the ground nodes.
other communication channel conditions. When the power of The estimation phase may need to be repeated periodically
the received signal is less than the sensitivity of the receiver, to serve two purposes: 1) to update the network’s estimated
successful decoding is not possible; but note that very weak size at each zone so as to account for dynamics in the network
signals will be undetected by the receiver on the satellite. In population and 2) to adjust communication parameters accord-
such cases, the CubeSat will erroneously assume that time slot ing to the changing requirements of the DtS-IoT network.
to be idle. Moreover, in this work, we call effective detection However, determining an optimal periodicity for the updates
the valid detection of actual transmissions from the IoT nodes, is out of the scope of this work.
while erasure happens when the CubeSat erroneously detects
an idle time slot. The latter resembles an on-off channel model
[34], [35].
Although all nodes under the footprint choose a slot upon B. Network Size Estimator Design
beacon reception, there can be slots that remain idle if no node
chooses that slot for a transmission attempt. Consequently, the Following the mathematical development presented by
satellite perceives three outcomes for a given slot: successful, Zanella in [14], let v = ⟨c, s, u⟩ be one observation of the set
collided, or idle, and can track the number of failures and V = ⟨C, S, U ⟩ which stores the number of collided, success-
successes within a frame. ful, and idle slots in the frame. The number of observations of
We consider the activity of the CubeSat divided into two the set V is numerically equal to the number of passes of the
phases: a) an estimation phase and b) an operational phase. nanosatellite over the same zone considered in the estimation.
During the estimation phase, the CubeSat travels its orbit, Then, the conditional probability of observing v given that n
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 6

probability, P [V = v|µ], of observing v given µ can then be

expressed as
Pµ (v) = µs e−µw (eµ − 1 − µ)c . (2)
Now, to find the value that maximizes that expression, the
derivative of (2) in µ is set to zero, which admits only one
non-negative solution
µw − s µ(eµ − 1)
= µ , (3)
c e −1−µ
which can be determined by the bisection search method [38].
The simplification made to get the expression of equation
(2) allows the computational and energy cost to be adequate
for a low-end device. However, the approximation restrains the
mechanism to Poisson-type traffic. Consequently, our proposed
estimation mechanism shares with [37] the design objective of
using computationally simple calculations but overcoming the
Poisson-type traffic restriction over the network traffic. We
introduce our proposed estimator in the following section.

C. Optimistic Collision Information-Based Estimator (OCI)

Considering that nk nodes are within the CubeSat footprint
at region k, if every node transmits in the frame choosing
a slot with probability w1 , we argue that a simple yet naive
way of estimating nk is by the consideration ϕ = s + 2c,
where ϕ is the estimation of nk , assuming no more than
two nodes are involved in every collision, and that the only
way a transmission fails is due to collisions. This assumption
is sufficiently accurate as long as nk ≤ w; otherwise, the
probability of triple, quadruple, and higher order collisions
becomes larger, and thus ϕ becomes an underestimation.
To assess the effectiveness of the naive estimator, we simu-
lated the communication between the CubeSat and nodes using
MATLAB. In this simulation, nodes are randomly distributed
within the nanosatellite footprint, and a total of K regions
are simulated with varying node population sizes. A 100%
Fig. 1: Characteristics of the DtS communication. effective detection is considered; that is, the outcome of a
transmission attempt can only be a success or a collision. We
employ a step-size of 10 to account for network sizes nk from
nodes transmit can be expressed as 10 nodes to 2000 nodes , reaching a total of K = 200 different
Pn (v) = P [V = v|n] (1) regions with different values of node population. For these
Pc c−j n−s−l simulations, k = 1, 2, ..., K is a variable designating fictitious
c c−j j



n! j=0 j (−1)c−j l=0 l (n−s−l)!
P
regions of increasing size.. The frame length is set to w = 128
=
−1 ,
wn ws w−s

as in [14]. At each step, the contention of nk nodes over the
c
w slots of an FSA frame is simulated, and the results in terms
where w indicates the length of the FSA frame in terms of of c, s, and u are stored. The naive estimates ϕ = s + 2c are
slots. The development of this probability mass distribution is calculated at each step and stored in the array Φ of length K.
further explained in [36]. The orange curve shown in Fig. 2 reflects the elements of Φ
The classical maximum likelihood estimator is the one that when plotted as a function of the number of nodes contending
returns the value of n that maximizes the probability shown for a slot. The curve converges to a value equal to twice the
in (1). However, such a calculation has several problems, number of slots in a frame due to high channel congestion for
such as the numerical instability of the computation of bi- greater network sizes. However, setting a longer frame length,
nomial terms, along with a high computational cost, which with w = 512, shows a better result, as depicted by the blue
makes the maximum likelihood estimator inappropriate for curve in Fig. 2. As long as triple and higher order collisions
implementation in constrained devices and much less in a are not considered, this naive estimation always converges to
nanosatellite. A simplification, suggested in [37], is to consider 2w, limiting its estimation capabilities for large network sizes.
the number of transmissions in each slot as an independent A way to solve this issue, without knowing the order of the
n
Poisson random variable of mean µ = w . The conditional collisions, is to create a function that relates each point of the
 IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 7

of these elements is
2000 Real value
Naive estimation w= 128 ϕ1 ϕ1q−1
 q
··· 1
   
p0 n1
Naive estimation w= 512
Network size estimation

1750 OCI w = 512

 ϕq ϕq−1
 2 2 ··· 1   1   n2 
  p  
..   ..  =  ..  (5)

