POULIQUEN.PDF

THE EFFECT OF SEABED BACKSCATTERING VARIABILITY
ON THE PROBABILITY OF DETECTION AND ON THE
PERFORMANCE OF SEABED CLASSIFICATION ALGORITHMS
E. POULIQUEN AND L. PAUTET
SACLANT Undersea Research Centre, Viale San Bartolomeo 400, 19138 La Spezia, Italy
E-mail: [email protected]
A.P. LYONS
The Pennsylvania State University, Applied Research Lab., P.O. Box 30, State College, PA 16804
E-mail: [email protected]
As the scattering properties of the seafloor affect not only the mean angular dependence
of backscatter but also cause a wide spread of scattered amplitudes, higher moment
statistics are essential for applications such as target detection and seabed classification
and characterization. The angular and frequency responses of the scattered amplitude
distribution are caused primarily by water-sediment interface roughness, upper sediment heterogeneity, patchiness and discrete scatterers. To quantify the impact of seabed
properties on higher moment statistics, scattered amplitudes were acquired at various
sites displaying a large spectrum of seabed properties. Concurrent measurements of the
seabed properties were conducted using two-dimensional stereo-photogrammetry, core
and grab samples, videos and seabed penetrometers. To complement this experimental
work, a temporal snapshot model based on the fourth order small slope approximation
for interface scattering and on the small perturbation theory for the volume has been
developed. Valid at high frequency and at all grazing angles, it allows the multiplication
of scattering scenarios and will help to understand and quantify the impact of complicated seabeds on detection performance and on seabed classification algorithms. To
illustrate high frequency acoustic variability, this paper presents selected experimental
and simulated results.
1 Introduction
At high frequency, seabed backscattering is a major component of sonar reverberation.
It is usually quantified in terms of averaged backscattering strength (BS) which relates
directly to an averaged intensity scattered from a unit surface at a unit distance. For very
large acoustic footprints compared to the wavelength, the BS is a meaningful and sufficient indicator for many detection and classification purposes. At high frequency where
the acoustic footprint is often reduced and approaches the dimension of the acoustic
wavelength, the problem becomes more complicated as scattering fluctuates greatly from
place to place even in apparently large-scale stationary environments. Higher moment
statistics and probability distribution functions (PDFs) provide enhanced information for
detection and classification applications. The dominant mechanisms affecting the PDFs
219
N.G. Pace and F.B. Jensen (eds.), Impact of Littoral Environmental Variability on Acoustic Predictions and
Sonar Performance, 219-226.
© 2002 Kluwer Academic Publishers. Printed in the Netherlands.
220
E. POULIQUEN ET AL.
are of three kinds: 1) scattering caused by the intrinsic nature of the seabed (i.e., interface roughness, volume heterogenities, patchiness, discrete scatterers), 2) the sensing
geometry (i.e., the beam pattern, the size of the scattering cell, the angle of incidence)
and 3) the pulse shape, spectrum and duration. In some cases, other phenomena such
as sea surface multipath and water column fluctuation may also significantly affect the
PDF. To understand and quantify the impact of each of these phenomena, in situ measurements are necessary. Modelling of higher order statistics of high frequency backscatter
is also required to allow a multiplication of scattering scenarios. Recent modeling work
by Abraham and Lyons [1, 2] and Pouliquen et al. [3–5] offers two complementary approaches to predict higher moment statistics, treating patchiness with Joint Characteristic
Functions (JCF) and using snapshot realizations, respectively. Processed signals acquired
at sea and snapshot model outputs will be presented in this paper. The interpretation
of the signals acquired is facilitated by in situ ground truth in terms of roughness (from
stereo-photographs) and histograms of grain size (from gravity cores and grab samples),
mainly. This paper discusses the effect of seabed scattering variability on the performance
of classification algorithms based on normal incidence monobeam echosounding. Examples of measured and simulated ampitude PDFs and Probability of False Alarm (PFA) at
off-normal incidence are also presented. It illustrates the highly variable nature of seabed
acoustic backscatter at high frequency and outlines the dominant role imposed by seabed
backscatter in target detection algorithm performance.
