JOE Manuscript v07

Embed Size (px)

Citation preview

  • 7/29/2019 JOE Manuscript v07

    1/28

    1

    A Combined Broadband Single-Beam and

    Narrowband Split-Beam SONAR System

    Jae-Byung Jung, Member, IEEE, Alexander B. Kulinchenko, Member, IEEE,

    Patrick K. Simpson, Senior Member, IEEE and James W. Tilley, Member, IEEE

    Abstract

    In this paper, we present a recent development in which the capabilities of a broadband single-beam and

    narrowband split-beam sonar system are integrated into a single system. This development not only provides an

    efficient tool to accommodate the benefits of both split-beam processing for target tracking (e.g. fish counting)

    and broadband processing for target identification (e.g. species identification), but also takes advantages of the

    synergism of both. Specifically, split-beam processing provides accurate angular information for targets within

    the ensonified coverage. The targets angular information is used in broadband processing to compensate for the

    irregularity of broadband spectrum caused by inherent uneven sensitivity of transducer across the effective beam

    angle. Simultaneously, exceptional range resolution derived by broadband pulse compression provides improvement

    over narrowband target resolution. Signal processing also allows a broadband sonar to emulate multiple narrowbandsonar systems by applying an instantaneous time-frequency representation. A series of tests under both controlled

    conditions at a lakeside acoustic test facility and in operational environments on rivers have been conducted.

    Index Terms

    Broadband, narrowband, single-beam, split-beam, fisheries, sonar, neural network, classification, time-frequency

    representation.

    I. INTRODUCTION

    A

    S the number of fish in rivers and streams diminishes and become threatened, endangered, or extinct,

    we see a growing concern from the local communities that is being met with increased funding and

    political attention. There is a need for better fish monitoring tools for the riverine environment.

    The primary tools for monitoring fish passage in an area where there are man-made impediments to

    travel are tags and sonars. There are many tagging solutions [48], ranging from sonar tags for actively

    following a fish [29], to clipping of a fin for later identification [2]. Sonar systems are used predominantly

    for assessing the quantity and behavior at key points along a fishs path [25][26].

    Split-beam sonar has a transducer that is typically divided into four quadrants. The target detection is

    determined by comparing the echoes received from all quadrants. Using the phase difference between the

    signals received by appropriate sections allows the target to be located within the beam [21].

    In the shallow water riverine environments, the sonars ping rate is set very high to provide multiple

    reflections from a single target and facilitate tracking. Each target detection is passed to a fish tracking

    routine that aggregates successive pings in the same or adjoining range cells into a track of the fishs path

    through the sonar beam [3][4][5]. Tracking fish is also the primary method of fish counting in riverine

    environments such as those found in Alaska and Canada. This information is the primary tool used for

    determining the amount of the fish runs [8][46][70][71].

    Although the narrowband split-beam systems can monitor the movement of fish tracks and estimate the

    size of targets by calculating target strength, target identification and spatial range resolution can be very

    limited with these systems.

    Jae-Byung Jung, Alexander B. Kulinchenko, and Patrick K. Simpson are with Scientific Fishery Systems, Inc., Anchorage AK 99516

    (email: [email protected]; [email protected]; [email protected]).

    James W. Tilley is with Alaska Native Technologies, LLC., Anchorage AK 99516 (email: [email protected]

  • 7/29/2019 JOE Manuscript v07

    2/28

    2

    The primary advantages of broadband sonar over the existing narrowband systems includes: (1) the

    broad spectral information for species identification, (2) the improved target detection, (3) more stable

    estimate of signal, and (4) improved spatial resolution.

    As a well-established theorem in information theory that holds that the information-carrying capacity of

    a communications channel depends on its bandwidth, the broadband acoustic data processing is expected

    to be a productive approach for species identification with the additional assistance of the appropriate

    analytical tools such as neural networks, expert systems, and new methods of discriminant analysis

    [42][43].

    Broadband systems also exploit time-bandwidth product of pulse compression to produce excellent

    spatial resolution as well as superior signal to noise ratio for improved target detection and stable signal

    estimate over conventional narrowband systems.

    Through the application of instantaneous time-frequency representation (ITFR) algorithms, it is possible

    to emuluate multi-channel narrowband spectral responses at every discrete sample from a broadband

    signal. Various kernels have been investigated to acquire better temporal and spectral resolution with less

    interference [9][10][28][44][72].

    Neural networks have been used to achieve over 80 % correct discrimination among frequency spectra

    for echoes from different species of fish [52] and zooplankton [68]. The results of several independent

    field tests have shown that broadband sonar provides a promising approach for discriminating various fish

    species using either (or both) spatial and spectral features [22] [31][32][33][34][35][51][56][60].

    In addition to the individual benefits of narrowband split-beam and broadband single-beam systems,

    the combination of these two capabilities in a single system supplies mutually advantageous cooperative

    functions. The accurate target location realtive to the Maximum Response Axis (MRA) provided by the

    split-beam portion of the system provides the information needed to compensate for the uneven broadband

    beam patterns. Conversely, the improved range resolution from broadband pulse compression helps split-

    beam tracks be distinguished clearly. Technical details of thsee synergistic combinations are discussed in

    the sections V and VI.

    From a practical point of view, especially for riverine fish migration monitoring, sonar systems are

    often deployed in very remote and secluded locations, often with great distances between the wet-end of

    the sonar where the transducer is housed, and the dry-end of the system where data is stored, power is

    provided, and communications with the system are established. Under such conditions, the wet-end (sonar

    head) and the dry-end (monitoring station) can be isolated by natural barriers and be required to cover

    a long distance. Therefore, analog sensor signal conditioning, such as band-pass filter and amplification,

    and analog-to-digital conversion as an initial step of data acquisition should be processed as early as

    possible in the wet-end sensor unit to prevent signal contamination from various noise sources during

    the transmission. Furthermore, the up front digitization and full digital signal processing in the wet-end

    unit allows digital data transfer in wired or wireless TCP/IP, providing additional flexibility in remote

    environments.

    The following sections discuss the overall architectural design of the system, details of system devel-

    opment, newly proposed data processing methods utilizing both narrowband and broadband processings,

    and various field tests for system evaluation.

    I I . FUNCTIONAL REQUIREMENTS

    A novel sonar system combining broadband single-beam and narrowband split-beam is designed to

    identify, count, and track migratory fish stocks for use in shallow water environments.

    The key requirements that were incorporated when combining the capabilities of each sonar system

    include the following:

    Switch between broadband and narrowband,

    Switch between single-beam and split-beam,

    Interleaving broadband single-beam and narrowband split-beam,

  • 7/29/2019 JOE Manuscript v07

    3/28

    3

    Programmable transmit power,

    Programmable receive gain,

    Programmable waveform,

    Wet-end signal conditioning and data acquisition,

    Network connectivity through TCP/IP, and

    Compatibility with existing softwares.

    When the sonar is operated as a single beam system, it can transmit broadband pulses or multi-frequencynarrowband pulses. This mode of operation is primarily used for determining species or size of the targets,

    such as fish, bubbles, and debris. When the sonar is operated as a split-beam system, the transmit frequency

    is fixed at the optimal frequency for transducer performance. This mode of operation provides split-beam

    processing for tracking targets as they move through the beam.

    Fig. 1. Interleaving ping patterns: 6 examples of flexible ping scheduling schemes are shown; narrowband pings only, 3 narrowband pings

    and 1 broadband ping, 2 narrowband pings and 2 broadband pings, 1 narrowband ping and 3 broadband pings, 1 narrowband ping and 1

    broadband ping, and broadband pings only, respectively.

    Furthermore, the system can change sonar parameters on a ping-by-ping basis allowing several shorter

    narrowband pulses in the split-beam mode to be interleaved with a smaller number of the longer broadband

    (or multi-frequency) pulses. Each sequence of interleaving pulses can be defined so that it provides

    adequate narrowband coverage for target tracking and provides additional broadband target identification.Fig. 1 includes six examples of ping scheduling schemes, including (1) narrowband pings only, (2) three

    narrowband pings and one broadband ping, (3) two narrowband pings followed by two broadband pings,

    (4) one narrowband ping followed by three broadband pings, (5) one narrowband ping followed by one

    broadband ping, and (6) broadband pings only, respectively.

    III. HARDWARE ARCHITECTURE

    The combined split-beam/broadband sonar system consists of three primary units: wet-end system; dry-

    end system; and underwater cable. The wet-end system is deployed in the water and is the main sonar

    unit with the computational resources and sensor module enclosed in an underwater housing. The dry-end

    system remains shore side. Its primary functions include power supply to wet-end system, communication

    with wet-end system, and wired or wireless TCP/IP with external computers or data storage. An underwatercable connects both wet-end and dry-end units together over a distance of as much as 100 m. Although

    all three units are needed and must function seamlessly together, we would like to focus our discussion

    on wet-end system in this section.

    The wet-end consists of three primary system segments: client PC, controller, and transducer as is

    illustrated in Fig. 2.

    A. Client PC Segment

    This is the main processing unit performing data acquisition, signal processing, and communication

    with the dry-end.

  • 7/29/2019 JOE Manuscript v07

    4/28

    4

    A low power single board computer (SBC) with capability of floating point computation executes

    signal processing tasks in coordination with an optional digital signal processing (DSP) board. One of the

    available serial ports (RS-232) in SBC transmits control commands to the Controller Segmentto all changes

    to the transmit power and receive gain. An integrated ethernet (TCP/IP) also allows communication with

    dry-end system.

    Data acquisition board passes on the waveform to Controller Segment so that it can be amplified

    according to the transmit power setup by SBC. At the same time, analog-digital conversions are executed

    in 4 separate channels simultaneously according to the receive gain setup by SBC.

    Once all the data processing tasks were completed, the data sets are either saved in the internal data

    storage or transferred to the dry-end system through TCP/IP in real-time.