 . .. ..
 ..
1500 . . .   .   . 
ϕqK ϕq−1
K ··· 1 pq nK
1250 In order to calculate the coefficients in P, the matrix X is
1000 decomposed on a orthogonal-triangular manner, also known
as QR decomposition, such that X = Q · R, where the R
750 factor is an K-by-(q + 1) upper triangular matrix and Q is
an K-by-K orthogonal unitary matrix, such that Q⊤ · Q = I.
500 Then, P is calculated as
250 P = R \ (Q⊤ · N). (6)
0 This process is valid if the polynomial degree q is less than
0 250 500 750 1000 1250 1500 1750 2000 the length of X. Once the coefficients are determined, they can
Actual number of nodes be stored in CubeSat’s ROM memory. Due to the processing
Fig. 2: Network size estimation with a naive approach (or- and energy costs of these calculations, a ground-based station
ange and blue curves) and with OCI estimator (green curve) is responsible for the polynomial regression. Architectures
for which the correction coefficients are [7.024e-09,-1.056e- with down-link and up-link channels have been developed for
05,0.006,-0.036,41.705] nanosatellites such as CubeSats [39]–[43], allowing these low-
cost devices to download and store data from an earth station.
A practical example of the latter is Swarm Technologies,
naive estimation curve to the corresponding real value. For which uses nanosatellite constellations for IoT communication
this to work correctly, bijectivity must be ensured between using a store and forward design [44].
These ground-based stations are also responsible for study-
the naive estimation and the real value of nodes. Using a
ing the conditions of the channel and readjusting the coeffi-
polynomial fit solves this matter as long as the frame length
cients in case of changes. The recalculation of the coefficients
w is long enough in terms of slots so that the naive estimation
is done by integrating the new channel characteristics to the
has no repeated values for the simulated domain, thus ensuring
simulation, followed by the polynomial regression. This work
injectivity. In a practical scenario, this function would be
represents these channel characteristics using the effective
provided to the CubeSat, in such a way that it only uses the
detection percentage.
naive estimate method and then calculates the image using the
The CubeSat’s detailed process at the end of each frame is
given function, keeping the estimation process as simple as
depicted in Algorithm 1. Step 1 is the calculation of the naive
possible. Accordingly, we propose to use a first polynomial fit
estimation ϕ using input parameters s and c. After declaring
of order seven to smooth the naive estimation curve, ensuring
y in step 2, steps 3 to 5 calculate y as the evaluation of ϕ
injectivity. Then, we calculate the coefficients required to
by the polynomial whose coefficients are present in vector
relate each point of the smooth curve to the real number of
P, in decreasing degree order. The calculation corresponds
nodes by using a second polynomial fit; the latter maps the
to Horner’s method of evaluating a polynomial, which has
elements of the naive estimates ϕ to the value of the real
proven to be optimal in terms of arithmetic operations [45]
number of nodes. The order of the second polynomial fit has
as it only requires some additions and multiplications equal
been chosen to be high enough with the purpose that the
to the order of the polynomial. Step 6 returns the estimate
correction oscillates less around the true values (i.e., the red
y(k,m) , which is the equally-weighted average of the estimate
curve in Fig. 2). In the simulation, the order of the second
of m passes if M > 1, or y if M = 1. The design variable M
polynomial fit is set to 4 by inspection; however, the impact
(i.e., the number of satellite passes devoted to the estimation
of this parameter is further discussed in Section V.
stage) that governs the iterations m, is finite and should be
The coefficients of a polynomial of degree q that best fit vastly smaller than the duration of the operation stage of the
the estimation y(x), y(x) = p0 xq + p1 xq−1 + · · · + pq x0 , in network. In step 7, the CubeSat overwrites y(k,m−1) with the
a least-squares sense, can be calculated from new estimate y(k,m) .
The result of applying Algorithm 1 to the naive estimates
X · P̃ = Ñ (4) associated with w = 512 and M = 1 are illustrated by the
green dots in Fig. 2.
where X represents the K ×q+1 Vandermonde matrix formed We call our proposed estimation mechanism the Optimistic
from the elements of the naive estimations array Φ, on a Collision Information (OCI) based estimator. The initial OCI’s
geometric progression. P is a row vector of length q + 1 estimation curve in Fig. 2 has a noisy behavior, which can
containing the polynomial coefficients in descending powers be refined by carrying out the estimation in multiple passes
to be calculated, and N is a K length vector containing the (i.e m > 1 in algorithm 1) and progressively averaging the
real value of nodes nk . Thus, the detailed composition of each estimations of each pass of the CubeSat over the same zone.
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 8

1


Algorithm 1: OCI algorithm where RR represents the response ratio, ψ = 1 − kL , k and
Input: s, c: number of successful and colliding slots L are parameters such that kL represents the FSA window
within FSA frame k on iteration m size, and n is the number of devices to be estimated.
Stored Data:
p[q + 1]: array of size q + 1 of the correction
coefficients in decreasing degree order
M : the number of satellite passes for the estimation IV. P ERFORMANCE E VALUATION
y(k,m−1) : estimated number of contenders for frame k
at the previous iteration (only if M ≥ 2) In this section, we provide an analysis of the performance
Output: y(k,m) : estimated number of contenders for of OCI. We selected the Zanella (henceforth Zan) collision
frame k at iteration m set size [14] and the sMMSE estimators [31], discussed in
1 ϕ = s + 2c;
Sections II and III, for comparison purposes. Furthermore, we
2 y = 0;
evaluate the impact of using network size estimators to provide
3 for a = 1 to (q+1) do
feedback to the MAC protocol in the scenario of study.
4 y = y · ϕ + p[a];
5 end
m−1 1
6 Return y(k,m) = y(k,m−1) · m + y · m ;
7 Replace in memory: y(k,m−1) by y(k,m)
A. Network Size estimator Performance