2
At sea measurement of seabed backscattering statistics
A prototype, hereafter refered to as ESP (Environmental Sensor Package) was designed
and assembled by SACLANTCEN to acquire temporal backscattered signals from the
seabed at various incident angles and frequencies. It was used at a number of sites near
Control/Acquisition
Cable
ESP
Pan/Tilt
Transducers
Seabed
Figure 1. Experimental configuration of signal acquisition using ESP in a driting mode. An additional pinger provides the vertical distance between the source and the seabed.
HIGH FREQUENCY SEABED BACKSCATTERING VARIABILITY
221
Halifax, Canada, during the MAPLE’2001 experiment [6]. ESP is a high frequency/multiincidence/multi-look system. Power, control commands and data collection are transmitted
through a 200 m cable (Fig. 1). The acquired signals allow computation of higher moment
statistics seafloor backscatter (i.e., mean levels of reverberation, variance, probability
distribution, probability of false alarm) and also provide seabed type information when
sounding at normal incidence. The system was designed to be operated in a drifting
mode (geographicaly referenced), in a stationary position, or mounted on a tower. The
latter allows the study of the effect of water column fluctuation on seabed backscattering
variability (assuming ping-to-ping stationary of the seabed response) [7]. The drifting
allows the study of the evolution of seabed response statistics as the environment changes.
The stationary mode allows the quantification of acoustic variability for given seabeds
assuming local stationarity and ergodicity. ESP was used at different depths (as low as 2
m above the seabed) and operated 3 different acoustic sources resonating at three different
frequencies (50, 100 and 140 kHz). Thirteen positions were established on a large variety
of seabeds with various levels of clutter often changing within less than a few metres.
Several acquisitions were also made as ESP was drifting. High frequency multibeam
systems, sidescan sonars, monobeam echosounders and advanced ground truth systems
were also deployed at the sites. Dense gridding of sediment sampling (cores and grabs),
expendable bottom penetrometers, videos and stereo-photos were used. Grabs allowed
quantification of the spatial variability of the sediment grain size distribution, which
provides information on acoustic impedance and heterogeneity affecting volume scattering.
Gravity cores provided vertical information on density and compressional velocity with
centimetric resolution at some locations. The core analysis (sound speed, density profile)
was done immediately using the multi-sensor core logger mounted vertically in order to
preserve the integrity of the upper structure of the core. Stereo-photographs [8] provided
two-dimensional digital information on the water-sediment interface roughness which has
an impact on interface scattering. One hundred pairs of stereo-photos (e.g., Fig. 2), 40
grabs, and 7 cores were taken during the sea trial.
3
Examples of seabed acoustic backscattering variability
3.1 Variability at Normal Incidence
Signals from echosounders provide point-to-point bathymetry and are also used to remotely classify the seabed. These algorithms base their analysis on the shape and energy
of narrow-band returned signals to provide segmentation and classification information
of surveyed areas. The shape and energy of the returned echoes strongly depends on the
interference structures produced by a heterogenous seabed volume bounded by a rough
interface. The sounding geometry particularly around the specular direction may change
from one ping to another. The source coordinates (x, y, z), pitch θ and roll φ work towards producing a snapshot of the temporal response from the seabed for a high frequency
transmit pulse of duration of the order of milliseconds. This ping-to-ping variability affects the performance of classification algorithms and may provide unreliable/ambiguous
classification. To quantify this variability, two typical classifying parameters were computed from signals acquired at-sea and simulated signals. As an example, 20000 time
series with various acquisition geometries were recorded on a rippled sandy seabed shown
in Fig. 2. The acoustic source movements were recorded. For normal incidence sound-
222
E. POULIQUEN ET AL.
Crest
Trough
Crest
600
mm
900
mm
50 m
Figure 2. Example of a sidescan sonar image featuring a strong ripple field. The red dots correspond
to the ESP acoustic acquisitions at a station. The yellow dots mark the locations where stereophotographs were taken. One stereo-photograph shows the presence of a ripple field with coarser
grains in the troughs and finer grains on the crests.
ings, the vertical distance between the seabed and the source showed a RMS variation of
20 cm around a mean value of 10.71 m. As the ship was anchored, the maximum horizontal translation during each acquisition series was less than 5 m (series of 1200 pings
in average were recorded at a ping rate of 4.1 pings/s). Pitch and roll RMS variations
were 0.26◦ and 1.50◦ around mean values of 0.69◦ and 0.19◦ , respectively. From each
received amplitude envelope, two typical classifying parameters ai and bi are computed
as follows:
◦
! t(θ1 =40◦ )
!
