Networking across Boundaries: Enabling Wireless

Communication through the Water-Air Interface

Francesco Tonolini and Fadel Adib
MIT Media Lab
ABSTRACT Radar
RF Rx Acoustic Tx/Rx
We consider the problem of wireless communication across Rx
medium boundaries, specifically across the water-air inter-
face. In particular, we are interested in enabling a submerged
underwater sensor to directly communicate with an airborne
node. Today’s communication technologies cannot enable Sound vibrates
Radio Sound reflects surface
such a communication link. This is because no single type of frequency
decays
wireless signal can operate well across different media and
most wireless signals reflect back at media boundaries.
We present a new communication technology, translational
RF Tx Acoustic Acoustic
acoustic-RF communication (TARF). TARF enables under- Tx Tx
water nodes to directly communicate with airborne nodes by (a) Radio (b) Acoustic (c) TARF
transmitting standard acoustic signals. TARF exploits the fact
Figure 1—Enabling Communications across the Water-Air Boundary.
that underwater acoustic signals travel as pressure waves, and (a) shows that a radio transmitter cannot communicate because radio signals
that these waves cause displacements of the water surface die exponentially fast in water. (b) shows that acoustic signals reflect off the
when they impinge on the water-air boundary. To decode the water surface. (c) shows that a TARF receiver employs a radar to sense surface
transmitted signals, TARF leverages an airborne radar which vibrations caused by acoustic pressure waves and use them for decoding.
measures and decodes these surface displacements. KEYWORDS
We built a prototype of TARF that incorporates algorithms Wireless, Subsea Internet of Things, Cross-Medium Commu-
for dealing with the constraints of this new communication nications
modality. We evaluated TARF in controlled and uncontrolled
ACM Reference Format:
environments and demonstrated that it enables the first prac- Francesco Tonolini and Fadel Adib. 2018. Networking across Bound-
tical communication link across the water-air interface. Our aries: Enabling Wireless Communication through the Water-Air
results show that TARF can achieve standard underwater bi- Interface. In SIGCOMM ’18: ACM SIGCOMM 2018 Conference,
trates up to 400bps, and that it can operate correctly in the August 20–25, 2018, Budapest, Hungary. ACM, New York, NY,
presence of surface waves with amplitudes up to 16 cm peak- USA, 15 pages. https://doi.org/10.1145/3230543.3230580
to-peak, i.e., 100, 000× larger than the surface perturbations
caused by TARF’s underwater acoustic transmitter. 1 INTRODUCTION
Underwater communication networks all face the same prob-
CCS CONCEPTS lem: they cannot directly communicate across the water-air
• Networks → Cyber-physical networks; Mobile networks; interface. Said differently, a deeply submerged sensor cannot
Sensor networks; directly communicate with another node above the water’s
surface [12, 13, 32, 52]. This is because wireless signals ex-
Permission to make digital or hard copies of all or part of this work for hibit different properties in different media making it hard
personal or classroom use is granted without fee provided that copies are not
to use any single type of signal for cross-medium communi-
made or distributed for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. Copyrights for components cations [12]. In particular, while radio waves can travel over
of this work owned by others than the author(s) must be honored. Abstracting long distances in air, they die exponentially fast in water (see
with credit is permitted. To copy otherwise, or republish, to post on servers or Fig. 1(a)). Conversely, while acoustic waves can travel over
to redistribute to lists, requires prior specific permission and/or a fee. Request long distances underwater, they reflect off the water’s sur-
permissions from permissions@acm.org.
face and hence cannot carry information across the water-air
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary
boundary (as shown in Fig. 1(b)).
© 2018 Copyright held by the owner/author(s). Publication rights licensed to
the Association for Computing Machinery. Yet, enabling communication across the water-air boundary
ACM ISBN 978-1-4503-5567-4/18/08. . . $15.00 would bring benefits to numerous applications. In particular,
https://doi.org/10.1145/3230543.3230580 offshore oil and gas exploration, ocean biological sensing,
1
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

and subsea Internet of Things (IoT) all need to establish com- displacement becomes even more shallow when the node
munication links between underwater sensors and airborne is deeply submerged in the ocean.
nodes [8, 15, 32]. Today’s state-of-the-art networks rely on au- • More importantly, these perturbations are easily masked by
tonomous underwater vehicles (AUVs) that act as data mules. ocean waves that disturb the water surface and are three to
These AUVs need to dive into deep sea to collect data from six orders of magnitude larger than them.
underwater sensors and continuously resurface to transmit col- • Finally, the underwater acoustic transmitter has no mecha-
lected data before diving back in. This makes the exploration nism of estimating the overall channel. This makes it pro-
process time-consuming and costly [18, 30], particularly in hibitive to choose the right modulation and coding schemes
offshore oil exploration which require scanning vast areas of to match the wireless channel quality. In particular, while
the seabed and where searching for and establishing a single the above design enables uplink communication, it remains
deep-sea well can cost more than $100 million [14]. Cross- elusive for the airborne sensor to send channel feedback to
medium communication also presents security challenges, the underwater node since RF signals are not mechanical
which are particularly problematic in military applications. and hence will not vibrate the water’s surface and translate
For example, to communicate with an airborne drone, a deeply into acoustic waves.
submerged submarine needs to surface, compromising its lo- To overcome these challenges, TARF co-designs the
cation to an adversary [17, 23, 49]. transceiver architecture with the communication protocols. At
A common approach to work around this problem has a high level, it leverages a highly accurate RF-based sensor
been to deploy relays that are partially submerged in wa- that can measure minute reflections and introduces new algo-
ter [37, 40, 57, 59]. The relays collect information from un- rithms that can decode and eliminate unwanted interference.
derwater nodes using acoustic links and relay it to nodes We highlight the different system components below:
above the surface using radio signals. However, such relays • First, we employ a millimeter wave sensor as a receiver
can easily drift away with waves, severing communication to to capture and decode the RF reflections from the water
underwater sensors [59]. Moreover, this workaround leaves surface. Specifically, TARF’s airborne sensor transmits sig-
out submarines, which cannot rely on stationary deployed nals whose wavelength is few millimeters and measures
relays as they need to roam vast areas of the ocean. the phase of their reflection. Due to the small wavelength,
We present TARF, the first system that enables deeply sub- even surface displacements of few microns can lead to de-
merged underwater nodes to directly communicate across tectable phase changes of few degrees, allowing TARF to
the water-air boundary by leveraging standard acoustic links. sense and decode very minute surface vibrations. We fur-
TARF’s design exploits the fundamental physical properties ther incorporate the millimeter wave sensor into an FMCW
of acoustic waves, as demonstrated in Fig. 1(c). In particular, (Frequency-Modulated Carrier Wave) radar, which allows
an acoustic signal emitted by a sound transducer travels as a it to focus its beam on the water surface, and mitigate noise
pressure wave. When the pressure wave hits the water surface, and interference from undesired reflections.
it causes a perturbation or displacement of the surface due to • Second, we realize that ocean waves can be treated as struc-
its mechanical nature. To pick up these signals, TARF relies tured interference in our context due to their mechanical
on an airborne Radio Frequency (RF) sensor. The sensor trans- nature, and design filters that can mitigate their impact on
mits an RF signal and measures its reflection off the water the received signal.
surface. These reflections vary due the surface displacement • Third, we discover unique properties of this new communi-
caused by the impinging acoustic signals from an underwater cation modality, which arise from the translation between
transmitter. TARF’s receiver analyzes the variations in RF pressure and displacement at the water-air interface. For
reflections and uses them to decode the bits communicated by example, we show that the channel’s frequency-selective
an underwater node. We call this phenomenon translational fading is inversely proportional to the transmit acoustic
acoustic-RF (TARF) communication, as it enables communi- frequency. Our transmitter and receiver take these prop-
cation by leveraging a translation between acoustic signals erties into account to design a power- and rate-optimal
and the RF reflections. modulation scheme across the operational bandwidth.
This new communication modality presents unique con- • Finally, to select an appropriate bitrate, TARF can incor-
straints due to the entanglement of both electromagnetic and porate a pressure sensor as a proxy for the channel. The
mechanical nature of the resulting links, as well as unique pressure sensor can be used to infer the distance to the sur-
environmental challenges. As a result, translating this com- face and estimate the dominant pathloss component. This
munication paradigm into a practical networked system still would allow a TARF transmitter to perform rate adaptation
faces multiple challenges: despite the lack of receiver feedback.
• First, the surface vibrations caused by acoustic waves are We built a prototype of TARF using underwater speak-
very minute – of the order of few to tens of microns. The ers and custom-made millimeter wave radars. We tested our
2
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

