This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE SYSTEMS JOURNAL 1

Floating Sensor Networks for River Studies

Andrew Tinka, Student Member, IEEE, Mohammad Rafiee, and Alexandre M. Bayen, Member, IEEE

Abstract—Free-floating sensor packages that take local mea- B. Environmental and Mobile Sensing
surements and track flows in water systems, known as drifters, The physical properties of large water systems can be
are a standard tool in oceanography, but are new to estuarial and
riverine studies. A system based on drifters for making estimates measured using several different sensor types and modalities.
on a hydrodynamic system requires the drifters themselves, Sensors are often categorized as Eulerian or Lagrangian (using
a communication network, and a method for integrating the terminology from fluid mechanics) according to whether they
gathered data into an estimate of the state of the hydrodynamics. observe the medium as it flows past a fixed location (Eu-
This paper presents a complete drifter system and documents lerian) or are embedded into the flow itself, measuring the
a pilot experiment in a controlled channel. The utility of the
system for making measurements in unknown environments medium while moving along a trajectory (Lagrangian). The
is highlighted by a combined parameter estimation and data canonical Lagrangian sensor is a small floating package that
assimilation algorithm using an extended Kalman filter. The transmits its location, and possibly other sensor measurements,
performance of the system is illustrated with field data collected as it is carried by the water current through the system.
at the Hydraulic Engineering Research Unit, Stillwater, OK. The oceanographic community calls such sensors drifters.
Index Terms—Data assimilation, hydrodynamics, Kalman While most infrastructural sensing in rivers and estuaries is
filters, sensor systems and applications. implemented using Eulerian sensors, the evolution of wire-
less sensor network technology has increased the interest in
I. Introduction novel Lagrangian sensor systems. The relative benefits of
Lagrangian sensors compared to Eulerian sensors can be clas-
A. Freshwater Systems sified into two categories: logistical benefits and information
The majority of the renewable freshwater available for benefits.
human use flows through rivers [1]. Human freshwater demand The logistical benefits of a Lagrangian sensor system derive
will increase significantly in the next 50 years, due mainly to from its flexibility and redeployable nature; in other words,
population increase, urbanization, and increased use of water- intrinsic benefits of self-contained devices designed for au-
intensive agriculture [2]. Modeling and monitoring the flow of tonomous operation. A fleet of drifters can be deployed, re-
freshwater, and the mixing and transport of constituents such covered, and redeployed in response to changing needs or new
as salt, can lead to improvements in water use efficiency and information. Their wireless communication allows them to be
can help balance supply and demand [3]. Specific examples of used in remote locations where power and communication
environmental management scenarios requiring understanding infrastructure may not be available. These advantages are not
of complex hydrodynamic systems include predicting the inherent to the Lagrangian or Eulerian distinction; it would be
movement of silt disturbed during dredging and underwater possible to build an Eulerian sensor that was battery-powered,
construction operations, planning reservoir release and gate communicated using wireless networks, and could be easily
control policies to affect the intrusion of salt water based on redeployed. Rather, these logistical advantages are between
specific local needs, and assessing vulnerabilities to contami- Lagrangian systems as they must be implemented compared
nant spills or other unforeseen events in critical water resource to Eulerian sensing as it is practised today.
regions. In each of these examples, high-quality hydrodynamic The information benefits of mobile sensing, however, are
models, based on data gathered from the actual system, can unique to the Lagrangian or Eulerian split. By following the
be crucial for responsible environmental policy and decision- flow of water, Lagrangian sensors determine the particle out-
making. comes of water in the system. An Eulerian sensor, observing
Manuscript received February 21, 2011; revised September 14, 2011; the water as it flows past, can (normally) not infer anything
accepted December 2, 2011. This work was supported by NSF Awards CNS- about the water’s history: where it came from, or where
0615299, CNS-0915010, and NSF CAREER Award CNS-0845076. The work it will end up. Tracking movement of water is particularly
of A. Tinka was supported by NSERC.
A. Tinka is with the Department of Electrical Engineering and Com- important for studying the movement of contaminants or other
puter Sciences, University of California, Berkeley, CA 94720 USA (e-mail: constituents, especially in regions with complex topology, such
tinka@berkeley.edu). as an estuary or a delta. Constituent transport is governed
M. Rafiee is with the Department of Mechanical Engineering, University of
California, Berkeley, CA 94720 USA (e-mail: rafiee@berkeley.edu). by processes including advection and diffusion [4]; the La-
A. M. Bayen is with the Department of Electrical Engineering and Computer grangian framework helps disambiguate the two, and allows
Sciences and the Department of Civil and Environmental Engineering, Univer- investigation into the precise location of interfaces or rapid
sity of California, Berkeley, CA 94720 USA (e-mail: bayen@berkeley.edu).
Color versions of one or more of the figures in this paper are available changes in concentration. One example of a hydrodynamic
online at http://ieeexplore.ieee.org. phenomenon of interest where Lagrangian drifters are relevant
Digital Object Identifier 10.1109/JSYST.2012.2204914 is tidal trapping, in which phase lags in tidal flow cause “dead
1932-8184/$31.00 
c 2012 IEEE
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE SYSTEMS JOURNAL

zones” where constituents can be trapped and released after a The first drifter that could actively communicate its position
delay [5], [6]. back to researchers was the “swallow float,” invented by
Lagrangian sensors do have some disadvantages in river J. Swallow in 1955 [15]. It was a neutrally buoyant float that
environments. Not all locations are suitable for deploying would drift approximately 1000 m underwater while transmit-
drifting sensors. Rapids and waterfalls have the potential to ting acoustic pulses that would be received by researchers’
damage these devices. Rivers can contain obstacles that can hydrophones. Development of drifters with acoustic commu-
capture drifters. The drifters must be retrieved at the end of nication capabilities continued in the 1960s and 1970s [16]. In
a deployment, which can be a difficult procedure if they are 1978, the introduction of the Argos satellite service [17] gave
scattered over a wide area (or snagged on different obstacles oceanographic researchers a global location and data uplink
over a long stretch of river). The suitability of an environment system, which lead to the development of oceanographic
for drifter studies must be assessed prior to drifter deployment. drifters that could communicate their position and sensor data
during the mission. Examples of oceanographic drifters that
C. Data Assimilation leverage the Argos system include the Davis, i.e., coastal
River hydraulics can be modeled with shallow water equa- dynamics experiment drifter [18], the Ministar, i.e., world
tions in one or two dimensions [7]. Shallow water equations ocean climate experiment drifter [19], and the low cost tropical
are a standard constitutive model used in the environmental drifter [20], each developed in the mid-1980s.
engineering community and the hydraulics community to Recent work in sensor networks for aquatic sensing mis-
model river flow; they are commonly used for simulation sions included the AMOUR Project at the Massachusetts
and control. When dealing with experimental measurements, Institute of Technology (MIT), Cambridge [21], the NEP-
algorithms are required to incorporate them into a model. TUS framework of AUVs at the Laboratório de Sistemas
One such technique is data assimilation, which is the process e Tecnologia Subaquática, Porto, Portugal [22], submersible
of integrating measurements into a flow model, and which pneumatic drogues built at the University of California, San
originated in meteorology and oceanography [8]. Diego [23], the Slocum underwater drifters at MBARI [24],
Most data assimilation methods can be placed into the and the SmartBay Sensor Network Project, Galway Bay,
historically named categories of variational or sequential Ireland [25].
assimilation methods [9]. Variational assimilation methods Although river and estuarine locations pose unique chal-
perform a single optimization step on all the observed data lenges, using drifting sensors in these environments is an
to minimize a cost functional. By contrast, sequential assim- emerging line of research. Other efforts include a low-cost
ilation methods, such as the Kalman filter and its extensions, floating GPS sensor [26] and a drogue carrying an acoustic
perform a series of update and analysis steps, blending the profiler [27]. Deployment scenarios for drifting sensors usually
observed data into the state estimate one step at a time. Several involve deploying them at a specific location, allowing them
extensions of the Kalman filter are applicable to nonlinear to propagate through the environment with the water currents,
systems. Examples include: the extended Kalman filter [10], and retrieving them at the end of the mission. The retrieval
which uses the Jacobian of the state update equation to update operation is usually assisted by the device transmitting its
the estimate of the mean and covariance of the state, the location to the research team.
ensemble Kalman filter [11], which tracks the evolution of The Floating Sensor Network (FSN) Project at the Univer-
a number of random samples in order to update the various sity of California, Berkeley (UC Berkeley) [28] designs and
estimates, and the unscented Kalman filter [12], which also builds drifters for riverine and estuarine environments. An ear-
tracks an ensemble of samples, but generates those samples lier generation with less developed capabilities was described
using a deterministic technique in order to accurately track in [29]. This paper gives a full system-level description of the
the mean and covariance with a minimal sample set. second generation system; the improvements in this version
This paper presents a data assimilation method based on include two reliable communication systems and new water-
the extended Kalman filter. Sequential assimilation methods quality sensing capabilities. The system described herein is
are well suited to real-time assimilation, which is one of the the first system built by the FSN Project that is capable of
future goals for this system. The extended Kalman filter is practical, unsupervised field deployments.
appropriate for nonlinear systems where the Jacobian is easy
to compute, which will be seen in Section III. E. Purpose and Organization of This Paper
This paper will describe the design and implementation of
D. Drifters in Oceanography and Hydrology a system for gathering data from multiple drifting sensors in
Although studies of flotsam drift (drawing inferences about a riverine or estuarial environment and assimilating it into a
currents from the observed movement of accidentally dropped model that can be used to estimate the state of the natural
material) can be found in antiquity, the first deliberate drifter environment. The design problem covers multiple domains,
study seems to be the work of G. Aimé circa 1845 [13]. including the mechanical design of the drifting sensor, the
His first drifters were drift bottles: sealed bottles containing selection and systems integration of the functional electronic
a message asking the eventual recipient to report the date components, an extensible and research-capable embedded
and location found. Drift bottle studies became a widely used computation capability on the individual sensors, the commu-
technique in European oceanography around the beginning of nication architecture for gathering the data from the field, and
the 20th century [14]. the software for the data assimilation problem on the back
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TINKA et al.: FLOATING SENSOR NETWORKS FOR RIVER STUDIES 3

