IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 1

Hybrid Low-Power Wide-Area Mesh Network for

IoT Applications
Xiaofan Jiang, Member, IEEE, Heng Zhang, Member, IEEE, Edgardo Alberto Barsallo Yi, Member, IEEE,
Nithin Raghunathan, Member, IEEE, Charilaos Mousoulis, Member, IEEE, Somali Chaterji, Member, IEEE,
Dimitrios Peroulis, Fellow, IEEE, Ali Shakouri, and Saurabh Bagchi, Fellow, IEEE,

Abstract—The recent advancement of the Internet of Things and smart homes. In the digital agriculture context, these
(IoT) enables the possibility of data collection from diverse devices may monitor various environmental conditions, such
arXiv:2006.12570v1 [cs.NI] 22 Jun 2020

environments using IoT devices. However, despite the rapid as soil moisture, nutrient quality, or microbial activity [1].
advancement of low-power communication technologies, the
deployment of IoT networks still faces many challenges. In The sensor data is collected and routed through the WSN
this paper, we propose a hybrid, low-power, wide-area network and sent to the cloud for further analysis and possible closed-
(LPWAN) structure that can achieve wide-area communication loop control. Our work addresses the development of a large-
coverage and low power consumption on IoT devices by utilizing scale WSN that is suitable for distinct application areas,
both sub-GHz long-range radio and 2.4 GHz short-range radio. namely: digital agriculture and smart and connected cities.
Specifically, we constructed a low-power mesh network with
LoRa, a physical-layer standard that can provide long-range Our study is based on real-world deployments and highlights
(kilometers) point-to-point communication using custom time- the practical design challenges and insights that arise from
division multiple access (TDMA). Furthermore, we extended the long-term unattended operation of those IoT systems. They
capabilities of the mesh network by enabling ANT, an ultra highlight a distinct set of challenges in terms of the wireless
low-power, short-range communication protocol to satisfy data networking capabilities.
collection in dense device deployments. Third, we demonstrate
the performance of the hybrid network with two real-world de- Digital agriculture refers to using modern technologies to
ployments at the Purdue University campus and at the university- increase the quantity and quality of agricultural products.
owned farm. The results suggest that both networks have superior Understanding the environmental conditions (e.g., tempera-
advantages in terms of cost, coverage, and power consumption ture, humidity, and soil fertility) is important for agricultural
vis-à-vis other IoT solutions, like LoRaWAN. management. Traditionally, it has been challenging to consis-
Index Terms—LoRa, WSN, TDMA, IoT, Mesh Network, Com- tently monitor large farmlands due to the lack of automation
munications, Smart City, Digital Agriculture. and inefficient labor. Recent agricultural industries have been
adopting automation technologies that can monitor the envi-
I. I NTRODUCTION ronment and optimize farming [1]–[4]. Some analytics can
As Internet of Things (IoT) devices, representing a network also be run on the back-end cloud computing resource and
of interconnected things, proliferate with an estimated popu- derive actionable knowledge from the raw data, such as, how
lation of 125 billion IoT devices in the next decade, systems to fertilize specific parts of the farm in a localized manner.
built out of these devices will play a role in many deployments, Such kinds of decisions do not require real-time (or, near real-
including those that need extended periods of unattended time) decision making.
operation. These things are essentially sensors and actuators, In addition to digital agriculture, smart city is another
fitted with a wireless network interface, and computing and important application in IoT. It refers to using different types
storage units. IoT provides the connectivity to physically of sensors in urban areas to collect data and then use the data
distributed devices, home appliances, and even devices in to manage urban assets and resources efficiently. According
more critical sectors, such as healthcare, public utilities (e.g., to the United Nations Department of Economic and Social
electric grids), environmental monitoring, and transportation. Affairs, it is estimated that by 2050 there will be 2.5 billion
These IoT devices sense, compute, and communicate, often people living in urban areas. WSNs are becoming part of a
in resource-limited deployments, forming a Wireless Sensor smart city infrastructure that can combat the strain of city
Network (WSN). In the smart city context, the IoT devices growth by providing access to real-time data and can be
may monitor energy and utility distribution (e.g., smart grid); realized through a robust network.
enable intelligent transportation systems, building automation, However, there are several technical challenges to be solved
for the successful deployment of such a WSN, such as
X. Jiang, H. Zhang, N. Raghunathan, D. Peroulis, A. Shakouri and S. harsh environmental conditions, communication range, quality
Bagchi are with the School of Electrical and Computer Engineering, Purdue of service (QoS), deployment cost and energy consumption
University, West Lafayette, IN, 47906 USA (e-mail: [email protected].)
E. Yi is with the Department of Computer Science, Purdue University, West (battery life). These technical challenges are further discussed
Lafayette, IN, 47906 USA. in our technical report [5]. To handle the above challenges,
S. Chaterji is with the School of Agricultural and Biological Engineering, we designed and fabricated a new custom wireless IoT hard-
Purdue University, West Lafayette, IN, 47906 USA.
Manuscript received April XX, XXXX; revised August XX, XXXX. ware platform that is connected to temperature, humidity,
(Xiaofan Jiang and Heng Zhang are co-first authors.) and nitrate soil sensors and additional I/O pins for enabling
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 2

(Throckmorton-Purdue Agricultural Center–TPAC) and urban

(West Lafayette campus at Purdue university) areas.
The rest of this paper is organized as follows. In Section II,
we provide a summary of popular WPANs and LPWANs
technologies and discuss related works of LoRaWAN. In
Section III, we present our solution with the network structure
as well as the new hardware platform. The evaluation results
are presented in Section IV and the paper is concluded with
the discussions of limitations and future work in Section VI.

II. BACKGROUND AND R ELATED W ORKS

A. Summary of Current Technologies in WSN
Low power communication technologies for wireless IoT
communication can grossly fall within two categories (Fig.
2):
• Wireless Personal Area Networks (WPANs): typically
Fig. 1. Hybrid Network Architecture. Long Range Mesh Network (LRMN)
consists of several Short Range Star Network (SRSN) and a few LoRa communicate from 10 meters to a few hundred meters.
nodes. Each node is capable of communicating in a WPAN protocol (ANT This category includes Bluetooth, Bluetooth Low En-
in this case) or an LPWAN protocol (LoRa in this case). SRSN is used to ergy (BLE), ANT, ZigBee, etc., which are applicable di-
satisfy dense sensing deployment while individual LoRa nodes are for sparse
deployment. rectly in short-range personal area networks or if designed
in a mesh topology and with higher transmit power, larger
area coverage is possible.
the connections to other sensors. We choose LoRa [6] as • Low-Power Wide Area Networks (LPWANs): have a
the communication protocol because of its long range and communication range greater than one kilometer. Each
low power consumption (Section II-B). LoRa utilizes chirp gateway could communicate with thousands of end-
spread spectrum (CSS) modulation and operates in the sub- devices. This category includes LoRaWAN, Sigfox, NB-
GHz ISM band to void penetration capability and heavy in- IoT, etc. A summary of these technologies are sum-
band interference. Furthermore, we proposed a lightweight, marised in Table I [7]–[9].
hybrid network combining the advantages of LoRa’s wide
area coverage and ANT’s ultra-low power consumption by
integrating them into a mesh network with following design
goal:
• Must be low-cost in system level, meaning not only
hardware of the node is low-cost, the receiver (gateway)
must be low-cost as well.
• Should be able to cover a large geographic area. For
example, in our farming deployment, the network covers
2.2 km2 of a farm and our campus deployment covers
around 1.2 km2 of Purdue campus.
• Has to be robust enough to survive the harsh environ-
mental conditions and the network should able to handle
node failures and recover from it.
• The deployment procedure for the IoT devices should Fig. 2. A summary of current wireless commutation technologies, where
trade-off between communication range (x-axis) and power/speed (y-axis) are
not require any specific domain knowledge. Regular users clearly represented.
should be able to simply install batteries to the IoT device
and the device will work properly. A more detailed discussion on the technologies can also be
• In addition to LoRa, the proposed network should also found in the technical report [5].
incorporate ultra-low power radio such as ANT to im- In conclusion, we choose LoRa in our deployment because
prove the performance and efficiency of the network in of the following advantages: i) the number of LoRa-enabled
dense deployments. deployment is increasing continuously while, on the other
• The proposed IoT device must be able to adopt to hand, few initial NB-IoT deployments have been already
additional sensors. deployed; ii) LoRa operates in the ISM band whereas cellular
To evaluate the performance of the proposed hybrid net- IoT operates in licensed bands; this fact favors the private
work, we conduct a series of experiments. First, we conduct LoRa networks without the involvement of mobile operators;
several in-lab tests to show the power efficiency, deployment iii) LoRaWAN, a cloud-based medium access control (MAC)
feasibility as well as the reliability of the network over layer protocol based on LoRa, has growing backing from
time. Next, we deploy two real-world testbeds in both rural industry, e.g.loRa Alliance, CISCO, IBM or HP, among others.
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 3

TABLE I
OVERVIEW OF LPWAN TECHNOLOGIES : S IGFOX , L O R A , AND NB-I OT.