prototype in synthesized water tanks and a swimming pool development of SONAR systems, which leverage sound and
(in the presence of practicing swimmers). Our experimental ultrasonic signals for submarine communications and for de-
evaluation demonstrates the following results: tecting icebergs and U-boats [28, 35]. The appeal of acoustic
• Our prototype achieves cross-medium throughput of hun- communication arises from their low attenuation in water in
dreds of bits per second in scenarios where existing com- comparison to RF signals. However, none of the early systems
munication technologies cannot establish any link. could communicate across the water-air boundary [35].
• TARF can decode the transmitted packets even in the pres- Interest in underwater communication and sensing resurged
ence of waves by up to 8 cm of height (16 cm peak-to-peak), during the Cold War [25, 27]. The US and Soviet navies devel-
i.e., 100, 000× larger than the (µm) displacement caused by oped ELF (extremely low frequency) communication systems
the transmitted acoustic signals. which operate at 30-300 Hz and are capable of communicat-
• We empirically evaluate the communication link with dif- ing across the air-water boundary [9, 41]. The key challenge
ferent modulation schemes (BPSK, QPSK, 16-QAM, etc.), with these systems is that, due to their very long wavelengths,
and we demonstrate that TARF’s channel-aware rate and they require kilometer-long antennas, which make them in-
power allocation algorithm can consistently outperform feasible to incorporate into underwater vehicles [41, 56]. As a
flat modulation schemes. Moreover at low SNRs, TARF’s result, most of the deployment of these systems remained lim-
adaptation scheme can improve the throughput up to 10× ited to restricted point-to-point anchors deployed in specific
compared to flat modulation schemes. locations [9, 41].
While these results are promising, we believe they only Over the past two decades, there’s been mounting interest
represent a first demonstration of TARF’s capability as a in underwater networking for ocean exploration as well as oil
cross-medium communication technology, and our design and gas mining [14, 42, 54]. To overcome the water-air barrier,
still exhibits multiple limitations. First, because our system these systems rely on nodes that incorporate two communica-
cannot sustain a communication link in the presence of waves tion modules: acoustic and RF [40, 41]. To send information
with amplitudes larger than 16 cm, it cannot operate under across the air-water boundary, these nodes dive deep into
all weather conditions. In particular, it is resilient to capillary the water to communicate with underwater sensors, typically
waves – which consist the dominant ocean surface wave on deployed on the sea bed, collecting information from them
calm days – but not to wind waves. Another key limitation using acoustic signals and re-surfacing frequently to relay this
arises from the need to have the transmitter and the receiver information using RF signals for in-air communication, be-
relatively aligned along a vertical axis, since the throughput fore diving again to collect more data [36, 37, 48]. Significant
decays rapidly when they are misaligned (as we quantify research in the robotics community has focused on how to
in §8). Despite these limitations, we hope that this work can perform this process efficiently with robotic swarms or how
motivate researchers to explore and develop TARF to enable to place partially-submerged relay nodes to optimize cover-
truly ubiquitous cross-medium communication, and allow age [42, 48, 57]. Similarly, the military has deployed such
underwater computing devices to seamlessly communicate relay nodes in permanent points of interest in the ocean [34].
with the outside world. However, these systems still suffer from the ability to scale,
and are not feasible for submarines as surfacing would com-
Contributions. TARF is the first communication technology
promise their location. In contrast, TARF does not suffer from
that enables a deeply submerged underwater node to directly
these problems as it enables submerged nodes to directly com-
communicate with a compact airborne node. We present the
municate through the water-air interface.
design, prototype implementation, and evaluation of this tech-
Finally, recent research has explored other means of under-
nology demonstrating that it can achieve standard underwater
water communication, including optics [31, 55] and quantum
data rates in scenarios where past technologies cannot estab-
entanglement [26]. In contrast to TARF, the former has the
lish any communication throughput.
same drawbacks of RF waves in its limited range [31, 55] and
the latter is theoretical or still in the proof-of-concept phase.
2 RELATED WORK
TARF builds on past literature in two main areas: underwater Wireless Sensing. Over the past few years, the networking
communication networks and wireless sensing, as we detail community has taken much interest in using communication
below. In contrast to past work in these areas, TARF intro- signals for sensing purposes, e.g., sensing human locations,
duces the first system that leverages sensing as a means for gestures, and vital signs [6, 7, 39]. Similarly, the radar com-
communication across the water-air boundary. munity has explored wireless for sensing coarse water surface
Underwater Communication. The sinking of the Titanic levels and surface currents [16]. TARF is inspired by these
in 1912 and the start of World War I spurred interest in un- recent advances but differs in its goals, technique, and ca-
derwater communication and sensing [35]. This led to the pabilities. Specifically, in contrast to past work on sensing,
3
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

mmWave • Deep-sea Exploration: Deployed underwater sensors could

FMCW
Radar perform continuous monitoring and leverage TARF to send
their collected information to the outside world. A drone
Phase may fly over large areas and collect information from a
Surface Modulation
Motion network of deployed underwater nodes.
• Submarine Communication: Submarines could leverage
Pressure
Wave
TARF to communicate with airplanes without the need for
surfacing or compromising their locations.
• Search and Recovery: Finally, uplink communication can
contribute to solving the long-standing problem of find-
Acoustic
Uplink ing vehicles that go missing underwater (e.g., missing air-
Figure 2—Surface Vibrations Translate into Phase Modulation. The planes). In particular, TARF would enable these vehicles to
phase of the wireless reflection changes with minute surface vibrations. continuously send distress signals to the surface, which can
be picked up from the air, enabling rapid airborne search
TARF introduces a new technique that leverages sensing for for lost or malfunctioning vehicles.
communication, particularly to enable communication across In what follows, we first explore the unique properties of
the water-air boundary. In terms of capabilities, due to its TARF’s wireless channel in §4, then describe our design of
wavelength of operation, TARF can extract displacements TARF transmitter and receiver in §5 and §6 respectively.
of the order of few microns, i.e., at a scale three orders of
magnitude finer than the millimeter-scale movements of past 4 UNDERSTANDING THE TARF
work [7, 10]. And finally, TARF builds on its basic idea of COMMUNICATION CHANNEL
acoustic-RF translational communication to develop a full
system that can address practical constraints including ocean We start by analyzing TARF’s communication channel. The
waves and coupled RF-acoustic channels. channel consists of three components: underwater propaga-
tion, the water-air interface, and in-air propagation. Since the
underwater and in-air propagation components follow stan-
3 TARF OVERVIEW
dard communication channels [33, 50], we focus our discus-
TARF is a new communication technology that allows sub- sion on the water-air interface then incorporate our analysis
merged underwater nodes to wirelessly communicate directly into the end-to-end channel.
with nodes over the water’s surface. The communication link
naturally consists of three components shown in Fig. 2: 4.1 The Water-Air Interface
• Transmitter: A TARF underwater node sends packets using Recall that a TARF underwater transmitter sends packets
a standard acoustic transducer (e.g., underwater speaker). using acoustic signals. These signals travel in the medium as
The transmitter leverages signals in the 100-200 Hz fre- pressure waves P(r , t), which vary in time t and range r , and
quency range, which are typically used for underwater can be expressed as [33]:
communications by submarines and AUVs due to their low P(ω, t) = A(ω)e jω(t −r /vw ) (1)
attenuation and long travel distances in water [44, 44, 45].
where A is the amplitude, ω is the angular frequency, and vw
• Channel: The acoustic signal travels as a pressure wave
is the velocity in water. Note that the amplitude A is also a
inside the water. When the pressure wave hits the water
function of distance r , but we omit it for simplicity.
surface, it causes a surface displacement that is proportional
Below, we first quantify the amount of surface displacement
to the pressure wave.
caused by these pressure waves, then describe how TARF can
• Receiver: TARF’s receiver consists of a millimeter-wave
measure these displacements.
FMCW (Frequency Modulated Carrier Wave) radar. The
radar transmits a wideband signal (centered around 4.1.1 How much surface displacement do acoustic
60 GHz) and measures its reflection off the water’s surface. pressure waves create?
As the water surface vibrates due to the acoustic pressure Acoustic pressure waves are longitudinal waves. As they prop-
waves, these vibrations modulate the phase of the reflected agate in a medium, they displace the medium’s particles along
signal. The radar receiver extracts these phase changes and their same direction of travel. (Such particle displacement is
decodes them in order to recover the transmitted packets. similar to how particles of a spring move as it compresses and
Scope. TARF focuses on the problem of uplink wireless com- relaxes due to a pressure wave traveling through it.) Hence,
munication between underwater and airborne nodes. Enabling when a pressure wave hits the surface of water, it also causes
such communication opens up capabilities in several areas: a surface displacement δ . This displacement can be computed
4
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