To measure the performance of the estimators, we calculate

D. Additional Benchmark Scheme the Root-Mean-Square Error (RMSE), commonly used to rate
how a regression model fits a dataset. For this, we compare the
In [31], sMMSE is presented as a two-phase algorithm: a) elements of the array of the real number of nodes N present in
an adaptation phase where the FSA frame size is dynamically the satellite footprint to the ones of the array of the estimated
and iteratively increased to be adapted to the number of size with the equation
devices and b) the estimation phase with a minimum square
error function. The adaptation phase consists of increasing the v
uK
frame length if the percentage of slots where transmission was uX (y(h,M ) − nh )2 )
received is larger than a predefined threshold. The process εRM SE =t (9)
K
is repeated until the response rate RR is lower than the h=1
threshold, meaning that the frame length is adapted to the node
population size to be estimated. An optimal set of parameters RMSE is heavily penalized when an estimate deviates far
is provided in [31] so that the frame is doubled when 40% of from the actual observed value. For this reason, it is used
its slots are occupied. The starting frame size of the algorithm in cases where significant errors are undesirable. Zan and
is 128 slots in [31], resulting, as an example, in a length sMMSE estimations are calculated for benchmarking purposes
of 4096 slots for a 1500 node population, representing a to compare to the performance of OCI. In previous work,
considerable resource efficiency disadvantage in comparison we have adapted and evaluated the Zan estimator to DtS-IoT
to other schemes with a fixed frame size. Another downside scenarios [16]. All simulations are developed using MATLAB
is the number of iterations needed to adapt the frame, which R2020a, with a fresh implementation of OCI and sMMSE, and
implies an additional adaptation cost. To adapt sMMSE to the using Zan’s code as provided by its author [46]. To ensure fair
requirements of our scenario, the frame is started at an initial comparisons, sMMSE’s frame size is set at 512 as its starting
length of 512 slots, which reduces the number of iterations. point.
The explanation for this choice is found in Section III.D.
To evaluate the performance, the first set of simulations
The estimation phase starts right at the end of the adaptation
considers nk nodes distributed randomly within the nanosatel-
phase. This phase consists of calculating the estimated number
lite footprint, ranging from 10 to 2000 with steps of 10,
of nodes n̂ that minimize the distance between the observed
for a total of K = 200 different node populations. At each
tag responses vf and the expected value Vf , according to
size-step, nk nodes contend for a random slot in the FSA
frame, and the estimations of Zan and OCI are calculated
HsM M SE (vf ) = arg min Pn (vf ), (7) accordingly, using the record of successes and collisions, s and
n∈N
c. Effective detection of 100% is considered. Furthermore, for
which, using a binomial distribution of the probability of node each k = 1, 2, ..., K, we consider multiple satellite passes M ,
occupancy and Chebyshev’s inequality [31], is expressed as ranging from 1 to 200. For each pair (k, M ), the estimation
results of OCI and Zan are denoted as y(k,M ) and z(k,M ) ,
respectively. Note that the estimate y(k,M ) is equal to the
Pn (vf ) = ||vf − Vf ||2 (8) average of all estimates y(k,m) for m = 1, 2, ..., M . The
n 2 n 2
matrices Ȳ and Z̄, expressed in (10), are constructed using
= (RR − 1 + ψ ) +(1 − RR − ψ ) , the aforementioned OCI and Zan estimates, respectively.
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 9

Pm y(k,h)
y(k,m) = h=1 m
Pm z(k,h) 102
z(k,m) = h=1 m
 
y(1,1) ··· y(k,1) ··· y(K,1)
 .. .. .. .. 

RMSE
 .
 . . . 
 y(1,m)
Ȳ =  ··· y(k,m) ··· y(K,m)  10'
 . .. .. .. 
 .. . . .  (10)
y(1,M ) ··· y(k,M ) ··· y(K,M ) --+- _...,_
sMMSE, 100%EDR OCI, 95%EDR
····•.,,,
···•·· Zan, 100% EDR -+- sMMSE, 75%EDR
........
--+- OCI, 100% EDR Zan, 75%EDR
-+- OCI, 75%EDR
 
z(1,1) ··· z(k,1) ··· z(K,1) _...,_ sMMSE, 95%EDR
·· •··· Zan, 95%EDR
 .. .. .. .. 
 . . . .  10 ° -'-r----------�-�----------�-------<
  100 10' 102 2xl02

 z(1,m)
Z̄ =  ··· z(k,m) ··· z(K,m)  Satellite passes
 . .. .. ..  Fig. 3: RMSE (log scale) obtained from the simulations of the
 .. . . . 
z(1,M ) ··· z(k,M ) ··· z(K,M ) three estimators under different effective detection rates (EDR)
and up to 200 satellite passes in the estimation phase. Initial
Each row of Ȳ and Z̄ compose a different estimation phase frame length w = 512 slots for Zan, sMMSE (assuming the
of m passes. We calculate the RMSE between rows and the adaptation phase is finished), and OCI.
real values of nodes that were set up in the simulations,
resulting in the 200 points plotted for each mechanism in
Fig. 3. The adaptation phase of sMMSE is not considered for Fig. 3 where the RMSE lowers asymptotically. Overall, the
the count of satellite passes. This means that the first satellite results on RMSE are good for all the estimators since, when
pass is considered to be carried out when the adaptation has confronted with a large number of nodes, the error between
already been made in all regions. Because of this, the different the estimate and the real number is slight.
estimates of each pass result from the numerical solution of These results show the adaptability of the OCI mechanism
(7) with a fixed frame size but a different response rate (RR) over different scenarios, maintaining an RMSE lower than Zan
at the end of the communication window.The simulation is and sMMSE when there is a nonzero percentage of signal
then repeated, considering three different effective detection loss. Space-ground channels are more prone to packet errors
percentages: 100%, 95%, and 75%. and low signal-to-interference-plus-noise ratio (SINR) values
Results in Fig. 3 show that OCI outperforms Zan and than ground IoT links. With this in mind, OCI’s effectiveness
sMMSE in scenarios with more realistic effective detection under imperfect detection ratios makes it a better candidate
rates, lower than 100%. Zan’s RMSE increases when the for DtS-IoT.
An important aspect to compare is the efficiency of channel
effective detection percentage decreases, since the Poisson
resources. In an environment such as the DtS-IoT, it is
approximation of the traffic is no longer valid. The Zan
desirable to minimize the number of wasted channel resources,
mechanism underestimates the number of nodes, given that
as they may negatively impact the network’s performance.
a lower effective detection percentage means that a portion of
Fig. 4 shows a comparison between the frame length of each
nodes does not manifest their presence in the frame. Some-
scheme. Note that the low RMSE in sMMSE’s estimation
thing similar happens with sMMSE as the response rate (RR)
goes along with low channel resource efficiency, as it has to
is decreasing due to reduced detections, and thus sMMSE
increase the frame length to keep an occupancy rate lower than
underestimates the number of nodes. On the contrary, OCI’s
40%. Furthermore, note that we have omitted the adaptation
correction coefficients implicitly have this information, which
phase of sMSSE in Fig. 3; however, when the two phases
explains why the resulting RMSE remains at a low value with
are considered, the FSA frame adaptation phase of sMMSE
lower effective detection rates. Additionally, OCI’s RMSE
requires four satellite passes to reach a 512-frame length as the
decreases as the effective detection ratio is reduced. When
starting point for the estimation. In such a case, the green curve
the collision probability decreases, the probability of multiple
of sMMSE in the figure would shift to the right, rendering a
collisions (triple and further) also decreases, rendering OCI’s
performance similar to Zan’s. In addition, considering the two
approximation more accurate. On average, Zan’s RMSE is 4
phases, sMMSE could only provide an estimate at the 5th
times higher than OCI’s considering a 95% effective detection
pass, whereas OCI and Zan can start the estimation process
scenario, and 38.7 times higher on a 75% effective detection
from the first pass.
scenario. Zan’s RMSE is close to sMMSE’s when imperfect
detection ratios are considered.
When a higher number of satellite passes is considered for B. FSA with Network Size Feedback
the estimation phase, the RMSE decreases progressively, but at In [13], the Slotted Aloha Game is proposed to control ac-
a decreasing rate for the three estimators. Nevertheless, none cess to the medium in a satellite network. The proposal focuses
of the RMSEs reached a point of convergence, as shown in on determining the transmission probabilities of a group of
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 10

and the viability of letting the node population grow without

re-adjusting the polynomial coefficients and frame length.

