JOURNAL OF LATEX CLASS FILES, VOL. 14, NO.

8, AUGUST 2022 1

Spatial Deep Deconvolution U-Net for Traffic

Analyses with Distributed Acoustic Sensing
Siyuan Yuan, Martijn van den Ende, Jingxiao Liu, Hae Young Noh, Robert Clapp, Cédric Richard, Biondo Biondi

Abstract—Distributed Acoustic Sensing (DAS) that transforms Our work addresses these problems by leveraging an
city-wide fiber-optic cables into a large-scale strain sensing array emerging fiber-optic sensing technology called DAS. This
has shown the potential to revolutionize urban traffic monitor- technology turns optical fibers into dense-sampling seismic
ing by providing a fine-grained, scalable, and low-maintenance
monitoring solution. However, the real-world application of DAS recording arrays, offering several benefits. First, DAS records
is hindered by challenges such as noise contamination and inter- absolute signals of human-induced deformation that cannot
ference among closely traveling cars. In response, we introduce be easily tied to any individual. Second, DAS monitoring
a self-supervised U-Net model that can suppress background can be cost-efficient for city-scale monitoring by leveraging
noise and compress car-induced DAS signals into high-resolution ubiquitous pre-existing telecommunication infrastructures as
pulses through spatial deconvolution. Our work extends recent
research by introducing three key advancements. Firstly, we sensors. Finally, the system relies on a single optoelectronic
perform a comprehensive resolution analysis of DAS-recorded device known as an interrogator to be deployed in a secure
traffic signals, laying a theoretical foundation for our approach. and easily accessible location. The DAS approach is in stark
Secondly, we incorporate space-domain vehicle wavelets into contrast to the conventional methods having numerous indi-
our U-Net model, enabling consistent high-resolution outputs vidually powered instruments scattered across a city, exposed
regardless of vehicle speed variations. Finally, we employ L-2
norm regularization in the loss function, enhancing our model’s to meteorological conditions and detrimental interactions with
sensitivity to weaker signals from vehicles in remote traffic lanes. humans, plants, and animals.
We evaluate the effectiveness and robustness of our method
through field recordings under different traffic conditions and
A. Background on DAS
various driving speeds. Our results show that our method
can enhance the spatial-temporal resolution and better resolve DAS-based monitoring is performed by connecting an in-
closely traveling cars. The spatial deconvolution U-Net model terrogator to one end of a standard telecommunications-grade
also enables the characterization of large-size vehicles to identify optical fiber. The interrogator sends short laser pulses into the
axle numbers and estimate the vehicle length. Monitoring large-
size vehicles also benefits imaging deep earth by leveraging the optical fiber and measures the subtle phase shifts of Rayleigh
surface waves induced by the dynamic vehicle-road interaction. scattered light returning to the detector at a predicted two-
way travel time [9], [10]. In this way, the strain field induced
Index Terms—Traffic monitoring, intelligent transportation,
Distributed Acoustic Sensing, deconvolution, U-Net. by natural processes (e.g. earthquakes) and urban activities
(e.g. moving vehicles, construction, pumps) acting on the
fiber coupled to the Earth can be sampled at a meter-scale
I. I NTRODUCTION spatial resolution over tens of linear fiber kilometers. Notably,
DAS technology has proven effective on the pre-existing
T RAFFIC monitoring systems, which automatically and
continuously detect, track, and characterize vehicles in
moving traffic, provide valuable information for urban man-
“dark fiber” telecommunication infrastructure for earthquake
monitoring [11], infrastructure monitoring [12], [13], and near-
surface imaging [14], [15], significantly reducing installation
agement, maintenance, and planning. Conventional monitoring
cost. Furthermore, by incorporating state-of-the-art interroga-
systems include vision-based [1]–[3] and pavement sensing
tors, such as OptaSense QuantX [16], DAS can record full
technologies (e.g., inductive loops [4]–[6] and piezoelectric
waveform signals over distances up to 50 km, with a gran-
sensors [4], [7], [8]). These approaches are well-developed
ular 1 m channel spacing, which makes DAS promising for
but have several drawbacks. For example, camera systems
continuous city-wide sensing [17].
bring individual-privacy concerns and are sensitive to weather
Despite the aforementioned advantages, analyzing data from
conditions; point pavement sensing systems provide spatially
a DAS array located in an urban environment has been
sparse sampling and are challenging to maintain. An additional
challenging because it records a complex mixture of inher-
drawback of these point-sensing systems is their requirement
ently unlabeled signals [18]. As such, researchers have been
for on-site installation. Recent mobile sensing methods over-
developing algorithms that try to detect vehicles accurately
come the on-site installation and maintenance challenge, but
and automatically from DAS recordings.
such systems require cellular connectivity and users to opt-in
location tracking.
B. Existing vehicle tracking and detection algorithms
Siyuan Yuan, Jingxiao Liu, Hae Young Noh, Robert Clapp, and Biondo [19] applied a common seismological method, the short-
Biondi are with Stanford University, CA, U.S.A.
Martijn van den Ende, and Cédric Richard are with the Université Côte time-average through long-time-average trigger (STA/LTA). To
d’Azur,, OCA, UMR Lagrange, France. exploit the array geometry of DAS, beamforming algorithms
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 2