20 70 Measured Amplitude

Amplitude (μm)
20
Amplitude (μm)

1/ω fit

SNR (dB/Hz)
0 50
10
-20
0 -40 30
-10 -60
10
-20
-80
0.68 0.7 0.72 0.74 110 130 150 170 190 100 200 300 400
Time (s) Frequency (Hz) Frequency(Hz)
(a) Displacement vs Time (b) Two Frequencies Transmitted (c) Displacement vs Freq
Figure 3—Understanding the Surface Displacement as a Function of the Acoustic Pressure Wave. (a) shows the displacement over time when a single
frequency is transmitted, at a frequency of 120H z and at a frequency of 180H z. (b) shows the absolute value of the fourier transform of the power amplitude
when the same two frequencies are transmitted simultaneously. (c) shows the amplitude of the displacement as a function of the frequency of the acoustic signal.

by solving the boundary conditions of the wave equation. In are amenable to different modulation schemes (AM, FM,
the interest of brevity, we include the solution below and refer BPSK, OFDM, etc.) and can be estimated with preamble
the interested reader to [51] for a derivation. Assuming the symbols and inverted for reconstruction and decoding.
incident wave is orthogonal to the surface, we can derive: • The amplitude of the displacement is inversely proportional
P(ω, t) to the frequency of the transmitted acoustic signal: This can
δ (ω, t) = (2)
ρw ωvw be seen through the 1/ω decay in Fig. 3(c), which matches
where P is the overall pressure created by the acoustic wave the expected behavior in Eq. 2. This property implies that
and ρw is the density of water. lower frequencies are more desirable for TARF communi-
To better understand this relationship, we perform experi- cation as they will cause a larger displacement, and hence
ments with an underwater speaker. We use the Electro-Voice a larger signal-to-noise ratio (SNR). It also implies that
Underwater Speaker [1], place it about half a meter below the signals at different frequencies experience very different
surface of water, and point it upward toward the surface in a attenuation and that an optimal communication protocol
setup similar to that shown in Fig. 2. The speaker transmits should account for this unique feature of the channel.
an acoustic signal, and we measure the displacement at the
surface of the water.1 4.1.2 Why can’t we rely on acoustic signals alone?
We perform three types of experiments. First, we transmit a Since the acoustic wave hits the surface and causes a displace-
single tone from the speaker, first at a lower and then a higher ment, the displacement itself can generate a pressure wave
frequency, and plot the measured displacement in Fig. 3(a). that travels in air. Hence, we ask whether it would be more
Next, we transmit two tones simultaneously from the speaker efficient to directly leverage the generated pressure wave in
and plot the fourier transform of the resulting displacement in the air for communication.
Fig. 3(b). And finally, we run an experiment where we vary There are multiple reasons why such an approach is unde-
the frequency of the transmitted tone over time and plot the sirable. First, while part of the pressure wave indeed crosses
peak-to-peak displacement in Fig. 3(c). the boundary and travels in air, the majority of the incident
Based on these figures, we observe the following: pressure wave reflects off the water-air interface. In particular,
• The displacement caused by the pressure wave is very by solving the sound wave equation for a wave incident at a
minute: Fig. 3(a) shows that the peak-to-peak displacement boundary between two different media, we obtain the follow-
is of the order of a few µm to a few tens of µm, even though ing relationship between the amplitude of the reflected wave
the underwater transmitter was only submerged half a meter Ar and the amplitude of the incident one Ai [33]:
v a ρ a − vw ρ w
below the water’s surface. Ar = Ai . (3)
• The water-air interface acts as a linear channel in the con- v a ρ a + vw ρ w
text of TARF communication: In particular, the frequency where va and vw are the speeds of sound in air and water
of the surface displacement matches the frequency of the respectively and ρ a and ρw are the air and water densities.
transmitted acoustic signals by the underwater speaker in Due to the large difference between the constants for air and
Fig. 3(a)-(b). Such behavior is in line with Eq. 2, which water, the reflected amplitude is almost equal to the incident
shows that the displacement is directly proportional to the one (i.e., Ar ≃ Ai ). And, by the law of conservation
q of energy,
pressure wave. This means that the water-air interface acts the amplitude of the transmitted signal At = A2i − A2r . Using
as a linear (and time-invariant) channel. Such channels standard values for velocity and density [60], we can show
1 Notethat for measuring the displacement, we use the millimeter-wave radar that pressure waves crossing into air attenuate by around 30dB
we built as described in §7. solely because of reflection at the boundary.
5
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

Second, aside from the attenuation at the boundary, acous- Given the above breakdown, the overall pathloss (PL) in dB
tic waves experience exponential attenuation when traveling is linear in depth r and logarithmic in height d 0 , density ρw ,
in air [33]. This makes them an unsuitable means for wireless frequency ω, and velocity vw . Since ρw and vw are known,3
communication over the air. Indeed, this is why wireless com- estimating the overall attenuation requires estimating only
munication systems like WiFi and cellular employ RF signals r and d 0 . Further, since the path loss increases linearly in r
instead of ultrasonic/acoustic signals. but logarithmically in d 0 , the dominant unknown path loss
component is expected to be r . In §5, we explain how TARF
4.1.3 Why can’t we leverage the water-air interface for
can estimate this component.
downlink communication?
So far, our discussion has focused on uplink communication. 5 DESIGNING A TARF TRANSMITTER
A natural question is: why can’t we use the same technique to In this section, we describe how TARF’s acoustic transducer
enable an in-air node to communicate with an underwater hy- encodes and modulates its transmissions by taking into ac-
drophone. In principle, an acoustic signal transmitted from an count the properties of the TARF communication channel.
airborne speaker should also cause a vibration of the water-air
interface that can be picked up by an underwater hydrophone. 5.1 What is the right modulation scheme?
The answer lies in the nature of interference between the
incident and reflected pressure waves at the water-air bound- Recall that TARF’s channel is amenable to various modula-
ary. Specifically, these waves constructively interfere when tion schemes since it is linear and time-invariant. The chan-
they hit the boundary of a less dense medium (i.e., when nel, however, is highly frequency selective, as can be seen
traveling from water to air), but destructively interfere when in Fig. 3(c). Such frequency-selective fading leads to inter-
they hit the boundary of a more dense medium (i.e., when symbol interference, which complicates the receiver design.
traveling from air to water). Since the displacement is directly To deal with such frequency-selective fading, TARF em-
proportional to the overall pressure as per Eq. 2, the displace- ploys Orthogonal Frequency Division Multiplexing (OFDM)
ment is maximized for underwater pressure waves, but it is as an encoding scheme at its transmitter. OFDM is widely
nulled for acoustic signals arriving from the air. Hence, while used in WiFi and LTE systems. In what follows, we briefly
this mechanism enables underwater-to-air communication, it describe how OFDM works and refer the interested reader
cannot enable an air-to-underwater communication link. to [60] for more information.
Instead of encoding the transmitted bits directly in the time
4.2 End-to-end TARF Channel domain, an OFDM transmitter encodes symbols in the fre-
Now that we understand the water-air interface, we would quency domain. For example, if we consider each frequency
like to quantify the impact of each of the channel components in Fig. 3(c) as a subcarrier, an OFDM transmitter can treat
on the overall signal attenuation: each frequency as an independent channel and transmit flows
• Underwater Propagation. The attenuation of acoustic sig- on all of them concurrently. The OFDM encoding scheme
nals traveling underwater can be described by e −γ r /r where is attractive because decoding can be done in the frequency
r is the depth and γ quantifies the absorption. This equation domain without the need for complex channel equalizers.
shows that the amplitude of the acoustic pressure wave
decays exponentially as it travels underwater. 5.2 What is the optimal power allocation?
• Water-Air Interface. The attenuation at the water-air inter- Next, we ask how should a TARF transmitter divide its power
face is given by Eq. 2 in terms of pressure. Assuming that across the different subcarriers? According to Fig. 3(c), a
the received power is proportional to δ (ω, t)2 , and knowing TARF channel has high SNR at lower frequencies and lower
that the transmitted power is proportional to P(ω, t)2 and SNR at higher frequencies. With this knowledge, it is clear
inversely proportional to ρw and vw [11], we can express that distributing the power evenly across the different subcar-
the sensed power at the water-air interface as: riers would result in sub-optimal performance. Conversely, a
Pincident
Psensed ∝ (4) power allocation strategy that concentrates all the available
ρ w vw ω 2 power into the lowest-frequency subcarrier would maximize
• In-Air Propagation. A standard radar signals attenuates the SNR, but also result in sub-optimal performance since it
as 1/d 02 , where d 0 is the distance between the transmitter forgoes much of the available bandwidth.
and the receiver [47].2 However, because water is specular Optimal power allocation is a well-studied problem in in-
at the wavelengths of RF signals (i.e., it reflects back all formation theory [50]. The generic solution for this problem
the impinging RF signals) [19], we can approximate the
overall signal attenuation as 1/(2d 0 ). 3 Note that these parameters depend on the water salinity and temperature,
2 Power decays as 1/d 04 , but the signal amplitude attenuates as 1/d 02 . which we assume the underwater sensor can directly estimate or infer.
6
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

