structured sparse methods for active ocean observation...

IEEE Communications Magazine • November 201588 0163-6804/15/$25.00 © 2015 IEEE

Urbashi Mitra, SunavChoudhary, NaveenKumar, ShrikanthNaryanan, and GauravSukhatme are with theUniversity of SouthernCalifornia.

Franz Hover and RobertHummel are with theMassachusetts Institute ofTechnology.

Milica Stojanovic is withNortheastern University.

ABSTRACT

Actuated sensor networks enabled by under-water acoustic communications can be efficientlyused to sense over large marine expanses thatare typically challenged by a paucity of resources(energy, communication bandwidth, number ofsensor nodes). Many marine phenomena ofinterest admit sparse representations, which,coupled with actuation and cooperation, cancompensate for being data starved. Herein, newmethods of field reconstruction, target tracking,and exploration-exploitation are provided, whichadopt sparse approximation, compressed sens-ing, and matrix completion algorithms. Theneeded underlying structure (sparsity/low-rank)is quite general. The unique constraints posed byunderwater acoustic communications and vehiclekinematics are explicitly considered. Resultsshow that solutions can be practically imple-mented, even over large ocean spaces.

INTRODUCTIONFuture advances in ocean monitoring, offshoreindustry, and basic marine sciences will rely heav-ily on our ability to jointly consider communica-tion, actuation, and sensing in a unified systemthat includes remote instruments, underwatervehicles, human operators, and sensors of alltypes. These tasks will require methods to detectand track large-scale ocean phenomena such asalgal blooms, oil spills, ocean currents, andhydrothermal vents, as well as man-made signalssuch as those emanating from an airplane’s blackbox. We envision a scenario as depicted in Fig. 1,where multiple autonomous underwater vehicles(AUVs) interact and coordinate via acousticcommunications with a network of sensors todetect and track a phenomenon of interest. Theunderlying network architecture includes bothstatic, communication-enabled sensor nodes, aswell as actuated nodes in the form of AUVs.Thus, our system needs to control and movesome of the nodes to achieve its sensing and

communication goals. Moreover, the choicesmade regarding communication, control, andsensing are interdependent [1].

These goals require active consideration ofthe underlying problem of exploration andexploitation. A critical feature of this problem issparsity: our sensors will be sparsely deployeddue to cost and complexity considerations; ourfleet of AUVs will be modest in number. As aresult, our observations of a vast ocean can onlybe a scant sampling of the phenomena of inter-est. Satellite remote sensing and traditional shipoperations in a large-scale underwater observa-tion system are similarly limited, and manyocean measurements are made point-wise orconfined to a local area. Furthermore, we oftenneed to first find (explore for) a feature andthen track (exploit) it, as illustrated in Fig. 2.

While there has been strong attention paidto exploiting the sparse nature of underwateracoustic channels (e.g., [2]), our problems ofinterest are very different. We wish to examinehow to task sensors, static and/or mobile,toward achieving an underwater mission goalthat will be dependent on the sensor network’sunderstanding of its environment. Prior workrelated to field estimation and target trackingsuggests that sparse representations exist formany features of interest, and therefore, wemay not be irreparably data-starved. At thesame time, a mission has to be accomplished bycoordinated sensing and action, typicallyenabled by acoustic communication betweenenergy-constrained and kinematically con-strained agents. In this article, we review howsparsity can be leveraged in the design of explo-ration-exploitation methods tailored to under-water sensing systems. Our particular approachdoes not require detailed prior informationabout the fields or targets of interest.

In particular, we tackle three related prob-lems: field reconstruction [1, 10, 11, 12], targettracking [5, 13], and methods for exploration-exploitation [14, 15], which build on the insightsderived from the previous approaches. Field

UNDERWATER WIRELESS COMMUNICATIONS ANDNETWORKS: THEORY AND APPLICATION

Urbashi Mitra, Sunav Choudhary, Franz Hover, Robert Hummel, Naveen Kumar, Shrikanth Naryanan,

Milica Stojanovic, and Gaurav Sukhatme

Structured Sparse Methods for Active Ocean Observation Systems withCommunication Constraints

MITRA_LAYOUT.qxp_Author Layout 10/30/15 3:10 PM Page 88

IEEE Communications Magazine • November 2015 89

reconstruction is an important issue for manyscientific, industrial, and tactical applications.Examples include geographical, chemical, andtemperature map building. In some cases, recon-struction is the end goal (e.g., climate record-ing), while in other cases it represents anintermediate step to support the mission athand, for example, one in which mobile assetsneed to learn about the environment in whichthey operate and adapt to it. Sensor measure-ments of the field are typically correlated insome domain, and this correlation forms thebasis of compressive sensing theory: when thefield to be reconstructed is sparse (with respectto some basis), it suffices to collect data only atsome of the points to reconstruct the full view ofthe field.

In recent years, there has been a surge of newresults in the area of sparse field reconstructionusing sensor networks. Although harsh environ-ments, energy management, and limited commu-nications have been considered, most of thesemethods have been optimized for terrestrial con-ditions. Underwater systems pose new challengesand constraints. For example, sparse methodshave been developed predominantly ignoring thecost of sampling, while in oceanographic fieldapplications, constraints such as limited energyand maneuverability can dominate, and the num-ber of samples may be quite low. The interplaybetween such resource constraints and the funda-mental sparse algorithms is a general problem werefer to as kinematically constrained sparseapproximation (KCSA) [10, 11].