0.40
0.35
0.30
0.25 FSA Theoretic Bound

Throughput
Pure FSA
0.20 Zan's feedback
sMMSE's feedback
0.15 OCI's feedback
0.10
0.05
0.00
Fig. 4: Frame length comparison between sMMSE, Zan, and 0 1000 2000 3000 4000
OCI versus the number of estimated nodes. Node Count
(a) Throughput comparison with 100% effective detection
independent terminals (nodes) that share the communication 0.6
channel and transmit via the FSA protocol. The probability of FSA Theoretic Bound
Pure FSA
transmission of each node depends on the number of active 0.5 Zan's feedback
nodes, in such a way that if this number is less than or equal to MMSE's feedback
the number of slots in the frame, they are allowed to transmit
0.4 OCI's feedback
with a probability equal to 1. However, if the number of
Efficiency

active nodes exceeds the number of slots, the probability of

transmitting decreases in such a way that, on average, only 0.3
a number of nodes equal to the size of the frame transmit.
The evaluation of the Slotted Aloha Game shows to allow the 0.2
communication of a large number of nodes, maintaining an
approximately constant throughput, with a low access delay 0.1
and a low power consumption [13]. Similarly to [12], the
authors assume that the number of active nodes is perfectly 0.0
known. 0 500 1000 1500 2000 2500 3000 3500 4000
In this section, we use the network size estimates to set Node Count
the threshold employed by an FSA protocol following the
aforementioned Slotted Aloha Game. Two metrics are calcu- (b) Energy efficiency comparison with 75% effective detection
lated to evaluate the performance: throughput and normalized Fig. 5: Comparison of FSA and FSA with size estimation
energy efficiency. Both pure FSA and FSA using the Slotted feedback, considering the average of 10 passes and a 512 slots
Aloha Game approach fed with network size feedback are window size for Zan and OCI
implemented in the simulation for comparison purposes. First,
a scenario with 100% effective detection is simulated, where Analyzing the 100% effective detection scenario, in Figs. 5a
Fig. 5a shows the throughput calculated as the number of and 5b, it can be seen that Pure FSA’s throughput decreases
successful transmissions in a frame divided by the time once the number of nodes contending exceeds the number of
length of the frame (i.e. w = 512), considering one pass slots in the frame and its energy efficiency decreases with
of the satellite for the estimation (M = 1). Then, Fig. 5b the number of nodes. When either of the three estimators’
shows the normalized energy efficiency, which is calculated feedback is applied, these two metrics have the same behavior
considering that when a sensor node transmits, it consumes until the network’s size exceeds the frame length. From that
one arbitrary unit of energy as in [13]; then, the energy point, throughput and energy efficiency oscillate close to 0.37.
consumed by successful transmissions is divided by the total However, with a number of nodes higher than 2000, OCI
energy consumed during the entire process for transmissions. starts to show slightly worse performance than sMMSE and
The results represent the averaged results of 30 repetitions for Zan, while the latter shows discontinuities when the network’s
each estimator. The results show the performance for twice size is close to 4000 nodes. These results allow us to affirm
the number of nodes that OCI was designed for. This is to that the feedback from the three estimators benefit the MAC
appraise the scalability of the designed estimation mechanism, protocol performance even if the number of nodes for which
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 11

its parameters were adjusted is exceeded. Zan and sMMSE

show the best result for this scenario.
0.40
On the one hand, the discontinuities that Zan presents 0.35
(seen by the red dots at abscissa 0), which translate into 0
throughput, and the slightly worse performance of OCI are 0.30
explained by the number of nodes for which the estimators’
0.25

Throughput
parameters were chosen. The algorithm designed by Zanella
in [14] returns an estimation equal to infinity once the number 0.20
of nodes increases past the point where all the slots in a frame
collide. On the other hand, OCI’s polynomial coefficients are 0.15 FSA Theoretic Bound
adjusted for a setting with a maximum of 2000 nodes, for
0.10 Pure FSA
which a frame length of w = 512 slots is suitable. Once the Zan's feedback
number of nodes increases considerably beyond the original
0.05 sMMSE's feedback
maximum, the estimator progressively losses estimation ca- OCI's feedback
pability until it converges to an estimate of 2415 nodes for a 0.00
population size of 4000 nodes. Nevertheless, the throughput is 0 1000 2000 3000 4000
maintained close to the theoretic bound by allowing 21.2% of Node Count
the present nodes to transmit [13], which is the percentage that
(a) Throughput with 75% effective detection
w = 512 represents to a 4000 node population. To improve
the performance for larger network sizes, the frame length 0.6
should be increased, and the coefficients for OCI’s correction FSA Theoretic Bound
recalculated for the new network size. However, depending on 0.5 Pure FSA
the characteristics of the application, it might be unsuitable Zan's feedback
to adopt a number of slots past a certain length. sMMSE’s sMMSE's feedback
estimation needs first an adaptation period which explains why
0.4 OCI's feedback
Efficiency