1 t(θ3 =25 )
si (t)dt
bi =
si (t)dt,
(1)
ai = C ∗
ai t(θ2 =18◦ )
t(θ0 =0◦ )
with C being a normalizing factor depending on the acoustic source and i being the ping
number. The amplitude envelope si (t) is power corrected (i.e., transmission loss and
footprint size are removed) and time normalized so that the echo streching is independent
of the source-seabed distance. The ai parameter is mostly sensitive to the water/sediment
impedance contrast (i.e., “hardness”) whereas the bi parameter depends on both volume
heterogeneity and interface roughness. Given the variability of the signals, the pings are
HIGH FREQUENCY SEABED BACKSCATTERING VARIABILITY
223
often aligned in time and averaged to provide a more stable result:
1
Ai =
N
i+N/2
"
j=i−N/2+1
aj
1
Bi =
N
i+N/2
"
bj ,
(2)
j=i−N/2+1
b
with N being the number of pings being averaged. The value of N is usually between 5 and 10. Figure 3 shows a high spread of a, b, A and B computed from the
measured signals. Given the consistent texture of the sidescan sonar image (Fig. 2), stationarity of the seabed properties for each acquisition series is a reasonable assumption.
Nevertheless, the non-averaged parameters a and b vary significantly from one ping to
another. Even the averaged parameters A and B (with N = 10) display a large spread
wich corresponds for this particular sediment type to uncertainty in roughness of about
±1.5 cm RMS and to relative acoustic impedance variation of about ±15%. Similar
a
b
Figure 3. Example of classification parameter variability computed from signals acquired by ESP
at Rose Bay (Fig. 2). The carrier frequency is 50 kHz, the beam pattern is 26◦ , the pulse length is
0.5 ms.
a
Figure 4. Example of classification parameter variability using BORIS-SSA. The acquisition geometry (source/sediment distance, roll and pitch) and the transmit pulse are identical to the ones of
Fig. 3 but the geo-acoustic properties are different.
224
E. POULIQUEN ET AL.
observations can be made (Fig. 4) from simulated data using the snapshot model BORISSSA [3–5]. The same geometry and source motion was used in the simulation. Volume
scattering was not included because the heterogeneity structure of the sediment was not
measured at this site. Horizontal isotropy of the water/sediment interface was assumed
and the RMS roughness was estimated from non-processed stereo-photos (σ " 3 cm).
The spread of the non-averaged a and b parameters is greater than the measured ones.
The averaging process to obtain parameters A and B is more effective as the vertical
distance source-sediment is known exactly but also displays a high inherent variability.
Variability observed on measured and simulated echoes illustrates the limits of classification algorithms based on monofrequency echosounder signals. Recent experiments have
shown that multi-frequency sensing of the seabed is a way to improve the classification
algorithm performance. As volume and interface backscatter is frequency dependent a
multi-frequency approach provides orthogonal information that can significantly reduce
ambiguity in the classification process [6].
3.2 Variability at Oblique Incidence
High frequency and low grazing angle backscattering from cluttered seabeds is an important factor limiting target detection. Compared to near normal incidence, the often abrupt
changes of the reflection and transmission coefficients around critical angle and shadowing (e.g., in Fig. 2) are additional physical effects producing complicated backscattered
signals. With the intent of understanding the formation of probability density functions,
off-normal incidence signals were also acquired by ESP at several sites displaying a large
spectrum of seabed types with various levels of complexity. Acquisitions for different
geometries and transmitted signals were made. Figure 5 displays an example of a sequence of acoustic envelopes backscattered from the same seabed (a mixture of silt and
gravel). Despite apparent seabed stationarity, the signals are highly variable. In Fig. 6,
their PDF and related PFA show a large spread of backscattering strength. A heavy tail of
the PDF or the slow decay of the related probability of false alarm (PFA) versus detection
threshold reveal a particular difficulty in detecting targets. The presence “patchiness” of
shells and gravel may be the cause. Similarly, Fig. 7 shows a PDF and a PFA computed
from synthetic signals obtained using BORIS-SSA [5]. The simulations were made on
a on a large number of snapshots on the same interface having well defined statistical
properties. The physical interface and volume properties, geometry of acquisition and the
source characteristics can be chosen and allow a multiplication of scenarios. Measurements and simulations illustrate the high variability of the seabed response and the critical
need to consider not only the mean backscatter as an indication of detection probabilities
but also higher moments as well as the probability density function. The comparison
between acquired and simulated statistics is promising and will be the object of further
studies.