Sigfox [7] LoRaWAN [8] NB-IoT [9]

Modulation BPSK CSS QPSK
Unlicensed ISM bands (868 MHz in Unlicensed ISM bands (868 MHz in Licensed LTE frequency
Band
Europe, 915 MHz in North America) Europe, 915 MHz in North America) bands
Bandwidth 100Hz 125 or 250kHz 200kHz
Maximum data rate 100bps 50kbps 200kbps
Bidirectional Limited/Half-duplex Yes/Half-duplex Yes/Half-duplex
Maximum messages/day 140(UL),4(DL) Unlimited Unlimited
Maximum payload length 12 bytes (UL), 8 bytes (DL) 243 bytes 1600 bytes
1 km (urban), 10 km
Range 10 km (urban), 40 km (rural) 5 km (urban), 20 km (rural)
(rural)
Interference immunity Very high Very high Low
Authentication & encryption Not supported Yes (AES 128b) Yes (LTE encryption)
Adaptive data rate No Yes No
End-devices join a
Handover End-devices do not join a single base station End-devices do not join a single base station
single base station
Localization Yes (RSSI) Yes(TDoA) No (under specification)
Allow private network No Yes No
Standardization No LoRa-Alliance 3GPP

B. Related Studies in LoRaWAN Furthermore, since we are talking about multi-hopping, the
The LoRaWAN network relies on the hub-and-spoke topol- energy consumption for RX has to be considered as well and
ogy in which LoRaWAN gateways relay messages between can be calculated as follows (Prx = 15.2 mW [12]):
end-devices and a central network server. This approach in-
troduces two main problems: cost and power consumption. Prx · Tonair 15.2 · 36.096
SF = 7, Erx = = = 8.57uJ/bit
Deploying multiple gateways for a large LoRaWAN network 8 8∗8
(2)
is expensive, since LoRaWAN gateways normally cost from Hence, the total energy consumption for 3 hops is:
hundreds to thousands of dollars. In addition, LoRaWAN
gateways require internet access to communicate with the ET otal = 3 · Etx + 2 · Erx = 0.67mJ (3)
server, which for many applications, like smart agriculture,
internet access might not be available. In such cases, we will
have to rely on cellular network which increases the network
development cost.
The second issue is the power consumption. To achieve
optimal transmission, LoRa utilises configuration parameters:
the carrier frequency, the spreading factor, the bandwidth and
the coding rate [10]. The combination of these parameters
affects energy consumption and transmission ranges. Taoufik
et al. [11] calculate the theoretical maximum range that can be
achieved at given output power (PT x ) level with at different
spreading factors (SF). In addition, they also proposed a energy
consumption model based on these parameters as following:

Pcons (PT x ) · (NP ayload + Np + 4.25) · 2SF

Etx = (1)
8 · P L · BW
where Etx is the energy consumed per bit, Pcons (PT x )
is the total consumed power which depends on transmission
power (PT x ), P L is the payload size and BW is the bandwidth.
To achieve long communication ranges with LoRaWAN ( >
Fig. 3. LoRa Time-on-Air vs different SF with 8 bytes payload (CR = 4/5,
10 km), high spreading factor (SF) are required with a transmit BW = 125kHz and 8 preamble symbols)
power greater than 20 dBm (assuming a path-loss exponent
equal to 3). However, a similar range can be achieved with 3 Similar observations could be found for SF9, SF10, SF11
continuous hops using 3 different nodes with SF = 7. Lowering and SF12. As a conclusion, for a battery operated LoRa net-
the spreading factor consumes significantly less energy. Total work covering a large range (> 10km), a dynamic multi-hop
energy consumption for these two scenarios can be calculated mesh network could be much more efficient than LoRaWAN.
based on Equation 1 for SX1262 LoRa transceivers at 20 dBm The power consumption is also distributed across multiple
(Pcons (20dBm) = 389.4 mW [12]) with 8 bytes payload. Fig. nodes, resulting in overall better life span for “battery driven”
3 shows the Etx for all spreading factors (6 to 12). WSN compared with LoRaWAN.
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 4

In addition, LoRaWAN’s asynchronous, ALOHA-based pro- infrastructure. The key feature of their network is the use of
tocol limits its scalability and reliability [13]. Capacity of intermediate repeater nodes (RN) that allow the formation of
LoRaWAN networks are simulated and discussed in [14], individual linear multi-hop network with clusters of sensor
which indicates LoRaWAN network has very limited capacity nodes (SN). Although this network works for monitoring
due to desiccation and duty-cycle restrictions. Varsier and underground infrastructure, their implementation comes with
Schwoerer [15] found that PDR reduced to 25% due to some limitations. The bare-bone underlying structure is still
packet collisions for a virtual large-scale application with high LoRaWAN. As discussed in Section II-B, the limitations of
node densities. To overcome the limitations of LoRaWAN, LoRaWAN still remain unsolved. In addition, the maximum
more recent studies describe time-slot-based medium access number of the child-nodes is limited to 5 SNs due to the
mechanisms. While Piyare et al. [16] describe an asynchronous inherent payload restriction of the LoRaWAN standard. The
time division multiple access (TDMA) with a separate wake- author also observed that the RNs consume about twice as
up radio channel (range of wake-up radio tested in lab envi- much energy as SNs because of the additional LoRaWAN
ronment, not multi-hop within sub-net), Reynders et al. [17] communication with the gateway. This means the RNs will
suggest using lightweight scheduling that needs an adoption drain faster than SNs in battery life, resulting in a dispropor-
of the LoRaWAN network. tionate failure rate. When RNs fail, all connected SNs will lose
the network connection. Thus these limitations could result in
C. Concurrent Transmission (CT) with LoRa higher failure rates in certain nodes and non-optimal power
Chun-Hao Liao et al. [18] developed a concurrent transmis- consumption.
sion (CT) flooding based multi-hop LoRa network with low
III. O UR S OLUTION
collision rate by introducing random delay. They demonstrated
a successful deployment of 18 sensors between multiple build- To enable the data collection with varying sensors as well
ings across 290 m x 195 m area. However, their approach falls as to support wide area coverage with low energy consump-
short in the following two design rules of WSN. tion, we proposed a hybrid network with short and long-
range communication links and designed our own sensor node
by integrating low-power micro-controller with dual wireless
communication interfaces (915MHz and 2.4GHz) to support
the proposed network.