    Fig. 2. The wet-end hardware architecture: component modules consist of client PC segment, control segment, and transducer segment.

    DC power lines and category-5 lines are fed from dry-end via underwater cable.

    B. Controller Segment

    This segment consists of custom electronics boards that have been developed exclusively for digital

    control, power distribution, transmit waveform amplification, and receive signal amplification.

    A field programmable gate array (FPGA) communicates with the SBC to control 16 programmable

    transmit power levels and to adjust 16 programmable receive gain levels to remain flexible for eachdeployment environment. Based upon the software configuration, the embedded analog power management

    module distributes the appropriate DC power levels to each of the functional modules across the entire

    wet-end system.

    Analog transmit (TX) module sends an amplified analog waveform to the transducer for signal projection

    in the water column. The analog receive (RX) module receives the raw signals in four separate hardware

    channels from each of the transducers quadrants and then executes the initial signal conditioning in

    hardware including preamplification, band pass filtering, and the application of signal gain.

  • 7/29/2019 JOE Manuscript v07

    5/28

    5

    C. Transducer Segment

    The transducer has four equally split quadrants that provide the ability to perform split-beam processing.

    Although each of the channels are recorded independently at reception, all quadrants simultaneously

    transmit the same waveform at the same power level for a composite conical beam projection.

    IV. SIGNAL PROCESSING

    The data processing paths vary depending on the mode of operation for that ping: broadband single-beam or narrowband split-beam. Each of these data processing paths has its own processing flow as well

    as assistance from its counter part. Fig. 3 shows the block diagram of the overall data processing chain

    for the received signal.

    Fig. 3. Overall data processing chain: First 4 blocks in each quadrant represent hardware implemented functions for initial common

    processing and the remainder of the blocks are software processed functions.

    The processing can be partitioned into five functional divisions: (1) the common hardware processing

    upon the signal reception, (2) channel mixer, (3) narrowband split-beam processing, (4) broadband single-

    beam processing, and (5) classifiers.

    A. Common Hardware Processing

    Each of four transducer quadrants are independently processed for signal conditioning, signal gain, and

    analog-digital conversion. Each of the four hardware channels have identical configurations for the initial

    common signal processing.

    1) Signal Conditioning: A fixed level of preamplification is applied to the incoming raw analog signal

    from each quadrant. A band pass filter passes only frequencies within the broadband range andattenuates frequencies outside that range.

    2) Signal Gain: A programmable gain amplifier further amplifies the analog signal with 16 discrete

    levels is controlled by an FPGA that receives commands from the SBC through RS-232.

    3) Analog-Digital Conversion: 4-channel simultaneous data acquisition is attained at high speed (500 kHz)

    and high sampling resolution (16 bits/sample).

    B. Channel Mixer

    Four individual discrete time series are fed into the channel mixer to produce one composite signal and

    four split-beam signals.

  • 7/29/2019 JOE Manuscript v07

    6/28

    6

    For the broadband processing, the composite signal is formed by the sum of the four discrete time series

    from each of the four quadrants. This composite signal is used mostly for broadband signal processing

    chain, although it can be also used for target strength calculations to help more accurate split-beam

    detection module.

    For narrowband processing, the sum of quadrants 1 and 2 are combined to represent upper half of the

    transducer, the sum of quadrants 2 and 3 are combined to represent left half of the transducer, the sum

    of quadrants 3 and 4 are combined to represent lower half of the transducer, and the sum of quadrants 4

    and 1 are combined to represent right half of the transducer, respectively.

    C. Narrowband Split-Beam Processing

    Narrowband split-beam data processing begins with phase comparison between upper and lower halves

    for vertical offset and between left and right halves for horizontal offset. Once the phase differentials

    are attained, a decimated version of the split-beam data and broadband target strength data are used for

    split-beam detection to provide target tracks for tracking and track classification.

    1) Phase Computation: Four discrete time series representing each half of the transducer configuration

    go through carrier removal, and then relative phase differences of the two opposite halves are

    calculated to provide horizontal phase offset and the vertical phase offset. Also, this data will be

    decimated from the original sample rate to a lower, more manageable, data rate before it is exported.2) Phase-Angle Conversion: Two phase differentials are converted into Cartesian angles using the

    physical configuration of the transducer.

    3) Split-Beam Detection: Broadband target strength calculations and narrowband phase/angle calcula-

    tions are used to build a robust target detector. A threshold value of the target strength and variance

    of the phase/angle differentials are the primary parameters used target detection. Also, a list of

    locations of individual targets are provided for tracking.

    4) Tracking: An - tracking system is used to track detected targets [3]. Track tuning to determine theappropriate number of missed detections, association windows, and track initiation and termination

    criteria will be conducted during data analysis. Also, more complicated tracking algorithms such as

    joint probability data association (JPDA) and multi-hypothesis tracker (MHT) can be applied when

    the complexity of the tracking increases [4].

    D. Broadband Single-Beam Processing

    Broadband single-beam processing starts with the composite signal from all four quadrants. Transmis-

    sion loss over the data collection range is compensated by time-varying gain, pulse compression provides

    excellent range resolution, and the spectral estimation offers broad spectral features for classification.

    1) Target Strength: The composite broadband signal is used to calculate target strength. Unlike the

    narrowband target strength with frequency dependency, broadband target strength can provide more

    reliable estimation by aggregating broadband response. Accurate target strength helps split-beam

    detection module in split-beam processing chain.

    2) Time-Varying Gain: The transmission loss is computed from a combination of spherical spreadingloss (20log R) and absorption (R), where R is the range [m] and is the attenuation coefficient[dB/m]. The application of the transmission loss results in a range-compensated discrete time seriesthat is propagated through the remainder of the system.

    3) Pulse Compression: A matched filter is used to produce the compressed pulse, which provides im-

    proved spatial resolution for signals when a sufficiently large bandwidth is available. Mathematically,

    the matched filter is computed by taking the inverse Fourier transform of the product of the Fourier

    transform of the transmit signal with the complex conjugate of Fourier transform of the received

    signal as described below.

    xMF = F1{F(xTX) F

    (xRX)} (1)

  • 7/29/2019 JOE Manuscript v07

    7/28

    7

    where xMF is a compressed pulse time series, xTX is a transmit time series, xRX is a received timeseries, F() is Fourier transform.The effective pulse width of the compressed pulse, TMF, becomes reduced by a factor of the time(T)-bandwidth(B) product. For example, a pulse duration of 667 s is compressed to 16.7 s for the asonar with 60 kHz of bandwidth.

    TMF = T 1

    T B=

    1

    B(2)

    A few detection criteria including peak threshold, peak window size, and peak separation are applied

    to the pulse-compressed data series, producing a list of locations (in range) of the start of each

    detected target. A peak detector is used on the detection signal to locate echoes in the return

    time series. The peak detector is the technique which first make an estimate about the background

    signal level in that neighborhood (e.g., by taking an average around the point) and then comparing

    the sample in the center of the neighborhood with the background signal estimate. When a peak

    is located, it signifies that a portion of the return has matched up with the filter and therefore

    represents a likely reflector in the water. If there is significant separation between last detected echo

    and current detection, then a detection is declared and echo extraction occurs. These three sonar

    configuration echo detection settings, as well as target strength, are used as coefficients for matched

    filter detection process.

    4) Spectral Estimation: Spectral estimation is performed using the Fast Fourier Transform (FFT). For

    an FFT of size N, the number of frequency bins is N/2. Frequency bin width is the bandwidthof each of the contiguous bins across the usable bandwidth. Frequency bin width is determined by

    the size of the FFT and the sample rate. For the classification purposes, the derived echo spectrum

    must undergo normalization to accommodate differences in frequency response between different

    transducers. This normalization process allows the classifier to be transducer independent

    5) Beam Compensation: Uneven broadband beam patterns and sensitivity can be adjusted by compen-

    sating for differences in the beam at various frequencies depending on the targets location relative

    to the maximum response axis. Details of this beam compensation process are discussed in section

    V.

    E. Classifiers

    Three individual classifiers (track, size, spectral) make independent decisions based on features derived

    from track, size, and spectrum. Neural networks are trained using known signals to build each classifier,

    and then tested with unseen patterns to evaluate the performance of each classifier. Afterward, the final

    decision is made from the fusion classifier for target identification. Alternatively, in place of the fusion

    classifier, decision level inference using fuzzy logic or other artificial/computational intelligence algorithms

    can be implemented.

    1) Track Classifier: The kinematic features such as location, velocity, and movement are extracted

    from the - tracker and used to identify the class of underwater target. Also, statistical acousticfeatures such as target strength and spectral/temporal information from individual detections along

    each track are added in the pool of features for target identification.2) Size Classifier: The samples associated with a detected target will be used to estimate the size of a

    target. Targets that have elongated detection envelopes represent larger targets. As an example, large

    aggregations of closely spaced detections are indicative of a marine mammal and sparse detections

    of only a few detections within a range window are indicative of fish or other small targets.

    3) Spectral Classifier: The spectra produced from the spectral estimation process are used as features for

    classification. The classifier is iteratively improved using data from a variety of target assemblages.

    The classifier that is utilized is similar to the Parzen estimator that produces probability densityfunctions for each class. Classes will depend on the quality of the ground-truth data. The more

    diverse and varied the data set, the better the performance.

  • 7/29/2019 JOE Manuscript v07

    8/28

    8

    4) Fusion Classifier/Identification: The classification decisions produced by the track classifier, size

    classifier, and spectral classifier are merged to form the fusion classifier for the final target identi-

    fication.

    V. BEA M COMPENSATION

    One of the synergistic effects of interleaving narrowband and broadband pings is that is allows for

    the compensation of the broadband spectrum using the information from the split-beam processing ofprevious ping. In this section, we describe this compensation process and explain how the narrowband

    split-beam can help broadband spectrum adjustment to fully utilize the dynamic range of the directivity

    and sensitivity of the transducer.