have been applied to detect cars and measure their speed Our first contribution is a comprehensive analysis of the
[20], [21]. These simple methods performed well on roads resolution of DAS-recorded traffic signals, based on both
with relatively light traffic and without complicated traffic physics and field studies. This key element forms the theoreti-
patterns. [22] took this a step further by utilizing the Hough cal bedrock of our method, with a detailed discussion provided
transform, a technique derived from the field of image pro- in Section II.
cessing, to estimate the flow and average speed of vehicles. Secondly, we introduce a novel departure from the method-
[23] introduced data mining and signal processing methods, ology in [26]. Contrary to the use of a Ricker vehicle wavelet
such as clustering algorithms and Kalman filtering techniques, that remains stationary in the time domain, we integrate a
to identify and track vehicles. A novel contribution by [24] physics-based vehicle wavelet in the spatial domain into our
was the application of a real-time object detection algorithm space-domain DAE model. A significant benefit of employing
based on deep learning techniques for estimating traffic flow a spatial wavelet is that it enables our model to produce
and vehicle speed from DAS data, which was assessed along uniformly high-resolution outputs, irrespective of changes in
a 500-meter fiber length in the suburbs of Beijing. More the speed of the vehicle. A detailed discussion is provided in
recently, [25] develop a spatial-domain Bayesian filtering and the Methods Section.
smoothing algorithm to detect, track, and characterize each The third unique aspect of our study involves the opti-
vehicle to obtain fine-grained traffic monitoring with DAS. mization of the regularization term in the loss function. [26]
Despite their limited success, these methods gradually become used the L-1 norm to enhance the sharpness of the outputs.
inaccurate if many vehicles transit simultaneously close to the Similarly, we employed the L-1 norm to amplify the output
same segment of the fiber cable, as signals from different cars sharpness. However, we noticed that the L-1 norm might miss
start overlapping. out on the weak energy released by vehicles traveling on
far-off traffic lanes. Consequently, we investigated the use of
C. Recent advancement that improves the resolution of the L-2 norm regularization, which enabled us to establish
vehicle-induced signals optimal settings to identify weaker signals from vehicles on
A parallel research direction from [26] aims to improve the more distant traffic lanes.
resolution of vehicle signals recorded by DAS by reducing The rest of the paper is organized as follows. Section II
interference among closely traveling cars. The research is valu- demonstrates the resolution reduction of traffic-induced DAS
able as a preprocessing step for algorithms to further improve signals due to a combined effect of physics and the limitation
their detection and tracking ability. Specifically, [26] proposed of the sensors. Through the numerical characterization of vehi-
a self-supervised time-domain deconvolution Auto-Encoder cle impulse responses and observations from field experiments,
(time-domain DAE) with the vehicles’ impulse responses we design a spatial DAE model aiming to improve the signal
from the quasi-static recordings. Compared to a conventional resolution for speed-varying vehicles. Section III describes our
channel-wise deconvolution algorithm, the time-domain DAE spatial DAE model in detail. Section IV describes the training
model has the benefit of incorporating the spatial-temporal procedure and baseline approaches. Section V benchmarks the
characteristics of car signals, leading to much sharper and proposed DAE model with baselines for car motion and speed
more localized outputs. The authors applied a beamforming tracking. Section VI concludes our work and describes future
algorithm to the localized outputs rather than the original direction.
inputs. They showed significant improvements in terms of II. V EHICLE -I NDUCED DAS R ESPONSE
the resolution in car-speed estimation and detection accuracy. This section investigates the vehicle-induced DAS response
Furthermore, as shown in [26], the time-domain DAE model, to inform the design of our proposed space-domain DAE
once trained, can deconvolve 24-hour recordings in less than model. We first review two categories of vehicle-induced DAS
30 seconds, achieving > 400 times speedup compared to responses, quasi-static and surface-wave signals, as a basis for
a conventional iterative approach, which makes the method our study. Subsequently, we carry out a theoretical analysis
promising for real-time processing. and numerical simulations which reveal a resolution decline
However, the time-domain DAE model assumes a stationary in traffic signals due to physical constraints and sensor limita-
Ricker wavelet in the time domain as the vehicle’s impulse tions. This discovery drives the need for resolution enhance-
response. This assumption can undermine vehicle-tracking ment through our novel deconvolution algorithm. Finally, we
accuracy. As the following sections of this paper will show, recognize the speed-invariance of the spatial wavelet, a pivotal
vehicles’ temporal wavelets are non-stationary to speed varia- observation that motivates our development of the space-
tion. Due to the wavelet mismatch, a time-domain DAE model domain DAE model. By incorporating the spatial wavelet into
that works well with fast-speed traffic produces low-resolution our model, we are able to achieve better accuracy with speed
results for low-speed vehicles. variation and better resolution for vehicle tracking.

D. Novelties of our work A. Quasi-static v.s. dynamic response

This study overcomes the constraints of [26] by developing Permanently deployed fiber-optic cables are often buried
a new algorithm that delivers precise results for vehicles trav- (trenched) or placed within underground conduits. We assume
eling at differing speeds. The study’s contribution is threefold. here that the DAS fiber is deployed alongside a road at
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 3

some depth below the surface. When a vehicle passes near-

by the virtual sensors of the roadside telecom fiber cable, the
interaction between the vehicle and the road structure induces
the deformation of the telecom fiber cable. The signal pattern
of vehicle-induced telecom fiber deformation is a function of
the vehicle characteristics, fiber conduit properties, ambient
conditions, etc. There are mainly two components of signals
produced by moving vehicles: quasi-static signals (< 1 Hz)
resulting from the ground deformation due to the vehicle’s
weight, and propagating surface waves (2 to 20 Hz) caused
by the dynamic vehicle-road interaction resulting from the
roughness of the road (e.g., bumps).
Previous studies [19], [27], [28] have found that the quasi-
static component dominates the energy of vehicle-induced
telecom fiber vibration and is theoretically described by the
Flamant-Boussinesq approximation [29]. As a vehicle ap-
proaches the virtual sensor, ground deformation above the
sensor increases, and the fiber coupled to the earth is stretched,
resulting in increased tension in the fiber. As the vehicle moves
away, ground deformation near the virtual sensor and the
fiber tension decreases. Due to the relatively strong energy,
simplicity, and compactness compared to the surface-wave
component, quasi-static signals have been used for car tracking
and detection tasks [26]. For the same reason, our car tracking
method is also based on the quasi-static response.

B. Sensing resolution limited by physics and sensor limitation

We conduct synthetic experiments to demonstrate the res-
olution degradation of quasi-static signals due to physics and
sensor limitations. We model vehicular seismic sources using
a collection of vertical point forces located at the vehicle’s
wheel-road contacts [19], [27], [28], [30]. Define x as the Fig. 1. Simulation of quasi-static signals of a DAS array subjected to a point
load with various lateral offsets, y, to the fiber. (a), (b), and (c) shows the
horizontal distance along the fiber. The quasi-static or geodetic simulated wavelets for gauge lengths, L, of 0, 8, and 16 m, respectively.
strain (Ex ) DAS signal from a vehicle centered at x = 0
is equal to the change in displacement over the DAS gauge
length (L). Displacement in the direction of the fiber (Ux ) (equivalent to point sensor), 8, and 16 m in Fig. 1 (a), (b),
can be modeled using the Flamant-Boussinesq equation for a and (c), respectively. From the point sensor wavelets shown
point load applied to a half space with basic knowledge of in Fig. 1 (a), we can see that the wavelet is compact and sharp
the fiber and vehicle locations, the vehicle load (Fz ), the soil when the fiber is right beneath the load. As y increases, we
shear modulus (µ), and Poisson’s ratio (ν): can observe an increasing smoothing effect leading to wider
wavelets and lower resolution. By comparing Fig. 1 (a) with
Fz zx (1 − 2ν) x (b) and (c), we can see that increasing gauge length leads
Ux (x) = ( − ( )), (1)
4πµ r3 r+z r to a wider wavelet for 0 offset, but has a limited impact on
L larger y (15 and 25 m). Fig. 2 (a) and (b) show the simulation
U (x + − U (x − L2 )
2)
Ex (x) = , (2) of a two-axle car with 2.8 m axle spacing and a three-axle
L vehicle with 9 m axle spacing, respectively. We compare the
p
where z is fiber depth, r = x2 + y 2 + z 2 is the 3-D point senor recording to that obtained with L = 16 m and the
distance from the fiber position to the car, and y is the lateral fiber-road offset of 15 m and 25 m. We can see that the axles
fiber-road offset. From the equation, we can see that for fixed are resolved from the point sensor recordings with zero lateral
soil properties, the shape of spatial strain response depends offset, but become indistinguishable in recording with a larger
on the gauge length and the relative position between the load offset due to the smoothing effect.
and the fiber. To compute the modeled synthetic horizontal
strain signal, we assumed a fiber depth of 2 m, appropriate
C. Wavelet dependency on vehicle speed
values for sandy/clayey soils, such as a Poisson’s ratio of
0.4. Fig. 1 shows the simulated strain along the fiber for a This subsection studies the effect of vehicle speed on the
point load at x = 0 m and y = 0, 15, 25 m. We investigate DAS response in the temporal and spatial domain, which is
the gauge length effect by simulating wavelets with L of 0 important to guide the design of the DAE model. For a vehicle
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 4

Fig. 3. Numerical simulation of the time-domain quasi-static signals of a

two-axle car with constant speeds of 10, 20, 30, and 40 mph.