end. The interdependencies between these domains make this

a system design problem.
As outlined in Section I-D, drifter-based sensing has
precedents in oceanography and other environments, but is a
novel approach in the estuarine environment. The purpose of
this paper is to describe the design problem and the approach
taken for this implementation, and to demonstrate that a
drifter-based system in riverine and estuarine environments
is feasible and useful, based on a pilot study performed in a
controlled environment.
The remainder of this paper is organized as follows. Sec-
tion II describes the design and implementation of the sec-
ond generation FSN, including the sensor hardware, sensor
firmware, communication architecture, and central server soft-
ware. Section III presents a deployment performed in Novem- Fig. 1. Overview of the drifter hull. Left: closed. Right: open. (a),
ber 2009 at the USDA-ARS Hydraulic Engineering Research (c) Fiberglass components. (b) Aluminum components. (d) Polycarbonate
top cap. (e) Battery and water quality sensor. (f) Sensor interface board.
Unit (HERU), Stillwater, OK. The data assimilation techniques (g) Main electronics. (h) Antennas.
used to integrate the gathered data into a model of the system,
and the validation used to evaluate the system performance,
are presented. Finally, Section IV presents conclusions and
plans for future improvements to the FSN system.

II. System Description

A. Hardware (Sensor)
The overall form of the drifting sensor is a waterproof
floating hull that encloses an electronics package.
The design considerations for the hull form include the
following.
1) Present a symmetric, nonrectifying drag to planar cur-
rents.
2) Expose modular sensor packages to the water while
providing mechanical protection.
3) Minimize wind drag exposure while placing multiple
antennas above water surface.
4) Allow easy opening and closing while still providing
waterproofing.
5) Survive rough handling, collisions, being thrown from Fig. 2. Cross-sections of methods to seal two fiberglass cylinders at an
shore, bridges, boats. aluminum bulkhead. Left: radial. Right: axial.
6) Use materials and construction techniques to permit
assembly in small numbers in a university laboratory The waterproof seal is made using machined aluminum
at low cost. and rubber O-ring seals. Fig. 2 shows two sealing schemes
The simplest way to satisfy 1) is to be as symmetric as that were tested with prototypes. The axial sealing scheme is
possible about a vertical axis. A cylinder is a good choice generally preferred for two reasons: the O-rings are less likely
because a commercially manufactured pipe can be used for the to be damaged during the opening or closing operations, and
body, satisfying 6). A vertical cylinder can also easily satisfy redundant protection can be added by adding more O-rings in
goals 2) and 3). the major seal without major mechanical changes. However,
The hull is manufactured at UC Berkeley using low-cost, the radial sealing scheme is more robust to damage and does
small-run manufacturing techniques. The drifter has a vertical not require tight tolerances in the aluminum parts. The pro-
cylinder configuration in order to present a uniform profile to totype cylindrical sealing systems were prone to failure when
surface currents while also supporting the antennas a small the outer cylinder was knocked out of round through rough
distance above the waterline. The hull consists of four major handling. The radial sealing system was the final selection.
components, shown in Fig. 1: (a) a hand-cast fiberglass lower Fig. 3 illustrates the position and mass of the major com-
hull; (b) machined aluminum parts for the watertight seal; (c) a ponents in the drifter. When floating at the desired waterline,
commercially available fiberglass pipe for the upper hull; and the drifter displaces 2.8 L of water. Before the battery, the
(d) a vacuum-formed polycarbonate top cap. The lower hull is mass of the drifter is 1.98 kg. This determines the mass of
flooded so that water quality sensors mounted in the bulkhead the ballast, which must be located as low as possible in order
may contact the water but also be mechanically protected. for the center of mass to be below the center of buoyancy
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE SYSTEMS JOURNAL

Fig. 5. Module-level block diagram of drifter electronics.

an NPT thread, are available in the main bulkhead. These

holes can either be sealed with a plug or hold a bushing in
Fig. 3. Schematic of major mass components.
which a desired water quality sensor is embedded. The contact
side of the sensor is exposed to water in the flooded lower
hull while still being protected from impact and debris by
the lower fiberglass hull; the electronics side of the sensor is
in the dry upper hull. The Omega CDH222 temperature or
electroconductivity sensor was selected as the first sensor to
be carried on the drifter.
The electronics are mounted near the top of the cylinder. See
Fig. 5 for a block diagram of the major modules. The architec-
ture was inspired by the Gumstix and Robostix products sold
by Gumstix, Inc., Portola Valley, CA, in which low-level tasks
are handled by a subordinate microcontroller while nonreal-
time, higher-level tasks such as communications are governed
by an embedded computer.
Communication between modules uses a variety of serial
busses: standard RS-232 serial communications, an inter inte-
grated circuit (I2C) bus, and a serial peripheral interface (SPI)
bus. The GPS receiver, GSM module, and embedded computer
Fig. 4. Three examples of water quality sensors with the desired form factor.
are on the main electronics printed circuit board, labeled (g) in
Foxboro 871DO [30], Omega CDH222 [31], Sensorex S8000CD [32]. Fig. 1. Antennas for the GPS and GSM modules, and a short-
range 802.15.4 2.4 GHz radio, are located at the top of the
(a necessary condition for the drifter to float in the desired hull (h). The GPS, 802.15.4, and GSM modules communicate
vertical configuration). Two components make up the ballast: with the embedded computer with individual RS-232 ports. A
a lithium ion battery at the bottom of the dry hull, and an subordinate microcontroller for real-time tasks such as sensor
encapsulated cast lead weight at the bottom of the wet hull. management is located on a lower board (f). The microcon-
The standard configuration uses a 200 g battery and a 600 g troller and embedded computer communicate with each other
lead weight. The battery and water quality sensor are labeled using the I2C bus. The temperature and electroconductivity
(e) in Fig. 1. sensors are interfaced to an analog-to-digital converter, which
One of the primary functions of the device is to carry water delivers digitized sensor readings to the microcontroller via
quality sensors. There are a wide variety of sensor modalities the SPI bus. In every case except the I2C bus, the choice of
that are of interest to researchers; cost and mass limitations communication protocols was determined by the modules.
preclude carrying all possible sensors at once. The solution is The GPS receiver is the Magellan AC-12 OEM module.
to adopt a modular design that allows different sensors to be When uncorrected (not using differential correction, SBAS, or
loaded into the body of the drifter. post-processing), its circular error probable range is 1.5 m. In
Many commercially available sensors are available in form addition to GPS coordinates, it also provides pseudorange and
factors designed either for installation into process pipes or carrier phase data, which can be used for post-processing or
handhold laboratory use. Examples can be found in Fig. 4. differential GPS techniques [33].
In all three cases, the sensors can be roughly described as Long-range communication with the server is performed
cylinders of diameter less than 25 mm and length of approxi- using the Motorola G24 GSM module. In areas with GSM
mately 120 mm. This representative sensor size was used for coverage, the General Packet Radio Service (GPRS) can be
the modular design. Two holes, 25.4 mm (1 in) diameter, with used to send TCP or UDP packets to servers on the Internet.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TINKA et al.: FLOATING SENSOR NETWORKS FOR RIVER STUDIES 5