A. Network Architecture
Our hybrid network architecture is shown in Fig. 1. Our
network topology is a mesh of multiple smaller star-topology
sub-networks. We use LoRa to build a Long-Range Mesh
Network (LRMN) and ANT to build a Short Range Star
Network (SRSN) for each individual sub-network. SRSN
Fig. 4. Demonstration of LoRa Concurrent Transmission Problem [18] can cover a circular area of a radius of about 30 meters.
SRSN works in the hub-and-spoke mode with a single hub
First, the CT-flooding approach is not applicable for WSN node receiving data from multiple spoke nodes. There are
due to high-power consumption. Figure 4 shows a basic two reasons behind the hybrid architecture. First, the IoT
relay map of a CT-LoRa network with 18 sensors. For the network should enable data collection in a wide area. Though
source node to transmit one data package to the destination LoRaWAN is capable of providing end-to-end communication
node, this CT-LoRa will have a total of 17 transmission of several miles, it suffers from high energy cost and the
and 17 receiving windows across the network. While in a inadaptability of dynamic environments in applications like
very optimized network only 5 transmission and receiving agriculture. Therefore, we utilize the long communication
windows are required for transmitting the same package. This ability of LoRa to design the LRMN while we use the much
is especially true for any “battery driven” LoRa based network more energy efficient SRSN for near-neighbor communica-
since LoRa transmission is extremely expensive in terms of tion.Second, for certain applications, there may be some areas
power when operating at high SF and transmission power. with dense deployments of sensor nodes, in which LoRa is
Second, collision still exists with a high density of end an overkill and will cause network contention. Therefore, we
devices. The authors studied the packet reception rate (PRR) design SRSN to collect data in such subareas for energy
with different numbers of relays for each node. The results conservation.
show that the PRR degrades significantly as the number of
relays increases. These drawbacks limit the general scalability B. LoRa Mesh
of their work in WSN. Our LoRa Mesh network supports the dynamic addition and
removal of sensor node without causing other nodes to stop
D. Synchronous LoRa mesh Network functioning or other manual efforts to reconfigure the network.
Recently, Ebi et al. [19] proposed a mesh network approach During deployment, the new node only needs to be placed
to extend the capability of LoRaWAN by integration with a lin- in the location of interest and it will join the mesh network
ear mesh network with multi-hopping to monitor underground automatically (Setup Phase in Algorithm 2).
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 5

1) TDMA Scheduling Algorithm: One important require- Algorithm 1 LoRa Communication Mode Scheduling
ment for the mesh network is to ensure the data is successfully 1: procedure B UILD S CHEDULE ON THE H UB N ODE
uploaded to the cloud, no matter how far the sensor node is Input: nodelist
away from the LoRa gateway. Since the sensor node that is Output: schedule
out of the communication range of a LoRa gateway needs to 2: do
find a intermediate node for routing its data, the intermediate 3: slot = new Slot();
nodes and the sensor node need to coordinate the time window 4: send coll list = [];
so that the intermediate node is in reception mode while the 5: recv coll list = [];
sensor node is sending its data. A trivial solution to this 6: while nodelist.length > 0 do
coordination problem is to always open the reception channel 7: node = nodelist[0];
of the intermediate node. However, an always-on reception 8: if (node.recv == 0
channel is energy inefficient (approximately 10 mA current for 9: and
the device in LoRa reception mode vs 2 µA in sleep mode, 10: node.packet > 0
5000 times increase in power). 11: and
Therefore, we adopt a Time-Division Multiple Access 12: !send coll list.contains(node)
(TDMA) scheduling algorithm [20] and customize it to our 13: and
system. The pseudo-code is shown in Algorithm 1. The 14: !receive coll list.contains(node)) then
input, nodelist is the list of nodes in the network sorted 15: slot[node] =0 SEN D0 ;
in the descending hops to the LoRa gateway. Therefore, the 16: slot[node.dest] =0 RECV 0 ;
algorithm starts with the node which is the furtherest from 17: send coll list.insert(node.dest.nbrs)
the gateway (line 7 in Algorithm 1) and ends with the node 18: send coll list.insert(node.dest)
which is closest to the gateway. There are three operating 19: recv coll list.insert(node.nbrs)
modes (Receive / Send / Sleep) of LoRa. The scheduling 20: node.dest.recv − −;
algorithm coordinates the sending and receiving actions for 21: node.packet − −;
all nodes in the mesh network so that the data is transmitted 22: if (node.packet == 0) then
without collisions with any other neighboring nodes that are 23: nodelist.remove(node);
also sending data at the same time. Additionally, customization 24: schedule.insert(slot)
puts the nodes mostly in Sleep mode. The difference between 25: while slot.length > 0
the original algorithm [20] and Algorithm 1 is that the original 26: return schedule
algorithm assumes a node can send all the data, including its
own sensor data as well as the data received from other nodes,
in the same packet in one timeslot. However, each node can follows the schedule to send or receive data. In the third phase
only send data with fixed size. Therefore, if the data to be (line 8 in Algorithm 2), the node sleeps to save energy since
sent is too large, it has to be fragmented into multiple packets it knows no one will send data to it.
and sent at multiple timeslots. This customization is due to the
packet size limitation in a single LoRa transmission. SX1262
LoRa transceiver [12], which we used in our deployment, has
only 256 bytes of the transmission buffer. Thus, assuming the
size of a single fragmented data packet is 256 bytes, if a node
has received 3 packets from 3 neighbors plus 1 packet of its
own sensor data, it will need to send 4 packets with 1 KB
size of data. The 1 KB data cannot be fulfilled in one timeslot
due to the SX1262 buffer limitation. Line 22 of Algorithm 1
checks the remaining packets (including those received from Fig. 5. Three phases of the nodes in LoRa mesh. Duty cycle consists of the
initial Setup phase and any Data Passing phases. One period cycle consists
other nodes as well as the packets generated by itself) of a of one Data Passing and the followed Sleep phase.
node and only after all of its packets have been scheduled, it
will be removed from the nodelist so that the algorithm will
not schedule it for the future timeslots. Section III-C gives a
concrete example to explain how the mesh network is built up C. Why centralized mesh protocol?
using Algorithms 1 and 2 The communication in our mesh network must guarantee
2) Our Mesh Protocol: The LoRa node in our mesh net- there is no collision where two nodes send data to a third
work has three phases as shown in Figure 5, namely Setup, node at the same time. Therefore, when creating the schedule
Data Passing, and Sleep. The first phase (line 2 to 6 in for each node to communicate data, we need to understand the
Algorithm 2) is the setup phase where the nodes build up a whole network structure so as to avoid such communication
routing table and the LoRa hub node builds a communication collisions. Therefore, we design a centralized approach where
schedule that indicates the time window for which nodes the hub node in the LoRa mesh is in charge of collecting
to send data (Algorithm 1). The second phase (line 7 in information from other nodes to understand the whole network
Algorithm 2) is the data passing phase where each node structure as well as creates the schedule according to the
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 6

Algorithm 2 LoRa Mesh Network Protocol during the next TDMA cycle, and will switch to reception
1: procedure O PERATIONS OF E ACH N ODE mode for further instructions. Eventually, all nodes will not
Setup Phase transmit anything, and the hub node detects there is a failure.
2: Each node builds local routing table using distance The hub node will issue a reset beacon message and floods it
vector routing protocol. to all node with similar techniques discussed in Section III-D.
3: Each node broadcasts its own routing table and for- After receiving the beacon, the network will repeat the setup
wards others’ routing tables, for the purpose of gathering phase and recovers from the node failure.
all routing tables at the Hub node.
4: Hub node builds the connectivity table for the whole
network after collecting all routing tables from other
nodes.
5: Hub node executes Algorithm 1 to build the communi-
cation schedule of each node and floods the schedule out;
all the other nodes forward the schedule after receiving it.
Data Passing Phase
6: Each node sends / receives / sleeps based on the
schedule.
Sleep Phase
7: All nodes sleep until the next Data Passing Phase.
Fig. 6. Flood hello message to build routing table in each sensor node. After
routing tables are built and collected by Node 1, a connectivity table is created
to reflect the connectivity of the whole network.
network structure. It first collects the routing table from each
LRMN node so that based on the neighborhood information
of each node, it can construct the whole network structure, TABLE II
namely the connectivity table referred in line 5 of Algorithm 2. T HE SCHEDULE BUILT FOR THE NETWORK IN F IGURE 6. A FTER TIMESLOT
6, ALL NODES SWITCH TO S LEEP MODE .
Therefore, the scheduling algorithm is centralized and we
empirically choose the node closest to the LoRa gateway as Node / Timeslot 1 2 3 4 5 6
the LoRa hub to do those jobs. 1 Rx Rx Rx Tx Tx Tx
2 Rx Tx Tx Sleep Sleep Sleep
Build individual routing table 3 Tx Sleep Sleep Sleep Sleep Sleep
We used a simple version of routing information proto- 4 Tx Sleep Sleep Sleep Sleep Sleep
col [21] to create the routing table. Figure 6 shows how
the routing table is built. The LoRa hub node broadcasts a
hello message and whoever receives the hello message will
rebroadcast it. The hello message includes the information of D. Time Synchronization
who the sender is and the shortest distance from the sender to Time accuracy is crucial for any TDMA based collision-
the hub node. Eventually, all nodes will hear the hello message free network. Un-synchronized time across the network will
from their neighbors and build an individual routing table. result in data loss and network failure. In addition, without
Collect individual routing table an external time source, the micro-controller often relies on
The individual routing tables are sent to the hub to create crystal oscillator to record time. However, the crystal oscillator
the connectivity table, which is then referred by the TDMA drifts over time. Therefore, periodic time synchronization over
algorithm to create routing schedules. Every node will broad- the network is required for stable operation. Ebi et al [19].
cast its routing table to its neighbors. Additionally, when a employ an external time source module (GPS) in their network
node receives a routing table that it has not received yet to acquire the coordinated universal time (UTC) at RNs [19].
will rebroadcast that routing table. Eventually, after all the This time will transmit to the connected SNs by a ”beacon
routing tables are collected at the hub node, the hub node flooding” with TDMA scheduling. However, using TDMA
will build a connectivity table that reflects the whole network in a LoRa mesh network for down-link communication will
structure. Figure 6 also shows the connectivity table for that significantly increase the overhead of the network. On the
mesh network. other hand, concurrent flooding addressed the need of the
Create schedules for all the nodes smaller overhead at a cost of higher chances of package
The hub node will refer to the connectivity table as well as collision [18]. Fig. 7 demonstrates the possible collision that
the customized TDMA scheduling algorithm in Section III-B1 could happen in such approach. To overcome this issue, we
to build the schedule of each node as shown in Table II. The insert a random delay between the flooding messages to
schedule is used by each node in the Data Passing phase to minimize the possibility of package collision similar to Liao
either send or receive data. et al.’s work [18].
Recover from node failure Fig. 7 shows the detailed time synchronization process, for
In case of a node failure, the associated nodes that normally each synchronization cycle, the center node (Node 1 in Fig.
receive data from the failed node will immediately discover 7) will initiate a 5 bytes beacon package containing source
the failure (no downlink communication from the failed node) of this beacon (Nsource ), number of hops (Nhops ) from the
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 7