its behavior is maintained close to the theoretic bound even

for large node population sizes. 0.3
We also carried out a simulation with up to 200 satellite
passes; the results showed that a single satellite pass devoted to 0.2
the estimation process is enough to improve the performance,
and thus a plot of these results is not shown in this manuscript. 0.1
Making use of more passes narrows the resulting points in the
graph but does not bring any substantial improvements.
To account for more realistic scenarios, we simulated with
0.0
0 1000 2000 3000 4000
a 75% effective detection at the satellite. Fig. 6a shows a Node Count
slightly worse performance, in terms of throughput, of FSA
with Zan’s and sMMSE’s feedback compared with using OCI’s (b) Energy efficiency with 75% effective detection
feedback. Due to the 25% of ineffective detection, the frame
is less congested, which translates to higher probabilities of Fig. 6: Comparison of FSA and FSA with size estimation
successful messages. This latter dampens the effects of the feedback, considering the average of 10 passes and windows
underestimation of Zan and sMMSE and explains the differ- of 512 slots for Zan and OCI
ence in performance. When comparing energy efficiency, Fig.
6b shows a bigger difference in favor of OCI’s feedback since
according to the traffic arrival process; (ii) the hardware
the better estimate allows better control on node transmissions,
requirements relative to computational costs and memory
mitigating energy waste. After the node count exceeds 2000,
capacity; and (iii) considerations related to the projection of
the number for which OCI was designed, the performance
the satellite trajectory over the Earth surface.
decreases towards matching Zan’s and sMMSE’s feedback
performance at 4000 nodes. Overall, the results of Figs. 5a
and 6a show that the peak throughput is achieved at a network A. Arrival Process
size equal to the frame length. Nevertheless, when using OCI, One advantage of the OCI estimator is its independence
the throughput is maintained at a high level for large network from the nature of the data arrival process at the satellite.
sizes of up to 4000 nodes. Hence, the supported network size is For the Zan estimator to be applicable in restricted IoT
eight times the number of nodes where the throughput started gateways, it is necessary to use the simplification in which
to decrease for the Pure FSA scheme. the number of transmissions in each slot of the frame is
considered to be an independent Poisson random variable.
V. P RACTICAL C ONSIDERATIONS Such a restriction is not present in OCI. However, the OCI
In this section, we discuss relevant practical aspects involved estimator is limited to be predefined for a precise frame length
in the estimators’ implementation, such as (i) the applicability used during the estimation process. To mitigate this limitation,
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 12

when the number of slots in the contention window is expected 100 OCI
to change, a lookup table can be stored in CubeSat’s ROM, Zan
containing the coefficients needed for the algorithm adjusted
to different window sizes. 80

B. Computational Cost

Time [%]
60
Computational cost is an area where differences in estima-
tors can be observed. In the case of the Zan estimator [14],
it increases with the frame length w, according to log2 w. 40
Instead, OCI’s computational cost is independent of the frame
length but grows with the order of the polynomial function
used for the correction process depending on the polynomial 20
evaluation method. For instance, Horner’s method [45] has
a complexity O(q), where q represents the order. sMMSE’s
0
complexity depends on the numerical method used to solve 0 250 500 750 1000 1250 1500 1750 2000
its equation. To this end, and to validate the complexity Frame length
analysis, OCI and Zan estimators were empirically tested on
a Raspberry Pi Zero W board, which is currently used as the Fig. 7: Normalized computation time required for both esti-
on-board computer (OBC) of CubeSat implementations such mators varying the frame length.
as in [47]. The hardware characteristics of the emulated OBC
are specified in Table V. The test is based on calculating the TABLE V: Main characteristic of the emulated CubeSat on-
average time needed to compute 105 repetitions of estimations board computer
using both methods and considering a contention window Component Characteristic
w = 512 slots. The experiment concluded with both estimators Processor Broadcom BCM2835 ARM1176
being suitable for a CubeSat in terms of computational cost: Architecture ARMv6 (32bit)
Speed 1.0 GHz
OCI shows a time requirement of 8.13 · 10−5 s while Zan RAM 512 MB
shows a requirement of 3.35 · 10−4 s. The reported times OS Raspbian 10 (buster) Linux 5.10.17
demonstrate the advantage of OCI in terms of computations
required over Zan, as OCI needed only 24.3% of the time
demanded by Zan’s method while achieving very similar or remains at the same values except for small random variations.
better performance. However, the sMMSE simulations could Furthermore, the Zan estimator is proven to be more time
not be run on the same hardware platform, so to fairly compare efficient than the best-performing estimators for RFID tags
the three estimators in terms of computation time, the same up to the year of its publication [14], which means that OCI
experiments were run on a computer. Now, the average time also outperforms methods like [48] in terms of computational
used to compute 105 repetitions of estimations was calculated cost. In [49], a look-up table is proposed to store all the
using Python 3.10.7. The results, summarized in Table IV, possible estimations for different combinations of FSA frame
show 1.22 · 10−3 s on average for sMMSE, 5.48 · 10−6 s results (only two among s, c and u due to the constraint
for Zan, and 7.59 · 10−7 s for OCI. For this test, only the of s + c + u = w) of Vogt’s method [48] to improve time
estimation phase was considered for sMMSE, using k = 128 delays. For OCI, a look-up table could also be implemented.
as done in [31] and different values for RR and L calculated This would imply negligible delays but there is a need for
by independent simulations of the first phase. These results storing the polynomial evaluations of all possible frames,
confirm OCI as the fastest scheme since OCI needed only which results with a fixed w.
13.9% of the time required by Zan’s method and 0.06% of We also compared the memory needed for both estimators.
the time required by the sMMSE method. Since the satellite travels at approximately 7.5 km/s in its
TABLE IV: Computational Costs orbit, it goes around the world in 90 minutes, i.e., 5400
seconds. It is common in the literature that the length of
Scheme Avg. Compute Time Ratio a slot in a communication protocol such as FSA lasts an
sMMSE 1.22 · 10−3 s 100 %
Zan 5.48 · 10−6 s 0.45% amount of time equal to the time of packet transmission. In the
OCI 7.59 · 10−7 s 0.06% existing literature, applications with packet lengths of around
1000 bits can be found [50]; the packet transmission times
To further compare Zan and OCI in terms of time require- vary depending on the IoT application. If the transmission of
ments, we carried out a simulation varying the frame length a packet takes for example 5 ms [51], then a frame takes
from 128 to 2048 slots with steps of 128. The results are 512 · 5 = 2560 ms, that is 2.56 seconds. Then, with that frame
shown in Fig. 7 where time is normalized to the maximum length, there could be 2109 contention windows, i.e. positions,
value reached by the Zan estimator when a frame of 2048 in a turn around the world. In the first satellite pass, OCI
slots is used. As can be seen, Zan’s computation cost increases needs to store two integers of 3 digits maximum (the number
along with the frame length, whereas OCI’s time requirement of successes and collisions among the frame slots) and the
 IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 13

175 maintenance techniques [53], it unlocks the possibility of

learning from reiterative passes over the same region of the
150 network. The longitudinal rotation after a certain period of
time of a planet is
125 T
∆L1 = −2π , (11)
TE
100
RMSE

while the quantization of the effect of the nodal precession

44%
75 can be expressed as

64% 3πJ2 Re2 cos(i)

50 77% 98% 95% 98%
∆L2 = −
a2 (1 − e2 )2
, (12)