The extension from field reconstruction totarget tracking is a natural one, and we providemethods that explicitly consider the dynamics ofthe targets and the nature of the underwateracoustic communication channel. An example israndom access compressed sensing (RACS) [12,13], which notably does not require synchroniza-tion. Target detection is a related applicationthat addresses whether a specific phenomenonof interest has occurred. The target usually can-not be observed directly, however, and the taskthen is to infer the presence or location of thetarget from indirect observations, with an eye tothe cost of acquiring each sample, thus resultingin the exploration-exploitation [14, 15] optimiza-tion trade-off.

We summarize key challenges of underwatersensing and communication networks that areenabled by actuation, including the unique fea-tures of the acoustic channel and the constraintsof underwater vehicles. We lay out the key ideasin compressive sensing, sparse approximation,and matrix completion. We examine how toreconstruct a large field when limited to a fewmobile vehicles operating with constraints over alarge space. We change the perspective by con-sidering field reconstruction when using staticunderwater sensors. Here, the impact of usingthe underwater acoustic medium to transmit sen-sor data to a fusion center introduces uniquecomplexities; and the work is extended to thetracking of mobile sources. Finally, the insightsof the prior two research areas are employed toderive novel methods based on matrix comple-tion to solve a realistic exploration-exploitationproblem.

THE CHALLENGES OF UNDERWATERACTUATED NETWORKS

The boundaries between communications, net-working, navigation, and sensing are blurred inunderwater actuated sensor networks, wheremultiple levels of dynamics exist: moving (possi-bly actuated) nodes, time-varying communica-tion channels, and evolving phenomena to bedetected, identified, and tracked.

UNDERWATER ACOUSTICCOMMUNICATION CHANNELS

Challenges in the design of underwater acousticcommunication systems include a severely limit-ed, range-dependent bandwidth, extensive time-varying multipath propagation, and longpropagation delays caused by the low speed ofsound underwater (1500 m/s) [3]. Absorptionleads to exponential path losses with respect tofrequency. Propagation delays result in signifi-cant delay spreads (on the order of hundreds ofsymbols). There are very short channel coher-ence times that are exacerbated in mobile envi-ronments and orders of magnitude shorter thanin RF systems. In terms of the channel, we focuson the sparse nature of multipath propagation.

UNDERWATER SENSORS AND VEHICLESOcean vehicle systems face obstacles not sharedin the terrestrial and aerial domains. Most of theocean remains uncharted, with detailed informa-tion about the sea floor available only at specificand scientifically important sites, or areas of spe-cific interest such as oil and gas fields. Underwa-ter systems are expensive to build andinstrument, and ship time also introduces a hugecost that encourages remote operations. Unlikeland-based vehicles, underwater vehicles areexposed to extreme pressure, corrosion, andfouling, and strong variable wave, current, andwind disturbances. Such disturbances make oper-ating a vehicle difficult, if not impossible. At the

Figure 1. The envisioned network architecture for future ocean observingsystems, comprising both fixed and moving nodes capable of advancedsensing and wireless communications.


IEEE Communications Magazine • November 201590

same time, it should be noted that currents cansometimes be exploited opportunistically, sinceenergy consumption is a major concern from thepoint of view of propulsion, acoustic communi-cations, and hotel load. Aerial vehicles face simi-lar energy issues, but usually have two keyadvantages over underwater systems: low-powercommunication over an RF channel with othervehicles or a base station, and the ability to mea-sure their own location with precision usingGPS. Underwater positioning methods are oftenacoustic and subject to the challenges notedabove for acoustic communications. Today,ocean vehicles typically break the surface forprecision localization via GPS — a hazardousand time-consuming operation.

As one specific example of these challenges,we carried out an experiment with anautonomous surface vehicle (ASV) station-keep-ing via commands sent through an acousticmodem. The effects of wind in such a setting areto both push the boat off the reference pointand degrade the acoustic channel because of sur-face chop. Indeed, we observed a full order ofmagnitude increase in positioning error as thewind speed increased from 1 to 8 m/s [4]. Anoth-er experiment coordinating vehicles throughacoustic communication is detailed in [5]. Over-all, the costs of actuation, communication, andsensing can be comparable for marine systems[3], and this parity strongly affects how underwa-ter acoustic networks are to be optimized in sup-port of a mission goal [1].

SPARSE APPROXIMATION ANDMATRIX COMPLETION REVIEW

Here we review key elements of sparse approxi-mation and matrix completion needed for under-standing the methods presented in this article.

Any family of signals with a linear algebraicrepresentation requiring many more parametersthan the actual number of unknowns in a giveninstance is called a low-dimensional signal. Twolow-dimensional signals that have been the sub-ject of active research are sparse vectors andlow-rank matrices.