Fig. 2. Simulation of quasi-static signals of a DAS array with zero and 16 m

gauge length, L, loaded with (a) a two-axle car with an axle spacing of 2.8
m; (b) a three-axle vehicle with an axle spacing of 9 m with various lateral
offsets, y, to the fiber.

traveling from an initial position, x0 , with a constant speed of

c, the measured strain response at fiber location x is,

εx (x, t) = Ex (x − ct − x0 ). (3)

The temporal wavelet at a constant time x0 and the spatial

wavelet at a constant location t0 can be written as ϕ(t) =
Fig. 4. Map view of a roadside section of the Stanford DAS-2 Array.
εx (x = x0 , t) and ψ(x) = εx (x, t = t0 ), respectively. From Distances along the fiber are labeled on the map
equations (1) to (3), we can see that the car speed, c, acts as
a scaling factor of ϕ(t), determining the compactness of the
temporal wavelet. This can be confirmed through numerical broader bandwidth (narrower temporal wavelet). We can also
experiments. Fig. 3 shows the simulation of the temporal observe that the wavenumber components remain relatively
wavelet of a two-axle car traveling at 10, 20, 30, and 40 mph. invariant to car speeds, implying that the spatial wavelet for
We can see that faster car speed leads to sharper temporal different speeds is stationary.
wavelet. On the other hand, we can see from the equations
that the shape of the spatial wavelet is speed-independent. The III. M ETHODS
speed term only introduces a space shift accounting for the The previous section brought attention to a key issue:
different vehicle positions at time t0 for different speeds. the degradation of resolution in quasi-static vehicle signals,
Our observation that the shape of the space-domain rather than primarily due to the substantial smoothing effect induced by
the time-domain wavelet is speed-invariant is confirmed using large fiber-road offset and gauge length. To combat this, we
field experiments at Sand Hill road monitoring by the Stanford propose a spatial-domain DAE model that enhances resolution
DAS2 array [28]. Figure 4 shows the map of a subsection of through spatial-axis deconvolution.
the roadside fiber labeled with distances along the fiber. The Our advanced spatial-domain DAE model evolves from the
horizontal distances from the centers of the south- and north- original time-domain DAE model that introduced in [26].
bound traffic lanes to the fiber are around 15 m and 25 m, The time-domain model proves effective under conditions
respectively. We drove a test car southward to understand the of roughly constant vehicle speed over time and when the
car’s impulse responses for various driving speeds, including temporal wavelet aligns with their proposed Ricker wavelet.
10, 20, 30, and 40 mph. Figure 5 shows the quasi-static However, as underscored in the previous section, the speed-
signals of our car indicated from DAS recordings using orange dependent variation in the vehicle’s time-domain impulse
arrows. The bottom panels show the corresponding Frequency- response can cause wavelet mismatch in the time domain,
Wavenumber spectra. It can be seen that the frequency com- leading to suboptimal deconvolution outcomes, a topic further
ponents vary with speeds, i.e., a higher car speed leads to a explored in Section V.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 5

Fig. 5. (Top) DAS recordings of the test car driving with constant speeds of 10, 20, 30 and 40 mph. The time axes are scaled differently. (Bottom) The
corresponding F-K spectra of the DAS recordings above. We can see that the frequency range becomes broader with increasing speed, whereas the wavenumber
components are invariant to car speeds.

This section begins with an outline of the network architec- The primary distinction between our DAE model and the
ture, followed by a detailed explanation of our novel extension conventional time-domain DAE model in [26] lies in the fact
to the original time-domain model. As will be elaborated in that we use a simulated wavelet in the spatial domain as
Section V, our innovative extension enables a more resilient opposed to a Ricker wavelet in the time domain. A principal
model, adept at generating high-quality outcomes that are advantage of this approach is that the impulse response in the
unaffected by speed variations. space domain is not affected by speed and can be associated
with the number of axles and the wheelbase. As a result, our
A. Network Architecture innovative application of a spatial-domain wavelet allows us
Our space-domain DAE model shown in Figure 6 is a 2-D to yield reliable results for vehicles with varying speeds, and
fully convolutional U-Net adapted from [26]. The inputs are retrieve the number of axles, and estimate the wheelbase for
quasi-static traffic recordings, a set of Nx = 256 (256 meters) larger vehicles.
consecutive waveforms of Nt = 1024 time samples (20.48 s) Additionally, our simulated wavelet is estimated based on
in duration, organized in an Nx × Nt matrix. The outputs are the physics defined in Equations (1) through (3). This ground-
sharp deconvolution results with the same shape as the inputs. ing in physics facilitates the integration of the gauge length
The U-Net model comprises 3 convolutional layers, followed and lateral fiber-road offset smoothing effects, which results
by 3 encoder blocks containing a downsampling (max pooling) in a more accurate wavelet and subsequently enhances the
layer and 3 convolutional layers. The kernel sizes for the con- performance of our model.
volution layers are 3 × 5. The number of convolutional filters When dealing with a batch of quasi-static inputs,
is initialized at 8 and gets doubled after each downsampling y1 , y2 , . . . , yNb (Nb being the batch size, Nb = 128), the
operation. The maxpooling operation downsamples the data by loss function is formulated as a combination of the L-2 norm
a factor of 2 along the DAS sensor axis and by a factor of 4 of the difference between the reconstructed input and the
along the time axis (i.e, the maxpooling kernel and strides are original input. Because of the use of the spatial wavelet,
of size 2 × 4). The decoder reverses the encoding operations our reconstructed input is obtained with spatial convolution
with 3 blocks of bilinear upsampling. The U-Net contains skip- between the space-domain impulse response and network
connections, which directly connect the output of one encoder output. This deviates from the temporal convolution in [26].
block with the corresponding decoder block. Lastly, the output In contrast to the exclusive use of the L-1 norm regularization
layer is a single convolutional layer with 1 output channel term in previous work, we explore both the L-1 and L-2
and ReLU activation, which enforces positivity in the model norm of the outputs. This extension allows us to determine
output. Our model is a semi-supervised algorithm, in the sense optimal settings to more accurately extract the weaker signals
that no ground truth deconvolution is required as labels to of vehicles traveling in distant lanes. Our loss function is
train the model. Weak supervision comes from the spatial car defined as follows:
impulse response kernel shown as the red curve in Figure 6. Nb
1 X
L= (||[k ∗ xi ]d − yi ||22 + ρ||xi ||1,2 ), (4)
Nb i=1
B. Distinguishing Features of Our space-domain DAE Model
where xi and yi denote the i-th deconvolution output and
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 6