the network is stabilized, during the first TDMA cycle, each

node will be aware of the RSSI from its previous transmission.
Then each node will adjust its PT X based on collected RSSI
value. After the adjustment, a new RSSI will be updated to
verify the quality of each LoRa link (RSSI -120dBm). This
process will increase the energy efficiency of each LoRa link
without degrading link quality.

F. ANT Hub-and-Spoke Network

Fig. 7. Left figure represents the possible package collision caused by time We use ANT to build the Short-Range Network (SRSN).
sync with concurrent flooding. Right figure shows the solution of the time ANT auto shared channel (ASC) is a communication channel
synchronization process with random delay. specified by ANT [23], and it is used to build reliable bi-
directional communication in a hub-and-spoke topology. The
ASC communication structure is shown in Figure 8. An ANT
center node and the random delay (Tdelay ) before transmitting
hub node receives data from other spoke nodes. All spoke
this beacon from the center node. Any node that receives this
nodes share a single channel to communicate with the hub.
beacon will wait for Tdelay that is smaller than the LoRa
ASC supports up to 66K spoke nodes. By default, ASC
symbol time Tsymbol and then immediately re-transmit this
requires a user specified channel master node to establish the
beacon. This approach minimizes the package collision as
network, we added a layer on top of that called Adaptive Mode
well as the down-link overhead since the flooding beacon
Switching (AMS) to dynamically select the hub node. AMS
is only 5 bytes. After the beacon has been received by the
is based on a round robin (based-on battery level) campaign
nodes, the relative time elapsed Tpast since the previous node’s
to select the hub node. During the setup phase, each node
transmission can be calculated as follows:
will broadcast its ID and battery level and actively listens for
other near-by ANT broadcast, node with the highest battery
Tpast = Tbeacon + Tnode + Tdelay (4) level will become the hub node. If multiple nodes have the
where Tnode is the time that is need for node to process same battery level, the node with the lowest ID will become
and re-transmit the beacon, Tbeacon is the time on air of the the hub node. Once the hub node was selected, the ASC uses
package. an ANT proprietary shared channel topology to establish the
However, because of the variability of the Tnode due to network, with the hub node being channel master and the rest
the SPI communication between the micro-controller and of the nodes become shared slaves [23].
LoRa transceiver and the imperfect time synchronization, we
manually expand the receiving windows by 5ms to compensate
for the inconsistency. This process will synchronize the timing
across all the node in the network to avoid the time drifts over
a long period of time.

E. Adaptive LoRa Link (ALL)

Fig. 8. ANT shared channel hub-and-spoke network structure. H represents
One of the benefits of adapting the LoRa technology is the hub node (Channel Master) and S represents the spoke node (shared slave)
the “degrees of freedom” at the physical layer. Ochoa et
al. study suggested that the the potential of an adaptive In each SRSN, all the spoke nodes send data to a hub
LoRa solution (i.e., in terms of spreading factor, bandwidth, node via ANT and the hub node is in charge of uploading
transmission power, and topology) could greatly optimize the aggregated data to the cloud. The hub node has two
energy consumption without sacrificing the communication methods to upload the aggregated data. First, if the SRSN
range [22]. In our proposed LoRa mesh network, we provide is a standalone network and associated with an ANT gateway,
the proof of such concept by adopting the Adaptive LoRa it does not need to enable the AMS functionality but directly
Link (ALL) to further improve the energy performance of the sends data to the ANT gateway. Second, if the SRSN is part
network. Although there are many physical parameters that can of a LRMN (the SRSN clusters in Figure 1), the hub node will
affect the energy of the LoRa link, SF and transmission power switch to LoRa mode to route the aggregated data to the LoRa
have the most impact in term of the energy consumption. In gateway. Clearly, the hub node consumes more energy than
this work, as a proof of concept, we are only focusing on the spoke nodes. Therefore, AMS is used for balancing the
adjusting the transmission power of the network. However, energy consumption. Figure 9 demonstrates this idea. When
similar methodologies will apply for adjusting the SF. a hub node drains to a lower-than-threshold battery level, it
During the setup phase, in addition to building a routing will issue a AMS message and the spoke node with highest
table, each node will note the Received Signal Strength battery and within the same SRSN will be selected as the new
Indicator (RSSI) which is a estimated measure of the signal hub. The original hub node then works as a spoke. This AMS
power level from each of the LoRa packages it received. Once feature guarantees no single node in SRSN is significantly
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 8

drained and ensures the SRSN cluster still remains in the same The hardware platform used in this study builds on the
LRMN network. hardware platform that we reported in our earlier work [24]
as shown in Fig. 11 and Fig.12(a). It utilizes the HMAA-1220
wireless transceiver module (Fig.12(b)) from HuWoMobility
[25]. The HMAA-1220 wireless transceiver was powered by
nRF52832 chip from Nordic Semiconductor [26] and SX1262
LoRa transceiver from Semtech [27]. nRF52832 features a
low power 32-bit ARM Cortex-M4F processor with a built-
in-radio that operates in the 2.4 GHz ISM band and supports
ANT, BLE and Bluetooth 5 wireless protocols with up-to +4
dBm transmit power. It is also equipped with 64 kB RAM and
512 kB of flash storage which can be used for storing data for
in-node data analysis.
Fig. 9. Adaptive Mode Switching (AMS) flowchart