Data rates depend on the GSM base station configuration, but

are at least 8.0 kbit/s upload and download [34].
Short-range communication between drifters, and between
drifters and field personnel, is performed with the Digi XBee-
PRO ZB module. Using the IEEE 802.15.4-2006 protocol [35],
these devices can form ad hoc mesh networks. The PRO
module can transmit with 50 mW (17 dBm) of power [36];
connectivity at distances of up to 1 km in river environments
has been observed when using these modules.
The embedded computer is a Gumstix Verdex Pro XM4,
a 20 mm × 80 mm single-board computer with a Marvell
PXA270 400 MHz processor and 64 MB of RAM. The
PXA270 is an applications processor designed around the
ARMv5 architecture [37]. Fig. 6. Software modules. Data are produced by the sources, coordinated by
the data− central process, and transmitted by the data sinks. Other processes
One relevant characteristic for designers of embedded sen- provide debugging and maintenance functions, as well as local storage.
sor systems is that the PXA270 does not have the hardware
floating point capability, which may make it difficult to effi- The software was decomposed by function into software
ciently implement intensive signal processing or other com- modules, as shown in Fig. 6. Each module was implemented as
putations. The Verdex is developed to run an OpenEmbedded a separate process, written in C. Each process was designed to
Linux distribution. For real-time tasks, such as collecting data run indefinitely; each one was launched from a shell script that
from sensors, an Atmel ATmega128L microcontroller [38] is would restart it in the event that it did terminate (which would
used. Following the Robostix architecture, an I2C link carries represent a bug). Many-to-one interprocess communication
data between the embedded computer and the microcontroller. (IPC) (represented by green links in Fig. 6) was implemented
In retrospect, this was a poor choice for the drifter implemen- using message queues, a seldom-used System V feature [39].
tation, due to the large distance between the upper and lower One-to-many IPC (represented by blue links in Fig. 6) was
boards. An RS-232 link would have been a better choice. implemented using fanout first in, first outs (FIFOs), a custom
The design of any field-deployed sensor system must take kernel module [40]. Many modern embedded software projects
power consumption into account. The size of the battery and in Linux environments use a Unix socket or TCP sockets
the power consumption of the electronics set the maximum for IPC [41]; message queues and fanout FIFOs, however,
mission time. For this reason, the computational capacity of have one very attractive feature: messages are not lost if
the device is in tension with the overall mission time of the the receiving process crashes. In the event that the receiving
system. The design of this system prioritized research flexi- process has terminated, messages are buffered (up to an OS-
bility and capabilities over long mission times; accordingly, a defined limit); if a new receiving process starts and connects
highly capable embedded computer was chosen, even though to the message queue or fanout FIFO, it can recover all the
the basic functionality of a Lagrangian sensor could be accom- queued messages that accumulated during the downtime.
plished with a much simpler, low-power microcontroller. The Data are stored on a 1 GB MicroSD card installed on the
estimated mission lifetime under moderate power discipline Verdex. Data are stored in two formats: a text-based log that
is 48 h, which is sufficient for most estuarial environments. records events and debugging information, and an sqlite3
Section IV-C describes some of the advanced capabilities database [42] for random access during missions. Section II-C
that might one day be possible with Lagrangian sensors describes how the sensor data are transmitted in real time to
with powerful onboard computation, which will hopefully be back-end servers; the local storage in this case is merely a
explored in successive studies. backup.
The following is a list of software modules or processes
B. Software (Sensor) described in Fig. 6.
The Verdex embedded computer has vastly more computa- 1) logrecorder: It receives generic event messages from
tional power than is needed for a record-and-transmit mission. other modules and records them on the flash storage.
The long-term vision of the FSN project is for a fleet of 2) parse− gps: It receives raw National Marine Electronics
sensors that not only measure the physical phenomenon but Association feed from the AC-12 GPS module, extracts
also perform real-time distributed computations with the data; useful information, performs basic conversions, and pro-
the computational requirements of the device were laid out vides location information.
with this goal in mind. The major design criterion for the 3) data− central: It receives data from any module that
software architecture was reliability. The development cycle generates it, combines all data into a single stream, and
precluded using formal methodologies for verifying software exports it to the database and to any other module that
correctness, and so the development team proceeded under uses information. If a publish or subscribe system [43]
the assumption that the software they developed would always were used instead, this module would be the dispatcher.
contain uncorrected bugs. Their approach was to exploit the 4) salinity− i2c: It communicates with the Atmega mi-
Linux operating system features for as many architectural crocontroller over the I2C bus, receives raw temper-
features as possible. ature and conductivity data from the water quality
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE SYSTEMS JOURNAL