C C(ω) power distribution with respect to the noise frequency profile

discussed above. The TARF transmitter uses this information
μ to assign power to its subcarriers according to this computed
distribution at the center frequency of each subcarrier.
P(ω)
5.3 How to modulate the subcarriers?
Recall that in OFDM-based systems, we can treat each subcar-
rier as a separate flow with its own modulation (BPSK, QPSK,
ωmin ωmax ω etc.). After TARF determines the optimal power allocation, it
Figure 4—TARF’s Waterfilling. The noise C(ω) increases with frequency proceeds to bitrate selection on a per-subcarrier basis.
ω. The level µ determines the optimal power allocation (shaded region), Specifically, knowing the power allocation P(ω) and the
where P (ω) denoting the power at every frequency. noise function C(ω), TARF can estimate the expected SNR
is called waterfilling. In what follows, we describe this con- at the receiver and choose the appropriate bitrate based on
cept in the context of a TARF communication channel and its estimate. In particular, it can leverage higher modulations
highlight why it is particularly interesting in this context. (e.g., 64-QAM) at lower-frequency subcarriers (which have
Fig. 4 plots the noise power C(ω) in blue as a function of higher SNRs) and lower modulation schemes (e.g., BPSK) at
frequency ω. As per Eq. 4, we can express C(ω) = ρw vw ω 2 /a, higher-frequency subcarriers (which have lower SNRs).
where a is a real positive constant which depends on the We note few more points about TARF’s bitrate selection:
transmitted signal power, the distance attenuation, and the • The exact SNR at which TARF should switch between the
receiver noise floor. The high level idea of waterfilling is different modulation schemes can be determined both ana-
that we can solve for a water level µ, depicted by the yellow lytically and empirically. In §8, we describe how TARF’s
line in Fig. 4. Specifically, the optimal power allocation is empirical evaluation matches the analytical solution.
the difference between µ and the noise power C(ω). We can • Our discussion above focused on performing rate adapta-
express the optimal power allocation as: tion by only changing the modulation scheme. In practice,
( the discussion can be extended to adapting the coding rate
µ − C(ω), if µ − C(ω) ≥ 0
P(ω) = (5) (e.g., 1/2-rate or 3/4-rate coding) as well [38].
0, otherwise • Finally, in order for a receiver to decode transmitted pack-
So how can we find µ? To solve for µ, we use the total ets, it needs to know the modulation scheme employed by
power constraint, which states that the total power across all every subcarrier. Such information is typically embedded
the subcarriers (i.e., the integral of the power densities) must in the packet header which is sent via BPSK modulation.
equal the total power of the transmitter P 0 .
∫ ∞ 5.4 How to adapt the bitrate?
P(ω)dω = P0 (6) So far, our discussion has assumed that TARF’s transmitter
ωmin
has perfect knowledge of the noise function C(ω). Unfortu-
In our context, ωmin is the lowest frequency at which the nately, however, TARF does not have direct access to channel
underwater speaker or acoustic transducer can operate. information. This is because TARF can only perform one-
In general, because of the non-linear nature of Eq. 5, the way communication; hence, the receiver is unable to send the
water filling problem is solved numerically. However, in the channel estimates as feedback to the transmitter. To accommo-
context of a TARF channel, the function P(ω) is continu- date for channel uncertainty and frequency-selective fading,
ously decreasing, meaning that the above integral can be one-way communication systems are typically conservative:
computed without the non-linearity over the interval in which They choose modulation schemes with very low bitrate and
it is positive. Such interval spans from ωmin to the frequency large redundancy. For example, a GPS transmitter spreads
at which the power density P(ω) is equal to zero, ωmax as every bit over 1024 chips and repeats each symbol 20 times.
shown in Fig. q 4. Setting Eq. 5 to zero and solving for ω we To overcome this challenge, a TARF transmitter can lever-
aµ
get ωmax = ρw vw . Using this maximum frequency and the age known properties of the channel and combine them with
total power constraint of Eq. 6, we obtain the following third side-channel information. In particular, recall from §4.2 that
√ the only unknown components of the attenuation are the
degree polynomial in µ:
3 height above the water d 0 and the depth of the TARF transmit-
ρw vw ωmin
r
2 a 3
µ 2 − ωmin µ + − P0 = 0. (7) ter r . Hence, if TARF can estimate these components, then it
3 ρ w vw 3a would be able accurately estimate the overall SNR.
The real positive root of this polynomial gives the level µ To estimate the depth underwater, a TARF transmitter can
which allows us to obtain an analytical form for the optimal employ a pressure sensor. In particular, underwater pressure
7
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

Algorithm 5.1 Transmitting through a TARF Channel 32

POWER ALLOCATION 39
▷ Path Loss Estimation

Distance (cm)
46 Reflection from
Estimate depth; r ← p/ρ w g
Estimate path-loss PL(ω) from §4.2 53 Water Surface
▷ Power Distribution 60
Solve for level µ from Eq. 7
67
Compute power allocation: P(ω) ← (µ − C(ω))+
1.6 3.2 4.8 6.4 8.0 9.6 11.2 12.8 14.4 16.0
MODULATION Time (s)
▷ SNR Estimation Figure 5—Capturing the Surface Reflection. The FMCW spectrogram
Estimate SNR per subcarrier: SNR(ω) ← P(ω) × 10PL(ω)/10 plots the power at each distance bin over time. The yellow line indicates the
▷ Modulation high power reflection arriving from the water surface.
if SN R(ω) <= SN R 1 minute surface displacements, then we discuss how it cancels
Mod(ω) ← BPSK interference caused by the ocean waves, and finally how it
elseif SN R 1 < SN R(ω) <= SN R 2 can decode the filtered reflection.
Mod(ω) ← QPSK
elseif SN R 2 < SN R(ω) <= SN R 3 6.1 How can TARF capture the minute
Mod(ω) ← 16QAM surface displacements?
else Recall that TARF’s receiver employs a radar to capture the
Mod(ω) ← 64QAM surface vibrations caused by the acoustic pressure waves. The
TRANSMISSION radar transmits an RF signal and measures its reflection off
▷ Add preamble, cyclic prefix, CRC the water surface. Given the very minute (µm-scale) displace-
▷ Transmit ment at the surface of the water, leveraging time-of-flight
based techniques to directly measure the displacement would
can be directly mapped to depth (through P = ρv дr , where ρv require few T Hz of bandwidth (since bandwidth is inversely
is the density and g is the gravitational field strength). In fact, proportional to the resolution).4
today’s off-the-shelf pressure sensors have millimeter-level Instead of trying to directly estimate the distance, TARF
precision in measuring underwater depth [53]. measures the change in distance by estimating the phase of
This leaves TARF only with the height of the receiver as the reflected signal. In particular, the phase of the reflected
an unknown. In practical scenarios, the transmitter may have radar signal ϕ(t) can be expressed as:
prior knowledge of the receiver’s height. For example, under- d 0 + δ (t)
ϕ(t) = 4π (8)
water submarines trying to communicate with airplanes can λ
have reasonable estimates on the altitude at which airplanes where d 0 is the distance between the radar and the water
fly based on standard flight patterns. Alternatively, the plane surface (in the absence of vibrations) and λ is the wavelength
may decrease its altitude to improve its SNR to an underwater of the radar’s transmitted signal.
submarine communicating with it via TARF. In the case of The above equation reveals three important observations:
subsea IoT, the expected height can be provided to a sensor • First, TARF’s ability to track the surface displacement is
prior to deployment. We summarize the overall procedure of strongly impacted by its choice of the wavelength λ. On
a TARF transmitter in Alg. 5.1. one hand, a relatively large wavelength (e.g., few centime-
Finally, one might wonder whether TARF’s transmitter ters as in WiFi or cellular) would result in very minute
could employ rateless codes instead of its bitrate adap- variations in the phase, making it less robust to noise. On
tation scheme. Unfortunately, rateless codes still require the other hand, choosing a very small wavelength (e.g.,
feedback from the transmitter (in the form of acknowledg- sub-µm as in THz or optical frequencies) would result in
ments) [20, 21], which is still not possible given the uplink- rapid phase wrapping, precluding the ability to track the
only constraint on a TARF communication link. In contrast, surface vibrations.
TARF’s transmitter can adapt its bitrate by exploiting side • Second, the choice of wavelength λ also impacts TARF’s
channel information despite this constraint. ability to adapt to ocean waves in the environment. In
particular a very small wavelength will suffer from rapid
6 DESIGNING A TARF RECEIVER phase rotation even in the presence of very small waves.
In this section, we describe how we design a TARF receiver.
We start by describing how the receiver can measure the 4 The resolution is c/2B where c is the speed of light and B is the bandwidth.
8
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

4 Wrapping 20

Phase (milliradians)
20
Phase (radians)

Phase (radians)
2 0
10
2cm
0 -20 Waves 0
-2 -40 -10
-4 -60 -20
1 1.2 1.4 1.6 1.8 2 1.2 1.4 1.6 1.8 2 1.1 1.3 1.5 1.7 1.9
Time(s) Time (s) Time (s)
(a) Raw recorded phase (b) Unwrapped phase (c) Filtered phase
Figure 6—Phase Extracted by the FMCW Receiver. (a) phase as extracted at the distance bin of interest, (b) phase after unwrapping and (c) phase after
applying a band pass filter to isolate the (communication) frequencies of interest.