Considering sparse vectors first, let Œ Rn×n

be an orthonormal basis so that any signal f ŒRn can be written as f = x for a coefficient vec-tor x Œ Rn. x is called s-sparse if it has at most s

nonzero elements, and if a sparse x suffices, thesignal f is said to admit a sparse representationin . Intuitively, an s-sparse signal has only sunknowns, many fewer than the full representa-tion [6]. f is called compressible if it is wellapproximated — but not constructed exactly —by a sparse x. Many natural phenomena are con-sidered to be compressible in domains such asfrequency, wavelets, or total variation. For low-rank matrices, a key classical result is that any m× n matrix X admits the singular value decompo-sition X = USV, and that the number of non-zero elements in S denotes the rank of X. TheEckart-Young-Mirsky result says further thatreduced-rank approximations to S, obtained sim-ply by keeping only the r largest elements in S,are optimal in the Frobenius norm [7]. Low-rankmatrices arise frequently in second-orderdatasets (Netflix) and data generated from bilin-ear observations [8].

Incoherence is a critical requirement for effi-ciently sampling low-dimensional models, andmeasures the dissimilarity between the signalbasis and the measurement basis. Informationmeasured in one basis will “spread out” whenmeasured in a second (incoherent) basis, enablingefficient reconstruction even if few measure-ments are made. In point-wise compressed sens-ing, the measurement y = Rf involves a “fat”matrix R, where R is a fat projection (a row-reduced identity matrix), and an orthonormalmeasurement basis. Coherence of the measure-ment and signal bases is defined as

where fj and k denote the jth and kth atoms of and , respectively, and <◊, ◊> denotes aninner product. For a 2D field represented in thediscrete cosine basis, and sampled with deltafunctions, m(, ) = 2, a highly incoherent basispairing is indicated. The reconstruction task canbe formulated as a convex optimization [6]

x̂ = argmin z1 subject to y = Rz,

where ◊1 is the one-norm. Pulling togethersparsity and incoherence, a key theorem of com-pressed sensing states that if f is s-sparse in ,then x̂ = x with high probability if the number ofmeasurements (i.e., the number of rows in R)exceeds Cm2(, )s log n for some positive con-

μ φ ψΦ Ψ < >≤ ≤

n( , )= max , ,j k n

j k1 ,

Figure 2. Exploration-exploitation at sea: autonomous agents first explore in a coordinated way to identify phenomena of interest(left); as the field evolves, the AUV network differentiates to allow individual agents to follow particular targets (exploitation).



stant C. Coherence, and thus the associated sam-pling rate, determine the feasibility of the recon-struction.

The measurement ideology in low-rank matrixcompletion bears a strong resemblance to that ofcompressed sensing. Here y=(X), where denotes a small set of indices, and hence themeasurement basis is the set of delta functionsover the vector space of m × n matrices. Inco-herence is defined as the sum of the inner prod-ucts, squared, between delta functions and theleft and right singular vectors [7]. The recon-struction is formulated as the convex problem

X^ = argmin Z* subject to y = (Z),

where ◊* is the sum of singular values. Akinto compressed sensing, if is sampled uniformlyat random, X^ = X with high probability whenthe number of measurements exceeds Cv(X)rnlog2n, where r is the rank of X, and n > m.

Turning to the selection of points via R, ran-dom sampling gives strong stability guaranteesfor both compressed sensing and low-rank matrixcompletion in the presence of measurementnoise as well as model mismatch. Mismatchoccurs when the unknown signal is compressibleinstead of sparse, or is approximately low-rankinstead of exactly low-rank. In compressed sens-ing, we appeal to the restricted isometry proper-ty (RIP) of the measurement matrix of order s,by which the measurement process approximate-ly preserves pairwise distances between s-sparsevectors. In low-rank matrix completion, a weakercondition called restricted strong convexity isdesirable, and similarly sufficient to show stabili-ty, given r << n [9].

Random sampling may be infeasible orunsuitable, and it is desired in such cases to findnew projection matrices that:• Result in efficient reconstruction• Can be implemented efficiently by the sam-

pling system (e.g., mobile robots in a largedomain)

In general, it is not easy to show that structuredprojections satisfy the RIP, although there havebeen recent efforts directed toward the design ofdeterministic projection matrices. Theseapproaches control the structure of the sensingmatrix R; however, in our case we can only

control the measurement matrix R. To the bestof our knowledge, there are no techniques forstructured random construction of the measure-ment matrix, especially given the motion con-straints of robots.

FIELD RECONSTRUCTION WITHMOBILE SENSORS VIA KCSA

Since kinematic constraints are one of our maindivergences from standard sparse approximation,a natural step is the construction of a cost relat-ed to resources. We can identify three keyaspects of such a cost:• The Euclidean length of a transit between

two sample sites• A heading change (maneuvering) required

to move between two successive sites• The resources required for an individual

sample itselfWorking with a large sea-surface temperaturedataset and the discrete cosine transform basis,we optimized statistics of randomly drawn transitlengths and maneuvers to minimize the sparseapproximation reconstruction error in a simplerandom walk scheme [10]. This optimized ran-dom walk, surprisingly, was outperformed insparse approximation error by an even simplerprocedure of randomly choosing a set of sitesand connecting them with a nearest-neighbortraveling salesperson problem (TSP) solution.Figure 3 depicts an experimental test with thismethod. More generally, reconstruction errorscan be bounded by ensuring that the RIP in thesampling basis is achieved, suggesting a secondclass of lightweight KCSA methods. For a givendictionary and a set of sample points, restrictedisometry is trivial to compute incrementally, so avehicle can be instructed to move to randomlyselected, as yet unvisited points based simply onwhat the next point will do to the RIP. This is apurely greedy approach that typically offers animprovement of 10–20 percent relative to pointsselected without considering the RIP [11].