To address the potential failures of the hub node, we use a

heuristic method, a timer (Failure Detection Timer in Fig. 10)
to detect the failure of the hub node. As shown in the left
part of Figure 10, during normal operation, the hub node
periodically sends an update request to each spoke node to
request for new sensor and battery data. When a spoke node
finishes data upload, the spoke node will reset the failure
detection timer. This timer is set to be much longer than the
data upload periodicity (e.g. 5 × periodicity) to count for any
possible data loss. In the scenario of hub node failure (the
right part of Fig. 10, the timers on the hub nodes will expire
and all the spoke nodes will reset itself, resulting in the entire
SRSN to be re-initialized. The remaining nodes will form a
new SRSN without the failed node. After the new SRSN is
initialized, the new hub node will switch on LoRa reception
mode and waits for the next TDMA cycle to join back to the Fig. 11. System block diagram of the hardware platform. Adopted from [24]
LoRa mesh network.
SX1262 from Semtech [12] is the new sub-GHz radio
transceivers which is ideal for long range wireless applications.
It supports both LoRa modulation for LPWAN applications
and FSK modulation for legacy use cases. In addition, SX1262
also complies with the physical layer requirements of the
LoRaWAN specification released by the LoRa Alliance [28]
and the continuous frequency coverage of SX1262 from 150
MHz to 960 MHz allows the support of all major sub-GHz
ISM bands. SX1262 was designed for long battery life with
current consumption of 4.6 mA in active receive mode and
600 nA in sleep mode. With the highly efficient integrated
power amplifiers, SX1262 can transmit up-to +22 dBm while
having a high sensitivity down to -148 dBm. Along with
the co-channel rejection of 19 dB in LoRa mode and 88
Fig. 10. SRSN Normal Operations and Failure Tolerance. dB blocking immunity at 1 MHz offset, SX1262 provides
a maximum of 170 dB link budget which is ideal for long
distance communication.
Figure12(a) shows the printed circuit board (PCB) with its
G. Hardware Sensing Platform in Our Deployment packaging. The HMAA-1220 (Fig.12(b)) module was mounted
There are three main challenges when designing our plat- on a “motherboard” with 4 LEDs and some pinouts for
form for deployment: building a sufficiently low weight, low connecting different “daughter board” (Fig.12(c)). This design
cost and energy efficient hardware capable of mass production, choice allow us to extend the flexibility of our hardware plat-
incorporating numerous subsystems to facilitate various appli- form to facilitate different applications. The entire PCB was
cations (e.g.smart agriculture and smart city), and protecting enclosed in an IP67 Industrial-grade packaging for protection
the electronics from harsh environmental conditions. against environmental factors.
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 9

meaning from a network stack point of view, this lab test will
be able to simulate the real-world conditions.
Table III shows the LoRa configuration of the lab experi-
ment. To fully review the performance as well as the stability
of the network, we kept the network running with each node
programmed to transmit one 64 Bytes data package per minute
(1.8kBps) for an entire week. As the performance indicator, the
receiver records all of the network traffic from all nodes. RSSI
are not evaluated in this test since we intentionally lowered
the TX power of all LoRa nodes. We evaluated the reliability
and stability of data packet delivery for each individual node
using the packet deliver rate (PDR), i.e. the ratio between the
number at the center node (# RECEIVED) and the number of
packets that should have been received (# EXPECTED). With
the help of the traffic monitoring receiver, two sets of PDRs
Fig. 12. The hardware platform: (a) Motherboard PCB and battery holder
with its packaging, (b) Motherboard PCB with two daughter boards, (c)the were can be calculated, P DRi and T otal P DRd :
HMAA-1220 module P
#RECEIV EDi
P DRi = P (5)
TABLE III #EXP ECT EDi
L O R A CONFIGURATION FOR THE IN - LAB TEST AND FIELD DEPLOYMENT
Where P DRi is the packet delivery rate of node i during
Parameter In-lab test Field Deployment the entire week.
Spreading
SF 7 7
Factor
P
#RECEIV ED
Bandwidth BW 125kHz 125kHz T otal P DRd = P d (6)
Preamble d #EXP ECT ED
8 8
length
Transmission Where T otal P DRd is total the packet deliver rate of all
PT x +0dBm Variable
Power the node during the #dth day.
Coding Rate CR 4/5 4/5
Head & Head & The estimation of the number of expected packets are based
CRC checking on each specific node with constant transmission interval (1
Payload Payload
min). Because of the nature of multi-hop network, counting
the number of packets arriving at the receiver will counts
IV. R ESULTS for both node-specific performance as well as the multi-hop
A. In-lab Evaluation route from that specific node to the center node. In contrast,
To verify the functionality and stability of our network, we T otal P DRd provides an overview of the network stabil-
conducted a series of both in-lab tests as well as real-world ity over time. Instead evaluating node-specific parameters,
deployments. T otal P DRd provides an overview of the system stability
1) Network Stability Test: We performed laboratory exper- of time synchronization as well as the TDMA routing.
iments with 9 sensor nodes with 1 node acting purely as Fig. 14(a) shows the results of the P DRi of our one-week
standalone receiver monitoring the entire network to check test, all nodes show more than 99% PDR except for node 4.
the stability of the network. Fig. 14(b) shows the results of the T otal P DRd for the same
a) Network Setup: Fig. 13 shows the configuration of test, it suggests our network stability is very strong and is not
the network structure. All 8 nodes are placed together on a time dependent.
lab bench along with the receiver. Node 1 is the center node
for this test. Testing the network structure in this condition is
nontrivial since all sensors are in close proximity, the long-
range capability of LoRa will not able to form the structure
that we desired since all node are in range with each others
no matter how we configure the LoRa parameters. To resolve
this issue, we created a filter in the low-level firmware (LoRa
driver) to block connections from un-wanted sensor nodes. For
example, in the structure shown in Fig 13, Node 4 should only Fig. 13. Network structure for the in-lab system stability test.
receive data from Node 3 and Node 5 in the desired structure.
However, in reality node 4 is able to receive data from all 2) Power Consumption: The power consumption was mea-
nodes because of the close proximity between sensors. The sured for one node under real-life conditions for a network
firmware filter will filter any data from nodes other than node structured with four nodes as shown in Fig 15. In addition,
3 and node 5. All other data will be ignored to simulate the each sensor is connected with HDC2010 temperature and
desired network structure. The biggest benefit of the approach humidity sensor. For each cycle, each node will transmit a
is that the mesh-layer of the network is completely un-touched, package of 64 bytes including one temperature and humidity
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 10

TABLE IV
M EASURED P OWER PROFILE AND E XPECTED BATTERY L IFE WITH 2 X AA
BATTERY

Average Current Draw (Iavg ) Expected Battery Life

Node 1 33 µA 9 years
Node 2 74 µA 4 years
Node 3 56 µA 5 years
Node 4 25 µA 11 years

batteries) of each node. Comparing between nodes 1,2, and

3, the energy consumption of different nodes is determined
by the number of receive/transmits windows. Note that Node
4 consumes significantly lower power. This shows that ANT
radio is superior in terms of energy efficiency compared with
LoRa. In conclusion, even though energy consumption is
highly dependent on the network structure and while some
nodes do consume high power, our network still has a very
Fig. 14. Results (PDR and T otal P DR) for the in-lab system stability test. acceptable expected battery life across all nodes.

reading. Due to the nature of our mesh network, energy

consumption within the network varies depending on i) the
position of the participating node in the hierarchy of the mesh
network, and ii) the topology type of the network. As shown
in Fig. 15 node 1, 2, and 3 are connected with the proposed
LoRa mesh protocol. Node 3 also communicates with node 4
via ANT. The goal of this configuration is to route the data
from Node 4 (ANT), Node 3 (ANT + LoRa), and Node 2
(LoRa) to Node 1 (LoRa). For each TDMA scheduling cycle,
Node 4 will transmit one package via ANT to Node 3 and
Node 3 will forward this package plus its own package to
Node 2 via the LoRa. Then, Node 2 will transmit 3 packages
(2 received + 1 own) to Node 1. The LoRa configuration for
this test is identical to the field deployments with +18 dBm
transmission power.

Fig. 16. Measured current profile for Node 3 during one transmission cycle
where blue region represents LoRa’s Tx and Rx windows and red region
represents the ANT transmission window.