    Unlike a narrowband-only system with fixed directivity pattern and sensitivity in a fixed frequency, a

    broadband transducers directivity pattern and sensitivity are a function of frequency. Fig. 4 shows the

    directivity patterns of the broadband transducer that was installed in this system. These directivity patterns

    were calibrated for eight separate narrowband waveforms in a controlled environment. Note that the higher

    the frequency, the narrower the effective beam angle. Also, when a Liner Frequency Modulated (LFM)

    waveform is transmitted, the target response with higher frequency component tends to return lower energy

    if the target is slightly off MRA.

    Fig. 4. Transducer directivity patterns for 8 different narrowband frequencies from 150 kHz to 200 kHz at 10 kHz intervals.

    Split-beam processing provides an accurate bearing angle of a target off MRA. Lets assume the z-axis

    is the MRA, the x-axis is the horizontal reference, and the y-axis is the vertical reference. Note that theaxes x and y are not in polar domain but in Cartesian coordinate. When a target is detected along thetarget vector, T, and the split-beam processing provides the horizontal and vertical angular offsets, and, relative to the transducer as illustrated in Fig. 5, we can calculate the bearing angle of T relative toMRA as described below in (3).

    = arctan1

    tan2 + tan2 (3)

    Although the measured directivity patterns are not perfectly symmetrical, especially after first side lobes

    outside 8 13 depending on frequencies, we can focus on the main lobes of the directivity patterns

  • 7/29/2019 JOE Manuscript v07

    9/28

    9

    Fig. 5. Split beam target detection and relationship with transducer geometry. See text for a description of the variables.

    on each frequency to construct the directivity pattern compensation factor as a function of frequency and

    bearing angle off MRA as illustrated in Fig. 6.

    As split-beam processing provides accurate bearing angle of the target off MRA, and the broadband

    pulse compression and instantaneous time-frequency representation (ITFR), which is discussed in the

    following section, provides precise frequency component of target return, we can estimate how much

    deviation of sensitivity we lose by interpolation from calibration points.

    Fig. 6. Transducer directivity patterns (3D regression). Using this data and knowing the position of the target relative to MRA, it is then

    possible to provide a uniform target response across the entire bandwidth.

  • 7/29/2019 JOE Manuscript v07

    10/28

    10

    Another nonlinear impact on the broadband spectrum is the transducer sensitivities from transmit voltage

    response (TVR) and open circuit voltage response (OCVR). The Figure of Merit (FOM) is the sum of TVR

    and OCVR, and is useful in providing a relative gauge of the performance of piezoceramic transducers.

    Therefore, the normalized total sensitivity can be calculated by combining TVR and OCVR, and then

    take out its maximum value as described in (4), which represents the relative deviation of sensitivity

    when the frequency component deviates from the resonance frequency of the transducer. Fig.7 shows the

    normalized total transducer sensitivity that was also measured in a controlled environment as a function

    of frequency.

    SN = F OM F OMmax. (4)

    Fig. 7. Normalized transducer sensitivity of high frequency version of SciFish2100-D system. Overall sensitivity (FOM) is normalized at

    the center (main resonance) frequency at 165 kHz.

    Therefore, we can combine both directivity-related adjustment factor and the sensitivity-related factor

    together and build the total beam compensation as a function of frequency and target angle off MRA as

    shown in Fig.8. In other words, if we can compensate for the transducer-inherent sensitivities using beam

    compensation, we will have more target-subjective frequency distribution instead of sensor-subjectiveversion.

    V I. EXTENSION TO MULTI-F REQUENCY SONAR

    As a broadband signal provides various perspectives within the specific spectral range, it is sometimes

    very useful to convert it into a finite number of narrrowband time series in various frequencies. The

    biggest advantage in this processing is to emulate multiple narrowband sonars using a single broadband

    sonar, although the increased processing burden for real-time implementation must be considered. This

    section explains the process of extending a broadband sonar to its equivalent multiple narrowband sonars.

  • 7/29/2019 JOE Manuscript v07

    11/28

    11

    Fig. 8. Total beam compensation (3D regression) considering target bearing angle and frequency response. Transducer-inherent non-linearity

    is cancelled out by taking out the total beam compensation and the generalized target response can be achieved.

    A. Time-Frequency Representation

    In general, Fourier Transform converts a temporal signal, x(t), to spectral information as described in(5). This is a very useful tool when analyzing a time series from the spectral perspective, but the temporal

    information is lost after the conversion.

    Fx(f) =

    x(t)ej2ftdt (5)

    One of the possible methods to retain the temporal information after the transformation is a Short Time

    Fourier Transformation (STFT). As described in (6), the STFS uses a sliding time window, w(t), to retaintemporal information while applying the Fourier Transformation. One of the big drawbacks of the STFT

    is that it has a fixed overall resolution, because the window function determines whether there is good

    frequency resolution or good time resolution. A wider time window provides better frequency resolution

    but poorer time resolution whereas a narrower window gives better time resolution but poorer frequency

    resolution.

    Sx(t, f) =

    x()w(t )ej2fd (6)

    When a precise frequency response needs to be achieved instantaneously or at very localized instant

    in time, a Time-Frequency Representation (TFR) can be applied. The analytically appealing Cohens

    Generalized Time Frequency Representation (GTFR) of a temporal signal, x(t), is shown in (7) whenseeking a performance improvement of the spectrogram and making instantaneous representation of both

    temporal and spectral information available.

    Cx(t, f) =

    (, )x

    t

    2

    x

    t +

    2

    ej2dd (7)

  • 7/29/2019 JOE Manuscript v07

    12/28

    12

    It is now straightforward to state some of the commonly used constraints imposed on the GTFR and

    their corresponding interpretation as kernel constraints using the kernels listed in (8).

    (, ) =

    1, W igner V illee(2)

    2/, Choi Williamsej, Kirkwood Rihaczekej2||, P agecos(), M argenau Hillsinc(), Born Jordong()||sinc(2a), Z hao Atlas Marks

    (8)

    where , a, and g() are kernel parameters selected to best suit the application.

    B. Real-Time Implementation of Instantaneous Time Frequency Representation (ITFR) for Multi-Frequency

    Sonar

    Performance of the GTFR depends on the choice of kernel in different applications. Those kernels

    listed above are well known to reduce interference (cross-term) during the conversion. However, complex

    kernels reduce also the feasibility of implementing the algorithm performing in real-time, especially when

    a high ping rate is required to monitor migratory fish schools in riverine environment. After reviewing each

    of those algorithms, the Wigner-Ville Distribution (WVD) was best determined to be the most practical

    choice for instantaneous time-frequency representation.

    The WVD, shown in (9) can be seen as a special case of GTFR, Cx(t, f), when the kernel, (, ), is 1.Although there is the potential not to be able to remove the interference without a more effective kernel,

    computational complexity of Wigner-Ville distribution is much lower than those of the listed kernels

    with temporal or/and spectral constraints. When considering these trade-offs, the benefit for real-time

    implementation with reasonable computational burden overshadows the aforementioned drawbacks.

    W Vx(t, f) =

    x

    t

    2

    x

    t +

    2

    ej2fd (9)

    An example of field data collected in Ship Creek, Anchorage, AK, on August 12, 2004 illustrates the

    ITFR process with WVD. Fig. 9 shows a snapshot of a ping recorded in a single composite broadband

    channel at sampling frequency of 500 kHz. Because of the protection circuit in hardware electronics,

    all signals were clipped at 1 V. Thus, the actual LFM transmit waveform from 135 kHz to 195 kHzwith amplitude of 100 Vpp appears at 0667 sec, followed by a relatively low transducer ringing until840 sec, but they appear clipped in the collected time series. Meanwhile, a clear target return of aChinook Salmon (Oncorhynchus tshawytscha) shows up between 3.72 msec and 4.40 msec.

    When we focus on the target return of the time series in Fig. 9, a power spectrum of the target can

    be calculated using the discrete version of (5) with FFT size of 512. In this calculation, all discrete

    samples between 3.72 msec and 4.40 msec were included to produce the target power spectrum shown in

    Fig. 10, which coincides with the frequency range of the transmit waveform. The shape of the distribution

    within the broadband frequency range between 135 kHz and 195 kHz can be used as features for target

    identification in spectral classifier.

    One of the properties of a spectrogram produced from a STFT in (6) is the ability to retain the temporal

    information while providing spectral features. Fig. 11 shows the spectrogram of a whole time series using

    an FFT size of 128 and time window size of 125 samples (=250 msec). The spectral features show up

    along the time axis except both edges of the time series caused by time window, but smearing features

    dilute the resolution in both time and frequency axes.

    To retain finer frequency resolution at every time interval with realistic computational expense, the

    WVD ITFR is used as shown in Fig. 12. The spectral features show up along the time axis at every

  • 7/29/2019 JOE Manuscript v07

    13/28

    13

    0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

    x 103

    1.5

    1

    0.5

    0

    0.5

    1

    1.5Received Time Series (Composite Broadband Signal)

    Time [second]

    SignalLevel[V]

    Fig. 9. An example of field data collected in Ship Creek, Anchorage, AK, on August 12, 2004. LFM from 135 kHz to 195 kHz for 667sec

    pulse length was transmitted and is seen at the beginning of the time series. A clipped version of transmit waveform that appears between 0

    and 667sec was recorded, although the actual transmit signal level was approximately 100 Vpp. Transducer ringing followed from 667 sec

    and faded away at about 840sec. The target return from a King Salmon shows up clearly between 3.72 msec and 4.40 msec.

    instance without black-out sections on both edges, and much sharper resolution was achieved in both time

    and frequency axes.

    A spectrum is available at every time interval and extracting an horizontal strip at a certain frequency

    from Fig. 12 represents a narrowband-equivalent time series.