Maintaining restricted isometry and minimiz-ing path length could be tackled as well in amore formal optimization setting. With the spikebasis, however, the projection (sampling) designis a binary programming problem. Furthermore,

Figure 3. Kinematically-constrained sparse approximation in action at sea. The left image shows two “random TSP” paths generat-ed and executed by an autonomous surface vehicle (center); the right image shows a depth-field reconstruction based on sparsesamples and the discrete-cosine transform.

East-West0

20

Nor

th-S

outh

40

0

60

80

100

20 40 60 80 100

Set 1Set 2



we have found that when the number of samplesis small, there is virtually no advantage in thedesigned plan over random points; with a largernumber of samples, the designed points have acoherence typically 20–30 percent lower than therandom points. The formal sampling problemcan also be merged with the TSP, but complexityscaling is very poor since they are not well cou-pled. An alternative is to generate a pool ofoptimal sample sets without attention to thepath, and then apply strong TSP solvers to mem-bers of the pool. Applying such a process, wefind that a pool typically identifies a Pareto-likefrontier of TSP cost vs. coverage (e.g., as charac-terized by the star discrepancy), and thus atrade-off space between sampling efficacy for thepurpose of reconstruction and the path length.This multi-way trade-off is developed furtherlater.

FIELD RECONSTRUCTION IN ANUNDERWATER SENSOR NETWORK

As noted earlier, in underwater acoustic systems,key physical constraints include limited band-width and the high rate of packet loss. More-over, synchronization is difficult, oftenmandating a design that does not rely on accu-rate timing. To address these issues, we draw onthe sparse (compressible) nature of the sensingfield, through a design that capitalizes on ran-dom channel access and compressive sensing(RACS), using multiple sensors interacting witha fusion center. In this approach, a sensor takeslocal field measurements at random instants intime [12]. Each measurement is encoded into adata packet, which is immediately transmitted tothe fusion center. The fusion center collects allthe packets over a certain interval of time, andthen attempts to reconstruct the field.

The key (and perhaps counterintuive) ideabehind combining random access and compres-sive sensing is that the fusion center does notcare which sensors provide the data packets aslong as the sensors are selected uniformly at ran-dom, and there are sufficient packets. With ran-dom access, packet collisions are inevitable.However, they occur randomly and thus do notalter the way in which the FC perceives packetarrivals. Consequently, the fusion center cansimply discard the packets that have collided (aswell as those that were received in error due tonoise, fading or interference), and perform fieldreconstruction using the remaining useful pack-ets. Because of the random nature of this archi-tecture, a probabilistic approach to system designbecomes necessary. Our approach relies on thenotion of sufficient sensing probability, the prob-ability with which full field reconstruction isachieved in a certain interval of time. Settingthis probability to a desired target value, systemoptimization under a relevant criterion (e.g.,minimum energy per bit or minimum reconstruc-tion time) yields the necessary parameters: thepacket generation rate (per-node sensing rate),required bandwidth, and power. Many fields ofinterest are sparse with respect to fairly generaldictionaries; thus, the reconstruction method isrelatively non-parametric.

Figure 4. Target tracking without field reconstruction for three mobilesources. Instaneous observations of the noise-free (top) and noisy (center)fields; (bottom) true and estimated paths.

Time (T)

Field strength

200100

Loca

tion

90

80

70

60

50

40

30

20

10 2.4

2.2

2

1.8

1.6

1.4

1.2

1

0.8400 600 800 1000 1200 1400 1600 1800 2000

Time (T)

Field strength

200100

Loca

tion

90

80

70

60

50

40

30

20

10 3.5

3

3.5

2

2.5

1

1.5

0

-0.5400 600 800 1000 1200 1400 1600 1800 2000

Time (T)2000

-40

Targ

et lo

catio

ns, t

rue

and

estim

ated

-20

0

20

40

60

80

100

400 600 800 1000 1200 1400 1600 1800 2000



The concept of random access compressivesensing can also be extended to the tracking ofvehicles each of which emanates a particular sig-nature [13]. Based on the knowledge of signa-tures, a reconstruction algorithm is designed thatside-steps the field reconstruction, and insteadfocuses directly on object tracking. Figure 4 pro-vides an example of the multi-vehicle trackingcapabilities of RACS. Clearly, there is correla-tion in the map data (i.e., the field is sparse).This particular spatio-temporal field is createdby three moving objects, each having an expo-nentially decaying signature; the level of sparsityis three. Shown on the right are the objects’ truetrajectories (solid) and estimates obtained by thetracking algorithm (dashed), applied to samplesof the noisy field (–20 dB signal-to-noise ratio,SNR, 30 percent packet loss) obtained from 10sensors distributed uniformly along a track. Aswith field reconstruction, effective target track-ing based on RACS does not require complexmodels of vehicle behavior (vs. field behavior).As such, our methods are highly robust.