Fig. 6. Conceptual overview of the spatial DAE model. The input is the quasi-static response of DAS to cars, which can be viewed as a matrix, 256 channels
(256 meters) x 1024 time steps (20.48 seconds). The output is the deconvolution results. The loss is computed with the input and the reconstructed input is
obtained through a spatial convolution of the output with a known car impulse response in the spatial domain.

quasi-static input of the U-Net model, respectively. ρ is a worth of traffic recordings of the Sand Hill DAS fiber. We
weighting term that promotes sparseness in the deconvolved split the 2-hour recordings into a training and evaluation set
results. [∗]d refers to convolution along the sensor axis. k with a ratio of 80% to 20%.
stands for the spatial wavelet. As we will see in the follow-
ing sections, unlike conventional channel-independent linear B. Evaluation baselines
filters, the 2-D deconvolution operations incorporate spatial- As a baseline, we trained a time-domain DAE model with
temporal features in the DAS recordings. The non-linear nature the same dataset. The temporal wavelet used for training is
of the U-Net introduces high frequencies that are not present simulated with a car speed of 30 mph (shown in Fig. 3)
in the inputs, producing sharp and localized outputs. matching the speed limit of the Sand Hill road.
C. Spatial kernel estimation We also benchmark our spatial DAE model using a con-
In practice, the spatial kernel can be estimated either through ventional spatial deconvolution algorithm with an objective
numerical simulation described in the previous section or function:
by performing statistical averaging of responses of multiple 1
passing cars assuming that the spatial impulse response is x̂q = argmin{ ||[k ∗ xq ]d − yq ||22 + ρ||xq ||1 } (5)
xq 2
constant in time at each fiber location. The statistical averaging
approach requires detecting several passing cars in a subsec- Note that [k ∗ xq ]d stands for convolution in space between a
tion of the fiber. The detection can be achieved through manual known spatial impulse response kernel, k, and the underlying
inspection. Herein, we apply a find-local-maximum algorithm impulse model, x̂q stands for the deconvolution output. yq
from the SciPy library [31] to the recordings at quiet midnight represents the input. k and ρ have the same meaning as in
to detect an average waveform from isolated cars. equation (4). One commonly used algorithm to solve this
optimization problem is the Iterative Shrinkage Thresholding
IV. T RAINING AND EVALUATION BASELINES Algorithm (ISTA; [32], [33]). For this study, we adopt an
accelerated version of ISTA (Fast-ISTA or FISTA) due to the
This section aims to outline the steps involved in training reason described in [34], which exhibits faster convergence
our space-domain DAE models and introducing the conven- guarantees. With FISTA performing spatial deconvolution,
tional methods used for evaluation. signals at each time index are processed independently.

A. Training procedure V. R ESULTS

We trained our proposed space-domain DAE models with We exhibit the evaluation outcomes of our proposed space-
the simulated spatial wavelets of a point load at y = 15 and domain DAE model across a range of applications, including
y = 25 m (targeting traffic at the south- and north-bound lanes, enhancing the precision of car tracking, improving the resolu-
respectively) with a DAS gauge length of 16 m shown as the tion for vehicles traveling closely, and providing accurate axle
orange and green curves in Fig 1 (c). The choice of y and count and wheelbase estimates for large vehicles. Additionally,
gauge length match the real settings of the Sand Hill Road a potential limitation of our approach is discussed, along with
DAS experiments. The training dataset we used is 2-hours’ a proposed solution to overcome it.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 7

A. Car tracking

To test and benchmark the performance of the proposed

spatial DAE model, we conducted controlled driving experi-
ments under a rare traffic condition where we drove a test car
equipped with a speed sensor and a GPS receiver southward
along the subsection of the Sand Hill road shown in Fig. 4.
In Fig. 7, we focus on a case where our car is first speeding
up and then slowing down. Fig. 7 (a) shows the quasi-static
signal of our car. The car speed is inversely proportional to
the slope in the time-space coordinate. We can see in (a) that
when the car speed is low (< 10 s and > 30 s), the wavelet is
“stretched” in time. When the car speed is relatively higher,
the time-domain wavelet is “compressed”, agreeing with our
observation in Section II. Fig. 7 (b), (c), and (d) show the
deconvolution results from the proposed space-domain DAE
model, the original time-domain DAE model, and the space-
domain FISTA algorithm. Both the space- and time-domain
DAE models are trained with L-1 norm regularization loss to
promote sparsity in the outputs. We can see that the proposed
space-domain DAE model yields the sharpest and the most
localized results regardless of car-speed variation. Meanwhile,
we can see that the background noise is suppressed by the Fig. 7. (a) DAS recording of the quasi-static signal of a controlled driving
DAE model. In contrast, the time-domain DAE model yields experiment where we first speed up and then slow down; Deconvolution
results that are dependent on the speed, e.g. the output is more results: (b) the proposed space-domain DAE model, (c) the time-domain DAE
model; (d) spatial deconvolution via the FISTA algorithm. The amplitude picks
compact in space with a higher speed but becomes stretched in each panel are shown as a light blue curve reflecting the car trajectory.
out when the speed is lower. The spatial deconvolution via the
FISTA algorithm yields results that are speed-invariant but are
not as compact as the spatial-DAE results. Signals < 5 seconds case study was intended to accurately assess our method’s
are oversuppressed by the L-1 norm term in the equation 5. ability to extract coherent signals from cars varying in speed.
We can also observe artifacts shown in black in the FISTA Although sparse traffic was used for this evaluation, our system
outputs. is generally applicable to any traffic conditions.
Using the signals in Fig. 7 (a) to (d), we can track the
car movement by picking the amplitude at each time step. B. Traffic pattern monitoring
The estimated trajectory for each panel is shown as a thin This section investigates the performance of the space-
blue curve in Fig. 7. Car speed estimates can be obtained domain DAE model under dense traffic conditions, which we
efficiently through a local beamforming algorithm applied intentionally selected to test the robustness of our model under
to either the quasi-static signals or the deconvolution results the most challenging circumstances with numerous closely
(Figure 8): The beamforming spectrum at each time step is traveling vehicles. Fig. 9 (a) shows a 200-second quasi-static
computed using a local 2-D window (3 s x 30 m) following recording in heavy traffic. Vehicles transiting southward are
the estimated trajectories. In Fig. 8, a brighter color indicates closer to the fiber and thus generate stronger quasi-static
higher stack energy in each panel. The red curves represent signals than the northbound traffic. In the plot, we label several
speed measurements from an onboard speed sensor. Our speed interesting events. We can see that when multiple cars trail
estimates, indicated with the black curves, are obtained by closely, their quasi-static signals interfere with each other caus-
picking the maximum amplitude of the beamforming spectrum ing complicated patterns that pose ambiguity for individual car
at each time step. We can see that our beamforming estimates identification. Fig. 9 (b) and (c) show the FISTA and time-
using the space-domain DAE output match the red curve the domain DAE (L-1 norm regularization) results, respectively.
best. Using the Controller Area Network (CAN) bus reading We can see that FISTA generates poorly resolved results and
from the vehicle as the ground truth, the root-mean-squared contains evident sidelobe artifacts shown in black. The time-
errors (RMSE) for the original recording, space-domain DAE, domain DAE model yields sharp results with fast car speed.
time-domain DAE, and FISTA are 2.56, 2.14, 4.55, and 2.48 However, as can be seen from the boxes in Fig. 9 (c), the
mph, respectively. Our space-domain DAE model achieves the results get blurred when the car speed is low, e.g. cars stopping
lowest RMSE thanks to the sharp deconvolution outputs and for a red light and restarting. Fig. 9 (d) shows the results of
its highest noise suppression performance. The large error of the proposed space-domain DAE model trained with the near-
the time-domain DAE model is mainly due to the falsely high- lane wavelet (y = 15 m) and L-1 norm regularization loss.
speed event around 33 s as shown in Fig.7 (c) and the wiggled We can see that the model yields sharp results regardless of
trajectory pickings when the car speed is low (t >33 s). speed changes. However, signals of the northbound traffic in
The selection of a rare traffic scenario for our car tracking a further lane are oversuppressed. To recover the further lane
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 8