3) Hardware durability: To fully test the performance and

durability of our hardware platform against harsh environ-
ments, we deployed 4 units equipped with our smart agri-
culture interface at Throckmorton-Purdue Agricultural Center
Fig. 15. Network setup for the energy consumption test (TPAC) at Purdue University. All four units were equipped
with temperature and humidity sensors to monitor the envi-
Power consumption was measured using a N6705B DC ronmental conditions at the farm and were placed 1 meter
power analyzer from Agilent Technology [29]. Each sensor above the ground. The deployment locations are shown in
node is powered with 3.3V DC by the DC power analyzer. Fig. 17. Two of the units with printed thin-film nitrate sensors
Fig. 16 shows part of the current profile of Node 3 where each were installed in the stream to measure the nitrate pollutants
state of operation is clearly marked. The hardware consumes from fertilizer runoffs in the stream. The other two units were
around 25 µA during sleep, 10 mA during ANT TX, 12.5 mA interfaced with four independent ECH2O 5TE Soil sensors
during LoRa RX and 72.5 mA during LoRa TX. Therefore, to monitor the soil temperature, conductivity, moisture, and
the average current consumption with 10 minutes TDMA dielectric at four different depths [30]. All four units were
cycle delay for Node 3 is 56 µA which can be translate to programmed to send data every 10 minutes and the data
5.2 years of expected battery life with standard AA alkaline received at the receiver is uploaded to the data server and
batteries (2500 mAh). Table IV shows the average current displayed on the web. As of January 2020, these units have
consumption and the expected battery life (with 2 AA alkaline been continuously operating for more than a year without
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 11

21(a)). The red node represents the location of the center node.
The green dot represents the receiver, a laptop connected with
SX1272DV K1CAS (LoRa development kit) from Semtech
[31]. Each solid line represents the actual LoRa link for the de-
ployed network and the dotted line represents available LoRa
links that were not being used. The node junction represents
two nodes (node 12 and node 13) that were deployed on
purpose at close range and were communicating via ANT
instead of LoRa (Fig.21(b)). In addition, three paths that are
highlighted in Fig. 18 represent three distinguished data flows
that are formed by the mesh network. For instance, path
number 1 represents the path 9 → 8 → 7 → 1. All nodes
(including the center node) are located 1 meter above ground
level as shown in Fig. 21 and the receiver is placed on the
3rd floor inside of an office building facing south-east. The
furthest node is placed 1.5 km away from the receiver and
Fig. 17. Deployment of the proposed mesh network on Purdue University across more than 18 buildings in between.
West Lafayette campus. Each blue dot represents a LoRa node and the black
dash line represent a stable LoRa link Each node was programmed to transmit 64 Byte package at
a fixed time interval (2 minute) with the LoRa configuration
shown in table III. In addition, two furthest nodes 6 and 9 will
major failure. The highest temperature recorded is above send an additional package with SF12 outside of the TDMA
110 o F and the lowest temperature is -40 o F . This proves cycle as comparison with transitional ALOHA based network
our hardware is capable of withstanding harsh environmental such as LoRaWAN. Furthermore, since the LoRa radio of node
conditions while being in situ in an unmonitored outdoor 13 in the node junction was not active (it only communicates to
environment. node 12 via ANT), it was enabled to broadcast (in parallel with
the mesh network, acting like a LoRaWAN node) with SF12
B. Large-scale Deployment directly to the receiver used as the comparison for evaluation.
After each transmission cycle, the center node will forward
1) Farm deployment: Our proposed network was first tested
all incoming packages to the receiver which will upload all
in TPAC to evaluate the multi-hop performance of the mesh
data to the cloud for future analysis. All data packages are
network for covering long-ranges. Five mesh nodes were
checked with 16-bit Cyclic Redundancy Check(CRC) to verify
deployed in linear hop formation in addition to the existing
their integrity, corrupted data package will be marked and
4 sensing nodes equipped with soil moisture sensors and
stored. Fig. 20 shows the TDMA schedule of the deployed
flexible nitrate sensors. The newly deployed mesh nodes were
network with each slot set to 125 ms for transmitting a 64
equipped with temperature and humidity sensors as well as
Byte package. The three paths represent the three continuous
nitrate sensors and powered with 4 AA batteries. Fig. 17 shows
LoRa links as shown in Fig. 18. Each section represents one
the map of the farm deployment: The green dot represents the
different LoRa hop and are color coded for clarity. The delay
receiving computer; The red dot represents the center node
inserted after the 3rd hop is required to avoid package collision
of the mesh network; The blue dots represents the 5 mesh
with node 7 from path 1. SF12 reference represents the time it
nodes; The yellow dots represent the four previously deployed
took for one 64 Bytes transmission with SF12 as comparison.
sensor nodes that were not part of the mesh network. The LoRa
It is clear that even for the longest path (Path #2) with 5 hops
configuration of each node is shown in Table III with SF7 and
in between, the overhead of the furthest mesh node (Node 6)
TX power = 15 dBm. Each node was programmed to send one
is smaller than the single-hop LoRaWAN node. However, it
64-byte packet every 10 minutes. Once each cycle is complete,
is worth noticing that the overheads are greater for the nodes
the center node will forward the packages to the receiver for
that were closer to the center node (Node 7, 10, and 2) and it
upload. With 4 linear hops, the proposed mesh network is able
is a necessary trade-off between power efficiency and network
to cover 3 km in farmland with more than 98% PDR across
robustness.
all nodes with SF7. This experiment confirmed that our mesh
network is able to cover long distances with low spreading Over the entire evaluation period of two weeks, each node
factor (SF7). transmitted one 64 Byte package every two minutes, which
2) Campus deployment: In the campus-scale deployment, means a total of 9360 packages were expected from each node
we placed 13 LoRa nodes, distributed in a 1.1 km by 1.8 (13 days of operation were considered, network was taken
km area of Purdue University campus. All 13 nodes were down by one day for evaluation). As in the previous section,
deployed and continuously operated for a period of two weeks. the network integrity is evaluated by analyzing PDR of all
Fig. 18 shows the the complete map of our campus-scale nodes. In addition, the Package Miss Rate (PMR) and Package
experiment where a complete mesh network is established. Error Rate (PER) are analyzed as well. Similar to PDR, PER
Each blue dot represents a network node which are randomly is calculated based on the number of the package that were
and evenly distributed across the entire Purdue campus (Fig. marked as corrupted and PMR are calculated based on number
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 12

Fig. 18. Map of the network deployment on Purdue campus with 13 nodes. The red dot represents the center node (CN), the network nodes are represented
as blue dots, the blue square represents the node junction (Node #12 and #13) and the green dots represents the receiver.

of missed packages: However, from the observation based on the results from Fig.
P 19, although both of the PMR and PER did trend to increase as
#ERRORi number of hops increased, these effects are minimal comparing
P ERi = P (7)
#EXP ECT EDi with the PDR. Furthermore, node 12 and 13 showed much
P P higher PER (> 4%), we suspect that this might be due to
#EXP ECT EDi − #RECEIV EDi higher interference in the sub-GHz ISM band since both of
P M Ri = (8)
those two nodes were located in one of the most populated
P
#EXP ECT EDi
areas on campus. On the other hand, the LoRaWAN reference
Fig. 19 shows the end-to-end PDR, PMR and PER based on
node 13* shows a very low PDR (51.5%) and very high PER
the total expected number of packages and received packages
and PMR, when compared to our mesh node deployed at the
for all 13 nodes including the SF12 reference, where the
same location (node 12). This confirms that our mesh is able
blue bars represent all nodes in the mesh network, and the
to provide better quality of service which further supports our
orange bar represents the SF12 reference. Over the entire
purposed network structure.
deployment period of two weeks, the proposed mesh network
achieves more than 96% PDR except for node 12 and node 13
comparing to 51.5% of the SF12 reference node. In addition, V. L IMITATIONS
from Fig. 19 (b) and (c), both the PMR and PER for all As Ochoa et al. [22] point out, the energy consumption
nodes are significantly lower than the SF12 reference node. of LoRa mesh nodes can be further optimized by exploiting
This confirms that the proposed network provides much better different radio configurations and the network topology (e.g.,
quality of service particularly for large area networks. Where the number of hops, the network density, the cell coverage).
higher spreading factors are necessary for transitional star net- For sparse networks, higher SF is necessary along with higher
work to cover such as in the ALOHA protocol in LoRaWAN’s transmitted power. The Adaptive LoRa Link in our implemen-
approach, it is worth noticing that the PMR increases as the tation did not include the functionality to change the spreading
number of hops increases, this is expected since the time factor as the network topology in our deployment did not
flooding will degrade as the number of hops increases. Imper- change over time. One aspect of our future work is to include
fect time synchronization will cause time slot mismatch and such adaptivity in our implementation.
therefore results in either missed packages (TX/RX window Another important limitation of the network occurs during
miss match) or package collision (TXs windows miss match). the flooding for the TDMA scheduling. In our current con-
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 13

(a)

(b)