Due to the time-varying nature of the field,one needs to optimize the sensing duration —the key is to acquire a sufficient number of sam-ples for reliable reconstruction, while ensuringthat the sensing duration is short enough so thatthe observed field does not change significantlywithin a single sensing period. From the per-spective of target detection using the acquiredimage, there is another notion of sensing rateoptimality corresponding to the trade-offbetween detection accuracy and energy efficien-cy. One must also consider the rate at which thefield is changing due to the speed of the vehiclesto be tracked. To address these issues, a feed-back scheme can be employed, resulting in adap-tive sensing. The feedback signal is generated byperforming target detection on the acquiredfield, which is then compared to a model of thefield, resulting in a model-based error. Figure 5depicts the evolution of the sensing rate (T)based on such an algorithm. If there is a largemodel-based error, the sensing rate is increased(state A). To compensate for too high a sensingrate, a motion-dependent error measure is alsocomputed (state D). If the vehicles move slowly,the fusion center can respond by decreasing thesensing rate to conserve energy. Thresholds onthe two error measures are optimized to maxi-mize performance while constraining the sensingenergy. States B and C correspond to conflicts inthe decisions resulting from the two differentsensing metrics.

TACKLING THE INTEGRATEDEXPLORATION-EXPLOITATION

PROBLEM

Our prior work on field reconstruction shows thatkinematic and communication constraints canreadily be employed to develop practical recon-struction methods exploiting new results in sparseapproximation. Furthermore, a large family offields admits the needed sparsity. We can extendthese notions to the exploration-exploitationproblem. As a concrete example of integrated

search with sparse methods, we consider anunderwater target that emits a field of observableintensity, as in Fig. 6a; close to the target, theintensity is high, but the intensity decays as wemove away from the target location. The aggre-gate field shown here is a side-scan sonar image,and the target signature could be due to an acous-tic signal from a flight data recorder lost in somesector of the ocean. A deployed AUV is restrict-ed to collecting a few samples along some naviga-tion path (Fig. 6b), with the hope that thesesamples contain enough information to localizethe target to a smaller region of interest (Fig. 6c).The intensity field in Fig. 6a is approximately sep-arable, however, and therefore admits a low-rankmatrix structure with respect to the image matrix.This property allows the reduction of the numberof samples required for target localization [14],subsequently leading to reduced cost of naviga-tion [15]. On the other hand, a decaying low-ranktarget field is also a very generic structuralassumption not specific to this particular dataset.Thus, a rich set of target signatures meet the twokey assumptions; additionally, we can extend ourresults to allow for multiple targets, differentadaptive and non-adaptive sampling approaches,and even distributed implementations using mul-tiple AUVs with communication costs. With ourapproach, we do not need to know the target sig-nature a priori.

We developed a new methodology for thisproblem based on low rank matrix completionand the high-dimensional properties of the trav-eling salesperson problem, and one appealingoutcome is a characterization of a multi-waytrade-off between sampling, navigation, compu-tation, and localization [15]. Our algorithm takesa hierarchical exploration-exploitation approachthat combines the strengths of low-rank matrixcompletion and binary search, viz. noise robust-ness and computational speedup. In each stage,based on the current field of view, a set of ran-dom locations is selected; the AUV determinesan optimized path covering these locations, atwhich sample observations are collected. Thefield is then reconstructed using low-rank matrixcompletion [7], computing the best rank oneapproximation. After post-processing, the targetis localized, a smaller field of view is deter-mined, and the steps are then repeated. Weassume that the algorithm terminates after k

Figure 5. Controlled sensing rate T using the proposed method; states Aand D correspond to undersensing and oversensing, respectively; states Band C correspond to the cases when the model error and sensing errormetrics lead to conflicting decisions. (Tcoh is the coherence time of thetime-varying field).

Time20

500B

1000

0

T coh

(in

s)

Stat

e

A

C

D

40 60 80 100 120 140 160 180 200

Tcoh System stateT



iterations, define the total search space of size n× n, and the field of view to be of size m × mrepresenting a uniformly decimated sub-matrixof the total search space. With these definitions,theoretical results on low rank matrix comple-tion and Euclidean TSP can be used to showthat our approach takes O(km log2m) samplesand O(km3.5) computational time and achievesO(k

—m log m) cost of navigation with high prob-

ability, where the number of iterations increaseswith noise level and satisfies logm n k n/m.

A design parameter is the field of view. Intu-itively, the size of the field of view is choseninversely proportional to the spread in the targetfield since a larger spread enables collecting sam-ple intensities above the noise floor farther fromthe target location, thus reducing samplingrequirements. Decaying separable fields admitgood rank-one approximation via singular vectorsthat are also decaying. The decaying property ofthe singular vectors can be used for further pro-cessing via unimodal function fitting (a form ofisotonic regression), which results in greater noisereduction and improved target localization.Numerical results show that there is an optimumspreading factor leading to the best detection per-formance. One can additionally characterize thebest balance between existence of a significantgradient in the signature (after noise corruption)and the target signature occupying a larger foot-print; that is, fields with large support of the gra-dient show better performance for a fixed samplecomplexity. We can also provide a theoreticalanalysis that provides the trade-off between ourability to localize the target within a stage and ourability to reconstruct the current windowed field.Finally, the overall multi-way trade-off can becharacterized by including navigation and kine-matic costs (or constraints) as well as the incorpo-ration of the randomness induced by packetdropping in the underwater acoustic communica-tion channel. Thus, the features of our fieldreconstruction, tracking, and exploration-exploita-tion work can be seamlessly combined.