25 where TE a sidereal day, J2 is the Earth’s second dynamic

form factor, Re is the Earth’s radius, i is the orbital inclination,
a is the orbit’s semi-major axis, e is the orbital eccentricity,
0
1 2 3 4 5 6 7 and T represents the elapsed time, which can be expressed
Order of the polynomial fit for the correction as a function of the q standard gravitational parameter for the
3
Fig. 8: RMSE of OCI as a function of the order of the Earth µ, like T = 2π aµ .
polynomial fit used for the correction. Therefore, a repeat orbit sees these two effects canceling
each other after a set of orbital revolutions j and sidereal
days k, such that j|∆L1 + ∆L2 | = k2π. Then, the required
estimation made with these numbers, for each position. The condition for the orbit to repeat can be written as
simplest way of storing a number requires the use of 10 bits q
3
for a 3-digit number; thus 2 · 10 · 2109 = 42.180 bits are 2π aµ 3πJ2 R2 cos(i)
needed in the first pass, plus the estimate. According to the j −2π − 2 e 2 2 = k2π. (13)
TE a (1 − e )
experiments, each position’s size estimate is an integer of up to
4 digits, which means 14 · 2109 = 29, 526 bits are also needed Replacing the variables with the parameters of the Earth and
in the first pass. OCI does not require storing the results of an using similar parameters to those used in a known Chilean
FSA frame in long-term memory, but only the estimates and CubeSat named Suchai [54], [55], it is possible to identify
the progressive averaging of these results, which is a shared different repeat orbits, as shown in Fig. 9. In this figure, dots
feature with Zan. of different colors represent a set of orbital inclination and
For OCI to work properly, the set of coefficients required orbital height parameters that holds the equation (13). Darker
for the correction is to be stored in memory. These coefficients colors represent shorter repetition periods, and bigger markers
are numbers with up to 15 digits that may require up to 64 bits indicate higher accuracy in the resolution of the equation
each, which then implies that 64 · C bits are required, with C (13). It should be noted that the parameters of the Suchai
being the polynomial order. Fig. 8 shows the impact that the nanosatellite are misaligned since its orbit is not repetitive.
order of the polynomial correction has on the RMSE of OCI
under the same experiment carried out in Section III. In Fig.
7
8 each label represents the error percentage of each point with 800
respect to the previous order. It can be seen that a coefficient
750 6
of order 4 is sufficient to have a low RMSE compared to lower
Repetition period [sideral days]
orders. If a 4 order is chosen, then 256 bits would be needed. 700
5
Orbital Height

650
C. Repeating Orbits
As mentioned in Section II, one important issue in estimat-
600 4
ing the size of the network beneath a passing by satellite is that 550 Suchai
the ground track does not necessarily repeat. In the literature, 3
orbits that repeat the exact same ground track are known as
500
Earth-repeat orbits [18], [19]. Such Earth-repeat orbits exploit 450 2
the nodal precession effect (a phenomenon provoked by the
400
non-spherical nature of the Earth) to shift the orbital plane such 1
that the ground track matches a previous one [52]. In other 70 75 80 85 90 95 100
words, the nodal precession is used to balance out the offset Orbital Inclination
of the revolution of Earth. Thus, the same ground track can be
covered after a controlled number of orbital revolutions in a Fig. 9: Parameters of orbits that hold equation (13) for
given time interval [18]. Although the efficient determination circular orbits with varying height and inclination. The current
of Earth-repeat orbits demands specific computation and orbit parameters for the Chilean Suchai satellite are highlighted.
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 14