• Third, because the phase of a reflection is not robust to largest radar cross section, and hence the highest reflection
interference, TARF requires a more sophisticated sensing power. In Fig. 5, this corresponds to the solid yellow line.
technology than a simple Doppler or phase-based radar. 6.2.2 Phase Extraction and Wave Elimination
To address these issues, the TARF receiver leverages Next, TARF zooms in on the phase of the range bin where it
a millimeter-wave Frequency Modulated Carrier Wave has identified the water reflection. Fig. 6(a) plots the phase of
(FMCW) radar. In the rest of this section, we describe how the that bin as a function of time. Note that the phase in this figure
receiver employs the radar and highlight why millimeter wave wraps around every 0.2 s. This indicates a phase displacement
frequencies offer a sweet spot for the operation wavelength. larger than 5 mm (i.e., the wavelength of our millimeter wave
radar). This phase wrapping arises from the waves at the
6.2 How does FMCW extract the information surface of the water, whose presence masks the µm-scale
of interest? vibrations from the acoustic transmitter.
To eliminate the impact of these waves, TARF first unwraps
In order to achieve high phase resolution while mitigating
the phase. We plot the output of the unwrapped phase over
interference from other reflectors in the environment, TARF
time in Fig. 6(b). The waves exhibit a peak-to-peak variation
leverages an FMCW-based wideband radar. At a high level,
of 50 radians. Given a wavelength of 5 mm, this corresponds
the wideband radar can filter the reflections coming from
to a 2cm peak-to-peak displacement, as per Eq. 8.
different distances into different bins. This enables it to isolate
Next, to eliminate the impact of the waves, TARF filters
the reflection off the water’s surface from other reflections
the unwrapped phase and plots the output in Fig. 6(c). Note
in the environment, and zoom in on its phase in order to
that in order to visualize the phase variations, the axis of this
decode the surface vibrations. In what follows, we describe
figure is zoomed in both in time and amplitude. Upon filtering
the operation of the receiver in three main steps: surface
the ocean waves, we can now see the single-tone transmitted
reflection identification, phase extraction, and decoding.
by TARF’s underwater speaker at 150 Hz. Note that TARF
6.2.1 Surface Reflection Identification can always filter out ocean waves since their frequency is
To explain the operation of TARF’s receiver, we run an ex- significantly lower than its range of operation. Specifically,
periment with the radar placed above the water’s surface in a ocean waves typically range between 0.1Hz − 3Hz [43] while
manner similar to Fig. 2 such that it can capture the reflection TARF’s transmitter operates above 100 Hz.
off the water surface. We configure TARF’s underwater acous- The above description demonstrates why using millimeter-
tic transmitter to transmit a single tone at 100Hz. The radar wave frequencies offers a sweet spot for TARF communica-
transmits a signal and measures its reflections. It can then tion. Specifically, they enable a TARF receiver to overcome
process these reflections to obtain the power of the reflections (unwrap and filter) the impact of ocean waves while at the
as a function of distance. (For a thorough explanation of how same time sensing surface displacements (of the order of few
it performs this processing, we refer the reader to [6].) µm) due to underwater acoustic pressure waves.
Fig. 5 plots the output of TARF’s FMCW processing as 6.2.3 Decoding
heatmap, where navy blue indicates low reflection power Our above experiment was conducted by configuring the un-
and yellow indicates high reflection power. The x-axis shows derwater speaker to transmit a single frequency. In practice,
time, while the y-axis indicates the distance. A horizontal line however, a TARF transmitter sends OFDM symbols over its
indicates a reflection arriving from a particular location. Note bandwidth of operation as described in §5. To decode these
that the different light blue patterns over time are due to noise. symbols, TARF’s receiver performs standard OFDM packet
To identify the reflection bin corresponding to the water detection, extracts the channel and the modulations from the
surface, TARF exploits the fact that the water surface has the header, and uses them to decode the packet payload.
9
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

7 IMPLEMENTATION & EVALUATION Millimeter-wave

FMCW radar
7.1 Implementation Millimeter-wave
Our prototype implementation of TARF consists of an un- 30cm FMCW radar
derwater acoustic speaker as a transmitter and an airborne
millimeter wave FMCW radar as a receiver.
(a) Acoustic Uplink. We implemented TARF’s uplink trans- 3.6 m
70cm
mitter using an underwater speaker, namely the Electro-Voice
UW30 Underwater Loudspeaker [1]. The speaker was con-
nected to the output audio jack of a Lenovo Thinkstation PC Acoustic Tx
Acoustic Tx
through a power amplifier. In our evaluation, we used two
types of amplifiers: the OSD 75W Compact Subwoofer Am- (a) Tank Setup (b) Pool Setup
plifier [2] and the Pyle 300W Stereo Receiver [3]. TARF’s Figure 7—Experimental Setups. (a) shows the tank setup, with the acous-
tic transmitter placed 70 cm below the surface. (b) shows the pool setup,
transmit power levels are comparable to standard low power where we experimented with different depths. In both setups, the millimeter-
acoustic transducers used in underwater communications [46]. wave radar was placed around 30 cm above the water surface and pointed
We configure the speaker to transmit signals over a bandwidth downwards to record the acoustic vibrations.
of 100Hz between 100Hz and 200Hz. Such bandwidth is typ-
ical for underwater communication systems [60].
TARF’s transmitter encodes its data using OFDM modula- an ethernet cable to a 64-bit machine running Ubuntu 16.04
tion. Each OFDM symbol consists of 64 subcarriers which for post-processing.
cover the available bandwidth. The transmitter performs per- We implemented TARF’s decoder in MATLAB. The de-
subcarrier power allocation and bit-rate adaptation as de- coder identifies the range bin corresponding to the water
scribed in §5. Each OFDM symbol is pre-pended with a surface as described in §6, then extracts the phase of the re-
cyclic prefix, as in prior proposals that perform per-subcarrier flection and performs unwrapping and filtering. To decode the
bitrate adaptation [38]. filtered phase signal, it performs packet detection, channel es-
Unless otherwise noted, in each experimental trial, we timation, and decoding similar to a standard OFDM decoder.
transmit 10 back-to-back OFDM symbols (two symbols act
as a preamble and 8 as payload). The transmitter can include 7.2 Evaluation
the modulation scheme for every subcarrier in its header, and
We evaluated TARF in controlled and uncontrolled settings.
a CRC for every subcarrier to determine whether the packet
We tested it in two environments: an enclosed water tank and
was received correctly.
a swimming pool during normal activity.
(b) Millimeter-Wave FMCW Radar. We implemented (a) Water Tank. Most of our evaluation was performed in
TARF’s receiver as a custom-built millimeter-wave FMCW a water tank of 1.3m depth and 3m × 5m rectangular cross
radar. To generate the desired millimeter-wave signals, we section, as shown in Fig. 7(a). In these experiments, we varied
first generate a reference FMCW signal using a design similar the height of the radar between 20cm and 40cm above the
to that implemented in [6]. The reference outputs a frequency water surface, and varied the depth of the speaker between
ramp with a center frequency of 8.65GHz and a bandwidth of 5cm and 70cm. We also experimented with different locations
500MHz. We feed the output of this FMCW signal generator across the tank and with different acoustic transmission levels.
into a 2× frequency multiplier, whose output is in turn fed as a To evaluate the robustness of TARF to waves, we manually
local oscillator to an off-the-shelf millimeter wave transmitter generated the waves with a floating object and measured
and receiver. This architecture enables transmitting and receiv- their peak-to-peak amplitude with a graded ruler at the water
ing an FMCW signal with a center frequency of 60GHz and a surface directly above the speaker.
bandwidth of 3GHz. This results in an effective range reso- (b) Swimming Pool. To evaluate TARF in a less controlled
lution of 5 cm, and a phase sensitivity of 1.25rad/mm. The environment and at greater depths, we performed experiments
millimeter wave transmit and receive boards are connected to in a swimming pool during normal activity. Fig. 7(b) shows
23 dBi horn antennas [4]. our experimental setup. In our experiments, we placed the
We programmed our FMCW generator to sweep its band- radar 30cm above the water surface and submerged the acous-
width every 80µs. The receiver captures and downconverts tic transmitter at depths between 90cm and 3.6m. During our
the reflected signals to baseband and feeds them into a USRP experiments, the water surface was continuously disturbed
N210 software radio [5] equipped with an LFRX daughter- by swimmers and circulation vents, enabling us to evaluate
board. The USRP digitizes the signals and sends them over TARF’s robustness to environmental challenges.
10
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