sensor, processes it into salinity data, and exports

it.
5) xbee− gumstix: It manages the XBee radio, periodically
transmits important state information (see Section II-C),
verifies incoming traffic, and pipes it to a shell.
6) g24− manage: It manages the G24 GSM module, con-
nects and disconnects the GSM module from the net-
work according to a mission-specific power management
policy, opens TCP connections to a central server, and
sends state information periodically.
7) cpu− track: It periodically gathers information on CPU
state (utilization, free memory, and so on) and exports
it.
8) log: It is a utility program to insert text manually into
the log stream and is useful for recording events from
scripts (such as program crashing and being restarted).
9) read− sensor: It is a utility program to print out all Fig. 7. Communication architecture. Drifters communicate amongst them-
current sensor information for debugging purposes. selves and with field teams via 802.15.4; both the field laptop and the drifters
can also communicate with the back-end servers using GPRS.
C. Communication Architecture
Any communication channel should make provision for
Sharing information between nodes in a network involves validation of the data being transmitted. In this case, the
a process called serialization: converting structured data into underlying protocols (TCP/IP and the ZigBee protocol) have
a stream of bytes suitable for transmission over a digital data integrity features. It was therefore unnecessary to include
communication link in a way that is efficient and can be such features into the serialization protocol.
unambiguously deserialized on the other end. In general, there There are three ways in which data collected in the field
are three approaches to serialization: make it back to the back-end servers for processing.
1) a custom, byte-compact binary format; 1) The drifter uses the GSM module to open a TCP
2) a generic framework such as XML [44], abstract syntax connection during the mission directly to the server and
notation (ASN) [45], or Google protocol buffers [46]; delivers data.
3) a custom, textual, human-readable format. 2) A field team monitors the drifter via the ZigBee network
These methods vary in data efficiency, ease of development, during the mission and uploads the collected data to the
ease of debugging, and complexity of supporting software. For server.
example, a custom byte-compact format will be very efficient 3) At the end of the experiment, the drifters are collected,
in terms of the length of messages needed to represent data, but and their logs and databases retrieved from their onboard
can be very difficult to debug. Generic frameworks with binary MicroSD cards and uploaded.
output, such as ASN, provide efficient messages, and the Methods 1) and 2) are illustrated in Fig. 7. Method 3) will
reliability of standardized packaged software, but may be too not be further discussed in this paper. The project goal is to
heavyweight for some applications. Given the relative simplic- build a system that delivers and processes data in real time;
ity of the data structures shared in this project, the developers the end-of-experiment uploading is useful for verification and
decided upon the third option, a custom human-readable tex- early research, but is not the project’s direction.
tual format. Messages were serialized in the following format A field team communicates with deployed drifters using
field− name/value/field− name/value a directional 2.4 GHz antenna connected to an XBee module.
The field laptop runs a program called xbee− netbook, very
with newline terminating a message. The possible fields are: similar to the xbee− gumstix running on the drifters. This
1) id: drifter ID number; program listens for the broadcasts of drifters, adds the mes-
2) ts: time stamp, seconds since Unix epoch; sages to a local version of the database, and synchronizes the
3) x− cm: UTM x (easting) coordinate, cm; database with the home server via a GPRS/GSM connection.
4) y− cm: UTM y (northing) coordinate, cm; Although the study described in Section III was performed
5) zn: UTM zone; over distances where ZigBee communications were easy, most
6) vel− x− cm: velocity, x component, cm/s; studies in open river environments will often not permit short-
7) vel− y− cm: velocity, y component, cm/s; range wireless communications. Both the distance between the
8) sats: number of GPS satellites locked; base station and the drifting sensors, and the attenuation due
9) sal: salinity recorded; to intervening obstacles such as vegetation or land masses,
10) temp: temperature recorded; mean that the low powered ZigBee radios may not succeed at
11) cpu− 1, cpu− 5, cpu− 15: CPU utilization over a 1, 5, 15 sending data. ZigBee radios are capable of mesh networking,
min window; meaning that a distant drifter could communicate with a base
12) mem− free: free memory, kB. station through a multi-hop path of other drifters, but the same
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TINKA et al.: FLOATING SENSOR NETWORKS FOR RIVER STUDIES 7

factors that make single-hop communications difficult can also

break connectivity over multiple hops. For this reason, the
drifter system includes both the short-range ZigBee radio and
the GSM module for connectivity over longer distances. The
elevation of the GSM infrastructural tower, and the higher
transmitting power, allow for GSM connectivity over much
longer distances. Although this method of communication
is dependent on the infrastructure of the GSM provider, in
practical terms the GSM channel is a much more reliable
method for transmitting data from the drifter for assimilation
as well as logistical (i.e., retrieval) purposes.
Both the field laptop and the home server maintain a
MySQL database [47] with all collected data from the drifters.
A single message from a drifter consists of some or all of its
sensor information for a particular time instant.

D. Back-End Architecture Fig. 8. HERU facility, with experimental channel annotated. Image courtesy
of USGS. (a) Drifter release point. (b) Drifter recovery point. (c) Downstream
Once the data have been uploaded from the field units gate.
(drifters or field netbook) to the MySQL database on the
UC Berkeley server, it can be used for real-time or post-
processing data assimilation and state estimation. The as-
similation software is usually implemented on a different
machine than the MySQL server. Remote query mechanisms
are used to fetch new data from the MySQL server to the
assimilation software. For the prototype implementation, the
assimilation software (described further in Section III-B) was
written in MATLAB and executed after the experiment was Fig. 9. Channel profile, including minimum and maximum water height.
completed (post-processing). As will be seen in Section III-B,
canal feeds a number of experimental units that are normally
the dominant computational bottleneck is the inversion of an
used for investigations into levee reliability, reservoir safety,
n×n matrix, where n is the assimilation state space. Therefore,
and spillway design [48]. For the experiment, drifters were
the computational time should scale with the third power of
deployed into the supply canal. The upstream boundary con-
the number of drifters plus the number of discretization points.
dition was the supply canal flow control, set to 1.42 m3 /s
Currently, the typical runtime for a complete assimilation job
(50 ft3 /s); the downstream boundary condition was a gate that
is 150 s to assimilate 450 s worth of data from six drifters,
could be raised or lowered to restrict the flow out of the
using a laptop with a 2.0 GHz Intel T2500 processor and
experimental region. Drifters were released at approximately
2 GB of RAM. This O(n3 ) scaling can easily be handled
30 s intervals near the upstream boundary in Fig. 8(a). After
on larger systems using computation clusters. This method
traveling through the canal for approximately 400 s, they were
is therefore feasible for real-time assimilation of appropriately
individually retrieved in Fig. 8(b). Fig. 8(c) marks the location
sized systems and drifter quantities.
of the downstream control gate.
Offline assimilation, as performed in the current system,
A total of 20 runs were performed, and divided into five
is a valuable tool for analyzing the hydrodynamics of a
cycles of four runs each. Each run in the cycle had a different
region and for identifying the phenomena that govern the local
operation of the downstream control gate. During the first
environmental behavior. Online assimilation, where data are
run, the gate remained open for the entire run. During the
processed as it becomes available to inform a real-time model,
second run, the gate was closed as soon as the sixth drifter
has additional applications including forecasting and real-
was released. During the third run, the gate remained closed.
time monitoring. Section IV-C discusses details of the FSN
Finally, during the fourth run the gate was opened as soon as
Project’s plans for the future scaling of the assimilation back-
the final drifter was released. The cycle was then repeated.
end, including development of online assimilation capabilities.
Fig. 9 shows the cross-section of the prismatic channel over
most of its extent.
III. Sample Deployment Because of the small experimental domain (and the low
probability of losing a drifter), the GSM modules were not
A. Mission Description activated in this experiment. Instead, GPS position and veloc-
In November 2009, an experiment was performed at ity readings were stored on a 1 GB MicroSD card installed
the USDA-ARS HERU, Stillwater, OK (see Fig. 8 for an on the Verdex and simultaneously transmitted over the XBee
overview). The HERU facility, located adjacent to Lake Carl radio to a nearby laptop, which uploaded them to the home
Blackwell, has a gravity-fed supply canal that can have a server using a database synchronization protocol over a single
controlled flow of up to 4.25 m3 /s (150 ft3 /s). The supply GSM link.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE SYSTEMS JOURNAL