CONCLUSIONSWe have argued that mobile underwater sensornetworks enabled with control and acousticcommunication will play a major role in futureocean observing infrastructure. Such autono-

mous systems need novel mechanisms for han-dling underwater communication constraints,and need to integrate sensing and classificationin providing solutions for key exploration-exploitation trade-offs. The unavoidable burdenof limited resources in these operations (e.g.,physical energy, time, communications, andaccurate synchronization) are in a sense bal-anced by the promise of compressed sensingand low-rank approximation theory over hugespatial domains, coupled with standard toolsincluding solvers for the traveling salespersonproblem. Thus, despite the paucity of resourcesand limited number of deployed sensors,whether they be static or mobile, the underlyingsparsity (or low rank property) of the processesof interest enables us to overcome the chal-lenge of being data-starved. Our structures ofinterest are general and do not require complexa priori model information for our fields or tar-gets. We have specifically considered con-straints inherent to our underwater system: thekinematic constraints of vehicles as well as theunique characteristics of the underwater acous-tic channel. These constraints have led to novelproblem formulations and attendant solutions.In fact, some of our most recent outcomes canbypass weaknesses of more classical approach-es to solving the exploitation-exploration prob-lem that are not flexible to incorporatingmultiple resource constraints such as naviga-tion, communication, and sensing while allow-ing for theoretical analysis. The successfulimplementation of these methods at sea willhave an impact on many applications andindustries, including environmental monitoring,aquatic ecosystems, ocean accident remedia-tion, surveillance for defense applications,homeland security, the oil and gas industry,aquaculture, geological and oceanographic sci-ence, and marine biology.

REFERENCES[1] G. Hollinger et al., “Underwater Data Collection Using

Robotic Sensor Networks,” IEEE JSAC, Special Issue onCommunications Challenges and Dynamics forUnmanned Autonomous Vehicles, vol. 30, no. 5, June2012, pp. 899–911.

[2] C. Berger et al., “Sparse Channel Estimation for MulticarrierUnderwater Acoustic Communication: From SubspaceMethods to Compressed Sensing,” IEEE Trans. Signal Pro-cessing, vol. 58 no. 3, Mar. 2010, pp. 1708–21.

Figure 6. (left) Original intensity map of an underwater target (sonar data); (center)initial sampling path of the AUV; (right) thelocalized region of interest after several stages of the matrix completion exploration-exploitation algorithm.



[3] M. Stojanovic and J. Preisig, “Underwater AcousticCommunication Channels: Propagation Models and Sta-tistical Characterization,” IEEE Commun. Mag., vol. 47,no. 1, Jan. 2009, pp. 84–89.

[4] F. Arrichiello et al., “Effects of Underwater Communica-tion Constraints on the Control of Marine RobotTeams,” Proc. IEEE Int’l. Conf. Robot Commun. andCoordination, Odense, Denmark, Mar. 2009.

[5] B. Reed et al., “Multi-Vehicle Dynamic Pursuit UsingUnderwater Acoustics,” Proc. Int’l. Symp. RoboticsResearch, Dec. 2013, Singapore.

[6] E. Candes and T. Tau, “Decoding by Linear Program-ming,” IEEE Trans. Info. Theory, vol. 51, no. 125, Dec.2005, pp. 4203–21.

[7] D. Gross, “Recovering Low-Rank Matrices From FewCoefficients in Any Basis,” IEEE Trans. Info. Theory, vol.57, no. 3, Mar. 2011, pp.1548–66.

[8] S. Choudhary and U. Mitra, “On Identifiability in Bilin-ear Inverse Problems,” Proc. IEEE Int’l. Conf. Acoustics,Speech and Signal Processing, May 2013, Vancouver,Canada, pp. 4325–29.

[9] S. Negahban and M. J. Wainwright, “Restricted StrongConvexity and Weighted Matrix Completion: OptimalBounds with Noise,” J. Machine Learning Research, vol.13, no. 1, Jan. 2012, pp. 1665–97.

[10] R. Hummel et al., “Mission Design for Compressive Sens-ing with Mobile Robots,” Proc. Int’l. Conf. Robotics andAutomation, May 2011, Shanghai China, pp. 2362–67.

[11] F. Hover et al., “One-step-Ahead Kinematic Compres-sive Sensing,” Proc. IEEE GLOBECOM Wi-UAV Wksp.,Dec. 2011, Houston, TX, pp. 1314–19.

[12] F. Fazel, M. Fazel, and M. Stojanovic, “Random AccessCompressed Sensing over Fading and Noisy Communi-cation Channels,” IEEE Trans. Wireless Commun., vol.12, no. 5, May 2013, pp. 2114–25.

[13] K. Kerse, F. Fazel, and M. Stojanovic, “Target Localiza-tion and Tracking in a Random Access Sensor Net-work,” invited paper, 47th Asilomar Conf. Signals,Systems and Computers, Nov. 2013, Pacific Grove, CA,pp. 103–07.

[14] S. Choudhary et al., “Active Target Detection with MobileAgents,” IEEE Int’l. Conf. Acoustics, Speech and Signal Pro-cessing, May 2014, Florence, Italy, pp. 4185–89.

[15] S. Choudhary et al., “Active Target Detection with Naviga-tion Costs: A Randomized Benchmark,” invited paper,52nd Annual Allerton Conference on Commun., Control,and Computing,Oct. 2014, Monticello, IL, pp. 109–15.