Up to this point, the estimation process has been considered [5] M. A. Diaz, J. C. Zagal, C. Falcon, M. Stepanova, J. A. Valdivia,
to be held by a single CubeSat. However, several studies on M. Martinez-Ledesma, J. Diaz-Peña, F. R. Jaramillo, N. Romanova,
E. Pacheco, M. Milla, M. Orchard, J. Silva, and F. P. Mena, “New
store and forward technology and its application to CubeSats opportunities offered by cubesats for space research in latin america: The
have been done over the years [39], [40], [42]. With store suchai project case,” Advances in Space Research, vol. 58, pp. 2134–
and forward, a CubeSat can store its estimate on a ground 2147, 11 2016.
[6] G. Boquet, P. Tuset-Peiró, F. Adelantado, T. Watteyne, and X. Vila-
terminal which can later be sent to another CubeSat of the josana, “LR-FHSS: Overview and Performance Analysis,” IEEE Com-
same constellation. In this way, the estimation process done munications Magazine, vol. 59, no. 3, pp. 30–36, 2021.
over multiple passes of a single CubeSat can be carried out [7] S. Cluzel, L. Franck, J. Radzik, S. Cazalens, M. Dervin, C. Baudoin,
and D. Dragomirescu, “3GPP NB-IOT coverage extension using LEO
by a single pass of multiple CubeSats. satellites,” IEEE Vehicular Technology Conference, vol. 2018-June, pp.
1–5, 2018.
VI. C ONCLUSIONS [8] I. Leyva-Mayorga, B. Soret, M. Röper, D. Wübben, B. Matthiesen,
A. Dekorsy, and P. Popovski, “LEO Small-Satellite Constellations for 5G
The massive and worldwide deployments of IoT networks and Beyond-5G Communications,” IEEE Access, vol. 8, pp. 184 955–
184 964, 2020.
face connectivity challenges related to the scalability, cover- [9] J. A. Fraire, S. Céspedes, and N. Accettura, “Direct-to-satellite iot - a
age, resilience, and ubiquity of these implementations. The survey of the state of the art and future research perspectives,” in Ad-
Direct-to-Satellite IoT paradigm allows cost-effective solutions Hoc, Mobile, and Wireless Networks, M. R. Palattella, S. Scanzio, and
S. Coleri Ergen, Eds. Cham: Springer International Publishing, 2019,
such as CubeSat nanosatellites in Low-Earth Orbit that serve pp. 241–258.
as orbiting gateways to remote IoT nodes. However, achieving [10] T. Ferrer, S. Céspedes, and A. Becerra, “Review and evaluation of mac
efficient use of temporal and spectral resources to overcome protocols for satellite IOT systems using nanosatellites,” Sensors, vol. 19,
the exponential connectivity demands remains one of the no. 8, pp. 1–29, 2019.
[11] C. Buratti, A. Conti, D. Dardari, and R. Verdone, “An overview on
main challenges of DtS-IoT. This efficient allocation of trans- wireless sensor networks technology and evolution,” Sensors, vol. 9,
mission resources is strongly favored when MAC protocols no. 9, pp. 6869–6896, 2009.
are aware of the number of contending nodes, but to the [12] K. Vogelgesang, J. A. Fraire, and H. Hermanns, “Uplink transmission
probability functions for LoRa-based direct-to-satellite IoT: A case
best of our knowledge, network size estimation mechanisms study,” in 2021 IEEE Global Communications Conference (GLOBE-
operating in IoT gateways deployed in continuously-moving COM). IEEE, 2021, pp. 01–06.
LEO CubeSats have not yet been addressed in the literature. [13] B. Zhao, G. Ren, and H. Zhang, “Slotted aloha game for medium
access control in satellite networks,” in 2019 IEEE/CIC International
To tackle this issue, in this work we proposed OCI, a novel and Conference on Communications in China (ICCC), 2019, pp. 518–522.
computationally efficient network size estimator for DtS-IoT. [14] A. Zanella, “Estimating collision set size in Framed Slotted Aloha
The results of the experiments carried out showed that OCI wireless networks and RFID systems,” IEEE Communications Letters,
vol. 16, no. 3, pp. 300–303, 2012.
has low estimation error and computational costs, and that the [15] R. Ali, S. S. Lor, and M. Rio, “Two algorithms for network size estima-
estimate can be used to assist a MAC protocol, achieving a tion for master/slave ad hoc networks,” 2009 IEEE 3rd International
stable throughput serving up to 4000 nodes without energy Symposium on Advanced Networks and Telecommunication Systems,
2009, pp. 23–25, 2009.
efficiency penalties. Channel impairments, such as erasures, [16] ANTS P. Ilabaca, S. Céspedes, and S. Montejo-Sánchez, “Network size estima-
can be integrated into OCI’s correction coefficients to adjust tion in direct-to-satellite iot,” in 2021 IEEE 8th International Conference
the estimates to the specific scenario for which it will be on Space Mission Challenges for Information Technology (SMC-IT),
2021, pp. 107–111.
implemented. This provides OCI with adaptability in low de-
[17] SES, “GEO, MEO, and LEO. How orbital altitude impacts network
tection ratio scenarios absent in the state-of-the-art estimation performance in satellite data services,” https://www.satellitetoday.com/
mechanisms used for benchmarking, further positioning OCI content-collection/ses-hub-geo-meo-and-leo/, 2020, accessed: 2021-04-
as a better candidate for DtS-IoT. 04.
[18] O. Montenbruck, E. Gill, and F. Lutze, “Satellite orbits: models, meth-
Regarding future work, we intend to design new approaches ods, and applications,” Appl. Mech. Rev., vol. 55, no. 2, pp. B27–B28,
that favor an equitable distribution of channel resources in 2002.
terms of medium access. Given the nature of a DtS-IoT [19] M. Lara, “Searching for repeating ground track orbits: a systematic
approach,” The Journal of the Astronautical Sciences, vol. 47, no. 3,
network with Frame Slotted Aloha, the work in progress pp. 177–188, 1999.
aims to provide terrestrial nodes with the capability to decide [20] J. Bustos-Jimenez, N. Bersano, S. E. Schaeffer, J. M. Piquer, A. Iosup,
in which frame to transmit, given the estimated potential and A. Ciuffoletti, Estimating The Size Of Peer-To-Peer Networks Using
Lambert’s W Function. Boston, MA: Springer US, 2008, pp. 61–72.
competition in each zone obtained with OCI estimations. [21] R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and
D. E. Knuth, “On the lambertw function,” Advances in Computational
Mathematics, vol. 5, no. 1, pp. 329–359, Dec 1996.
R EFERENCES
[22] J. Bustos-Jiménez, M. Hidalgo, and E. Schaeffer, “Estimating the size
[1] J. Greig, “IoT industry projected to have economic impact of $11 of natural networks using local information,” Proceedings - Interna-
trillion by 2025 - TechRepublic,” https://www.techrepublic.com/article/ tional Conference of the Chilean Computer Science Society, SCCC, no.
iot-industry-projected-to-have-economic-impact-of-11-trillion-by-2025/, November, pp. 307–313, 2010.
2019, accessed: 2021-03-19. [23] N. Accettura, G. Neglia, and L. A. Grieco, “The Capture-Recapture
[2] L. S. Vailshery, “Global IoT and non-IoT connections 2010- approach for population estimation in computer networks,” Computer
2025 Statista,” https://www.statista.com/statistics/1101442/ Networks, vol. 89, pp. 107–122, 2015.
iot-number-of-connected-devices-worldwide/, 2021, accessed: 2021- [24] S. L. Peng, S. S. Li, X. K. Liao, Y. X. Peng, and N. Xiao, “Estimation
04-19. of a Population Size in Large-Scale Wireless Sensor Networks,” Journal
[3] M. Centenaro, C. E. Costa, F. Granelli, C. Sacchi, and L. Vangelista, of Computer Science and Technology, vol. 24, no. 5, pp. 987–997, 2009.
“A Survey on Technologies, Standards and Open Challenges in Satellite [25] A. Achtzehn, L. Simić, M. Petrova, and P. Mähönen, “IEEE 802.11 Wi-
IoT,” IEEE Communications Surveys & Tutorials, 2021. Fi access point density estimation with capture-recapture models,” 2015
[4] I. F. Akyildiz and A. Kak, “The internet of space things/cubesats,” IEEE International Conference on Computing, Networking and Communica-
Network, vol. 33, no. 5, pp. 212–218, 2019. tions, ICNC 2015, pp. 153–159, 2015.
IEEE INTERNET OF THINGS JOURNAL, VOL. 1, NO. 1, NOVEMBER 2022 15

