Continental Shelf Research 25 (2005) 2273–2293 Acoustic seafloor discrimination with echo shape parameters: A comparison with the ground truth Paul A. van Walree a, , Jaroslaw T ˛ egowski b , Cees Laban c , Dick G. Simons a,d a TNO Defence, Security and Safety, Oude Waalsdorperweg 63, P.O. Box 96864, 2509 JG The Hague, The Netherlands b Institute of Oceanology, Polish Academy of Sciences, Powstancow Warszawy 55, 81-712 Sopot, Poland c TNO Netherlands Institute of Applied Geoscience, Princetonlaan 6, 3584 CB Utrecht, The Netherlands d Department of Earth Observation and Space Systems, Delft University of Technology, Kluyverweg 1, 2629 HS Delft, The Netherlands Received 19 July 2004; received in revised form 5 September 2005; accepted 9 September 2005 Abstract Features extracted from echosounder bottom returns are compared with the ground truth in a North Sea survey area. The ground truth consists of 50 grab samples for which the grain size distribution, and the gravel and shell contents were determined. Echo envelopes are analysed for two single-beam echosounders, operated at frequencies of 66 and 150 kHz. It is shown that a set of six energetic, statistical, spectral and fractal parameters carries useful information that can be exploited for seafloor characterization and classification purposes. A quantitative comparison of the individual features with the grab sample mean grain size reveals significant correlations. The echo features are also subjected to a principal component analysis in tandem with a cluster analysis. Four sediment classes with different geo-acoustic properties are examined and compared with the grab samples and existing geological information. A subtle difference between the two sounder frequencies is observed in the rendition of an isolated trench with a soft infill of clay and Holocene channel deposits. The acoustic transition from the valley to neighbouring sand and gravel fields occurs more rapidly at the lower of the two examined frequencies. A direct comparison of the acoustic sediment classes with the ground truth reveals that the main sediment types mud, sand, and gravel are more clearly separated at 150 kHz. The acoustic bottom classification scheme also appears particularly sensitive to the presence of gravel at this sounder frequency. r 2005 Elsevier Ltd. All rights reserved. Keywords: Echosounder; Seafloor characterization; Shape parameters; North Sea 1. Introduction Acoustic seafloor characterization has been long recognized as a useful tool for fast preliminary geological analysis. Acoustic methods for seabed identification and classification can be exploited in many fields, including marine geology, hydrogra- phy, marine engineering, environmental sciences, military sonar, and fisheries. The advantage over conventional bottom grabs is the nearly continuous versus sparse probing and a vast reduction in survey time and costs. Difficulties include echo to echo variations that may occur even for low sea states and homogeneous seabeds, and the need for identification of echo parameters that are best suited to discriminate between various seafloors. ARTICLE IN PRESS www.elsevier.com/locate/csr 0278-4343/$ - see front matter r 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.csr.2005.09.002 Corresponding author. Tel.:+31 70 374 0726; fax: +31 70 374 0654. E-mail address: [email protected] (P.A. van Walree).
Transcript
ARTICLE IN PRESS
0278-4343/$ - see
doi:10.1016/j.csr
�Correspondifax: +3170 374
E-mail addre
Continental Shelf Research 25 (2005) 2273–2293
www.elsevier.com/locate/csr
Acoustic seafloor discrimination with echo shape parameters:A comparison with the ground truth
Paul A. van Walreea,�, Jaroslaw Tegowskib, Cees Labanc, Dick G. Simonsa,d
aTNO Defence, Security and Safety, Oude Waalsdorperweg 63, P.O. Box 96864, 2509 JG The Hague, The NetherlandsbInstitute of Oceanology, Polish Academy of Sciences, Powstancow Warszawy 55, 81-712 Sopot, PolandcTNO Netherlands Institute of Applied Geoscience, Princetonlaan 6, 3584 CB Utrecht, The Netherlands
dDepartment of Earth Observation and Space Systems, Delft University of Technology, Kluyverweg 1, 2629 HS Delft, The Netherlands
Received 19 July 2004; received in revised form 5 September 2005; accepted 9 September 2005
Abstract
Features extracted from echosounder bottom returns are compared with the ground truth in a North Sea survey area.