    Besides, to attain spectral energy at a fixed frequency, an envelope of each time series needs to be

    calculated by shifting the spectrum by the constant frequency offset, fo. The relationship between timedomain and frequency domain in frequency shift is denoted in the equation below.

    e2f0tx(t) Fx(ej2(ff0)) (10)

    Thus, multi-channel time series from a single broadband signal can be derived to produce multi-

    frequency narrowband signals. Fig. 13 focuses on the time segment between 3.72 msec and 4.40 msec

    to show 16 narrowband time series extracted from a single broadband signal. In this example, narrowband-

    equivalent target returns are in frequencies 135.0 kHz, 138.2 kHz, 144.7 kHz, 147.9 kHz, 151.1 kHz, 154.4 kHz

    160.8 kHz, 164.0 kHz, 167.3 kHz, 173.7 kHz, 176.9 kHz, 180.2 kHz, 183.4 kHz, 189.8 kHz, 193.1 kHz, and

    196.3 kHz.

    Although some artifacts with negative energy are present after computation (a well-known side-effect

    of the WVD), which is not physically possible, some simple methods such as taking the absolute values

    or discarding negative values can easily cure these processing anomalies.

    VII. EXPERIMENTAL RESULTS

    There have been several field tests of the combined broadband/split-beam sonar system, including

    controlled experiments in pools and a lake and deployments along a river for the duration of a large

  • 7/29/2019 JOE Manuscript v07

    14/28

    14

    0 50 100 150 200 25040

    30

    20

    10

    0

    10

    20

    30Target Power Spectrum

    Frequency [kHz]

    Power[dB]

    Fig. 10. Power spectrum from target detection. All samples in the time window from 3.72 msec to 4.40 msec were extracted and the power

    spectrum was calculated to provide the distribution of energy in frequency domain. Most of the energy is concentrate within 135 kHz195kHz

    according to the transmit waveform of LFM.

    Fig. 11. Spectrogram with FFT size of 128 and time window size of 125 samples (=250 msec). The spectral features along the time axis

    are available, but the resolutions in time and frequency depend on the FFT size and time window size. Also, spectral features on both ends

    of time series are not available due to the use of time window.

  • 7/29/2019 JOE Manuscript v07

    15/28

    15

    Fig. 12. Wigner-Ville distribution for instantaneous time frequency representation. Improvement of resolution is noticeable in both time

    and frequency domains compared with STFT.

    Fig. 13. Sixteen individual time series are extracted from a broadband signal using ITFR with Wigner-Ville distribution. Each of the signals

    is equivalent to its narrowband time series at each of those 16 constant frequencies. Because the frequency of transmit time series (LFM) is

    linearly increasing in time, the lower frequency components arrive earlier than those of higher frequency components in the returned target

    echo.

  • 7/29/2019 JOE Manuscript v07

    16/28

    16

    salmon run. All of the experiments were conducted in Alaska, with pool tests being performed during

    the winter and lake and river testing during the summer. These experiments were conducted to verify that

    the combined sonar operates correctly in their individual split-beam and broadband modalities. In a later

    paper, we will present results of the beam compensated broadband performance in a series of controlled

    classification studies [36].

    Three tests are described in the sections that follow. The first of these tests focused on the ability of

    the split-beam sonar system to track and enumerate fish during high passage rates. The remaining two

    tests focused on the ability of the broadband sonar system to perform target identification. The first of

    these broadband tests was conducted in a lake focused on the ability to discriminate a swimmer from

    similar targets [35]. The second broadband test was conducted with Pacific Northwest National Lab for

    discriminating smolt-size Chinook salmon from aquatic macrophytes [34].

    A. Splitbeam Experiment

    The primary purpose of the field test in Wood River, Alaska, was to evaluate narrowband split-beam

    sonar processing. Wood River is known to have a large Sockeye Salmon (Oncorhynchus Nerka) return

    between late June and early July. The Alaska Department of Fish and Games (ADF&G) monitors this

    migration from several counting towers along the rivers shore. The tower crews visually count individual

    fish passing underneath the tower for 10 minutes every hour during the period of fish run when the majorityof the fish are migrating upriver and then estimates the hourly fish passage rate from the subsample.A combined broadband/split-beam sonar system was deployed in the Wood River near Aleknagik,

    Alaska, fully operating with minimum interruption for power and data backup from 6/26/06 to 7/4/06.

    Fig. 14 shows the combined sonar system deployed next to an ADF&G counting tower.

    Fig. 14. ADF&G counting tower and combined sonar system deployed to monitor Sockeye salmon migration in Wood river, AK.

    Throughout the test period, the sonar system was continuously transmitting 200 sec narrowband pulsesat 15 pings per second. Periodically, a tower count was made when the visibility in the river was adequate.

    The tower counts and the sonar counts were performed independently and the results were provided to a

    third party (Dr. Tim Mulligan). The comparative results were not reported until after the summer season

    was completed and Dr. Mulligan had sufficient time to perform his comparative evaluation.

  • 7/29/2019 JOE Manuscript v07

    17/28

    17

    The comparative evaluation was conducted using all visual tower counts. Each of the corresponding

    sections of recorded narrowband signals were extracted. Utilizing the EchoView application from Sonar-

    Data (now Myriax, www.echoview.com), the split-beam data was processed offline to count fish. The

    number of fish tracks in each ten-minute tower count session were logged and compared with that of

    visual tower count. Table I enlists date, start time, end time, visual fish count from counting tower, and

    fish track count from off-line sonar data analysis of each of the 20 sessions where tower counts were

    taken.

    TABLE I

    SPLIT-BEAM SONAR FIS H COUNT COMPARISON BETWEEN VISUAL TOWER COUNT AND SONAR SYSTEM TRACK COUNT

    Date Start Time End Time Tower Count Sonar Count

    6/29/06 09:21 09:31 441 430

    6/29/06 10:00 10:10 165 156

    6/29/06 11:11 11:21 868 654

    6/29/06 11:55 12:05 467 426

    6/29/06 14:03 14:13 373 348

    6/29/06 14:57 15:07 770 584

    6/30/06 00:02 00:12 0 6

    6/30/06 06:30 06:40 918 647

    6/30/06 15:24 15:34 505 491

    6/30/06 15:41 15:51 217 215

    6/30/06 18:40 18:50 430 397

    6/30/06 18:50 19:00 590 499

    6/30/06 19:00 19:10 651 478

    7/1/06 06:23 06:33 1284 902

    7/1/06 23:57 00:07 1074 839

    7/2/06 23:11 23:21 1023 744

    7/2/06 06:56 07:06 1210 923

    7/3/06 06:51 07:01 953 771

    7/3/06 11:23 11:33 330 304

    7/3/06 14:59 15:09 1710 939

    For the analytical comparison of each fish counting technique, each fish count was converted into an

    hourly passage rate and directly compared as shown in Fig. 15. The horizontal axis represent the visual

    fish count from counting tower, the vertical axis represents the number of fish tracks extracted from

    sonar data recorded at the same time, and the dotted diagonal line represents a virtual 1:1 mapping as a

    reference, respectively.

    Also, the 2nd order polynomial curve fitting, p(x), that is described in (11) can be derived from the 20data points and depicted in a solid curve in Fig. 15.

    p(x) = 0 + 1x + 2x2 + 3x

    3 (11)

    where 0=-2.5267x109, 1=-4.0365x10

    6, 2=0.8510, 3=160.8368, and x is a vector including 20 sonarcounts.

    Of the nine sessions where the hourly fish passage rate was under 3000 fish per hour, the sonar counts

    and tower counts closely agreed with a very small margin. Deviations are seen in the remaining sessions

    as the fish passage rate increased. There are two possible reasons for the sonar under-counting when the

    population increased drastically. One possible cause is the fish swimming out of sonar beam angle so that

    the system could not sufficiently ensonify those fish. The other cause could be that many fish would have

    traveled too close to each other to make multiple tracks appear to be a single track in the recorded data

    although they swam inside the beam angle.

    Considering the majority of the Salmon migrations were reported with less than 3000 hourly fish passage

    rate, the split-beam performance evaluation in Wood Rover was considered successful.

  • 7/29/2019 JOE Manuscript v07

    18/28

    18

    Fig. 15. Fish count comparison for hourly passage rate. The square data points represent the relationship between the visual fish count

    from counting tower (horizontal axis) and the number of fish tracks extracted from sonar data (vertical axis) along the dotted diagonal line

    for 1:1 mapping as a reference. The 3rd order polynomial fitting curve shows the trend of the performance (solid line).

    B. Broadband Experiment - Swimmer Identification

    A field test has been conducted to demonstrate the capability of the broadband sonar system to identify

    swimmers from other underwater objects. Classifiers were built and tested from data collected from an

    swimmer donned with mask, fins, and a scuba tank in Marion Lake, Alaska in 2005. The test results using

    low frequency version of the combined sonar system (60 kHz120 kHz) are described in the followingsubsections.

    Two known targets (air filled 2-liter plastic bottles) were placed in the middle of water column as

    reference targets emulating the approximate size and shape of human lungs. One bottle was placed at

    30 m and the other was placed at 71 m range. A swimmer swam in approximately a straight line along the

    MRA from a location 22 m from the transducer toward the first bottle at 30 m. After passing the first bottle,the swimmer came to the surface to adjust his swimming direction, and then swam in approximately a

    straight line toward the second bottle. The swimmer again surfaced to re-adjust the swimming direction

    at 65 m, and swam past the 2nd bottle up to 95 m range where he turned around and attempted to retrace

    the same route back. The swimmer surfaced several times to re-adjust his direction as he was coming

    back.

    Throughout the swimming sequence there is a significant number of air bubbles produced from the scuba

    gear and the swimming motion, especially when the swimmer was at close range and at the surface. The

    presence of these bubble clouds had the effect of diminishing the separation between the swimmer and

    the air-filled bottles for the classifier, as the air bubbles were necessarily included with the swimmer and

  • 7/29/2019 JOE Manuscript v07

    19/28

    19

    represented a significant portion of the backscatter from the swimmer.