with Q(x, t) = V (x, t)A(x, t) the discharge across cross-section

A(x, t), P the wetted perimeter, i.e., the perimeter of the wetted
portion of the cross-section, and m the Manning roughness
coefficient (sm−1/3 ).
The boundary conditions are usually taken to be the up-
stream flow Q(0, t) and the downstream stage H(L, t).
Applying the Lax diffusive scheme [51], [53] that is a first-
order explicit scheme to discretize the equations to (1) and (2),
the following set of finite difference equations are obtained:
1 t
Ak+1
i = (Aki−1 + Aki+1 ) − (Qk Qk ) (4)
2 2x i+1 i−1
1
Qk+1
i = (Qki−1 + Qki+1 )
2  k  2 k 
t Q2 Q
− + gAhc − + gAhc (5)
2x A i+1 A i−1
 k k 
Fig. 10. Two drifters in the HERU facility supply canal. φi+1 + φi−1
+t (6)
2
One disadvantage of this experiment was that there was no where
appreciable variation in water temperature, salinity, pH, or any
other water quality factor. The only interesting state variables φ = gA(Sb − Sf ). (7)
to measure and estimate were the velocity (equivalently, flow) This scheme is stable provided that the Courant–Friedrich–
and stage of the water. The Omega CDH222 salinity sensors Lewy condition holds, that is
were therefore not used in this experiment.
t
|V + C| ≤ 1 (8)
B. Assimilation Technique x
√
The goal of the data assimilation is to incorporate the drifter where C = gD is the wave celerity and V is the average
position data into a model of the flow in order to estimate the velocity.
state of the system, flow, and stage. The 1-D Saint–Venant The equations above may only be used for interior grid
equations are used as the model of the flow. After deriving a points. At the boundaries, these equations cannot be applied
nonlinear state-space model from the Saint–Venant equations, since there is no grid point outside the domain. Therefore,
the extended Kalman filter (EKF) is used to perform the another method needs to be used to compute the unknown
estimation. variables at the boundaries. The method of specified time
1) Derivation of the State-Space Model: The Saint–Venant intervals is used to compute these variables [53]. In this
model is among the most common models used for modeling method, after computing the characteristics, the boundary grid
the flow in open channels and irrigation systems [49], [50]. In point is projected backward to the previous time step along
the 1-D case, Saint–Venant equations are two coupled first- its corresponding characteristic curve. After computing the
order hyperbolic partial differential equations derived from variables at the projected point, which is usually done by using
conservation of mass and momentum. For prismatic channels linear interpolation, the characteristic equations are used to
with no lateral inflow, these equations can be written as compute the unknown variable at the boundary grid point at
follows [51]: the next time step.
The discretized equations obtained above can be used to
∂H ∂Q obtain a state-space model
T + =0 (1)
∂t ∂x
 2
∂Q ∂ Q ∂ xk+1 = f (xk , uk ) (9)
+ + (ghc A) = gA(S0 − Sf ) (2)
∂t ∂x A ∂x where xk is the state vector at time k
for (x, t) ∈ (0, L)× , where L is the river reach (m), Q(x, t)
+
xk = (Qk2 , . . . , QkN , H1k , . . . , HN−1
k
)T (10)
is the discharge or flow (m3 /s) across cross-section A(x, t) =
T (x)H(x, t), H(x, t) is the stage or water depth (m), T (x) is and the input uk contains the boundary conditions, i.e., the
the free surface width (m), D = A/T is the hydraulic depth m, upstream flow and downstream stage
Sf (x, t) is the friction slope (m/m), Sb is the bed slope (m/m),
g is the gravitational acceleration (m/s2 ), and hc is the distance uk = (Qk1 , HNk )T (11)
of the centroid of the cross-section from the free surface (m). where Qki and Hik are the flow and stage at cell i at time
The friction slope is empirically modeled by the Manning– kt, respectively, and N is the number of cells used for the
Strickler formula [52] discretization of the channel.
m2 Q2 P 4/3 Assuming that all model parameters are known, when
Sf = (3) measurements of the flow other than the boundary conditions
A10/3
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TINKA et al.: FLOATING SENSOR NETWORKS FOR RIVER STUDIES 9

where w is the channel width, d is the water depth, and Aq , Bq

and Cq are constants and κ = 0.4. Aq is commonly calculated
experimentally and (14) and (15) are used to compute Bq and
Cq . These equations are the constraint obtained from the facts
that FT (y) is zero at the sides of the channel and the average
FT (y) is equal to 1.
Denoting the collection of velocity measurements obtained
from the drifters at time step k by yk , the measurement model
Fig. 11. (a) Applied quartic function to the average velocity in the transverse.
(b) Applied Von Karman log function to the average velocity in the vertical
can be written as
(right). Figures adapted from [55].
yk = g(xk , k). (17)
are available, these measurements can be incorporated into Note that the observation operator g is time-varying since
the state-space model using one of the standard nonlinear the drifters are moving with the flow. Therefore, the cells at
filters, e.g., the extended Kalman filter. However, in practice, which the flow velocity is measured are changing over time.
it is sometimes impossible or expensive to obtain accurate
3) Stochastic State-Space Model and the Extended Kalman
values for one or more of these parameters. For instance, it
Filter: The effect of modeling uncertainties, as well as inaccu-
is usually a difficult task to obtain an accurate value for the
racies in measurements of the inputs, is commonly considered
bed slope of a channel. As will be shown in Section III-C,
as an additive noise term in the state equations (9) to obtain
the results of the model are very sensitive to the value of the
a stochastic equation
bed slope. If parameters are critical but a direct measurement
is not possible, proper experiments can be designed to obtain
xk+1 = f (xk , uk , wk ). (18)
measurements of the system and these measurements may be
used later to identify the unknown parameters. Nevertheless,
The noise wk is usually assumed to be zero-mean white
it is sometimes not possible to carry out these types of
Gaussian and
experiments beforehand due to time constraints, lack of proper
equipment, high costs, and so on. E[wk wTl ] = Qk δkl . (19)
In order to obtain estimates of the unknown parameters of
the system, a vector of unknown parameters vk is appended x0 ∈ Rm is the initial state that is also assumed to be Gaussian
to the state vector; these parameters are considered to have and
zero dynamics (vk+1 = vk ). A nonlinear filter can then be
applied to the augmented state-space model to simultaneously x0 = N (x̄0 , P0 ) (20)
estimate the parameters and the actual state of the system. As
the estimation of the state (that now includes the parameters as where x̄0 and P0 are the initial guesses for state and error
well) progresses, the estimates of the parameters get updated covariance.
by incorporating the new measurements. Note that in the case of combined state-parameter estima-
2) Measurement Model: The drifter velocity data provide tion, xk is the augmented state, i.e., the concatenation of the
local measurements of the flow velocity field. The relation actual state vector and the vector of unknown parameters.
between the drifter velocity and the flow at the corresponding Similarly, the errors and uncertainties in the measurements
cross-section relies on assumptions made about the profile of can be taken into account by adding a noise term to the
the water velocity. The surface velocity profile is assumed to measurement model (17) to obtain
be quartic, and the Von Karman logarithmic profile is assumed
in the vertical direction [54] as shown in Fig. 11. For a given yk = g(xk , ek , k) (21)
particle moving at a distance y from the center line and z
from the surface, the particle’s velocity vp (y, z) is related to where ek is the measurement noise of the sensors which is
the flow Q with the following equations: assumed to be zero-mean white Gaussian and
Q
vp (y, z) = FT (y)FV (z) (12)
A E[ek eTl ] = Rk δkl . (22)
with
 2  4 The process and measurement noises and the initial conditions
2y 2y are all assumed to be independent.
FT (y) = Aq + Bq + Cq (13)
w w In the EKF, the states of the system are approximated by
a Gaussian random variable and are propagated through a
Aq + B q + C q = 0 (14)
linearized approximation of the state equations. The prior
B q Cq
Aq + + =1 (15) mean of the state is fed into the state equations to yield the
3 5 prediction of the state. The posterior covariance matrices are
 
0.1   z  calculated for a linear model that is obtained from linearizing
FV (z) = 1 + 1 + log (16)
κ d the state equations around the current estimate [10].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE SYSTEMS JOURNAL

Fig. 12. Downstream stage (m).