response wavelet. We can see the three axles are recovered

as three strong-energy trajectories corresponding to the three
axles. The amplitude difference between the three trajectories
could imply a non-uniform weight distribution on different
wheel groups. The distance between the two most substantial
peaks is about 17 m agreeing with the bus length, indicating
the usage of our L-1 norm model for counting long-size
vehicle axles and length characterization. Figure 10 (b) shows
the results of the space-domain DAE model trained with L-2
norm regularization. We can see that axles are smoothed out
by the L-2 norm regularization. The L-2 norm model could
be used to track bus motion.
Monitoring large-size vehicles benefits not only traffic man-
agement but also non-intrusive shear-wave velocity imaging,
which can support city sustainability on applications including
sinkhole detection, excavation monitoring, and earthquake
hazard analysis. A shear-wave profile can be estimated through
an optimization algorithm using traffic-induced surface waves
as inputs. Fig. 11 (a) and (b) show the dynamic surface wave
components (> 1 Hz) excited by a moving bus and a car,
respectively. We can see that the bus excites much stronger
surface waves than the car. With the surface waves within
Fig. 8. Local beamforming spectra are performed following the trajectory
picks in Fig. 7 using (a) the quasi-static signal, (b) space-domain DAE results,
the two dashed lines, we compute the dispersion spectrum
(c) time-domain DAE results, and (d) FISTA results. The red curve in each reflecting the relationship between the phase velocity and
panel shows the speed measurements of an onboard speed sensor. The black frequency as shown in Fig. 11 (c) and (d) for the bus and
curves indicate the speed estimates picked from the amplitude of the spectra
at each time step.
car, respectively. We can see that large-size vehicles, such as
a bus, generates surface waves with frequency as low as 2.5
Hz, which is absent from the car-induced surface waves. Low
traffic, we experiment with a spatial DAE model trained with frequencies at 2.5 Hz have a wavelength of ∼ 200 m [28].
the wider far-lane wavelet (simulated with y = 25 m matching The maximum investigation depth of the shear-wave velocity
the fiber-northbound lane offset). Fig. 9 (e) shows the far-lane structure using surface waves is about half of the longest
DAE results. We observe fewer oversuppression issues for the wavelength [35]. Therefore, with buses, we can retrieve depths
northbound traffic and sharper signals for both near and far down to 100 m, whereas using only the car, results at depths
lanes. Nonetheless, artifacts that could be misinterpreted as below approximately 50 m are unreliable.
vehicles appear as circled out in the plot. In Fig. 9 (f), we
experiment with space-domain DAE models trained with L-2
D. Spatial non-stationarity
norm regularization loss using the near-lane wavelet. We can
see that with L-2 norm regularization, weak signals of the far- Our spatial model assumes that the car impulse response is
lane traffic are better preserved, which benefits the tracking of stationary in space. However, the impulse response is a func-
the northbound vehicles. Fig. 9 (g) shows the results of the tion of the recording system and the near-surface conditions
L-2 norm space-domain DAE model trained with the far-lane surrounding the fiber optic cable. Thus, the deconvolution of
wavelet. Although the signals are slightly sharper than that a DAS array covering heterogeneous near-surface conditions
shown in (f), the background noise level gets increased. using a single stationary spatial impulse response could be
suboptimal. Figure 12 (a) shows DAS recordings of traffic
in downtown San Jose City (SJC). (b) shows deconvolution
C. Large-size vehicle monitoring results of the proposed space-domain DAE model with a
Heavier and longer vehicles, e.g. buses, trucks, and trains, simulated car impulse response kernel. The deconvolution
generate quasi-static signals with much larger amplitudes and results look encouraging for most of the fiber, which can
wider spatial wavelets than cars, e.g., sedans and SUVs. The be contributed to the consistency of the assumed wavelet to
wider spatial wavelet can be viewed as the superposition of the real wavelet. However, we can see that the results in the
quasi-static signals of wheels at axles that are farther apart. red box are relatively poorly resolved. The low resolution
Figure 10 (a) shows a quasi-static recording of an 18-meter is more noticeable from the zoomed-in view of data and
three-axle bus transiting southward (video captured by our deconvolution results in the red box in Figure 13 (a) and
camera). Due to the smoothing effect of the road-fiber offset (b). The under-performance could be explained as the wavelet
and gauge length, the axles are indistinguishable from the inconsistency due to the non-stationarity of the near-surface
quasi-static signals. Figure 10 (b) shows the deconvolution properties. Assuming that the near-surface properties, and in
results of the space-domain DAE model trained with L-1 turn the car response, at each channel location is stationary
norm regularization loss using the simulated near-lane impulse in time. A possible solution would be to estimate the spatial
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 9

Fig. 9. (a) Quasi-static signals of passing vehicles in heavy traffic. (b) spatial deconvolution via FISTA. (c) temporal deconvolution via the time-domain DAE
model. Boxes indicate signals in lower resolution due to slower car speeds. (d) and (e) show the results from the proposed space-domain DAE model with
L-1 norm regularization using near and far lane wavelets, respectively. Artifacts that could be misinterpreted as transiting cars are circled out. (f) and (g) show
the results of the space-domain DAE model with L2 regularization using the near and far wavelets, respectively.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 10

Fig. 10. (a) Signals of an 18-meter bus and two regular-sized cars. Deconvolution results of (b) space-domain DAE model with L-1 regularization, and (c)
space-domain DAE model with L-2 regularization, respectively.