    Only a small portion of the collected data were extracted for spectral comparison using principal

    component analysis (PCA). Approximately 200 echoes from the second bottle at 71 m were compared with

    those of the swimmer at 6370 m range outward bound (flippers toward the sonar, generating bubbles).The extracted spectra within the frequency of interest (60 kHz120 kHz) were converted into principalcomponents.

    Fig. 16 shows the two largest principal components that provide the greatest separation between two

    target types (swimmer and bottle). The use of PCA is very useful as an intial analysis tool as it provides

    a preliminary indication of separability among different targets in high dimensional feature space. The

    PCA test was very promising, with only two principal components out of 63 distinct spectral energy bins

    indicating the possibility of separation among different target classes.

    Fig. 16. Principal component analysis: 2-dimensional presentation of 2 different targets. Two largest principal components that provide

    the greatest separation between two target types (swimmer and bottle) were plotted as a preliminary observation of separability. There are

    several points that appear to be overlapped, although two noticeable clusters can be distinguished.

    Since some degree of separation was observed between the swimmer and the bottle data, we anticipate

    that the artificial neural network would improve the discrimination between the two types, because (1)

    the artificial neural network utilizes the full degree of freedom of feature sets without data loss, and (2)

    the artificial neural network provides flexible non-linear separation.

    Having used PCA to demonstrate that separability can be achieved, an artificial neural network was

    trained to build a classifier. In this example, a multi-layered perceptrons was trained using 10% of

    the randomly extracted spectra from two different data sets: swimmer target and non-swimmer target,

  • 7/29/2019 JOE Manuscript v07

    20/28

    20

    respectively. The swimmer data set was collected when the swimmer was on the surface of the water

    for adjusting the swimming direction and the bottle data set was collected when the swimmer was not

    present.

    Fig. 17. Classification result for swimmer vs. bottles at Marion lake, AK. The swimmer track stands out with air bubbles that were caused

    by the swimming (kicking) motion. Two tracks of bottles at about 30 m and 70 m were classified differently from that of swimmer.

    The trained neural network was used to colorize target detections by class. Fig. 17 shows the final

    classification result. The black detections represent the swimmer and the gray detections represent thebottle. Although there are several intermittent misclassifications, the overall track of the swimmer appears

    obvious as well as that of the bottle. In fact, this plot does not simply show the testing of the classification,

    but includes both training and testing. However, using only a small fraction of the entire data set provided

    very good generalization capability.

    C. Broadband Experiment - Smolt Fish Identification

    The Pacific Northwest National Laboratory conducted a study of the broadband fisheries sonar systems

    ability to discriminate between Chinook Salmon smolt and aquatic macrophytes [34]. Like the counting

    study that was performed with the split-beam evaluation, a blind study was conducted in which a trusted

    third party (R. Johnson) held back the identity of the test subjects until after the system reported its results.Data collection was conducted at one of the rectangular ponds in 100K-basin area at the Hanford

    Site near Richland, WA, from August 31 through September 2, 2004. All data collection work was

    conducted with a horizontal acoustic projection across the narrow axis of the pond with the anticipation

    of discriminating Chinook salmon smolt, in dorsal or ventral aspect from various assemblages of aquatic

    macrophytes.

    The targets were located at 16.7 m from the transducer, and the back-wall was 1.7m beyond the target

    (18.4 m from the transducer). The water depth of the pond was approximately 5.1m, and the transducer

    and the acoustic targets were deployed in the middle of the water column approximately at 2.55 m from

    the bottom of the pond as illustrated in Fig 18.

  • 7/29/2019 JOE Manuscript v07

    21/28

    21

    Fig. 18. Geometry of the pond that field test was conducted in the Hanford Site near Richland, WA. Dimensions of the pond and the

    relative distance between target and transducer are illustrated.

    Juvenile fall Chinook salmon 130-150 mm in length were transported to the site live in a large aerated

    cooler. Just prior to being tested fish were anesthetized using MS 222. All fish were tested in good to

    excellent condition (i.e. minimal scale and fin loss). Macrophytes were collected at a marina in North

    Richland and consisted of water milfoil and elodea.

    A special tool was devised to control the aspect of targets. A long bar was tied across the top of the

    hand railings on either side of a walkway and another bar attached at one end of the long bar was rotated

    by a rope to change the aspect angle of the target. The fish and macrophytes were positioned in the centerportion of the sonar beam along the MRA. To suspend the fish, clear monofilament line was fastened to

    the bar and the fish was fastened to the line by threading line down one side of the mid portion of the

    fish. A lead weight was added approximately 4 ft below the fish to keep the line tight. To achieve the

    proper fish orientation a separate section of monofilament was attached the pivoting arm 4 ft from the fish

    line. The line was attached to mouth region and weighted at the lower end. By moving the pivoting arm

    the fish aspect could be changed. For all tests on either a dorsal or ventral aspect was used. Macrophytes

    were deployed by using a separate weighted line and attached using a small section of monofiliament to

    a loop in the line.

    The transducer was mounted at the end of a metal tube, which was attached to a wood block tied to

  • 7/29/2019 JOE Manuscript v07

    22/28

    22

    the bottom rails of the path to keep the transducer in the middle of the water column firmly pointing

    straight to the target under the changing weather condition.

    The dry end side of the combined sonar system, incuding a PC, monitor, and terminal box, were set

    up in a rented moving truck, which was about 50 m away from the wet-end deployment. 100 m-long

    SciFish2100 transducer cable connected between the transducer and the terminal box.

    The first 1.5 days were used for the data collection and the creation of the neural network classifier.

    The remainder of the time was used to collect blind test sets. In total, 8 different fish and 7 different

    macrophytes were used to collect training data. Also, 86 unknown targets were used to collect blind test

    datasets, whose identification would be released after the test results were provided.

    All 4,257 pings of the collected data were fed to train a neural network. Of the classifier samples,

    2419 samples are actual fish and the remaining 1838 samples are actual macrophytes. The classifier made

    probabilistic decision in terms of the degree of similarity between 0 and 1.0. If the value is close to 0,

    the sample is fish, and if it is close to 1.0, it is macrophyte. However a hard (deterministic) decision can

    also be made by applying a threshold value.

    A total of 86 blind test sets were collected during the 2nd and 3rd day, but the identifications of 5

    blind test sets (ID 20911, 20919, 21110, 21125, 21134) were released later and used for training the

    neural network classifier. Thus, those five data sets were not considered as true blind test data sets and

    they were not counted for testing. Since all blind test data sets were collected under in a more controlled

    environment and under slightly different weather conditions than the training data sets, adding those sets

    would make training data sets more representative of a varying realistic environment.

    A total of 81 unknown and 5 known data sets were tested to evaluate the performance of the classifier

    built by the known data sets. Table II shows the summary of the blind test results. The columns under

    Classified as were outputs from the neural network classifier. Each decision with higher probability

    above 50% was declared as corresponding class. The columns under Reality (Ground Truth) were

    actual truth table released after we had provided the outputs of classifier to Pacific Northwest National

    Lab. The column Result tells if the classifier made correct decision or not.

    TABLE II. CLASSIFICATION RESULT FROM BLIND TESTS BASED ON THE CLASSIFIER BUILT USING THE KNOWN SAMPLES.

    Order Target ID PingsClassified as Reality (Ground Truth)

    ResultFish Macrophyte

    Count % Count % Decision Actual Aspect

    1 11451 251 239 95.21 12 4.78 Fish Fish dorsal TRUE

    2 11456 161 81 50.31 80 49.68 Fish Fish dorsal TRUE

    3 11513 260 184 70.76 76 29.23 Fish Fish ventral TRUE

    4 11522 260 16 5.38 281 94.61 Macrophyte Macrophyte small/short TRUE

    5 11531 211 14 6.63 197 93.36 Macrophyte Macrophyte large/short TRUE

    6 11541 257 189 73.54 68 26.45 Fish Fish dorsal TRUE

    7 11548 248 19 7.66 229 92.33 Macrophyte Macrophyte long/skinny TRUE

    8 11554 137 125 91.24 12 8.75 Fish Fish dorsal TRUE

    9 11604 238 90 37.81 148 62.19 Macrophyte Macrophyte no entry TRUE

    10 11609 257 113 43.96 144 56.03 Macrophyte Fish dorsal FALSE

    11 11616 141 132 93.61 9 6.38 Fish Fish ventral TRUE

    12 11630 277 78 28.15 199 71.84 Macrophyte Macrophyte very small TRUE

    13 11642 233 211 90.55 22 9.44 Fish Fish dorsal TRUE

    14 11648 262 213 81.29 49 18.70 Fish Fish ventral TRUE

    15 11705 175 170 97.14 5 2.85 Fish Fish ventral TRUE

    16 11715 260 223 85.76 37 14.23 Fish Fish dorsal TRUE

    17 11723 279 249 89.24 30 10.75 Fish Fish dorsal TRUE

    18 11735 277 172 62.09 105 37.90 Fish Fish dorsal TRUE

    19 11741 279 263 94.26 16 5.73 Fish Fish ventral TRUE

    20 11746 219 2 0.91 217 99.08 Macrophyte Macrophyte small millfoil TRUE

    21 20829 223 107 47.98 116 52.01 Macrophyte Fish dorsal FALSE

    continued on next page

  • 7/29/2019 JOE Manuscript v07

    23/28

    23

    continued from previous page

    Order Target ID Pings

    Classified as Reality (Ground Truth)