BIOGRAPHIESURBASHI MITRA [S’88, M’98, SM’04, F’07] received B.S. andM.S. degrees from the University of California at Berkeleyand her Ph.D. from Princeton University. Prior to her Ph.D.studies, she was a member of technical staff at Bellcore.After a six-year stint at the Ohio State University (OSU), shejoined the Department of Electrical Engineering at the Uni-versity of Southern California (USC), Los Angeles, whereshe is currently a professor. She is the inaugural Editor-in-Chief of IEEE Transactions on Molecular, Biological andMulti-Scale Communications. She is a Distinguished Lec-turere for the IEEE Communications Society for 2015–2016.She was a member of the IEEE Information Theory Soci-ety’s Board of Governors (2002–2007, 2012–2014) and theIEEE Signal Processing Society’s Technical Committee onSignal Processing for Communications and Networks(2012–2014). She is the recipient of: a 2015 Insight Maga-zine STEM Diversity Award, 2012 GLOBECOM Signal Pro-cessing for Communications Symposium Best Paper Award,2012 U.S. National Academy of Engineering Lillian GilbrethLectureship, USC Center for Excellence in Research Fellow-ship (2010–2013), the 2009 DCOSS Applications & SystemsBest Paper Award, Texas Instruments Visiting Professor (Fall2002, Rice University), 2001 Okawa Foundation Award,2000 OSU College of Engineering Lumley Award forResearch, 1997 OSU College of Engineering MacQuiggAward for Teaching, and a 1996 National Science Founda-tion CAREER Award. She currently serves on the followingIEEE award committees: Fourier Award for Signal Process-ing, James H. Mulligan, Jr. Education Medal, and the PaperPrize Award. She has been/is an Associate Editor for IEEETransactions on Signal Processing (2012–), IEEE Transac-tions on Information Theory (2007–2011), IEEE Journal ofOceanic Engineering (2006–2011), and IEEE Transactionson Communications (1996–2001). She has co-chaired theTechnical Program Committees of the 2014 IEEE Interna-tional Symposium on Information Theory in Honolulu,Hawaii, 2014 IEEE Information Theory Workshop in Hobart,

Tasmania, the IEEE 2012 International Conference on Sig-nal Processing and Communications, Bangalore India, andthe IEEE Communication Theory Symposium at ICC 2003 inAnchorage, Alaska. She served as General Co-Chair of thefirst ACM Workshop on Underwater Networks at Mobicom2006, Los Angeles, California. She was Tutorials Chair forIEEE ISIT 2007 in Nice, France, and Finance Chair for IEEEICASSP 2008 in Las Vegas, Nevada. She has held visitingappointments at the Delft University of Technology, Stan-ford University, Rice University, and the Eurecom Institute.She served as co-Director of the Communication SciencesInstitute at USC from 2004 to 2007. Her research interestsare in wireless communications, communication and sensornetworks, underwater acoustic communication, biologicalcommunication systems, detection and estimation, and theinterface of communication, sensing, and control.

SUNAV CHOUDHARY is a Ph.D. candidate in the Ming HsiehDepartment of Electrical Engineering at USC. He receivedhis B.S. degree in 2010 from the Indian Institute of Tech-nology, Kharagpur. He is the recipient of the followingawards and honors: 2015 USC Center for Applied Mathe-matical Sciences Prize Winner, 2014 IEEE International Sym-posium on Information Theory Student Travel Grant, 2013Ming Hsieh Institute Electrical Engineering Research FestivalBest Poster Award, 2010–2014 USC Annenberg GraduateFellowship, 2006–2010 IIT Kharagpur Goralal SyngalMemorial Scholarship, and the 20016–2010 CBSE MeritScholarship for Professional Studies. His present researchinterests are in the field of sparse signal approximation andits applications to underwater acoustic communications.

FRANZ S. HOVER [M’92] received his B.S.M.E. degree fromOhio Northern University, and his M.S. and Sc.D. degreesfrom the Joint Program in Applied Ocean Physics and Engi-neering at the Woods Hole Oceanographic Institution andMassachusetts Institute of Technology (MIT), Cambridge.He was a consultant to industry and then a member of theresearch staff at MIT, where he worked in fluid mechanics,biomimetics, and marine robotics. His research has led tocommercial development of the HAUV platform forautonomous ship hull inspection, advances in computa-tional tools for power systems, and innovations in subseaflow control technology. He is currently an associate pro-fessor at the MIT Department of Mechanical Engineering,with research focusing on design methodologies and com-plex marine systems. He has authored or co-authored over100 refereed papers.

ROBERT HUMMEL [S’11] received his B.S. and M.S. degrees inMechanical Engineering from MIT in 2009 and 2012,respectively, with a focus on robotics, mechatronics, andautonomous sampling techniques. He experience includesdesigning and building autonomous vehicles, machines,and mechanisms, and various electro-mechanical systems.He has also worked with numerical simulations and opti-mization. Currently, he is co-founder and design engineerat ATSE in Northborough, Massachusetts, an engineeringconsulting firm specializing in electric motors, generators,and drives for industrial, defense, and renewable energyapplications.

NAVEEN KUMAR [S’10] received his B.Tech. degree in instru-mentation engineering from the Indian Institute of Tech-nology, Kharagpur, in 2009. He is currently workingtoward his Ph.D. degree at the Department of ElectricalEngineering, USC. His research interests include machinelearning and signal/image processing for applications tospeech and multimedia problems. He was awarded theViterbi School Dean’s Doctoral Fellowship at USC in 2009.He has also worked on teams that have won the Inter-speech 2012 Speaker Trait Challenge and the Northern Dig-ital Inc. Excellence Awards at the 2014 InternationalSeminar on Speech Production (ISSP).