Fig. 12. Deconvolution of traffic recording in downtown SJC via a space-

domain DAE model with a simulated impulse response stationary in space.
(a) Input data; (b) deconvolution results. The red box indicates results that
are poorly resolved due to the spatial non-stationarity of the car wavelets.
Fig. 11. (a) Surface-wave components of (a) a passing bus and (b) a car. (c)
and (d) show the dispersion image estimated from the surface waves in (a)
and (b), respectively
DAS array. To denoise the data and to reduce the interference
among closely traveling cars, we propose a self-supervised
car response at different parts of the fiber by averaging the convolutional U-Net model (space-domain DAE model) that
responses of multiple passing cars around each location. To can compress the quasi-static signals into sharp pulses and
verify this, we employ a local maximum finding algorithm as remove the background noises. The goal is achieved through
a simple car detector for the quasi-static signals. We average spatial deconvolution with an assumed spatial wavelet of the
the responses of six identified cars passing the area in the quasi-static signals, which is a major difference from the
red box of Figure 12 as the impulse response input to the U- previously proposed time-domain DAE model.
net model. We retrain the U-net using this estimated kernel,
which produces a sharper result in Figure 13, indicating the This paper shows that using the spatial instead of the tempo-
effectiveness of our approach. ral kernel is advantageous because it is invariant to car speed
variation. This leads to improved precision and robustness
VI. D ISCUSSION AND CONCLUSIONS of tracking cars with varying speeds, which is essential for
This paper focuses on the application of traffic monitoring driving behavior identification and accident detection. The
with car-induced quasi-static signals recorded by an urban benefits of our DAE model are more obvious for heavy-traffic
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 11

thank the Stanford School of Earth IT team for hosting the

interrogator in the Scholl computer room.

R EFERENCES

[1] K. Robert, “Video-based traffic monitoring at day and night vehicle

features detection tracking,” in 2009 12th International IEEE Conference
on Intelligent Transportation Systems, 2009, pp. 1–6.
[2] S. R. E. Datondji, Y. Dupuis, P. Subirats, and P. Vasseur, “A survey of
vision-based traffic monitoring of road intersections,” IEEE Transactions
on Intelligent Transportation Systems, vol. 17, no. 10, pp. 2681–2698,
2016.
[3] P. Reinartz, M. Lachaise, E. Schmeer, T. Krauss, and H. Runge,
“Traffic monitoring with serial images from airborne cameras,” ISPRS
Journal of Photogrammetry and Remote Sensing, vol. 61, no. 3,
pp. 149–158, 2006, theme Issue: Airborne and Spaceborne Traffic
Monitoring. [Online]. Available: https://www.sciencedirect.com/science/
article/pii/S0924271606001146
[4] N. K. Jain, R. Saini, and P. Mittal, “A review on traffic monitoring system
techniques,” Soft Computing: Theories and Applications, pp. 569–577,
2019.
[5] S.-T. Jeng and L. Chu, “A high-definition traffic performance monitoring
system with the inductive loop detector signature technology,” in 17th
International IEEE Conference on Intelligent Transportation Systems
(ITSC), 2014, pp. 1820–1825.
[6] T. Cherrett, H. Bell, and M. McDonald, “Traffic management
Fig. 13. (a) and (b) are respectively zoomed-in views of the data and parameters from single inductive loop detectors,” Transportation
deconvolution results in the red box shown in Figure 12. We can see artifacts in Research Record, vol. 1719, no. 1, pp. 112–120, 2000. [Online].
(b) due to the wavelet mismatch. (c) shows the better-resolved deconvolution Available: https://doi.org/10.3141/1719-14
results with a statistical estimate of spatial wavelet for this zoomed-in location. [7] J. Zhang, Y. Lu, Z. Lu, C. Liu, G. Sun, and Z. Li, “A new smart
traffic monitoring method using embedded cement-based piezoelectric
sensors,” Smart Materials and Structures, vol. 24, no. 2, p. 025023,
2015.
conditions where car signals interfere. With our space-domain [8] Z.-X. Li, X.-M. Yang, and Z. Li, “Application of cement-based piezo-
DAE model, compressing signals of individual cars to sharp electric sensors for monitoring traffic flows,” Journal of transportation
engineering, vol. 132, no. 7, pp. 565–573, 2006.
pulses, traffic patterns are much better revealed. Furthermore, [9] R. Posey, “Rayleigh scattering based distributed sensing system for
we show that our space-domain DAE model is robust to structural monitoring,” in Fourteenth International Conference on
deconvolve signals of large-size and heavy vehicles, such as Optical Fiber Sensors, A. G. Mignani and H. C. Lefèvre, Eds., vol.
4185, International Society for Optics and Photonics. SPIE, 2000, p.
a bus. Being able to identify large vehicles can be helpful in 41850E. [Online]. Available: https://doi.org/10.1117/12.2302157
extracting low-frequency surface waves to image the velocity [10] A. Masoudi and T. P. Newson, “Contributed review: Distributed
structure down to hundreds of meters beneath the fiber. Lastly, optical fibre dynamic strain sensing,” Review of Scientific Instruments,
vol. 87, no. 1, p. 011501, 2016. [Online]. Available: https:
we point out that a limitation of our spatial-domain DAE //doi.org/10.1063/1.4939482
model arises when car impulse responses at different locations [11] B. L. Biondi, S. Yuan, E. R. Martin, F. Huot, and R. G. Clapp, Using
along the fiber vary due to coupling heterogeneity and/or fiber Telecommunication Fiber Infrastructure for Earthquake Monitoring
and Near-Surface Characterization. American Geophysical Union
properties. We address the issue by extracting car wavelets us- (AGU), 2021, ch. 10, pp. 131–148. [Online]. Available: https://agupubs.
ing statistical averaging, and train separated networks. Future onlinelibrary.wiley.com/doi/abs/10.1002/9781119521808.ch10
work would be adapting the DAE model to different parts of [12] J. Liu, S. Yuan, B. Luo, B. Biondi, and H. Y. Noh, “Turning
telecommunication fiber-optic cables into distributed acoustic sensors
the fiber using location-dependent spatial wavelets. for vibration-based bridge health monitoring,” Structural Control
and Health Monitoring, vol. 2023, 2023. [Online]. Available:
https://doi.org/10.1155/2023/3902306
ACKNOWLEDGMENTS [13] S. Yuan, J. Liu, H. Young Noh, and B. Biondi, “Urban system monitoring
using combined vehicle onboard sensing and roadside distributed
This research was supported financially by the affiliates of acoustic sensing,” in SEG International Exposition and Annual Meeting,
vol. Day 1 Sun, September 26, 2021, 09 2021, d011S137R003. [Online].
the Stanford Exploration Project and the Stanford Sustain- Available: https://doi.org/10.1190/segam2021-3584136.1
ability Initiative and the UPS Foundation Endowment Fund. [14] E. R. Martin, C. M. Castillo, S. Cole, P. S. Sawasdee, S. Yuan,
Jingxiao Liu is supported by Leavell Fellowship on Sustain- R. Clapp, M. Karrenbach, and B. L. Biondi, “Seismic monitoring
able Built Environment at Stanford University. Martijn van den leveraging existing telecom infrastructure at the SDASA: Active,
passive, and ambient-noise analysis,” The Leading Edge, vol. 36,
Ende was supported by the French government through the no. 12, pp. 1025–1031, 12 2017. [Online]. Available: https:
3IA Côte d’Azur Investments in the Future project managed //doi.org/10.1190/tle36121025.1
by the National Research Agency (ANR) with the reference [15] G. Fang, Y. E. Li, Y. Zhao, and E. R. Martin, “Urban near-
surface seismic monitoring using distributed acoustic sensing,”
number ANR-19-P3IA-0002. The interrogator unit was loaned Geophysical Research Letters, vol. 47, no. 6, p. e2019GL086115,
to us by OptaSense Inc. We thank Martin Karrenbach, Victor 2020, e2019GL086115 10.1029/2019GL086115. [Online]. Available:
Yartsev, and Lisa LaFlame from Optasense, as well as the https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2019GL086115
[16] Optasense, a Luna company, “Quantx das interrogator: Optasense,”
Stanford ITS fiber team, and in particular Erich Snow, for 2022. [Online]. Available: https://www.optasense.com/technology/
crucial help with the Stanford DAS-2 experiment. We also quantx/
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 12