100 Empirical BPSK 8

Empirical QPSK Empirical BPSK
Empirical 16QAM Empirical QPSK

Throughput(bps)
10-1 Theoretical BPSK 6 Empirical 16QAM
Theoretical QPSK
Theoretical 16QAM Theoretical BPSK
BER

Theoretical QPSK
10-2 4 Theoretical 16QAM

10-3 2
-10 -5 0 5 10 15 20 25
SNR (dB)
0
Figure 8—Per-subcarrier BER vs SNR. The figure plots the per-subcarrier -40 -20 0 20 40
BER against SNR for BPSK, QPSK and 16QAM modulations. The dashed SNR(dB)
lines show the theoretically predicted behavior for each modulation. Figure 9—Per-subcarrier Throughput vs SNR. The figure plots the per-
subcarrier throughput against SNR for BPSK, QPSK and 16QAM. The
dashed lines show the theoretically predicted behavior for each modulation.
8 RESULTS
8.1 Performance (b) Per-subcarrier Throughput-SNR curves. Recall that
To evaluate TARF as a communication modality, we per- TARF multiplexes independent flows across OFDM subcar-
formed controlled experiments in the water tank setup de- riers as described in §5. We are interested in evaluating the
scribed in §7.2 and measured TARF’s throughput across the throughput versus per-subcarrier SNR. In our evaluation, we
water-air interface. compute the throughput as the average number of correctly
We performed 500 experimental trials in total. We varied decoded packets (multiplied by bits per packet) at each SNR.
the location and height of TARF’s transmitter and receiver as Fig. 9 plots the per-subcarrier throughput versus the SNR.
described in §7.2. We tested TARF in four configurations: the The figure shows empirical results for BPSK (in blue), QPSK
first three employ uniform power distribution and modulation (in red), and 16-QAM (in green) and compares them to the
across all the OFDM subcarriers using BPSK, QPSK and theoretical throughput-SNR curves (plotted with dashed lines).
16QAM. The final configuration incorporates TARF’s power We make the following remarks:
allocation and rate adaptation schemes. • Similar to our BER-SNR curves from §8.1(a), these per-
(a) BER-SNR curves. The performance of a wireless receiver subcarrier throughput-to-SNR curves follow a similar trend
can be evaluated through plots of the bit error rate to the to the theoretical ones. This further confirms TARF as a
signal to noise ratio, called the BER-SNR curves [22, 38]. We viable communication channel.
computed the BER as the fraction of correctly decoded bits • Also similar to the BER-SNR curves, there is a discrepancy
to total transmitted bits. We computed the signal power as the between the empirical and theoretical curves. We observe
squared channel estimate, and computed the noise power as that this discrepancy is more pronounced at lower SNRs, an
the squared difference between the received signal and the observation that can be explained by less perfect channel
transmitted signal multiplied by the channel estimate. estimation at the lower SNRs.
Fig. 8 plots the BER-SNR curves of TARF with BPSK (in • The figure shows that for lower SNRs, higher modulations
blue), QPSK (in red), and 16-QAM (in green) modulations can achieve higher throughput. This demonstrates the need
and compares them to the theoretical BER-SNR curves of for TARF’s rate adaptation technique.
the respective modulation schemes (in dashed lines) [22]. We
make the following observations: (c) Aggregate Throughput-SNR curves. Next, we evaluate
TARF’s overall throughput performance as a function of over-
• TARF’s BER-SNR curves follow a similar trend to the the- all SNR. The overall throughput is computed by summing
oretical ones. These trends demonstrate that TARF presents the per-subcarrier throughput across all the subcarriers. The
a viable cross-medium communication channel. overall SNR is computed as the total signal power across
• Similar to standard communication systems, conservative all subcarriers divided by the total noise power across the
modulation schemes (e.g., BPSK) maintain lower BER at subcarriers. For fair comparison to TARF’s power and rate
the same SNR. This is expected since more conservative adaptation scheme, we plot the achieved throughput as a func-
schemes allocate more power for every bit. tion of the SNR computed prior to TARF’s power allocation.
• There is a discrepancy between TARF’s performance and Fig. 10 plots TARF’s overall throughput for the flat modu-
a theoretically optimal decoder. This can be explained by lation schemes: BPSK (in green), QPSK (in red), 16-QAM (in
the fact that TARF’s decoder is not ideal and its channel blue) as well as with the adaptive modulation scheme from §5
estimation is not perfect. (in black). We make the following observations:
11
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

400 30
TARF with BPSK
TARF with QPSK
Throughput (bps)

TARF with 16QAM

300 TARF with Power and 20

SNR (dB)
Rate Adaptation
200
10

100
0
1 1.5 2 2.5 3 3.5 4
0 Depth (m)
-20 -10 0 10 20 30
Average SNR (dB) Figure 11—SNR vs Transmitter Depth. The figure shows that TARF’s
SNR decreases with depth (in orange) following a 1/r 2 trend (in blue). It
Figure 10—Throughput vs Average SNR. TARF’s power and rate adapta- outperforms pure RF links which decay exponentially in distance underwater.
tion achieves higher throughput than uniform power allocation and outper- Error bars indicate standard deviation.
forms flat modulations.
• The SNR decreases from around 25 dB at 90 cm to 14 dB
• TARF can achieve throughputs of 100 bps, 200 bps and at 3.6 m. The SNR trend follows a 1/r 2 curve (plotted in
400 bps for BPSK, QPSK and 16-QAM modulation respec- orange on the same figure) and matches the theoretically-
tively. This can be explained by the 100Hz bandwidth and predicted behavior of pathloss from §4.2.
the corresponding modulation schemes. These throughputs • The SNR trend with depth demonstrates the superiority of
are similar to standard communication rates for underwater TARF over a pure RF communication link. In particular,
acoustic communication links [46]. RF signals in the GHz range decay exponentially at a rate
• TARF’s power and rate adaptation consistently outper- of around 1000 dB/m in seawater [24, 29].
forms flat modulation schemes. This is due to two reasons: • TARF’s recorded SNR displayed some variation from the
optimal power allocation and adapting the modulation for expected trend. This is due to noise and interference from
each channel to the per-subcarrier SNR. waves caused by swimmers and water circulation cycles.
• TARF’s benefits are particularly significant in low SNR
regimes, which represent the prevailing scenario of under- (b) SNR vs Alignment. Our experiments so far have focused
water communication. Specifically, at an SNR of 0 dB, on scenarios where the transmitter and the receiver are aligned
TARF’s rate and power adaptation can achieve about along the same (vertical) axis to maximize the SNR. Next,
10× throughput increase over any of the flat modulation we evaluated TARF’s performance with varying degrees of
schemes. This emphasizes the need for power and rate misalignment. We performed this evaluation in the water tank
adaptation in a TARF communication link. setup, where we placed the radar 20 cm above the water sur-
face and placed the speaker 40 cm below the surface. To
8.2 SNR Microbenchmarks understand TARF’s performance as a function of different
alignments, we varied the speaker’s location at different hori-
Next, we are interested in studying the SNR trends of TARF as zontal displacements in the plane parallel to the water surface.
a function of different parameters. In particular, we would like Fig. 12 plots the computed mean and standard deviation of
to quantify the impact of depth (of the underwater transmitter) the SNR as a function of the horizontal displacement between
and alignment (between the transmitter and the receiver) on the underwater speaker and the FMCW radar. The figure
the SNR of the received signal. shows that the SNR decreases from 11 dB when the transmit-
(a) SNR vs Depth. To understand the impact of depth on ter and receiver are vertically aligned to around 3 dB when
SNR, we evaluated TARF in the swimming pool setup de- the horizontal misalignment is around 28 cm. Interestingly,
scribed in §7. We placed the FMCW radar 30 cm above the the degradation is less than 10dB within a disk of diameter
water surface at the edge of the pool, and varied the depth about half a meter. It is important to note, however, that this
of the underwater speaker between 90 cm and 3.6 m. We SNR degradation is dependent on the beam profiles of both
overpowered the speaker with the Pyle 300W Stereo receiver the acoustic transmitter and FMCW radar.
and configured it to transmit a single tone at 150 Hz, and we
measured the resulting surface vibration using the radar. We 8.3 Robustness to Waves
performed 45 three-second trials across the different depths. Lastly, we are interested in investigating TARF’s performance
Fig. 11 shows our results. The figure plots the mean and in the presence of surface waves. In particular, because TARF
standard deviation of the received signal as a function of communicates by measuring minute vibrations on the water
distance (in orange). We make the following observations: surface, we expect naturally occurring waves to interfere with
12
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