With the stochastic state-space model given in the previous Fig. 13. (a) Flow and (b) stage at the tenth cell for Sb = 0.000 (solid),
Sb = 0.001 (dashed), Sb = 0.002 (dot-dashed).
section and the following notations:
x̂k|k−1 = E[xk |y0 , y1 , . . . , yk−1 ] (23) channel are unavailable, the bed slope of the channel cannot be
x̂k|k = E[xk |y0 , y1 , . . . , yk ] (24) calculated. In order to determine the sensitivity of the model
Pk|k−1 = E[(xk − x̂k|k−1 )(xk − x̂k|k−1 )T |y0 , y1 , . . . , yk−1 ] (25) with the given boundary conditions to the value of the bed
slope, the forward simulation is run with three different values
Pk|k = E[(xk − x̂k|k )(xk − x̂k|k )T |y0 , y1 , . . . , yk ]. (26)
of bed slopes. In each case, the initial condition is chosen to
The iterations of the EKF can be summarized as follows: be the backwater curve (steady state) that is computed using
Time update the following equations:
∂Q
x̂k|k−1 = f (x̂k−1|k−1 , uk−1 , 0) (27) =0 (34)
∂x
Pk|k−1 = k−1 Pk−1|k−1 Tk−1 + Bk−1 Qk−1 Bk−1
T
. (28) ∂H gA(S0 − Sf )
= (35)
Measurement update ∂x T b + 2H
−Q2 2 + g(T b H + H 2
)
H (Tb + H)2
Kk = Pk|k−1 GTk (Gk Pk|k−1 GTk + Dk Rk DkT )−1 (29)
where Tb is the bottom width.
ŷk = Gk x̂k|k−1 (30)
Fig. 13 shows the flow and stage at the tenth cell, as a
x̂k|k = x̂k|k−1 + Kk (yk − ŷk ) (31) representative cell, for the three values of the bed slope. It
Pk|k = (I − Kk Gk )Pk|k−1 (32) is not surprising to see that the results of forward simulation
vary significantly with different values of the bed slope.
where To implement the data assimilation method, the measure-
 
∂f  ∂f  ments obtained from the five drifters are used. Next, the
k−1 = Bk−1 = . (33)
∂x x̂k|k−1 ,uk−1 ∂w x̂k|k−1 ,uk−1 velocity of the sixth drifter is estimated using the estimated
flow that is compared with its actual value obtained from
the sixth drifter. Two methods are implemented: the extended
C. Numerical Results Kalman filter with, and without, estimating the bed slope.
This section presents the results of the implementation of Figs. 14 and 15 show the flow and stage at a few different
the data assimilation method on the data collected from the cells predicted by the forward simulation (i.e., state-space
experiment performed at the USDA-ARS Hydraulic Engineer- model) assuming that the bed slope is zero, estimated flow
ing Research Unit, Stillwater, OK, in November 2009. The and stage by performing the data assimilation method while
measurements used for data assimilation are the positions and the bed slope is assumed to be zero, and estimated flow
velocities of the drifters. and stage by performing the data assimilation method and
Fig. 12 shows the stage at the downstream end of the estimating the bed slope as an unknown parameter. As can
channel corresponding to a run used for evaluating the method. be seen in Fig. 12, the downstream stage starts to decrease
As can be seen in this figure, the downstream stage is initially at around time step 150 due to the gate opening. As can be
1.33 m and it starts to decrease as the downstream gate is seen in Fig. 14, the flow increases as a result of opening
opened until it becomes 0.92 m. the gate. It can be seen in Fig. 15 that the stage reduction
The discretization is done by dividing the channel into 60 caused by opening the downstream gate propagates backward
cells, each of approximately 5 m length. The temporal step size through the channel. However, in the case of assuming the
is chosen as 1 s. Since data about the bottom elevation of the bed slope as an unknown parameter, this reduction stage is
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TINKA et al.: FLOATING SENSOR NETWORKS FOR RIVER STUDIES 11

Fig. 14. Flow (m3 /s) at the 10th, 20th, 30th, 40th cells, forward simulation (dot-dashed), EKF with zero bed slope (dashed), and EKF with estimating bed
slope (solid).

Fig. 15. Stage (m) at the 10th, 20th, 30th, 40th cells, forward simulation (dot-dashed), EKF with zero bed slope (dashed), EKF with estimating bed slope
(solid).

more moderate. In particular, at cell 10, which is close to value. The peak in the measurement graph corresponds to
the upstream end of the channel, no decrease in the stage is when the sixth drifter is thrown into the water from the channel
seen. This is due to the fact that for a nonzero bed slope, bank with an initial speed. As can be seen in this figure, the
the backwater curve (steady state) is not uniform. Since the data assimilation methods significantly improve the estimation
initial estimate of the bed slope is taken to be equal to zero, results. Also, it can be seen that considering the bed slope as an
the extended Kalman filter is initialized by a uniform steady unknown parameter and using the measurements to estimate
state corresponding to a zero bed slope. However, as the it improves the estimation results. In order to quantify the
estimated bed slope deviates from zero, the steady state of performance of the methods, we calculate the relative error of
the system deviates from the uniform steady state accordingly. the estimated velocity of the sixth drifter at each time step
The time evolution of the estimated bed slope is illustrated in using the following formula:
Fig. 16. While the values of flow and stage estimated by the
data assimilation methods seem physically more reasonable,
it is not possible to formally evaluate the performance of (v̂k − vk )2
the method by looking at these figures. In order to obtain E(k) = × 100% (36)
(vk )2
a more quantifiable assessment of the method, the velocity of
the sixth drifter is calculated using the estimated flow at the
corresponding cell. The same velocity profiles on the surface where vk and v̂k are the true and estimated values of the
and along the depth as described in Section III-B2 are used to velocity of the sixth drifter at time step k.
calculate the drifter velocity from the estimated flow. Fig. 17 The relative error is calculated for all cases, and Table I
shows the velocity of the sixth drifter predicted by the forward provides the average relative error per time step corresponding
simulation, and both data assimilation methods and its actual to each case.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE SYSTEMS JOURNAL

Kalman filter-based assimilation algorithm for incorporating

the data into a 1-D shallow water model. While the system
as currently implemented performs the data assimilation after
the deployment phase of the experiment is complete, the EKF
algorithm is designed to process data incrementally, and is fast
enough to permit a real-time assimilation on the data as it is
being gathered.
These pilot studies show that the mobile sensor technique
for studies in river channels is feasible and can yield useful
information about the state of the system. Moreover, the use
of a simultaneous parameter estimation and data assimila-
tion technique demonstrates that this method is feasible for
environments where the parameters of the system (channel
geometry, bed slope, friction constants, and so on) have not
yet been measured or estimated. This is a strong advantage
when studying new hydrodynamic systems for the first time,
Fig. 16. Estimated bed slope (black) compared to the assumed bed slope
(dashed). and supports the known advantage of Lagrangian sensing
techniques in terms of flexibility and ease of deployment in
remote, unexplored, and unfamiliar regions.

B. Limitations of Drifters as Sensors

The major limitation of using a drifting sensor package to
estimate the velocity of the water is the variation in the velocity
of the water at different positions within the channel (the
“profile” of the velocity). The 1-D shallow water equations
model the cross-sectional average of the water velocity; the
raw velocity observed by a drifter will be different from
the average value. Determining the best possible estimate of
the average velocity based on the drifter’s local velocity is
nontrivial. Some aspects of the profile are easy to work with;
for example, it is known that the drifter is floating at the
water surface, and logarithmic depth profiles are a reasonable
assumption. Dealing with the lateral position of the drifter
within the channel is more challenging.
Fig. 17. Velocity of the sixth drifter: forward simulation (dot-dashed), EKF
with zero bed slope (dashed), EKF with estimating bed slope (dotted), and C. Future Work
the actual drifter measurements (solid).
The future plans for the FSN Project include improvements
TABLE I to the sensor hardware as well as to the data assimilation
Error Summary for Cross-Validation Study back-end. The next generation of the hardware will feature
a propulsion system that will allow some modifications of the
Method Relative Error drifter’s trajectory. Naturally, a propelled drifter is no longer
Forward simulation 139.2% an ideal Lagrangian sensor, so the use of this feature will
EKF with zero bed slope 53.5% have to be carefully considered. On the software side, the
EKF with est. bed slope 22.9% next generation system will be capable of performing real-time
assimilation; the data coming in from the sensors via GSM
IV. Conclusion will be integrated within minutes to provide a live estimate
of the state of the hydrodynamic system. The domains of
A. Feasibility of Real-Time Assimilation Using Drifters in future experiments will have more complicated topologies;
Rivers networks of branching and merging channels will require more
The FSN Project has designed, implemented, and tested sophisticated models than the 1-D shallow water equation, and
a complete system for gathering data from estuarial and the assimilation techniques will need to be refined to match.
riverine systems using independent passive floating sensor The problem of extracting the average water velocity from
packages known as “drifters” and assimilating these data the local drifter measurements will be investigated in future
into a hydrodynamic model of the system being studied. In work. Possible approaches include making profile parameters
addition to the drifters, the system incorporates the commu- part of the estimation process, using the GPS position to
nication infrastructure to transfer the information gathered estimate the lateral position of the drifter, and placing several
by the drifters to the field team and to databases on the drifters close together in the channel to provide a diversity of
back-end server. The system also incorporates an extended measurements.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TINKA et al.: FLOATING SENSOR NETWORKS FOR RIVER STUDIES 13