[17] B. Biondi, R. G. Clapp, S. Yuan, and F. Huot, Scaling up to pp. 1413–1457, 2004. [Online]. Available: https://onlinelibrary.wiley.
city-wide dark-fiber seismic arrays: Lessons from five years of the com/doi/abs/10.1002/cpa.20042
Stanford DAS array project, 2021, pp. 3225–3229. [Online]. Available: [34] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algo-
https://library.seg.org/doi/abs/10.1190/segam2021-3594539.1 rithm with application to wavelet-based image deblurring,” in 2009 IEEE
[18] F. Huot, E. R. Martin, and B. Biondi, Automated ambient-noise International Conference on Acoustics, Speech and Signal Processing,
processing applied to fiber-optic seismic acquisition (DAS), 2018, pp. 2009, pp. 693–696.
4688–4692. [Online]. Available: https://library.seg.org/doi/abs/10.1190/ [35] C. B. Park, R. D. Miller, and H. Miura, “Optimum field parameters of
segam2018-2997880.1 an masw survey,” Japanese Society of Exploration Geophysics Extended
[19] N. J. Lindsey, S. Yuan, A. Lellouch, L. Gualtieri, T. Lecocq, and Abstracts, vol. 36, 2002.
B. Biondi, “City-scale dark fiber das measurements of infrastructure use
during the covid-19 pandemic,” Geophysical Research Letters, vol. 47,
no. 16, p. e2020GL089931, 2020, e2020GL089931 2020GL089931.
[Online]. Available: https://agupubs.onlinelibrary.wiley.com/doi/abs/10.
1029/2020GL089931
[20] X. Wang, Z. Zhan, E. F. Williams, M. G. Herráez, H. F. Martins, and
M. Karrenbach, “Ground vibrations recorded by fiber-optic cables reveal
traffic response to covid-19 lockdown measures in pasadena, california,”
Communications Earth & Environment, vol. 2, 2021.
[21] M. van den Ende, A. Ferrari, A. Sladen, and C. Richard, “Next-
generation traffic monitoring with distributed acoustic sensing arrays
and optimum array processing,” in 2021 55th Asilomar Conference on
Signals, Systems, and Computers, 2021, pp. 1104–1108.
[22] C. Wiesmeyr, C. Coronel, M. Litzenberger, H. J. Döller, H.-B.
Schweiger, and G. Calbris, “Distributed acoustic sensing for vehicle
speed and traffic flow estimation,” in 2021 IEEE International Intelligent
Transportation Systems Conference (ITSC). IEEE Press, 2021, p.
2596–2601. [Online]. Available: https://doi.org/10.1109/ITSC48978.
2021.9564517
[23] B. Chan, E. R. Korfanty, S. Nataj, J. Park, B. Peng, and
J. Zhang, “Traffic Monitoring with Distributed Acoustic Sensing,”
2020, [Online; accessed 13-June-2023]. [Online]. Available: https:
//m2pi.ca/project/2020/fotech-solutions/Fotech-final.pdf
[24] Z. Ye, W. Wang, X. Wang, F. Yang, F. Peng, K. Yan, H. Kou, and
A. Yuan, “Traffic flow and vehicle speed monitoring with the object
detection method from the roadside distributed acoustic sensing array,”
Frontiers in Earth Science, vol. 10, 01 2023.
[25] J. Liu, S. Yuan, Y. Dong, B. Biondi, and H. Y. Noh, “Telecomtm: A
fine-grained and ubiquitous traffic monitoring system using pre-existing
telecommunication fiber-optic cables as sensors,” Proc. ACM Interact.
Mob. Wearable Ubiquitous Technol., vol. 7, no. 2, jun 2023. [Online].
Available: https://doi.org/10.1145/3596262
[26] M. van den Ende, A. Ferrari, A. Sladen, and C. Richard, “Deep
deconvolution for traffic analysis with distributed acoustic sensing data,”
IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 3,
pp. 2947–2962, 2023.
[27] P. Jousset, T. Reinsch, T. Ryberg, H. Blanck, A. Clarke, R. Aghayev,
G. P. Hersir, J. Henninges, M. Weber, and C. M. Krawczyk, “Dynamic
strain determination using fibre-optic cables allows imaging of seismo-
logical and structural features,” Nature communications, vol. 9, no. 1,
pp. 1–11, 2018.
[28] S. Yuan, A. Lellouch, R. G. Clapp, and B. Biondi, “Near-surface
characterization using a roadside distributed acoustic sensing array,”
The Leading Edge, vol. 39, no. 9, pp. 646–653, 09 2020. [Online].
Available: https://doi.org/10.1190/tle39090646.1
[29] Y. Fung, “Foundations of solid mechanics prentice-hall,” Inc, New
Jersey, 1965.
[30] L. Li, S. Nimbalkar, and R. Zhong, “Finite element model of
ballasted railway with infinite boundaries considering effects of moving
train loads and rayleigh waves,” Soil Dynamics and Earthquake
Engineering, vol. 114, pp. 147–153, 2018. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0267726118303968
[31] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy,
D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J.
van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J.
Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W.
Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henrik-
sen, E. A. Quintero, C. R. Harris, A. M. Archibald, A. H. Ribeiro,
F. Pedregosa, P. van Mulbregt, and SciPy 1.0 Contributors, “SciPy 1.0:
Fundamental Algorithms for Scientific Computing in Python,” Nature
Methods, vol. 17, pp. 261–272, 2020.
[32] A. Chambolle, R. De Vore, N.-Y. Lee, and B. Lucier, “Nonlinear wavelet
image processing: variational problems, compression, and noise removal
through wavelet shrinkage,” IEEE Transactions on Image Processing,
vol. 7, no. 3, pp. 319–335, 1998.
[33] I. Daubechies, M. Defrise, and C. De Mol, “An iterative thresholding
algorithm for linear inverse problems with a sparsity constraint,”
Communications on Pure and Applied Mathematics, vol. 57, no. 11,
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 13