15

Normalised Throughput
10
SNR(dB)

0
0 5 10 15 20 25 30 0 2 4 6 10 16 22
Horizontal Displacement (cm) Waves Peak-to-Peak Amplitude (cm)
Figure 12—SNR vs Misalignment. The figure shows that TARF’s SNR Figure 13—Throughput vs Surface Wave Amplitude. TARF’s throughput
degrades as a function of the horizontal misalignment between the transmitter decreases in the presence of interference from surface waves. Error bars
and the receiver. Error bars indicate standard deviation. indicate standard deviation.

the communication signals. Hence, we would like to evaluate Below, we highlight some of TARF’s key limitations and
the effectiveness of TARF’s interference cancellation in deal- potential development opportunities:
ing with these waves. To this end, we performed controlled • One-directional Communication: TARF only enables up-
evaluation in the water tank, where we fixed the FMCW radar link communication from the underwater sensor to an air-
30cm above the water surface and fixed the speaker 30cm be- borne node. As described in §3, this limits the scope of the
low the surface. In order to quantify the effect of waves with design, but still opens up many exciting applications in the
our system, we emulate naturally occurring waves by creating underwater networking.
2Hz − 3Hz waves [58] at different peak-to-peak amplitudes • Ocean Wave Amplitude: Our current prototype can sustain
and perform TARF communication with BPSK modulation. a communication link in the presence of surface waves
Waves were generated manually as described in §7, and their with peak-to-peak amplitudes up to 16 cm. It is desirable to
amplitudes were measured with a graded ruler. We estimate extend TARF to operate with stronger waves and inclement
the uncertainty in our wave amplitude measurement to be of weather conditions. One promising approach is to actively
the order of 1 cm. track the surface waves by both the transmitter and receiver
Fig. 13 plots the mean and standard deviation of the and adapt the communication protocol accordingly.
throughput as a function of the peak-to-peak amplitude. We • Misalignment: Our evaluation has demonstrated that
make the following observations: TARF’s performance degrades when the transmitter and
receiver are misaligned. As a result, an airborne TARF re-
• TARF’s channel maintained minimal degradation up to ceiver will need to finely scan the water surface in order to
6 cm waves, which are 100, 000× larger than the surface localize the underwater transmitter. This limitation may be
vibrations caused by the underwater acoustic transmitter, as mitigated by innovative scanning solutions that adapt the
observed in Fig. 3(a). TARF’s ability to deal with this large beam profiles of both the acoustic and radar devices with
interference arises from its unwrapping and filtering stages, height, depth, and expected SNR.
which significantly mitigate the slower moving waves.
• For waves whose peak-to-peak amplitude is 22 cm or larger, Despite these limitations, we believe that TARF marks an
TARF’s throughput drops to zero. Multiple factors that important step toward practical and scalable cross-medium
contribute to this performance. First, the phase wraps too communications. It can enable many applications including
quickly for the unwrapping stage to reliably unwrap it. And submarine-to-drone communication, deep-sea exploration,
second, larger waves may deflect the radar reflection away and subsea IoT (Internet of Things). We hope that this paper
from our receiver (due to radio waves’ partially specular will motivate researchers to explore TARF as a means to
nature), leading to a reduction in the overall SNR. enable truly ubiquitous communication across boundaries.
Acknowledgments. We thank Nanette Wu, Alex Sludds, Harry Hsu,
9 LIMITATIONS AND OPPORTUNITIES and Ali Zartash for helping with the experiments, and thank Yunfei
This paper presents TARF, the first technology that enables Ma for his help in designing the FMCW radar. We also thank Thomas
underwater sensors to wirelessly communicate with com- Consi and Michael Sacarny from the MIT Sea Grant for their help in
pact airborne nodes. TARF transforms the water-air boundary, the setup, and Katy Croff Bell from the MIT Open Ocean initiative
which has been traditionally considered an obstacle for com- for helpful discussions. Finally, we thank Unsoo Ha, our shepherd
munication, itself into a communication interface. As such, Aaron Schulman, and the anonymous SIGCOMM reviewers for
it holds promise to solve the long-standing problem of cross- their helpful feedback on the manuscript. This research is partially
medium communications. supported by the MIT Media Lab and the NSF.
13
SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary Francesco Tonolini and Fadel Adib

REFERENCES ACM, 5–16.