(c)
Fig. 19. (a) Packet Delivery Rate (PDR), (b) Package Error Rate (PER), (c) Package Miss Rate (PMR) of the two weeks on-campus deployment. Blue bars
represent node 1 to 13 across the network. Orange bar represents the SF12 LoRaWAN reference from node 6, 9 and 13 respectively

figuration, one TDMA schedule table for the entire network based on the routing path. Only the necessary TDMA schedule
is flooded to each node for the simplicity of the design. will be flooded to each multi-hop path in the network. This
However, there are two inherent problems with this approach. approach will significantly reduce the overhead time during
First, flooding the entire table requires transmitting multiple the setup phase without sacrificing the network performance.
LoRa packages throughout the entire network. Not only is
While LoRaWAN allows for AES 128-bit encryption, in
this approach inefficient, but it also increases the overhead
the current phase of our work, no encryption mechanisms
for the setup phase. Second, due to the limitation of the
have been deployed. While we plan on deploying AES-128
maximum data package (255 bytes at SF7) of LoRa, the
encryption in future LoRaWAN deployments, a caveat of
maximum number of nodes in the network will be limited.
introducing secure network channels will be the reduction of
Although, this limitation can be patched with flooding multiple
the available payload size, which may further limit the number
schedule tables throughout the network, this approach will still
of supportable nodes in a sub-network. Thus, we will look into
be inefficient and will significantly impact the overhead during
deploying lightweight encryption mechanisms for IoT devices,
the setup phase. For our future work, instead of flooding a
such as ACES [32], [33].
entire network table to every node, we will divide the table,
As of the sensor data management, there are several other
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 14

Fig. 20. TDMA schedule of the campus deployed network, each Hop is clearly marked. Path 1 to 3 corresponds to the path from Fig. 18. The total length
represents the total overhead of the corresponding path.

consumption, we designed and manufactured our own sensor

node integrating a micro-controller, wireless communication
interfaces, and a hybrid network with short (2.4GHz) and long-
range (915MHz) communication links.
With our hybrid mesh network, we have shown a signif-
icant improvement in both power consumption as well as
communication range while comparing with traditional single
hop network like LoRaWAN. In addition, full-scale real-world
experiments on both Purdue Campus and agricultural farms
with more than 20 nodes further suggested that the proposed
network significantly improves the quality of service while
maintain long-term stability. We provide several areas of future
Fig. 21. Deployment at Purdue campus. (a) node #7 installed on a street lamp work motivated by our design and experiments on these large
post (b) node junction consisting node 12 and 13 next to a campus building
scale IoT testbeds, including sophisticated anomaly detection,
on-device computation, and network synchronization.
challenges about network latency and data analytics, which we
have more thorough discussion in [5] ACKNOWLEDGMENT
The work described in this paper is part of a project funded
VI. C ONCLUSION through the Wabash Heartland Innovation Network (WHIN)
The recent advancement of the Internet of Things (IoT) and the SMART Film consortium at Purdue University.
enables the possibility of data collection from diverse en-
vironments using IoT devices. However, despite the rapid R EFERENCES
advancement of low-power communication technologies, the
deployment of IoT network still faces many challenges. In [1] S. Chaterji, N. DeLay, J. Evans, N. Mosier, B. Engel, D. Buckmaster,
and R. Chandra, “Artificial intelligence for digital agriculture at scale:
particular, large-scale WSN such as digital agriculture and Techniques, policies, and challenges,” arXiv preprint arXiv:2001.09786,
smart and connected cities remains a major challenge in 2020.
terms of communication range, quality of service and power [2] K. Matthews. (2019, Nov.) 5 iot use cases that will shape the
future of agriculture. [Online]. Available: https://ubidots.com/blog/
consumption. agriculture-smart-farming/
This paper presents the design of a hybrid LPWAN mesh [3] M. Aleksandrova. (2018, Jun.) Iot in agriculture: 5 technology use
network for IoT application that delivers several-kilometers cases for smart farming (and 4 challenges to consider). [Online].
Available: https://easternpeak.com/blog/
with only low-power nodes while provides excellent QoS. [4] S. Chaterji, P. Naghizadeh, M. A. Alam, S. Bagchi, M. Chiang, D. Cor-
Our work addresses the development of large-scale WSN that man, B. Henz, S. Jana, N. Li, S. Mou et al., “Resilient cyberphysical
is suitable for distinct application areas with real world de- systems and their application drivers: A technology roadmap,” arXiv
preprint arXiv:2001.00090, 2019.
ployments. To enable the data collection with varying sensors [5] X. Jiang and H. Zhang. (2020, Apr.) Technical report. [Online]. Avail-
as well as to support wide area coverage with low energy able: https://lladzhang.github.io/heng.github.io/Technical report.pdf
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 15

[6] K. Mikhaylov, J. Petaejaejaervi, and T. Haenninen, “Analysis of capacity IEEE/IFIP International Conference on Dependable Systems and Net-
and scalability of the lora low power wide area network technology,” in works (DSN). IEEE, 2019, pp. 234–246.
European Wireless 2016; 22th European Wireless Conference. VDE, [33] A. A. Clements, N. S. Almakhdhub, S. Bagchi, and M. Payer, “ACES:
2016, pp. 1–6. Automatic compartments for embedded systems,” in 27th {USENIX}
[7] (2017, Feb.) M2m and iot redefined through cost effective and energy Security Symposium (USENIX Security 18), 2018, pp. 65–82.
optimized connectivity. SIGFOX. [Online]. Available: https://www.
sigfox.com/sites/default/files/1701-SIGFOX-White Paper Security.pdf
[8] L. Alliance. What is lorawan specification. [Online]. Available:
https://lora-alliance.org/about-lorawan
[9] “Lte evolution for iot connectivity,” White Paper, Nokia, Nov.
2016. [Online]. Available: https://www.open-ecosystem.org/assets/
lte-evolution-iot-connectivity Xiaofan Jiang is a Ph.D. Candidate at the School
[10] O. Khutsoane, B. Isong, and A. M. Abu-Mahfouz, “Iot devices and of Electrical and Computer Engineering, Purdue
applications based on lora/lorawan,” in IECON 2017-43rd Annual Con- University in West Lafayette, Indiana. He is advised
ference of the IEEE Industrial Electronics Society. IEEE, 2017, pp. by Dimitrios Peroulis. He received the B.S. degree
6107–6112. from Purdue University in 2015. His research inter-
[11] T. Bouguera, J.-F. Diouris, J.-J. Chaillout, R. Jaouadi, and G. Andrieux, ests include design and implementation of various
“Energy consumption model for sensor nodes based on lora and lo- embedded systems, wireless sensor networks and
rawan,” Sensors, vol. 18, no. 7, p. 2104, 2018. IoT devices for general and industry applications.
[12] “SX1261/2 Long Range, Low Power, sub-GHz RF Transceiver data
sheet,” Semtech, Camarillo, CA, USA.
[13] D. Zorbas, K. Abdelfadeel, P. Kotzanikolaou, and D. Pesch, “Ts-lora:
Time-slotted lorawan for the industrial internet of things,” Computer
Communications, vol. 153, pp. 1–10, 2020.
[14] F. Adelantado, X. Vilajosana, P. Tuset-Peiro, B. Martinez, J. Melia-
Segui, and T. Watteyne, “Understanding the limits of lorawan,” IEEE
Communications magazine, vol. 55, no. 9, pp. 34–40, 2017. Heng Zhang is a Ph.D. Student at the School of
[15] N. Varsier and J. Schwoerer, “Capacity limits of lorawan technology for Electrical and Computer Engineering, Purdue Uni-
smart metering applications,” in 2017 IEEE International Conference on versity West Lafayette, Indiana. He is advised by
Communications (ICC). IEEE, 2017, pp. 1–6. Saurabh Bagchi. He received the B.S. degree from
[16] R. Piyare, A. L. Murphy, M. Magno, and L. Benini, “On-demand tdma Shanghai Jiao Tong University in 2016. His research
for energy efficient data collection with lora and wake-up receiver,” in interests include edge computing, mobile sensing,
2018 14th International Conference on Wireless and Mobile Computing, wearable resilience, and IoT networking.
Networking and Communications (WiMob). IEEE, 2018, pp. 1–4.
[17] B. Reynders, Q. Wang, P. Tuset-Peiro, X. Vilajosana, and S. Pollin,
“Improving reliability and scalability of lorawans through lightweight
scheduling,” IEEE Internet of Things Journal, vol. 5, no. 3, pp. 1830–
1842, 2018.
[18] C.-H. Liao, G. Zhu, D. Kuwabara, M. Suzuki, and H. Morikawa, “Multi-
hop lora networks enabled by concurrent transmission,” IEEE Access,
vol. 5, pp. 21 430–21 446, 2017.
[19] C. Ebi, F. Schaltegger, A. Rüst, and F. Blumensaat, “Synchronous lora
mesh network to monitor processes in underground infrastructure,” IEEE Edgardo Barsallo Yi is a Ph.D. Candidate at the
access, 2019. Computer Science Department, Purdue University.
[20] N. A. Pantazis, D. J. Vergados, D. D. Vergados, and C. Douligeris, He is advised by Saurabh Bagchi. Before he joined
“Energy efficiency in wireless sensor networks using sleep mode tdma Purdue, he worked as a software engineer and a soft-
scheduling,” Ad Hoc Networks, vol. 7, no. 2, pp. 322–343, 2009. ware architect at the Indra Software Labs, Panama.
[21] C. L. Hedrick, “Routing information protocol,” 1988. He holds a Master in Software Engineering from
[22] M. N. Ochoa, A. Guizar, M. Maman, and A. Duda, “Evaluating lora UPSAM (Spain) and a B.S. in Computer Systems
energy efficiency for adaptive networks: From star to mesh topologies,” from UTP, Panama. His research interests include
in 2017 IEEE 13th International Conference on Wireless and Mobile databases, distributed systems, mobile computing,
Computing, Networking and Communications (WiMob). IEEE, 2017, and wearable resilience.
pp. 1–8.
[23] (2007, Jul.) Ant message protocol and usage. Dynastream Innovations
Inc. [Online]. Available: https://www.sparkfun.com/datasheets/Wireless/
Nordic/ANT-UserGuide.pdf
[24] X. Jiang, J. F. Waimin, H. Jiang, C. Mousoulis, N. Raghunathan,
R. Rahimi, and D. Peroulis, “Wireless sensor network utilizing flexible
nitratesensors for smart farming,” in IEEE SENSORS 2019. IEEE,
Nithin Raghunathan received his Ph.D in electrical
2019.
engineering from Purdue University, West Lafayette,
[25] (2019) Huwomoibilty. [Online]. Available: http://www.huwomo.com/ IN, USA, in 2014. His dissertation focused on
[26] (2019) nrf52832. [Online]. Available: https://www.nordicsemi.com/ the development on micro-machined g-switches for
Products/Low-power-short-range-wireless/nRF52832 impact applications typically in the ranges of 100
[27] (2018) semtech. [Online]. Available: https://www.semtech.com/products/ 60,000 gs. He worked as Post-Doctoral Research
wireless-rf/lora-transceivers associate from 2014 to 2015 and was involved in
[28] L. Alliance, “A technical overview of lora and lorawan,” White Paper, the development of wireless radiation sensors for
2015. dosimetry applications. He is currently a Staff Sci-
[29] (2019) nrf52832. [Online]. Available: https://www.keysight.com/en/ entist at the Birck Nanotechnology Center at Pur-
pd-1842303-pn-N6705B/dc-power-analyzer-modular-600-w-4-slots due University. He is currently working on Wabash
[30] (2019) Meter group ech2o 5te soil moisture sensor. [Online]. Available: Heartland innovation Network (WHIN) on the development of IoT sensors
https://metos.at/portfolio/decagon-5te-soil-moisture-sensor/ and network for Industrial and agricultural operations. His other interests
[31] (2019) Semtech sx1272dvk1cas development kit, sx1272, 915 mhz. include novel MEMS inertial devices, development of new microfabrication
[Online]. Available: https://www.semtech.com/products/wireless-rf/ techniques, wireless and flexible sensors and sensors for Lyophilization and
lora-transceivers/sx1272dvk1cas aseptic processing and also sensors for industrial and harsh environments.
[32] N. S. Almakhdhub, A. A. Clements, M. Payer, and S. Bagchi, “Benchiot:
A security benchmark for the internet of things,” in 2019 49th Annual
IEEE INTERNET OF THINGS JOURNAL, APRIL 2020 16