[26] C. H. Wei, G. Bianchi, and R. G. Cheng, “Modeling and analysis [49] A. T. H. Bui, C. T. Nguyen, T. C. Thang, and A. T. Pham, “A
of random access channels with bursty arrivals in OFDMA wireless Comprehensive Distributed Queue-Based Random Access Framework
networks,” IEEE Transactions on Wireless Communications, vol. 14, for mMTC in LTE/LTE-A Networks with Mixed-Type Traffic,” IEEE
no. 4, pp. 1940–1953, 2015. Transactions on Vehicular Technology, vol. 68, no. 12, pp. 12 107–
[27] S. Duan, V. Shah-Mansouri, Z. Wang, and V. W. Wong, “D-ACB: 12 120, 2019.
Adaptive Congestion Control Algorithm for Bursty M2M Traffic in LTE [50] J. Su, G. Ren, Q. Wang, and H. Zhang, “Randomly Pre-coded Packets
Networks,” IEEE Transactions on Vehicular Technology, vol. 65, no. 12, based Random Access Scheme for IoT-Oriented Satellite Networks,”
pp. 9847–9861, 2016. IEEE Access, vol. 8, 2020.
[28] L. Tello-Oquendo, V. Pla, I. Leyva-Mayorga, J. Martinez-Bauset, [51] D. Li, S. Wu, Y. Wang, J. Jiao, and Q. Zhang, “Age-Optimal HARQ
V. Casares-Giner, and L. Guijarro, “Efficient random access channel Design for Freshness-Critical Satellite-IoT Systems,” IEEE Internet of
evaluation and load estimation in LTE-A with massive MTC,” IEEE Things Journal, vol. 7, no. 3, pp. 2066–2076, 2020.
Transactions on Vehicular Technology, vol. 68, no. 2, pp. 1998–2002, [52] S. D. Vtipil and B. Newman, “Determining an earth observation re-
2019. peat ground track orbit for an optimization methodology,” Journal of
[29] J. Sun, R. Liu, and E. Paolini, “Detecting the Number of Active Users Spacecraft and Rockets, vol. 49, no. 1, pp. 157–164, 2012.
in IRSA Access Protocols,” in 2018 IEEE 29th Annual International [53] Y. He, M. Xu, X. Jia, and R. Armellin, “High-precision repeat-
Symposium on Personal, Indoor and Mobile Radio Communications groundtrack orbit design and maintenance for earth observation mis-
(PIMRC), 2018, pp. 1972–1976. sions,” Celestial Mechanics and Dynamical Astronomy, vol. 128, no. 2,
[30] M. Shahzad and A. X. Liu, “Fast and Accurate Estimation of RFID pp. 275–294, 2017.
Tags,” IEEE/ACM Transactions on Networking, vol. 23, no. 1, pp. 241– [54] N2YO, “SUCHAI Satellite details 2017-036Z NORAD 42788,” https:
254, 2015. //www.n2yo.com/satellite/?s=42788, 2017, accessed: 2021-06-1.
[31] L. Arjona, H. Landaluce, A. Perallos, and E. Onieva, “Scalable RFID [55] H.-A. G. Chris Peat, “SUCHAI - Orbit,” https://www.heavens-above.
Tag Estimator With Enhanced Accuracy and Low Estimation Time,” com/orbit.aspx?satid=42788, 2017, accessed: 2021-06-1.
IEEE Signal Processing Letters, vol. 24, no. 7, pp. 982–986, 2017.
[32] S. Cakaj, B. Kamo, A. Lala, and A. Rakipi, “The coverage analysis
for low earth orbiting satellites at low elevation,” International Journal
of Advanced Computer Science and Applications, vol. 5, no. 6, 2014.
[Online]. Available: http://dx.doi.org/10.14569/IJACSA.2014.050602
[33] J. A. Fraire, S. Henn, F. Dovis, R. Garello, and G. Taricco, “Sparse
Satellite Constellation Design for LoRa-based Direct-to-Satellite Internet
of Things,” in GLOBECOM 2020 - 2020 IEEE Global Communications
Conference, 2020, pp. 1–6.
[34] F. Tondo, S. Montejo Sánchez, M. Pellenz, S. Céspedes, and R. Souza,
“Direct-to-Satellite IoT Slotted Aloha Systems with Multiple Satellites
and Unequal Erasure Probabilities,” Sensors, vol. 21, p. 7099, 10 2021.
[35] A. Munari, F. Clazzer, G. Liva, and M. Heindlmaier, “Multiple-relay
slotted aloha: Performance analysis and bounds,” IEEE Transactions on
Communications, vol. 69, no. 3, pp. 1578–1594, 2021.
[36] J. Riordan, An Introduction to Combinatorial Analysis. Princeton
University Press, 1978.
[37] M. Kodialam and T. Nandagopal, “Fast and reliable estimation schemes
in RFID systems,” Proceedings of the Annual International Conference
on Mobile Computing and Networking, MOBICOM, vol. 2006, pp. 322–
333, 2006.
[38] H. A. Eiselt and C.-L. Sandblom, Linear programming and its applica-
tions. Springer Science & Business Media, 2007.
[39] S. N. M. Rahim, J. Johari, S. A. Ence Ab Rahim, and M. H. Jusoh, “Esti-
mation of communication link on Ground Sensor Terminal (GST) system
for nanosatellite (UiTMSAT-1) store-and-forward mission,” ICSET 2018
- 2018 IEEE 8th International Conference on System Engineering and
Technology, Proceedings, no. October, pp. 108–111, 2019.
[40] D. Abbasi and M. Abolghasemi, “Store & forward communication
payload design for leo satellite systems,” Majlesi Journal of Electrical
Engineering, vol. 10, pp. 7–18, 09 2016.
[41] A. Addaim, A. Kherras, and E. Zantou, “Design and analysis of store-
and-forward data collection network using low-cost leo small satellite
and intelligent terminals,” Journal of Aerospace Computing Information
and Communication - J AEROSP COMPUT INF COMMUN, vol. 5, pp.
1–11, 02 2008.
[42] A. Addaim, A. Kherras, and B. Zantou, “Design of store and forward
data collection low-cost nanosatellite,” in 2007 IEEE Aerospace Con-
ference, 2007, pp. 1–10.
[43] Q. Verspieren, T. Obata, and S. Nakasuka, “Innovative Approach to Data
Gathering in Remote Areas Using Constellations of Store & Forward
Communication Cubesats,” in 31st International Symposium on Space
Technology and Science (ISTS), 2017, pp. 1–6.
[44] “Swarm - Low cost, global satellite connectivity for IoT,” accessed:
2022-07-9. [Online]. Available: https://swarm.space/
[45] V. Y. Pan, “METHODS OF COMPUTING VALUES OF POLYNOMI-
ALS,” Russian Mathematical Surveys, vol. 21, no. 1, pp. 105–136, feb
1966.
[46] dgt Dei Gruppo Telecomunicazioni, “MATLAB source code,” http://dgt.
dei.unipd.it/pages/read/54/, accessed: 2020-10-01.
[47] ECE3SAT, “OBC: On Board Computer - ECE3SAT,” http://www.
ece3sat.com/cubesatmodules/obc/, 2018, accessed: 2021-07-10.
[48] H. Vogt, “Efficient Object Identification with Passive RFID Tags,” in
Pervasive Computing, F. Mattern and M. Naghshineh, Eds. Berlin,
Heidelberg: Springer Berlin Heidelberg, 2002, pp. 98–113.