    ResultFish Macrophyte

    Count % Count % Decision Actual Aspect

    22 20835 268 1 0.37 267 99.62 Macrophyte Macrophyte small poto TRUE

    23 20851 262 257 98.09 5 1.90 Fish Fish dorsal TRUE

    24 20857 276 111 40.21 165 59.78 Macrophyte Macrophyte long strand TRUE

    25 20906 292 0 0 292 100 Macrophyte Macrophyte short millfoil TRUE

    26 20911 136 136 100 0 0 Fish Fish dorsal TRUE

    27 20919 121 121 100 0 0 Fish Fish ventral TRUE

    28 20933 251 243 96.81 8 3.18 Fish Fish dorsal TRUE

    29 20939 159 22 13.83 137 86.16 Macrophyte Macrophyte no entry TRUE

    30 20945 268 158 58.95 110 41.04 Fish Macrophyte no entry FALSE

    31 20949 265 177 66.79 88 33.20 Fish Fish dorsal TRUE

    32 20953 218 214 98.16 4 1.83 Fish Fish dorsal TRUE

    33 20958 147 146 99.31 1 0.68 Fish Fish dorsal TRUE

    34 21005 259 223 86.10 36 13.89 Fish Fish ventral TRUE

    35 21009 207 179 86.47 28 13.52 Fish Fish dorsal TRUE

    36 21014 198 57 28.78 141 71.21 Macrophyte Macrophyte small TRUE

    37 21020 186 4 2.15 182 97.84 Macrophyte Macrophyte small TRUE

    38 21026 187 41 21.92 146 78.07 Macrophyte Macrophyte small/longer TRUE

    39 21033 230 22 9.56 208 90.43 Macrophyte Macrophyte small millfoil TRUE

    40 21039 154 140 90.90 14 9.09 Fish Macrophyte very small poto FALSE

    41 21046 165 55 33.33 110 66.66 Macrophyte Macrophyte small TRUE

    42 21050 232 226 97.41 6 2.58 Fish Fish dorsal TRUE

    43 21055 219 90 41.09 129 58.90 Macrophyte Macrophyte no entry TRUE

    44 21059 218 195 89.44 23 10.55 Fish Fish ventral TRUE

    45 21103 210 184 87.61 26 12.38 Fish Macrophyte small poto FALSE

    46 21110 123 0 0 123 100 Macrophyte Macrophyte med. Millfoil TRUE

    47 21116 233 203 87.12 30 12.87 Fish Fish dorsal TRUE

    48 21120 242 98 40.49 144 59.50 Macrophyte Fish ventral FALSE

    49 21125 239 0 0 239 100 Macrophyte Macrophyte large poto TRUE

    50 21129 228 221 96.92 7 3.07 Fish Fish dorsal TRUE

    51 21134 206 0 0 206 100 Macrophyte Macrophyte small poto TRUE

    52 21141 223 26 11.65 197 88.34 Macrophyte Macrophyte small poto TRUE

    53 21147 143 3 2.09 140 97.90 Macrophyte Macrophyte small millfoil TRUE

    54 2115 200 169 84.50 31 15.50 Fish Macrophyte small millfoil FALSE

    55 21158 186 17 9.13 169 90.86 Macrophyte Macrophyte small millfoil & poto TRUE

    56 21202 220 8 3.63 212 96.36 Macrophyte Macrophyte bigger bundle poto TRUE

    57 21209 205 69 33.65 136 66.34 Macrophyte Macrophyte poto TRUE

    58 21214 210 157 74.76 53 25.23 Fish Fish ventral TRUE

    59 21257 235 39 16.59 196 83.40 Macrophyte Macrophyte small millfoil TRUE

    60 21302 191 178 93.19 13 6.80 Fish Fish dorsal TRUE

    61 21309 207 43 20.77 164 79.22 Macrophyte Macrophyte med poto TRUE

    62 21314 157 41 26.11 116 73.88 Macrophyte Macrophyte med poto TRUE

    63 21319 196 126 64.28 70 35.71 Fish Macrophyte smaller millfoil FALSE

    64 21324 233 30 12.87 203 87.12 Macrophyte Macrophyte long poto TRUE

    65 21328 246 200 81.30 46 18.69 Fish Fish dorsal TRUE

    66 21333 228 15 6.57 213 93.42 Macrophyte Macrophyte long poto TRUE

    67 21338 209 94 44.97 115 55.02 Macrophyte Macrophyte small milllfoil TRUE

    68 21344 239 88 36.82 151 63.17 Macrophyte Macrophyte small millfoil TRUE

    69 21351 254 208 81.88 46 18.11 Fish Fish ventral TRUE70 21355 215 206 95.8 9 4.18 Fish Fish dorsal TRUE

    71 21359 211 91 43.12 120 56.87 Macrophyte Macrophyte small millfoil TRUE

    72 21404 247 221 89.47 26 10.52 Fish Fish ventral TRUE

    73 21410 258 122 47.28 136 52.71 Macrophyte Fish dorsal FALSE

    74 21413 249 234 93.97 15 6.02 Fish Fish ventral TRUE

    75 21417 241 94 39.00 147 60.99 Macrophyte Macrophyte small millfoil TRUE

    76 21422 192 114 59.37 78 40.62 Fish Macrophyte small poto FALSE

    77 21427 130 103 79.23 27 20.76 Fish Fish ventral TRUE

    78 21431 213 118 55.39 95 44.60 Fish Fish dorsal TRUE

    79 21436 182 157 86.26 25 13.73 Fish Macrophyte small poto FALSE

    continued on next page

  • 7/29/2019 JOE Manuscript v07

    24/28

    24

    continued from previous page

    Order Target ID Pings

    Classified as Reality (Ground Truth)

    ResultFish Macrophyte

    Count % Count % Decision Actual Aspect

    80 21443 230 59 25.65 171 74.34 Macrophyte Macrophyte long strand poto TRUE

    81 21451 264 234 88.63 30 11.36 Fish Fish ventral TRUE

    82 21501 210 103 49.04 107 50.95 Macrophyte Macrophyte med. poto TRUE

    83 21511 243 209 86.00 34 13.99 Fish Fish ventral TRUE

    84 21522 170 54 31.76 116 68.23 Macrophyte Macrophyte med. poto TRUE

    85 21528 253 102 40.31 151 59.68 Macrophyte Macrophyte long strand poto TRUE

    86 21532 210 236 97.92 5 2.07 Fish Fish dorsal TRUE

    Excluding the five known blind test sets, the results showed that 70 true decisions and 11 false

    decisions, resulting in (86.42% correct decision) The system had the most difficulty distinguishing fish

    from macrophytes when smaller portions of macrophytes were used. The confidence levels were near 50%

    in the majority of the cases.During 3-day data collection period, we had rough weather condition, especially, on 2nd and 3rd day

    due to strong wind. One of the possible reasons of the incorrect classification could be that the relative

    target-transducer geometry change often pushed the target out of the beam angle causing weak signal

    return and intermittent detections. In such a case the returned signals usually dont represent the whole

    target.Overall performance of broadband fish identification was satisfactory in discriminating smolt-size

    Columbia river Chinook salmon against aquatic macrophytes.

    VIII. CONCLUSION

    This paper has presented a recently developed sonar system that combines a broadband single-beam and

    narrowband split-beam into a single sonar system. By combining these two sonar modalities, it provides

    a research tool that can provide target tracking and target identification a capability that is not available

    in either system independently.Furthermore, synergistic impacts from both modalities provides exceptional performance. We discussed

    details of broadband spectrum adjustment using beam compensation from split beam bearing angle and

    sensitivity of transducer. Also we used instantaneous time frequency representation to create variouspseudo-narrowband signals for multi-frequency data streams.

    Field evaluations have been conducted in Alaskan lake and rivers to verify that each portion of the

    system narrowband and broadband provided the ability to track and identify targets.Efforts are currently underway to validate and automate the beam-compensation capability provided by

    the split-beam processing and improve the real-time performance of the multi-frequency portion of the

    sonar.

    APPENDIX A

    TECHNICAL SPECIFICATION OF COMBINED SPLIT-B EA M / BROADBAND SONAR SYSTEM

    Technical specification of high frequency version (135 kHz to 195 kHz) of the combined split-beam

    narrowband / single-beam broadband sonar system is described below. the low-frequency version sharesmost of the technical specification except the frequency range (60 kHz to 120 kHz). Fig. 19 provides a

    picture of the sonar system.

    ACKNOWLEDGMENT

    The authors would like to thank Alaska Department of Fish and Games (ADF&G) and Alaska Sus-

    tainable Salmon Fund (AKSSF) for their supports and thoughtful reviews during several rounds of

    development. They would also like to thank Donald Degan and Anna-Maria Mueller from Aquacoustics,

    Inc. and Timothy Mulligan for their valuable comments and reviews. Further, they thank the reviewers

    of this paper for very insightful comments and suggestions, which they believe have led to a remarkable

    improvement in quality and presentation of this paper.

  • 7/29/2019 JOE Manuscript v07

    25/28

    25

    TABLE III

    TECHNICAL SPECIFICATION OF HIG H FREQUENCY VERSION OF THE COMBINED SPLIT-B EA M/B ROADBAND SYSTEM

    Transmit Specification

    Beam Shape/Width 6 degrees conical at 165kHz

    Frequency, Broadband 135kHz to 195 kHz

    Frequency, Split-Beam 16510 kHz

    Frequency, Multi-Freq. Any non-overlapping 5 kHz bands between 135 kHz and 195 kHz

    Waveform, Split-Beam & Multi-Freq. Continuous Wave (CW)Source Level 214 dB at 165 kHz

    Detection Range (min/max) 12 to 200 m at 165 kHz

    Range Resolution, Broadband 1.2 Cm

    Power Settings 16, from 0 (listen only) to 15 (max)

    Pulse Length, Broadband & Multi-Freq. 15.40 sec (compressed)

    Pulse Length, Split-Beam 200 sec (adjustable)

    Ping Rate (min/max) 1 ping / 2 secs to 30 pings / sec

    Receive Specification

    Bandwidth, Broadband 60 k Hz

    Bandwidth, Split-Beam 10 k Hz

    Gain 16 settings (programmable between 20 d B and 60 d B

    A/D Conversion 500 Ksamples/sec per channel simultaneously

    Split-Beam Angular Resolution 0.1 degrees

    Processor SpecificationSingle Board Computer 1.1G Hz Pentium M

    A/D & D/A Converter 4 channels, 500 Ksamples/sec

    Network TCP/IP Ethernet (10/100/1000 M bps)

    Software Specification

    Data Storage 4 channels, raw data

    Export Formats Echoview (multi-freq, split-beam)

    Displays Sonar Grams, Split-Beam

    Dimensions Specification

    Sonar 22.8 Cm Diam, 53.8 Cm Length

    Cable 30 m Length, OD 1.45 Cm

    Fig. 19. Combined split-beam narrowband / single-beam broadband sonar system with terminal box and underwater cable.