The ground truth consists of 50 grab samples for which the grain size distribution, and the gravel and shell contents were
determined. Echo envelopes are analysed for two single-beam echosounders, operated at frequencies of 66 and 150 kHz. It
is shown that a set of six energetic, statistical, spectral and fractal parameters carries useful information that can be
exploited for seafloor characterization and classification purposes. A quantitative comparison of the individual features
with the grab sample mean grain size reveals significant correlations. The echo features are also subjected to a principal
component analysis in tandem with a cluster analysis. Four sediment classes with different geo-acoustic properties are
examined and compared with the grab samples and existing geological information. A subtle difference between the two
sounder frequencies is observed in the rendition of an isolated trench with a soft infill of clay and Holocene channel
deposits. The acoustic transition from the valley to neighbouring sand and gravel fields occurs more rapidly at the lower of
the two examined frequencies. A direct comparison of the acoustic sediment classes with the ground truth reveals that the
main sediment types mud, sand, and gravel are more clearly separated at 150 kHz. The acoustic bottom classification
scheme also appears particularly sensitive to the presence of gravel at this sounder frequency.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Echosounder; Seafloor characterization; Shape parameters; North Sea
1. Introduction
Acoustic seafloor characterization has been longrecognized as a useful tool for fast preliminarygeological analysis. Acoustic methods for seabedidentification and classification can be exploited in
front matter r 2005 Elsevier Ltd. All rights reserved
many fields, including marine geology, hydrogra-phy, marine engineering, environmental sciences,military sonar, and fisheries. The advantage overconventional bottom grabs is the nearly continuousversus sparse probing and a vast reduction in surveytime and costs. Difficulties include echo to echovariations that may occur even for low sea statesand homogeneous seabeds, and the need foridentification of echo parameters that are bestsuited to discriminate between various seafloors.
Fig. 1. Overview of the area, featuring the Cleaver Bank and the
Botney Cut. The red rectangle, north–west of the Netherlands
(NL), indicates the location of the acoustic surveys.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932274
For example, the presence of current ripples caninfluence certain echo features, but this is undesir-able when the purpose is to discriminate betweensediment compositions only.
The physical mechanisms behind the process ofecho formation are nontrivial and a theoreticaltreatment that includes specular reflection, surfacebackscattering and volume reverberation can beexceedingly complex, depending on the level ofcompleteness that is pursued. Roughly speaking,bottom classification algorithms can be divided intwo categories: model based and empirical. Model-based techniques attempt to characterize the seabedby translating echo signal properties directly intogeological quantities via a physical model. Empiri-cal approaches simply rely on the observation thatcertain features of echo signals are correlated withthe sediment type. Here, the idea behind the use of aparticular feature may very well be based ontheoretical assumptions or expectations, but thistheory is not used in the signal processing orclassification. Typically, two or more features arecombined by means of a cluster analysis, after whicheach identified cluster is associated with a particulartype of sediment.
Among the reported techniques to exploit asingle-beam echosounder for seafloor mapping wefind approaches as diverse as the determination ofthe energy ratio of the first and second bottomreturn (Chivers et al., 1990; Dyer et al., 1997; Heald,2001), a comparison between measured and theore-tically modelled echo patterns in the time domain(Pouliquen and Lurton, 1992; Sternlicht and DeMoustier, 1997) or in the frequency (wavelet)domain (Caiti and Zoppoli, 1998), and artificialneural networks (Alexandrou and Pantzartsis,1993). Hamilton et al. (1999) compare the commer-cially available bottom classification systems Rox-Ann and QTC.
Apart from simple echosounders, parametricsources (Caiti and Zoppoli, 1998; Gensane andTarayre, 1992), wideband ‘‘chirp’’ signals (leBlancet al., 1992; Maroni and Quinquis, 1997) and multi-beam echosounders (De Moustier and Matsumoto,1993) have been reported as potential bottomclassification sonars.