4
Summary
This paper presented a selection of measured and simulated signals with their related
higher statistical characteristics represented in terms of PDF and PFA. At normal incidence, acoustic signal variability causes a large spread of classification parameters which
reduces the classification algorithm performance. At oblique incidence, signal variability
is such that higher moment statistics are essential for target detection.
HIGH FREQUENCY SEABED BACKSCATTERING VARIABILITY
225
Figure 5. 36 successive amplitude envelopes acquired by ESP at 100 kHz over a silt+gravel area.
Incident angle: θ = 60◦ . Beam aperture is 16◦ . Ping rate is 4.1 pings/s.
Figure 6. PDF and PFA obtained from a series of 1200 pings acquired by ESP at 100 kHz over a
silt+gravel area. Incident angle: θ = 70◦ . Beam aperture is 16◦ . Ping rate is 4.1 pings/s.
Acknowledgements
We would like to thank everyone who took part to the MAPLE’2001 sea trial. Prototypes were designed and assembled by the SACLANTCEN Engineering & Technology
226
E. POULIQUEN ET AL.
Figure 7. PDF and PFA obtained from time series simulated by BORIS-SSA. Only seabed interface
scattering is considered. f = 100 kHz. The seabed is an isotropic filtered “power-law” power
spectral density with an 3.2 exponent. Incident angle: θ = 50◦ . Beam aperture is 16◦ .
Department. Special thanks to F. Cernich, A. Figoli, P. Franchi, P. Guerrini, R. Lombardi
and P.A. Sletner. The experiment was conducted from the R/V Alliance and the CFAV
Quest.
References
1. Abraham, D.A., Lyons, A.P., Novel physical interpretation of K-distributed reverberation,
IEEE J. Oceanic Eng. (submitted 2001).
2. Lyons, A.P., Abraham, D.A. and Pouliquen, E., Prediction of scattered amplitude statistics of
patchy seafloors. In Impact of Littoral Environmental Variability on Acoustic Predictions
and Sonar Performance, edited by N.G. Pace and F.B. Jensen (Kluwer, The Netherlands,
2002) pp. 211–218.
3. Pouliquen, E., Bergem, O. and Pace, N.G., Time-evolution modeling of seafloor scatter. I.
Concept, J. Acoust. Soc. Am. 105(6), 3136–3141 (1999).
4. Bergem, O., Pouliquen, E., Canepa, G. and Pace N.G., Time-evolution modeling of seafloor
scatter. II. Numerical and experimental evaluation, J. Acoust. Soc. Am. 105(6), 3142–
3150 (1999.)
5. Pautet, L., Pouliquen, E., Canepa, G., A study on ping-to-ping coherence of the seabed
response, In Impact of Littoral Environmental Variability on Acoustic Predictions and
Sonar Performance, edited by N.G. Pace and F.B. Jensen (Kluwer, The Netherlands,
2002) pp. 489–496.
6. Pouliquen, E., Trevorrow, M., Blondel, Ph., Canepa, G., Cernich, F. and Hollett, R., Multisensor analysis of the seabed in shallow water areas: overview of the Maple’2001 experiment. In Proc. 6th European Conf. on Underwater Acoustics, Gdansk, Poland, 2002.
7. Pautet, L. and Pouliquen, E., Experimental study of fluctuation in coherent backscattering. In
Proc. 6th European Conf. on Underwater Acoustics, Gdansk, Poland, 2002.
8. Lyons, A.P., Fox, W.L.J., Hasiotis, T. and Pouliquen, E., Characterization of the twodimensional roughness of shallow-water sandy seafloors, IEEE J. Oceanic Eng. (July
2002).
9. Orlowski, A., Application of multiple echoe measurements for evaluation of sea bottom type,
Oceanologia 19, 61–78 (1984).