SHRIKANTH S. NARAYANAN [S’88, M’95, SM’02, F’09] receivedhis M.S., Engineer, and Ph.D. degrees in electrical engineer-ing from the University of California Los Angeles in 1990,1992, and 1995, respectively. From 1995 to 2000 he waswith AT&T Labs-Research, Florham Park, New Jersey, andAT&T Bell Labs,Murray Hill, New Jersey, first as a seniormember and later as a principal member of technical staff.He is currently a professor of engineering at USC, where hedirects the Signal Analysis and Interpretation Laboratory.He holds appointments as a professor of electrical engi-neering, computer science, linguistics, and psychology, andis the founding director of the Ming Hsieh Institute, Ming

The successful imple-

mentation of these

methods at sea will

have an impact on

many applications

and industries,

including environ-

mental monitoring,

aquatic eco-systems,

ocean accident

remediation, surveil-

lance for defense

applications, home-

land security, the oil

and gas industry,

aquaculture, geologi-

cal and oceano-

graphic science, and

marine biology.



Hsieh Department of Electrical Engineering, USC. He haspublished more than 600 papers and has been granted 16U.S. patents. His research interests include human-centeredsignal and information processing and systems modelingwith an interdisciplinary emphasis on speech, audio, lan-guage, multimodal and biomedical problems, and applica-tions with direct societal relevance. He is a Fellow of theAcoustical Society of America and the American Associa-tion for the Advancement of Science, and a member ofTau Beta Pi, Phi Kappa Phi, and Eta Kappa Nu. He is alsoan Editor of the Computer Speech and Language Journaland an Associate Editor of IEEE Transactions on AffectiveComputing, APSIPA Transactions on Signal and InformationProcessing, and the Journal of the Acoustical Society ofAmerica. He was also previously an Associate Editor of IEEETransactions on Speech and Audio Proseccing (2000–2004),IEEE Signal Processing Magazine (2005– 2008), and IEEETransactions on Multimedia (2008–2011). He has received anumber of honors, including Best Transactions Paperawards from the IEEE Signal Processing Society in 2005(with A. Potamianos) and 2009 (with C. M. Lee), and selec-tion as an IEEE Signal Processing Society Distinguished Lec-turer for 2010–2011. Papers coauthored with his studentshave won awards at Interspeech 2014 Cognitive and Physi-cal Load Challenge, Interspeech 2013 Social Signal Chal-lenge, Interspeech 2012 Speaker Trait Challenge,Interspeech 2011 Speaker State Challenge, Interspeech2013 and 2010, Interspeech 2009 Emotion Challenge, IEEEDCOSS 2009, IEEE MMSP 2007, IEEE MMSP 2006, ICASSP2005, and ICSLP 2002.

MILICA STOJANOVIC [(SM’08, F’10] graduated from the Univer-sity of Belgrade, Serbia, in 1988, and received M.S. (’91)and Ph.D. (’93) degrees in electrical engineering fromNortheastern University, Boston, Massachusetts. She was aprincipal scientist at MIT, and in 2008 joined NortheasternUniversity, where she is currently a professor of electrical

and computer engineering. She is also a guest investigatorat the Woods Hole Oceanographic Institution and a visitingscientist at MIT. Dr. Stojanovic is the recipient of the 2015IEEE/OES Distinguished Technical Achievement Award. Herresearch interests include digital communications theory,statistical signal processing and wireless networks, and theirapplications to underwater acoustic systems. She is anAssociate Editor for the IEEE Journal of Oceanic Engineeringand a past Associate Editor for IEEE Transactions on SignalProcessing and IEEE Transactions on Vehicular Technology.She also serves on the Advisory Board of IEEE Communica-tion Letters, and chairs the IEEE Ocean Engineering Society’sTechnical Committee for Underwater Communication, Navi-gation and Positioning.

GAURAV S. SUKHATME [SM’05, F’11] is a professor of com-puter science at USC. He received his undergraduate edu-cation at IIT Bombay in computer science and engineering,and M.S. and Ph.D. degrees in computer science fromUSC. He is co-director of the USC Robotics Research Labo-ratory and director of the USC Robotic Embedded SystemsLaboratory. His research interests are in robot networkswith applications to environmental monitoring. He hasserved as Principal Investigator (PI) on numerous NSF,DARPA, and NASA grants. He is a co-PI on the Center forEmbedded Networked Sensing, an NSF Science and Tech-nology Center. He is a recipient of the NSF CAREER awardand the Okawa Foundation research award. He is one ofthe founders of the Robotics: Science and Systems confer-ence. He was Program Chair of the 2008 IEEE Internation-al Conference on Robotics and Automation and the 2011IEEE/RSJ International Conference on Intelligent Robotsand Systems. He is Editor-in-Chief of Autonomous Robotsand has served as Associate Editor of IEEE Transactions onRobotics and Automation and IEEE Transactions on MobileComputing, and is on the Editorial Board of IEEE PervasiveComputing.

MITRA_LAYOUT.qxp_Author Layout 10/30/15 3:10 PM 6

structured sparse methods for active ocean observation...

Documents