In this paper three families of echo parameters arecombined into an empirical seafloor characteriza-tion approach. The families are statistical moments(Van Walree et al., 2002), spectral moments(Tegowski and Łubniewski, 2002), and fractaldimensions (Tegowski, 2005). These approaches,
and the corresponding echo shape parameters, havebeen relatively underexploited. The analysis isapplied to echosounder data collected in the CleaverBank and Botney Cut areas, located in the NorthSea north–west of the Netherlands (Fig. 1). Theechosounder surveys were carried out specifically tocollect data for the evaluation of acoustic seafloordiscrimination methods. Extensive ground truth isavailable for an assessment of the algorithms.Between 1979 and 1991, a number of geologicalexcursions were undertaken on the Cleaver Bankand the Botney Cut with side-scan sonar andsubbottom profilers for the purpose of gravellocation and extraction. Subsequent sampling pro-grammes with Hamon grabs and a vibrocoreryielded detailed information on the seafloor surfacelayer. Moreover, up-to-date ground truth for thelocation pertinent to the present paper was estab-lished by 50 bottom grabs collected during the firstechosounder survey. The two available sets ofbottom reflected echosounder signals, in combina-tion with the historical information and the actualground truth, form an invaluable data set withopportunities to compare echo shape parametersbetween the two sounder frequencies on the onehand, and with the ground truthing on the otherhand.
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2275
This paper is organized as follows. Section 2describes the geology of the area, the bathymetry,the ground truth, and the acoustic surveys. A briefdescription of the physical processes that influencethe intensity and shape of echosounder bottomreturns is given in Section 3. The data processingand calculation of the echo shape parameters aretreated in Section 4. Qualitative and quantitativecomparisons between the echo features and theground truth are presented in Section 5. Finally, inSection 6 the extracted echo parameters areanalysed by multivariate statistical methods. Sedi-ment classes with different acoustic signatures areidentified by a principal component analysis fol-lowed by a K-means cluster analysis.
2. Trials area and seafloor geology
2.1. The Botney Cut and Cleaver Bank
An overview map of the extended trials area ispresented in Fig. 1. This part of the North Sea hasbeen the subject of extensive geological surveys. Tillhas been sampled in numerous boreholes and coresin the area of the Botney Cut and Cleaver Bank atthe south–east margin of the Dogger Bank in theDutch sector of the North Sea. It consists pre-dominantly of brown to greenish-grey, stiff, sandyclay (o63 mm varying from 47% to 75%) withmainly matrix-supported crystalline and chalkgravel (varying from 3% to 410%) together witha high percentage of flint. Sand layers and laminaeare locally present. The till is rich in calciumcarbonate. On seismic profiles the formation has achaotic to poorly ordered internal seismic reflectorconfiguration. In areas with a formation thicknesssmaller than 2m no reflectors are observed on theseismic profiles. The depth to the top of the forma-tion varies between 40 and 50m below the mean sealevel (MSL) and is strongly undulating, whichmakes it difficult to draw contour lines. The thick-ness of the formation varies between less than 2to 6m.
In the British sector west and south of the DoggerBank, extensive areas with medium-to-coarse-grained gravelly sand are present which overlie thetill of the formation. In the Dutch sector thedeposits are found around the Botney Cut andthe Cleaver Bank. The deposits are referred to theIndefatigable grounds formation (Harrison et al.,1987). The gravel content of this formation variesbetween 30% and approximately 70%. The thick-
ness of the layer ranges between 0.1m in areas witha flat seabed and 42m on top of and around ice-pushed knobs (Jeffery et al., 1989). Side-scan sonarimages indicate the occurrence of blocks of morethan 1m in diameter on the seabed. Duringsampling programmes in this area, blocks with adiameter of up to 0.7m have been collected. Northof this gravel field, small areas with thin gravellylayers are locally present.