The future version of the data assimilation system will [15] J. C. Swallow, “A neutral-buoyancy float for measuring deep currents,”
Deep Sea Res. (1953), vol. 3, no. 1, pp. 74–81, 1955.
include an online server that will assimilate the data in real [16] W. J. Gould, “From swallow floats to argo: The development of neutrally
time as it arrives; this system will be implemented on a buoyant floats,” Deep Sea Research Part II: Topical Stud. Oceanography,
large parallel computational cluster. This future system will vol. 52, nos. 3–4, pp. 529–543, 2005.
[17] D. D. Clark, “Overview of the Argos system,” in Proc. OCEANS, vol. 3.
assimilate the data into the river, estuary, and land model, a Sep. 1989, pp. 934–939.
new 2-D model of the San Francisco Bay and Sacramento-San [18] R. E. Davis, “Drifter observations of coastal surface currents during
Joaquin Delta currently under development by the California CODE: The method and descriptive view,” J. Geophys. Res., vol. 90,
no. C3, pp. 4741–4755, 1985.
Department of Water Resources and the Lawrence Berkeley [19] P. P. Niiler, R. E. Davis, and H. J. White, “Water-following character-
National Laboratory [56]. An even more ambitious plan for istics of a mixed layer drifter,” Deep Sea Res. Part A. Oceanographic
future assimilation work is to use the onboard computation of Res. Papers, vol. 34, no. 11, pp. 1867–1881, 1987.
[20] D. S. Bitterman and D. V. Hansen, “The design of a low cost tropical
the drifter sensors to perform a distributed data assimilation, drifter buoy,” in Proc. Marine Data Syst. Int. Symp. (MDS), 1986, pp.
as opposed to transmitting the sensor data to back-end servers 575–581.
for centralized assimilation. [21] C. Detweiller, I. Vasilescu, and D. Rus, “An underwater sensor network
with dual communications, sensing, and mobility,” in Proc. OCEANS
2007 Eur., 2007, pp. 1–6.
Acknowledgment [22] P. B. Sujit, J. Sousa, and F. L. Pereira, “UAV and AUVs coordination
for ocean exploration,” in Proc. OCEANS 2009 Eur., 2009, pp. 1–7.
The authors would like to thank G. Hanson and his col- [23] Y. Han, R. A. de Callafon, J. Cortés, and J. Jaffe, “Dynamic modeling
leagues, USDA Hydraulic Engineering Research Unit, Still- and pneumatic switching control of a submersible drogue,” in Proc. 7th
Int. Conf. Informatics Control Autom. Robot., 2010, pp. 89–97.
water, OK, as well as D. Resio and D. Ward, Coastal and [24] P. Bhatta, E. Fiorelli, F. Lekien, N. Leonard, D. Paley, F. Zhang,
Hydraulics Laboratory, U.S. Army Engineer Research and R. Bachmayer, R. Davis, D. Fratantoni, and R. Sepulchre, “Coordi-
Development Center, Champaign, IL, for their assistance and nation of an underwater glider fleet for adaptive ocean sampling,”
in Proc. Int. Workshop Underwater Robot., 2005 [Online]. Available:
accommodation during the experiments at the Stillwater HERU http://publications.iot.nrc.ca/documents/IR/IR-2005-44.pdf
Facility. The Stillwater Experiment was performed with the [25] J. Ryan, N. Aonghusa, and E. Sweeney, “SmartBay, Ireland: Design and
assistance of L. Anderson, K. Weekly, J. Reilly, and S. Amin. planning for a cabled ocean observatory off the west coast of Ireland,”
in Proc. OCEANS, 2008, pp. 1–4.
J. Thai helped with the data assimilation work. Many talented [26] J. Austin and S. Atkinson, “The design and testing of small, low-cost
undergraduate engineering students have applied their skills GPS-tracked surface drifters,” Estuaries, vol. 27, no. 6, pp. 1026–1029,
and enthusiasm to the Floating Sensor Network Hardware Dec. 2004.
[27] C. M. Schacht and C. J. Lemckert, “A new Lagrangian-acoustic Drogue
Development Project. This version of the drifter would not (LAD) for monitoring flow dynamics in an estuary: A quantification of
have been possible without the efforts of C. Oroza, C. Foe- its water-tracking ability,” J. Coastal Res., vol. 50, pp. 420–426, 2007.
Parker, D. Chan, and M. Holland. [28] J. Martinez. (2010). Floating Sensor Network [Online]. Available: http:
//float.berkeley.edu
[29] A. Tinka, I. Strub, Q. Wu, and A. M. Bayen, “Quadratic programming
based data assimilation with passive drifting sensors for shallow water
References flows,” Int. J. Control, vol. 83, no. 8, pp. 1686–1700, Aug. 2010.
[1] T. Oki and S. Kanae, “Global hydrological cycles and world water [30] 871DO-C Dissolved Oxygen Sensors and Sensor Accessories: Product
resources,” Science, vol. 313, no. 5790, pp. 1068–1072, 2006. Specifications, document PSS 6-9B1 A, Invensys Systems, Inc., Ash-
[2] S. L. Postel, “Entering an era of water scarcity: The challenges ahead,” burn, VA, 2005.
Ecol. Applicat., vol. 10, no. 4, pp. 941–948, 2000. [31] CDH222 Conductivity/TDS Meter Manual, 0212-YK-22CT, Omega En-
[3] R. B. Jackson, S. R. Carpenter, C. N. Dahm, D. M. McKnight, R. J. gineering, Inc., Stamford, CT, 2010.
Naiman, S. L. Postel, and S. W. Running, “Water in a changing world,” [32] S8000 Series Outline and Dimensions, Sensorex Corporation, Garden
Ecol. Applicat., vol. 11, no. 4, pp. 1027–1045, 2001. Grove, CA, 2008.
[4] K. R. Dyer, Estuaries: A Physical Introduction. New York: Wiley, 1973. [33] A12, B12, & AC12 Reference Manual, 630871 rev. D, Thales Navigation,
[5] R. Smith and C. F. Scott, “Mixing in the tidal environment,” J. Hydraulic Tulsa, OK, 2005.
Eng., vol. 123, no. 4, pp. 332–340, Apr. 1997. [34] Motorola G24 Developers’ Guide: Module Hardware Description,
[6] L. MacVean and M. Stacey, “Estuarine dispersion from tidal trapping: A 6889192V27-G, Motorola, Schaumburg, IL, 2007.
new analytical framework,” Estuaries Coasts, vol. 34, no. 1, pp. 45–59, [35] IEEE Standard for Information Technology—Telecommunications and
Jan. 2011. Information Exchange Between Systems—Local and Metropolitan Area
[7] A. Chadwick, J. Morfett, and M. Borthwick, Hydraulics in Civil and Networks—Specific Requirements. Part 15.4: Wireless Medium Access
Environmental Engineering. London, U.K.: Spon Press, 2004. Control (MAC) and Physical Layer (PHY) Specifications for Low-Rate
[8] F.-X. Le Dimet and O. Talagrand, “Variational algorithms for analysis Wireless Personal Area Networks (WPANs), IEEE Std., 2006.
and assimilation of meterological observations: Theoretical aspects,” [36] XBee/XBee-PRO ZB RF Modules, 90000976− C, Digi International,
Tellus, vol. 38A, no. 2, pp. 97–110, Mar. 1986. Minnetonka, MN, 2009.
[9] K. Ide, P. Courtier, M. Ghil, and A. Lorenc, “Unified notation for data [37] Intel PXA27x Family Design Guide, 280001-002, Intel Corporation,
assimilation: Operational, sequential and variational,” J. Met. Soc. Japan, Santa Clara, CA, May 2005.
vol. 75, no. 1B, pp. 71–79, 1997. [38] ATmega128/ATmega128L Datasheet, 2467R-AVR-06/08, Atmel Corpo-
[10] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Upper Saddle ration, San Jose, CA, 2008.
River, NJ: Prentice-Hall, 1979. [39] W. R. Stevens, Advanced Programming in the UNIX Environment.
[11] G. Evensen, Data Assimilation: The Ensemble Kalman Filter. Berlin, Reading, MA: Addison-Wesley, 1992.
Germany: Springer, 2006. [40] B. Smith, J. Hardin, G. Phillips, and B. Pierce, Linux Appliance Design:
[12] S. J. Julier and J. K. Uhlmann, “A new extension of the Kalman filter A Hands-on Guide to Building Linux Appliances. San Francisco, CA:
to nonlinear systems,” in Proc. Int. Symp. Aerospace/Defense Sensing No Starch Press, 2007.
Simul. Controls, vol. 3. 1997, pp. 182–193. [41] J. Gowdy, “A qualitative comparison of interprocess communications
[13] J. N. Carruthers, “Some oceanography from the past,” J. Navigation, toolkits for robotics,” Robotics Instit., Carnegie Mellon Univ., Pittsburgh,
vol. 16, no. 2, pp. 180–188, 1963. PA, Tech. Rep. CMU-RI-TR-00-16, Jun. 2000.
[14] J. N. Carruthers, “Further investigation upon the water movements in the [42] M. Owens, The Definitive Guide to SQLite. New York: Apress, 2006.
English channel: Drift-bottle experiments in the summers of 1927, 1928, [43] P. T. Eugster, P. A. Felber, R. Guerraoui, and A.-M. Kermarrec,
and 1929, with critical notes on drift-bottle experiments in general,” J. “The many faces of publish/subscribe,” ACM Comput. Surveys (CSUR),
Marine Biol. Assoc. United Kingdom, vol. 17, no. 1, pp. 241–275, 1930. vol. 35, no. 2, pp. 114–131, 2003.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE SYSTEMS JOURNAL