Charilaos Mousoulis received the Ph.D. degree in Ali Shakouri is the Mary Jo and Robert L. Kirk
Electrical and Computer Engineering from Purdue Director of the Birck Nanotechnology Center at Pur-
University, West Lafayette, IN, USA in 2012. Since due University. He received his diplome d’Ingenieur
2014 he is a Senior Research Scientist at the School in 1990 from Telecom ParisTech, France and his
of Electrical and Computer Engineering, Purdue Ph.D. in 1995 from California Institute of Technol-
University, West Lafayette. He was previously a ogy in Pasadena, CA. His group studies nanoscale
Postdoctoral Research Associate at the School of heat transport and electrothermal energy conversion
Biomedical Engineering, Purdue University, West to improve electronic and optoelectronic devices.
Lafayette. His research interests include microsys- They have also developed novel imaging techniques
tems for biomedical applications, silicon-based radi- to obtain thermal maps with sub diffraction-limit
ation sensors for occupational dosimetry, sensors for spatial resolution and 800ps time resolution. He is
food safety, flexible hybrid electronics, and IoT-based sensors for precision applying similar methods to enable real-time monitoring of functional film
agriculture and advanced manufacturing. continuous manufacturing. He is leading SMART industry consortium to
manufacture low-cost internet of thing (IoT) devices and sensor network. As
a part of Wabash Heartland Innovation Network (WHIN), they are developing
community IoT testbeds in advanced manufacturing and high tech agriculture.

Saurabh Bagchi is a Professor in the School of

Electrical and Computer Engineering and the De-
Somali Chaterji is an Assistant Professor in the partment of Computer Science at Purdue University
Department of Agricultural and Biological Engi- in West Lafayette, Indiana. He is the founding Direc-
neering at Purdue University, where she specializes tor of a university-wide resilience center at Purdue
in developing algorithms and statistical models for called CRISP (2017-present). He was elected to the
genome engineering, precision health, and digital IEEE Computer Society Board of Governors for the
agriculture. Dr. Chaterji got her PhD in Biomed- 2017-19 term and re-elected in 2019. He is a co-
ical Engineering from Purdue University, winning lead on the WHIN-SMART center at Purdue for IoT
the Chorafas International Award, College of En- and data analytics. Saurabh’s research interest is in
gineering Best Dissertation Award, and the Future dependable computing and distributed systems. He
Faculty Fellowship Award. She did her Post-doctoral is proudest of the 21 PhD students and 50 Masters thesis students who have
Fellowship at the University of Texas at Austin in graduated from his research group and who are in various stages of building
the Department of Biomedical Engineering, where her work was supported wonderful careers in industry or academia. In his group, he and his students
by an American Heart Association award. She followed this up with a Post- have way too much fun building and breaking real systems.
doctoral stint at Purdue Computer Science when she got her first NIH R01 on
computational metagenomics. Dr. Chaterji is a technology commercialization
enthusiast and has been consulting for the IC2 Institute at the University of
Texas at Austin, since Spring 2014.

Dimitrios Peroulis (S99M04SM15F17) is the Reilly

Professor and Michael and Katherine Birck Head of
the School of Electrical and Computer Engineering
at Purdue University. He received his PhD degree
in Electrical Engineering from the University of
Michigan at Ann Arbor in 2003. His research in-
terests are focused on the areas of reconfigurable
systems, cold-plasma RF electronics, and wireless
sensors. He has been a key contributor in developing
high quality widely-tunable filters and novel filter
architectures based on miniaturized high-Q cavity-
based resonators in the 1-100 GHz range. He is currently leading research
efforts in high-power multifunctional RF electronics based on cold-plasma
technologies. He received the National Science Foundation CAREER award
in 2008. He is an IEEE Fellow and has co-authored over 380 journal and
conference papers. In 2019 he received the Tatsuo Itoh Award and in 2014
he received the Outstanding Young Engineer Award both from the IEEE
Microwave Theory and Techniques Society (MTT-S). In 2012 he received
the Outstanding Paper Award from the IEEE Ultrasonics, Ferroelectrics, and
Frequency Control Society (Ferroelectrics section). His students have received
numerous student paper awards and other student research-based scholarships.
He has been a Purdue University Faculty Scholar and has also received ten
teaching awards including the 2010 HKN C. Holmes MacDonald Outstanding
Teaching Award and the 2010 Charles B. Murphy award, which is Purdue
University’s highest undergraduate teaching honor.