  • 7/29/2019 JOE Manuscript v07

    26/28

    26

    REFERENCES

    [1] W. L. Au and K. Benoit-Bird, Acoustic backscattering by Hawaiian Lutjanid snappers II: Broadband temporal and spectral structure,

    J. Acoust. Soc. Am. 114, 2767-2774, 2003.

    [2] R. E. Bailey, J. R. Irvine, F. C. Dalziel, and T. C. Nelson, Evaluations of Visible Implant Fluorescent Tags for Marking Coho Salmon

    Smolts, North American Journal of Fisheries Management, vol. 18, pp. 191196, 1998.

    [3] Y. Bar-Shalom, and X. Li, Estimation and Tracking: Principles, Techniques, and Software, Artech House, Boston, MA, 1993.

    [4] S. Blackman, and R. Popoli, Design and Analysis of Modern Tracking Systems, Artech House, Boston, MA, 1999.

    [5] E. Brookner, Tracking and Kalman Filtering Made Easy, Wiley Interscience, New York, 1998.

    [6] B. Brumley, R. Cabrera, K. Deines, and E. Terray, Performance of a broad-band acoustic Doppler current profiler, IEEE Journal ofOceanic Engineering, vol. 16, pp. 402-407, 1991.

    [7] W. S. Burdic, Underwater Acoustic System Analysis, Springer-Verlag, New York, 1991.

    [8] D. L. Burwen, D. E. Bosch, and S. J. Fleischman., Evaluation of hydroacoustic assessment techniques for chinook salmon on the Kenai

    River, Alaska Department of Fish and Game Fishery Data Series, no. 98-3, Anchorage, AK, 1998.

    [9] L. Cohen, Time-Frequency Distribution - A Review, Proc. IEEE, vol. 77, pp. 941-981, July 1989.

    [10] H. I. Choi and W. J. Williams, Improved Time-Frequency Representation of Multicomponent Signals Using Exponential Kernels,

    IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-37, no. 6, pp. 862-871, 1989.

    [11] D. Chu, T. Stantan, and P. H. Wiebe, Frequency dependence of sound backscattering from live individual zooplankton, ICES J. Mar.

    Sci., vol. 49, pp. 97-106, 1992.

    [12] D. Chu and T. Stantan, Application of pulse compression techniques to broadband acoustic scattering by live individual zooplankton,

    J. Acoust. Soc. Am., vol. 104, pp. 39-55, 1998.

    [13] C. Clay and H. Medwin, Acoustical Oceanography: Principles and Applications, Wiley Interscience, New York, 1977.

    [14] P. Dahl and O. Mathison, Measurement of fish target strength and associated directivity at high frequencies, Journal of the Acoustical

    Society of America, vol. 73, pp. 1205-1211, 1983.[15] P. Dahl and O. Mathison, Some experiments and considerations for development of Doppler-based riverine sonars, IEEE Journal of

    Oceanic Engineering, vol. OE-9, No. 3, pp. 214-217, 1984.

    [16] D. Demer, M. Soule, and R. Hewitt, A multiple-frequency method for potentially improving the accuracy and precision of in situ

    target strength measurements, J. Acoust. Soc. Am., vol 105, pp. 2359-2376, 1999.

    [17] O. Diachok, and C. Ferla, Measurement and simulation of the effects of absorptivity due to fish on transmission loss in shallow water,

    Proceedings of the Meeting of the Ocean Engineering Society , pp. 524-529, IEEE Press, Bellingham, WA, 1996.

    [18] J. Ehrenberg, A comparative analysis of in situ methods for directly measuring the acoustic target strength of individual fish, IEEE

    J. of Ocean Eng., vol. OE-4, pp. 141-152, 1979.

    [19] J. Ehrenberg, and T. Torkelson, Application of dual-beam and split-beam target tracking in fisheries acoustics, ICES Journal of Marine

    Science, vol. 53, pp. 329-334, 1996.

    [20] J. Ehrenberg, A comparative analysis of in situ methods for directly measuring the acoustic target strength of individual fish, IEEE

    Journal of Oceanic Engineering, vol. OE-4, pp. 141-152, 1997.

    [21] K.G. Foote, Summary of methods for determining fish target strength at ultrasonic frequencies, ICES Journal of Marine Science, vol.

    48, pp. 211-217, 1991.

    [22] J. Haralabous and S. Georgakarakos, Artificial neural networks as a tool for fish species identification, ICES Journal of MarineScience, vol. 53, pp. 173-180, 1996.

    [23] D. Holliday, R. Pieper, and G. Kleppel, Determination of zooplankton size and distribution with multifrequency acoustic technology,

    J. Const. Int. Explor. Mer., vol. 46, pp. 52-61, 1989.

    [24] D. Holliday and R. Pieper, Bioacoustical oceanography at high frequencies, ICES J. Mar. Sci., vol. 52, pp. 279-296, 1995.

    [25] J.K. Horne and C.S. Clay, Sonar systems and aquatic organisms: matching equipment and model parameters, Can. J. Fish. Aquat.

    Sci., vol. 55, pp. 1296-1306, 1998.

    [26] J.K. Horne and J.M. Jech, Models, measures, and visualizations of fish backscatter. In H. Medwin [ed.]. Sounds in the Seas: From

    Ocean Acoustics to Acoustical Oceanography, pp. 374-397. Academic, New York, 2005.

    [27] B. H. Houston, J. A. Bucaro, T. Yoder, L. Kraus, J. Tressler, J. Fernandez, T. Montgomery, and T. Howarth, Broadband Low Frequency

    Sonar for Non-Imaging Based Identification, OCEANS 2002.

    [28] D. L. Jones and R. G. Baraniuk, An Adaptive Optimal-Kernel Time-Frequency Representation, IEEE Transactions on Signal

    Processing, vol. 43, no. 10, pp. 2361-2371, Oct 1995.

    [29] E. Josse, P. Bach and L. Dagorn, Simultaneous observations of tuna movements and their prey by sonic tracking and acoustic surveys,

    Hydrobiologia, vol. 371/372, pp. 6169, 1998.

    [30] J.-B. Jung, Neural Network Ensonification: Training and Application, Ph.D, Dissertation, University of Washington, Seattle, WA.,

    2001.

    [31] J.-B. Jung, J. H. Jacobs, G. A. Dowding, and P. K. Simpson, Initial species discrimination experiments with riverine salmonids,

    Proceedings of the International Joint Conference on Neural Networks , 20-24 July 2003, vol. 2, pp. 1295-1300.

    [32] J.-B. Jung, J. H. Jacobs, G. F. Denny, and P. K. Simpson, Broadband sonar target classification: pool experiments, Proceedings of

    the International Joint Conference on Neural Networks , July 20-24, 2003, vol. 2, pp. 1307-1312.

    [33] J.-B. Jung, G. F. Denny, and P. K. Simpson, A Preliminary Study of Swimmer Detection and Identification, NSF Design, Service

    and Manufacturing Grantees and Research Conference, Dallas, TX, January 5-8, 2004.

    [34] J.-B. Jung, R. Johnson and P.K. Simpson, Demonstration of Broadband Sonar Technology for Columbia River Chinook Salmon Smolt

    Identification, Pacific Northwest National Lab, PNNL-14883, September.

    [35] J.-B. Jung, G. F. Denny, J. W. Tilley, A. B. Kulinchenko, and P. K. Simpson, Broadband Active Sonar Swimmer Detection and

    Identification, Proceedings of the International Joint Conference on Neural Networks, Vancouver, B.C., Canada, July 16-21, 2006, pp.

    2600-2605.

  • 7/29/2019 JOE Manuscript v07

    27/28

    27

    [36] J.-B. Jung, A. B. Kulinchenko and P. K. Simpson, Compensated broadband spectral measurements, in progress.

    [37] R. Love, Maximum side-aspect target strength of an individual fish, Journal of the Acoustical Society of America, vol. 46, pp.

    746-752, 1969.

    [38] R. Love, Dorsal-aspect target strength of an individual fish, J. Acoust. Soc. Am., vol. 49, pp. 816-823, 1971.

    [39] R. Love, Resonant acoustic scattering by swimbladder-bearing fish, J. Acoust. Soc. Am., vol. 64, pp. 571-580, 1978.

    [40] R. Love, A comparison of volume scattering strength data with model calculations based on quasisynoptically collected fishery data,

    J. Acoust. Soc. Am., vol.4, pp. 2255-2268, 1993.

    [41] A. Lovik and J. Hovem, An experimental investigation of swimbladder resonance in fishes, J. Acoust. Soc. Am., vol. 66, pp. 850-854,

    1979.

    [42] D. MacLennan, Acoustical measurement of fish abundance, Journal of Acoustic Society of America, vol. 87, pp. 1-15, 1990.

    [43] D. MacLennan, and E. Simmonds, Fisheries Acoustics, Chapman & Hall, London, 1992.

    [44] R. J. Marks II, Fourier Analysis and Its Applications, Oxford University Press, USA, October, 2008.

    [45] L. Martin and et. al., Acoustic classification of zooplankton, ICES Journal of Marine Science, vol. 53, pp. 217-224, 1996.