[44] C. F. Goldfarb and P. Prescod, The XML Handbook. Upper Saddle River, Mohammad Rafiee received the B.Sc. degree in
NJ: Prentice-Hall, 1998. mechanical engineering from the Sharif University
[45] Data Networks and Open System Communications OSI Networking and of Technology, Tehran, Iran, in 2005, and the M.Sc.
Systems Aspects: Abstract Syntax Notation One (ASN.1): Information degree in mechanical engineering and the M.A.
Technology, ASN.1 Encoding Rules: Specification of Basic Encoding degree in mathematics from the University of Cali-
Rules (BER), Canonical Encoding Rules (CER) and Distinguished En- fornia, Berkeley (UC Berkeley), in 2008 and 2011,
coding Rules (DER), International Telecommunication Union, Geneva, respectively, and the Ph.D. degree from the Depart-
Switzerland, 1995. ment of Mechanical Engineering, UC Berkeley, in
[46] Google, Inc. (2010, Oct. 2). Protocol Buffers Developers’ Guide May 2012.
[Online]. Available: http://code.google.com/apis/protocolbuffers/docs/ His current research interests include machine
overview.html learning, time series analysis, Monte Carlo meth-
[47] M. Kofler, The Definitive Guide to MySQL 5. New York: Apress, 2005. ods and particle filters, data assimilation and inverse problems, distributed
[48] S. L. Britton, G. J. Hanson, and D. M. Temple, “A historic look at the parameter systems, and convex optimization.
USDA-ARS hydraulic engineering research unit,” in Henry P. G. Darcy
and Other Pioneers in Hydraulics: Contributions in Celebration of the
200th Birthday of Henry Philibert Gaspard Darcy, G. O. Brown, J. D.
Garbrecht, and W. H. Hager, Eds. Reston, VA: American Soc. Civil
Engineers, 2003, pp. 263–276.
[49] V. Chow, Open-Channel Hydraulics. New York: McGraw-Hill, 1988.
[50] J. Cunge, F. Holly, and A. Verwey, Practical Aspects of Computational
River Hydraulics. London, U.K.: Pitman, 1980.
[51] T. Strum, Open Channel Hydraulics. New York: McGraw-Hill, 2001.
[52] X. Litrico and V. Fromion, Modeling and Control of Hydrosystems.
Berlin, Germany: Springer, 2009.
[53] M. Chaudhry, Open-Channel Flow, 2nd ed. Berlin, Germany: Springer,
2008.
[54] G. V. Bogle, “Stream velocity profiles and longitudinal dispersion,” J.
Hydraulic Eng., vol. 123, no. 9, pp. 816–820, 1997.
[55] Office of State Water Project Planning, “Methodology for flow and Alexandre M. Bayen (S’02–M’04) received the
salinity estimates in the Sacramento-San Joaquin Delta and Suisun Engineering degree from École Polytechnique,
Marsh,” California Dept. Water Resources, Sacramento, CA, Tech. Palaiseau, France, and the M.S. and Ph.D. degrees
Rep. 21, 2000 [Online]. Available: http://baydeltaoffice.water.ca.gov/ from Stanford University, Stanford, CA.
modeling/deltamodeling/annualreports.cfm He is currently an Associate Professor with the
[56] E. Ateljevich, P. Colella, D. T. Graves, T. J. Ligocki, J. Percelay, P. O. Department of Electrical Engineering and Com-
Schwartz, and Q. Shu, “CFD modeling in the San Francisco Bay and puter Sciences and the Department of Civil and
Delta,” in Proc. 4th SIAM Conf. Math. Ind., 2009, pp. 99–107. Environmental Engineering, University of Califor-
nia, Berkeley (UC Berkeley). He was a Visiting
Researcher with the NASA Ames Research Center,
Moffett Field, CA, from 2000 to 2003. He has been
Andrew Tinka (M’04–S’06) received the B.A.Sc.
the Research Director of the Autonomous Navigation Laboratory, LRBA,
degree in engineering physics from the University of
Ministere de la Defense, Vernon, France, where he has held the rank of Major.
British Columbia, Vancouver, BC, Canada, in 2002,
He has authored one book and over 100 articles in peer-reviewed journals and
and the M.S. degree in civil and environmental engi-
conferences.
neering (systems engineering) from the University of
Dr. Bayen was the recipient of the Ballhaus Award from Stanford University
California, Berkeley, in 2008, where he is currently
in 2004, the CAREER Award from the National Science Foundation in 2009,
pursuing the Ph.D. degree in electrical engineering
and was a NASA Top 10 Innovator on Water Sustainability in 2010. His
with the Department of Electrical Engineering and
projects Mobile Century and Mobile Millennium received the Best of ITS
Computer Sciences, focusing on the design and
Award for Best Innovative Practice at the ITS World Congress in 2008, and
applications of the floating sensor network.
the TRANNY Award from the California Transportation Foundation in 2009.
He has been with Powis Parker, Inc., Berkeley, CA,
He was the recipient of the Presidential Early Career Award for Scientists
where he worked on embedded systems engineering, and with the Center
and Engineers from the White House in 2010. Mobile Millennium has been
for Collaborative Control of Unmanned Vehicles, University of California,
featured more than 100 times in the media, including TV channels and radio
Berkeley, where he worked on systems engineering. His current research
stations (CBS, NBC, ABC, CNET, NPR, KGO, BBC), and in the popular
interests include Lagrangian sensor design, multivehicle control and planning,
press (Wall Street Journal, Washington Post, LA Times).
and data assimilation.