[1] 2017. https://www.performanceaudio.com/item/electro-voice-uw30- [22] Daniel Halperin, Wenjun Hu, Anmol Sheth, and David Wetherall. 2010.
underwater-loudspeaker/41594/. (2017). Predictable 802.11 packet delivery from wireless channel measure-
[2] 2017. https://www.outdoorspeakerdepot.com/75w-sub-amp-150w-in- ments. In ACM SIGCOMM Computer Communication Review, Vol. 40.
wall-subwoofwe.html. (2017). ACM, 159–170.
[3] 2017. http://www.pyleaudio.com/sku/PT270AIU/300-Watt-Stero- [23] Duo-Min He. 1999. High-power Nd: YAG-generated underwater sound
Receiver-with-Built-In-iPod-Docking-Station–AM-FM-Tuner,-USB- source for air-submarine communication. In Solid State Lasers VIII,
Flash-and-SD-Card-Readers-and-Subwoofer-Control. (2017). Vol. 3613. International Society for Optics and Photonics, 83–93.
[4] 2017. https://www.sagemillimeter.com/23-dbi-gain-wr-15-v-band- [24] EJ Hilliard Jr. 1960. Electromagnetic Radiation in Sea Water. Technical
rectangular-horn-antenna/. (2017). Report. Naval Underwater Ordnance Station Newport RI.
[5] 2017. usrp n210. http://www.ettus.com. (2017). ettus inc. [25] David Hoffman. 2009. The Dead Hand: The Untold Story of the Cold
[6] Fadel Adib, Zachary Kabelac, Dina Katabi, and Robert C. Miller. 2014. War Arms Race and Its Dangerous Legacy. Anchor.
3D Tracking via Body Radio Reflections. In Usenix NSDI. [26] Ling Ji, Jun Gao, Ai-Lin Yang, Zhen Feng, Xiao-Feng Lin, Zhong-Gen
[7] Fadel Adib, Zachary Kabelac, Hongzi Mao, Dina Katabi, and Robert C Li, and Xian-Min Jin. 2017. Towards quantum communications in
Miller. 2014. Real-time breath monitoring using wireless signals. In free-space seawater. Optics Express 25, 17 (2017), 19795–19806.
Proceedings of the 20th annual international conference on Mobile [27] Balakrishnan Kaushik, Don Nance, and Krish Ahuja. 2005. A re-
computing and networking. ACM, 261–262. view of the role of acoustic sensors in the modern battlefield. In 11th
[8] Ian F Akyildiz, Dario Pompili, and Tommaso Melodia. 2005. Under- AIAA/CEAS Aeroacoustics Conference. 2997.
water acoustic sensor networks: research challenges. Ad hoc networks [28] Peter Kimball and Stephen Rock. 2008. Sonar-based iceberg-relative
3, 3 (2005), 257–279. AUV navigation. In Autonomous Underwater Vehicles, 2008. AUV 2008.
[9] Steven L Bernstein, Michael L Burrows, James E Evans, AS Griffiths, IEEE/OES. IEEE, 1–6.
DA McNeill, CW Niessen, I Richer, DP White, and DK Willim. 1974. [29] Liu Lanbo, Zhou Shengli, and Cui Jun-Hong. 2008. Prospects and
Long-range communications at extremely low frequencies. Proc. IEEE problems of wireless communication for underwater sensor networks.
62, 3 (1974), 292–312. Wireless Communications and Mobile Computing 8, 8 (2008), 977–994.
[10] D Brumbi. 2000. Low power FMCW radar system for level gaging. In [30] Fill Youb Lee, Bong Huan Jun, Pan Mook Lee, and Kihun Kim. 2008.
Microwave Symposium Digest. 2000 IEEE MTT-S International, Vol. 3. Implementation and test of ISiMI100 AUV for a member of AUVs
IEEE, 1559–1562. Fleet. In OCEANS 2008. IEEE, 1–6.
[11] W. S. Burdic. 1991. Underwater acoustic system analysis. NJ: Prentice [31] Yingzhuang Liu and Xiaohu Ge. 2006. Underwater laser sensor net-
Hall. work: A new approach for broadband communication in the underwater.
[12] Xianhui Che, Ian Wells, Gordon Dickers, Paul Kear, and Xiaochun In Proceedings of the 5th WSEAS International Conference on Telecom-
Gong. 2010. Re-evaluation of RF electromagnetic communication in munications and Informatics. 421–425.
underwater sensor networks. IEEE Communications Magazine 48, 12 [32] Xavier Lurton. 2002. An introduction to underwater acoustics: princi-
(2010), 143–151. ples and applications. Springer Science & Business Media.
[13] Xianhui Che, Ian Wells, Gordon Dickers, Paul Kear, and Xiaochun [33] Xavier Lurton. 2002. An introduction to underwater acoustics: princi-
Gong. 2010. Re-evaluation of RF electromagnetic communication in ples and applications. Springer Science & Business Media.
underwater sensor networks. IEEE Communications Magazine 48, 12 [34] G Meinecke, V Ratmeyer, and G Wefer. 1999. Bi-directional communi-
(2010), 143–151. cation into the deep ocean based on ORBCOMM satellite transmission
[14] Steven Constable. 2010. Ten years of marine CSEM for hydrocarbon and acoustic underwater communication. In OCEANS’99 MTS/IEEE.
exploration. Geophysics 75, 5 (2010), 75A67–75A81. Riding the Crest into the 21st Century, Vol. 3. IEEE, 1405–1409.
[15] Caraivan Mitruţ Corneliu, Dache Valentin, and Sgârciu Valentin. 2012. [35] Michael V Namorato. 2000. A concise history of acoustics in warfare.
Deploying Underwater Sensors Safe-Net in Offshore Drilling Oper- Applied Acoustics 59, 2 (2000), 101–135.
ations Surrounding Areas Using Remote Operated Vehicles. IFAC [36] David Pearson, Edgar An, Manhar Dhanak, Karl von Ellenrieder, and
Proceedings Volumes 45, 6 (2012), 871–876. Pierre Beaujean. 2014. High-level fuzzy logic guidance system for
[16] Wang C. J. Wen B. Y. Ma Z. G. Yan W. D. and Huang X. J. 2007. an unmanned surface vehicle (USV) tasked to perform autonomous
Measurement of river surface currents with UHF FMCW radar systems. launch and recovery (ALR) of an autonomous underwater vehicle
Journal of Electromagnetic Waves and Applications (2007). (AUV). IEEE.
[17] Mari Carmen Domingo. 2011. Securing underwater wireless communi- [37] Jeffery J Puschell, Robert J Giannaris, and Larry Stotts. 1992. The
cation networks. IEEE Wireless Communications 18, 1 (2011). autonomous data optical relay experiment: first two way laser commu-
[18] Matthew Dunbabin, Peter Corke, and Gregg Buskey. 2004. Low-cost nication between an aircraft and submarine. In Telesystems Conference,
vision-based AUV guidance system for reef navigation. In Robotics and 1992. NTC-92., National. IEEE, 14–27.
Automation, 2004. Proceedings. ICRA’04. 2004 IEEE International [38] Hariharan Rahul, Farinaz Edalat, Dina Katabi, and Charles G Sodini.
Conference on, Vol. 1. IEEE, 7–12. 2009. Frequency-aware rate adaptation and MAC protocols. In Proceed-
[19] CR Grant and BS Yaplee. 1957. Back scattering from water and land at ings of the 15th annual international conference on Mobile computing
centimeter and millimeter wavelengths. Proceedings of the IRE 45, 7 and networking. ACM, 193–204.
(1957), 976–982. [39] Shobha Sundar Ram and Hao Ling. 2008. Through-wall tracking
[20] Aditya Gudipati and Sachin Katti. 2011. Strider: Automatic rate adap- of human movers using joint Doppler and array processing. IEEE
tation and collision handling. In ACM SIGCOMM Computer Communi- Geoscience and Remote Sensing Letters 5, 3 (2008), 537–541.
cation Review, Vol. 41. ACM, 158–169. [40] Mark Rhodes, Derek Wolfe, and Brendan Hyland. 2011. Underwater
[21] Aditya Gudipati, Stephanie Pereira, and Sachin Katti. 2012. AutoMAC: communications system comprising relay transceiver. (2011). US
Rateless wireless concurrent medium access. In Proceedings of the 18th Patent 7,877,059.
annual international conference on Mobile computing and networking. [41] H Rowe. 1974. Extremely low frequency (ELF) communication to
submarines. IEEE Transactions on Communications 22, 4 (1974),
14
Networking across Boundaries SIGCOMM ’18, August 20–25, 2018, Budapest, Hungary

371–385. [51] EJ Tucholski and S Traffic. 2006. Underwater Acoustics and Sonar.
[42] Manecius Selvakumar, Ramesh R Subramanian, AN Sathianarayanan, SP411 Handouts and Notes. Fall 2006. Physics Department, US Naval
D Harikrishnan, G Jayakumar, VK Muthukumaran, D Murugesan, M Academy 12 (2006), 11–1.
Chandresekaran, E Elangovan, Doss Prakash, et al. 2010. Technology [52] Lloyd Butler VK5BR. 1987. Underwater radio communication. Origi-
tool for deep ocean exploration-remotely operated vehicle. In Pro- nally published in Amateur Radio (1987).
ceedings of the 20th International Offshore and Polar Engineering [53] John G Webster and Halit Eren. 2017. Measurement, instrumentation,
Conference, Beijing, China. 206–212. and sensors handbook: spatial, mechanical, thermal, and radiation
[43] Robert H Stewart. 2008. Introduction to physical oceanography. Robert measurement. CRC press.
H. Stewart. [54] Louis Whitcomb, Dana R Yoerger, Hanumant Singh, and Jonathan
[44] Milica Stojanovic. 1995. Underwater acoustic communications. In Howland. 2000. Advances in underwater robot vehicles for deep ocean
Electro/95 International. Professional Program Proceedings. IEEE, exploration: Navigation, control, and survey operations. In Robotics
435–440. Research. Springer, 439–448.
[45] Milica Stojanovic. 2007. On the relationship between capacity and [55] T Wiener and Sherman Karp. 1980. The role of blue/green laser sys-
distance in an underwater acoustic communication channel. ACM tems in strategic submarine communications. IEEE Transactions on
SIGMOBILE Mobile Computing and Communications Review 11, 4 Communications 28, 9 (1980), 1602–1607.
(2007), 34–43. [56] S Wolf, J Davis, and M Nisenoff. 1974. Superconducting extremely low
[46] M. Stojanovic. 2007. On the relationship between capacity and distance frequency (ELF) magnetic field sensors for submarine communications.
in an underwater acoustic communication channel. In SIGMOBILE IEEE Transactions on Communications 22, 4 (1974), 549–554.
Mobile Computing and Communications Review. ACM, 34–43. [57] Robert Woodall, Felipe Garcia, and John Sojdehei. 2000. Magneto-
[47] Andrew G Stove. 1992. Linear FMCW radar techniques. In IEE Pro- inductive submarine communications system and buoy. (May 2 2000).
ceedings F (Radar and Signal Processing), Vol. 139. IET, 343–350. US Patent 6,058,071.
[48] Kuan Meng Tan, Tommie Liddy, Amir Anvar, and Tien-Fu Lu. 2008. [58] I. R. Young. 1999. Wind generated ocean waves (Vol. 2). Elsevier.
The advancement of an autonomous underwater vehicle (AUV) tech- [59] Ju Zhang, Xiaoping Zhu, and Zhou Zhou. 2010. Design of time delayed
nology. In Industrial Electronics and Applications, 2008. ICIEA 2008. control systems in UAV using model based predictive algorithm. In
3rd IEEE Conference on. IEEE, 336–341. Informatics in Control, Automation and Robotics (CAR), 2010 2nd
[49] Paul J Titterton, Frederick Martin, Dan J Radecki, and Robert W Cotter- International Asia Conference on. IEEE, 269–272.
man. 1991. Secure two-way submarine communication system. (Aug. 6 [60] Shengli Zhou and Zhaohui Wang. 2014. OFDM for underwater acoustic
1991). US Patent 5,038,406. communications. John Wiley & Sons.
[50] David Tse and Pramod Viswanath. 2005. Fundamentals of wireless
communication. Cambridge university press.

15