    [46] S. L. Maxwell and N. E. Gove, The Feasibility of estimating migrating salmon passage rates in turbid rivers using a dual frequency

    identification sonar (DIDSON), Alaska Department of Fish and Game Regional Information Report1 No. 2A04-05, March 2004.

    [47] M. Mazenksi and D. Alexandrou, Active wide band detection and localization on an uncertain multipath environment, J. Acoust. Soc.

    Am., vol. 101, pp. 1961-1970, 1997.

    [48] L. Miranda, R. Brock, and B. Dorr, Uncertainty of Exploitation Estimates Made from Tag Returns, North American Journal of

    Fisheries Management, vol. 22, pp. 13581363, 2002.

    [49] T. J. Mulligan and R. Keiser, A split-beam echo-counting model for riverine use, ICES Journal of Marine Science, 531:403-406,

    1996.

    [50] B. Pederson and M. Trevorrow, Continuous monitoring of fish in a shallow water channel using a fixed horizontal sonar, JASA, vol.

    105, no. 2, pp. 3126, 1999.

    [51] J.-B. Jung, M. A. Weiland, G. R. Ploskey, R. P. Mueller, and R. L. Johnson , Demonstration of Broadband Sonar Technology forColumbia River Chinook Salmon Smolt Identification, Pacific Northwest National Lab, PNNL-14883. September 2004.

    [52] N. Ramani, and P. Patrick,Fish detection and classification using neural networks - some laboratory results, IEEE J. Oceanic

    Engineering, vol. 17, pp. 364-368, 1992.

    [53] L. Richards, et. al., Classification of fish assemblages based on echo integration surveys, Can. J. Fish. Aquat. Sci., vol. 48, pp.

    1264-1272, 1991.

    [54] G. Rose, and W. Leggett, Hydroacoustic signal classification of fish schools by species, Can. J. Fish. Aquatic. Sci., vol. 45, pp.

    597-604, 1988.

    [55] C. Scalabrin and et al., Narrowband acoustic identification of monospecific fish shoals, ICES Journal of Marine Science, vol. 53, pp.

    181-188, 1996.

    [56] E. Simmonds, F. Armstrong, and P. Copland, Species identification using wide-band backscatter with neural network and dicriminant

    analysis, ICES Journal of Marine Science, vol. 53, pp. 189-195, 1999.

    [57] P. K. Simpson, Artificial Neural Systems: Foundations, Paradigms, Applications, and Implementations, Pergamon Press, New York,

    1990.

    [58] P. K. Simpson, Fuzzy min-max neural networks - part 1: Classification, IEEE Trans. on Neural Networks, vol. 3, no. 5, pp. 776-786,

    1992.

    [59] P. K. Simpson, Fuzzy min-max neural networks - part 2: Clustering, IEEE Trans. on Fuzzy Systems, vol. 1, no. 1, pp. 32-45, 1993.

    [60] P. K. Simpson, Technology Update Series: Neural Networks Theory, Technology, and Applications, IEEE Press, Piscataway, NJ,

    2005.

    [61] M. Soule, I. Hampton, and M. Barange, Potential improvements to current methods of recognizing single targets with a split-beam

    echo-sounder, ICES Journal of Marine Science, vol. 53, pp. 237-243, 1996.

    [62] T. K. Stanton,Sound scattering by cylinders of finite length. I. Fluid cylinders, J. Acoust. Soc. Am., vol. 83, 55-63, 1988.

    [63] T. K. Stanton, D. Chu, and P. H. Wiebe, Acoustic scattering characteristics of several zooplankton group, ICES J. Mar. Sci., vol. 53,

    pp. 289-295, 1996.

    [64] T. K. Stanton, D. Chu, and P. H. Wiebe, Sound scattering by several zooplankton groups. II. Scattering models, J. Acoust. Soc. Am.,

    vol. 103, pp. 236-253, 1998.

    [65] T. K. Stanton, D. Chu, P. H. Wiebe, L. V. Martin, and R. L. Eastwood, Sound scattering by several zooplankton groups. I. Experimental

    determination of dominant scattering mechanisms, J. Acoust. Soc. Am., vol. 103, pp. 225-235, 1998.

    [66] T. K. Stanton, D. Chu, P. H. Wiebe, R. L. Eastwood, and J. D. Warren, Acoustic scattering by benthic and pelagic shelled animals,

    J. Acoust. Soc. Am., Vvol. 108, pp. 535-550, 2000.[67] C. Thompson and R. Love, Determination of fish size distributions and areal densities using broadband low-frequency measurements,

    ICES Journal of Marine Science, vol. 53, pp. 197-201, 1996.

    [68] L. Traykovski, R. Driscoll, and D. McGehee, Effect of orientation on broadband acoustic scattering of antarctic krill Euphasia superba:

    Implications for inverting zooplankton spectral acoustic signatures for angle of orientation, J. Acoust. Soc. Am., vol. 104, pp. 2121-2135,

    1998.

    [69] R. J. Urick, Principles of Underwater Sound, Peninsula Pub, 1983.

    [70] Y. Xie, G. Cronkite and T. J. Mulligan, A split-beam echosounder perspective on migratory salmon in the Fraser River: A progress

    report on the split-beam experiment at Mission, B.C., in 1995. Pacific Salmon Commission Techical Report, no. 8, 1997.

    [71] Y. Xie, A range-dependent echo-association algorithm and its application in split-beam sonar tracking of migratory salmon in the

    Fraser river watershed, IEEE Journal of Oceanic Engineering, vol. 25, pp. 387-398, 2000.

    [72] Y. Zhao, L. E. Atlas, R. J. Marks II, The Use of Cone-Shaped Kernels for Generalized Time-Frequency Representations of Nonstationary

    Signals, IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 38, no. 7, pp. 1084-1091, July 1990.

  • 7/29/2019 JOE Manuscript v07

    28/28

    28

    [73] L. Zedel, R. Hay, R. Cabrera, and A. Lohrmann, Performance of a single beam, pulse-to-pulse coherent Doppler profiler, IEEE

    Journal of Oceanic Engineering, vol. 21, pp. 290-297, 1996.

    Jae-Byung Jung (S01-M02) received the B.S. and M.S. degrees in electronics engineering from Hanyang University,

    Korea, in 1993 and 1995 respectively. He was an Assistant Research Engineer at LG Industrial Systems, Co., Ltd.,

    Korea, from 1995 to 1996. He earned the Ph.D. in electrical engineering from the University of Washington, Seattle,WA, in 2001. From 2001 to 2002, he was a Post-Doctoral Research Associate at the Department of Electrical

    Engineering, the University of Washington. In 2002, he joined Scientific Fishery Systems, Inc., where he is currently

    a Senior Engineer responsible for research and development of underwater acoustic instruments.

    His areas of interest include neural networks, fuzzy systems, evolutionary algorithms, signal processing, computer

    vision, robotics, and software engineering.

    Alexander B. Kulinchenko was born in Lviv, Ukraine in 1971. In 1993, Mr. Kulinchenko received Electro-Mechanical

    Engineering Degree in State University Lvivska Polytechnika, Lviv, Ukraine. Continuing education Mr. Kulinchenko

    received Master of Science Degree in Electro-Mechanical Engineering in State University Lvivska Polytechnika,

    Lviv, Ukraine, in 1994. In 1995 Mr. Kulinchenko moved to Detroit, MI where he worked as draft assistant at Lear

    Seating Corporation, Plymouth, MI. In 1997, Mr. Kulinchenko moved to Anchorage, AK. In December 2001, Mr.Kulinchenko graduated from University of Alaska in Anchorage with Bachelor of Science Degree in Computer Science.

    Since August 2001, Mr. Kulinchenko has been working as software engineer at Scientific Fishery Systems, Inc. Mr.

    Kulinchenko has experience in software applications development, modifications and testing. Good knowledge of Java,

    Enterprise Java Beans (EJB), Extensible Markup Language (XML), C++, Intel Image Processing Library (IPL), Open

    Source Computer Vision Library (Open CV), Visual Basic (VB), MatLab and Neural Networks. Knowledge of Microsoft Visual Developer

    Studio (C++, VisualBasic, SourceSafe), Borland JBuilder (Java).

    Mr. Kulinchenko has also experience in electro-mechanical engineering including knowledge of AutoCAD.

    Patrick K. Simpson received his Bachelor of Arts degree in Computer Science in 1986 from the University of

    California at San Diego. Soon after graduating, Mr. Simpson has distinguished himself in the application of neural

    networks, fuzzy systems, and artificial intelligence to difficult defense related problems in areas such as electronic

    intelligence, radar surveillance, sonar signal identification, and various aspects of automated diagnostics.

    Early in his career, Mr. Simpson wrote several archival papers, taught several short courses, wrote a text book

    that has been used in college courses around the United States, and lectured on the theory and application of neural

    networks and fuzzy systems world wide. Mr. Simpson served as President of the IEEE Neural Networks Council in

    1994 and was the General Chairman of the 1998 IEEE World Congress on Computational Intelligence.

    In 1992, Mr. Simpson founded Scientific Fishery Systems, Inc. (SciFish) to develop technologies for more efficient

    fisheries. Through initial funding from the Small Business Innovative Research (SBIR) program, and subsequent funding from commercial

    sales, government sales, and continued research and development, SciFish has developed several technologies that have been applied to

    various aspects of the fishing industry.

    James W. Tilley received the B.S. degree in electrical engineering from the University of Nevada, Las Vegas in 1998.

    Upon graduation, Mr. Tilley moved to the Seattle area and began work in the field of electromagnetic compatibilitytesting. In January of 2005, Mr. Tilley began working for Scientific Fishery Systems on their split-beam sonar system.

    Later that year, he transferred to sister company Alaska Native Technologies, LLC. where he is currently working on

    unmanned underwater systems.