Siyuan Yuan is a Ph.D. candidate in Geophysics Hae Young Noh is an Associate Professor in the
at Stanford University. His research involves digital Department of Civil and Environmental Engineering
city applications of Distributed Acoustic Sensing at Stanford University. Her research focuses on
(DAS) repurposing pre-existing telecommunication indirect sensing and physics-guided data analytics to
fibers as large-scale sensors for urban traffic mon- enable low-cost non-intrusive monitoring of cyber-
itoring, near-surface imaging with vehicle-induced physical-human systems. She is particularly inter-
surface waves, car-based fiber mapping, and struc- ested in developing structures to be self-, user-, and
tural health monitoring. Additionally, he is interested surrounding-aware to improve users’ quality of life
in Machine Learning based speech enhancement and provide safe and sustainable built environment.
technologies. Before his Ph.D. study, he received a The results of her work have been deployed in a
M.S. degree in Civil Engineering from Stanford, and number of real-world applications from trains, to
a B.S. degree in Civil Engineering from Tongji University, China. the Amish community, to eldercare centers, to pig farms. Before joining
Stanford, she was a faculty member at Carnegie Mellon University. She
received her Ph.D. and M.S. degrees in Civil and Environmental Engineering
and the second M.S. degree in Electrical Engineering at Stanford University.
She earned her B.S. degree in Mechanical and Aerospace Engineering at
Cornell University. She received several awards, including the Google Faculty
Research Awards (2013, 2016), the Dean’s Early Career Fellowship (2018),
the NSF CAREER Award (2017), and various Best Paper Awards from ASCE,
ASME, ACM, IEEE, and SEM conferences.

Robert Clapp grew up in the northwest, spending most of his childhood

in Eugene, Oregon. In 1993 he received my B.Sc. (Hons.) in Geophysical
Engineering from Colorado School of Mines. He then began graduate studies
Martijn van den Ende received his BSc and MSc in the Geophysics Department at Stanford University. He joined the Stanford
degrees in Earth Sciences in 2011 and 2013, respec- Exploration Project (SEP) headed by Jon Claerbout and Biondo Biondi.
tively, from Utrecht University, the Netherlands. In He received his Masters in 1995 and his Phd in 2000. The title of his
2018 he received a PhD degree in fault mechanics thesis was Migration velocity analysis with geologic constraints. Since 2000
from the High Pressure and Temperature laboratory he has held various titles (PostDoc, Research Associate, and now Senior
at Utrecht University. Since then, he has joined Research Engineer) within the Geophysics Department. He now works half
the Géoazur laboratory (Université Côte d’Azur) time with SEP and half the time with the Center for Computational Earth and
to work on earthquake cycle numerical modelling, Environmental Science (CEES). His current research interests include velocity
seismic array processing, and Distributed Acoustic analysis, imaging, computational interpretation, high-performance computing,
Sensing (DAS), aided by Deep Learning. His latest hardware accelerators, and parallel computing. In 2006 he received the J.
research involves the identification and analysis of Clarence Karcher award from the SEG. He is a member of the SEG and
(micro)earthquakes recorded with submarine DAS along the Chilean margin, AGU.
paving the way for the world’s first DAS-based earthquake early warning
system.

Cédric Richard (Senior Member, IEEE) received

the Dipl.-Ing. and the M.S. degrees in 1994, and
the Ph.D. degree in 1998, from the Compiègne
University of Technology, Compiègne, France, all
in electrical and computer engineering. He is a Full
Professor with Université Côte d’Azur, Nice, France.
He is the author of more 300 journal and conference
papers. His current research interests include statis-
tical signal processing and machine learning.
Dr. Richard is the Director of the French federa-
tive CNRS research association ISIS (Information,
Signal, Image, Vision). In 2010–2015, he was distinguished as a Junior
Member of the Institut Universitaire de France.
Jingxiao Liu is a post-doctoral fellow of Geophysics Dr. Richard, since 2020, has served as a Senior Area Chair for the IEEE
at Stanford University. He received his Ph.D. in the Signal Processing Letters and, since 2019, as an Associate Editor for the
Department of Civil & Environmental Engineering IEEE Open Journal of Signal Processing. In 2015–2018, he was also a Senior
with a Ph.D. minor in Electrical Engineering at Area Chair for the IEEE Transactions on Signal Processing, and an Associate
Stanford University. His research focuses on struc- Editor for the IEEE Transactions on Signal and Information Processing over
tural health monitoring, smart infrastructure systems, Networks. He was an Associate Editor for the IEEE Transactions on Signal
and smart city applications integrating structural dy- Processing (2006–2010). He is the Vice-Chair of the IEEE Signal Processing
namics, signal processing, physics-guided machine Theory and Methods Technical Committee. In 2019–2020, Prof. Richard
learning, mobile sensing, and fiber-optic sensing served as the Director-at-Large of Region 8 (Europe, Middle East, and Africa)
techniques. He received his M.S. in Civil Engineer- of the IEEE Signal Processing Society (IEEE-SPS) and as a Member of the
ing from Carnegie Mellon University, and his B.S. Board of Governors of the IEEE-SPS.
in Civil Engineering from Central South University, China. He received the
Leavell Fellowship on Sustainable Built Environment and various best paper
and presentation awards from ASCE, ASME, and ACM conferences.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2022 14

Biondo Biondi is the Barney and Estelle Morris

Professor of Geophysics at Stanford University. He
is director of the Stanford Exploration Project (SEP).
SEP is an academic consortium whose mission is to
develop innovative seismic imaging methodologies
and to educate the next generation of leaders in
applied seismology.
Biondo and his students devise new algorithms
to improve imaging of active and passive seismic
data. Images obtained from seismic data are the
primary source of information on the structural and
stratigraphic complexities in Earth’s subsurface and on many subsurface
dynamic processes. These images are constructed by processing seismic
wavefields recorded at the Earth’s surface and generated by either active-
source (e.g., vibroseis trucks) experiments or by natural (e.g., ocean waves)
and anthropogenic (e.g., vehicle traffic) sources. Because our datasets are
enormous, and wavefield propagation needs to be accurately modeled to
achieve high-resolution imaging, we need to harness the power of the
latest computational hardware to test our methods on field data. Therefore,
mapping imaging algorithms into high-performance architecture is an essential
component of our research. The amount and quality of information that
we can extract from seismic data are directly linked to the temporal and
spatial sampling of the sources and the receivers. In the past several years,
we have been working on methods to process data recorded by using fiber
cables as seismic sensors. Fiber-optic seismic recording promises to enable
cost-effective continuous seismic monitoring at a large scale. A particularly
exciting possibility is leveraging preexisting telecommunication infrastructure
to record seismic data with dense arrays in urban environments continuously.
In 2016 we pioneered that idea by recording data under the Stanford campus.
Since then, we recorded data in San Jose and on a 48-km array under Stanford
and neighboring cities.