Most of the gravel has been reworked andredeposited by the action of strong tidal currentsin this area during recent times. Side-scan sonarimages show the occurrence of small scale currentripples between overlying sand ribbons. Accordingto observations by divers the ripples are up to0.15m in amplitude and consist of gravelly sand.Between the Botney Cut and the Cleaver Bank aseries of more or less north–south trending knobsare found with a height up to 13m. Several corestaken in the knobs showed that they are made up ofthe same till as sampled around the knobs (Laban,1995). They appear to represent ice-pushed struc-tures. In contrast, the east–west trending CleaverBank is built up of marine sand and was mostprobably formed during the Holocene transgres-sion.
A pattern of braided and isolated valleys has beeneroded into and through the formation. Theorientation of the Botney Cut, one of the largestvalleys, is south–east. The base of the valleys isirregular. Almost all valleys are buried. The infillconsists partly of soft meltwater clays at the baseoverlain by Holocene channel deposits. The BotneyCut however is partly open. The base of the valleylocally reaches a depth of 100m below MSL and isfilled to a depth of between 50 and 70m belowMSL. The water depth of the surrounding area isabout 40m. The infill of the partly buried subglacialvalley of the Botney Cut consists at the base ofcoarse gravelly sediments which have been sampledand which have also been observed on seismicprofiles. Boreholes indicate that the coarse infill isoverlain by greyish-brown, soft to stiff, Pleistocenesilty clay. On seismic profiles draped structures areobserved over the coarse basal infill. The soft clay iscovered by Holocene silty sand and clay.
2.2. Acoustic survey area
A subsection of the Cleaver Bank and the BotneyCut has been surveyed with single beam echosoun-ders. This area is indicated by the red rectangle in
ARTICLE IN PRESS
Table 1
The textural names of the Folk class abbreviations in Fig. 2
Abbreviation Textural name
G Gravel
sG sandy Gravel
msG muddy sandy Gravel
mG muddy Gravel
gS gravelly Sand
gmS gravelly muddy Sand
gM gravelly Mud
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932276
Fig. 1. The trials area was selected with the purposeto achieve as much diversity in sediment type aspossible in this part of the North Sea. A geologicalmap of the trials section is shown in Fig. 2. Thearea, which measures approximately 10� 10 nau-tical miles, fulfils the requirement of sedimentarydiversity, as evidenced by the legend (after Folk,1954). According to the 1987 map, the sedimentranges from soft and flat (‘sandy Mud’) to hard andrough (‘sandy Gravel’). The Folk categorization is
Fig. 2. Historical geological map of the trials area (Harrison et
al., 1987). The map results from a sampling programme inspired
by side-scan sonar images. The legs of the 150 kHz survey are
drawn in black and the grab positions of the October 2000
ground truth are indicated by white markers. The legend of the
map is furnished by the triangle below the map, which gives the
textural names for mixtures of gravel, sand, and mud. (After
Folk, 1954). See Tables 1 and 2 for the meaning of the
abbreviations.
(g)S slightly gravelly Sand
(g)mS slightly gravelly muddy Sand
(g)sM slightly gravelly sandy Mud
(g)M slightly gravelly Mud
S Sand
mS muddy Sand
sM sandy Mud
M Mud
Table 2
Particle-size divisions of the sediment categories mud, sand, and
gravel
Grain size D Primary class j ¼ �log2ðDÞ(D ¼ particle
diameter in mm)
42mm gravel jo� 1
0.0625–2mm sand �1ojo4
o0.0625mm mud (silt+clay) 4oj
based on a triangle divided into 15 partitions andaddresses mixtures of gravel, sand, and mud. Thetextural names that belong to the abbreviations inthe legend of Fig. 2 are given in Table 1 and theparticle-size divisions of the chief sediment cate-gories gravel, sand, and mud are listed in Table 2.Often a logarithmic size parameter j is used insteadof the actual particle size (third column).
2.3. Acoustic surveys
Acoustic data were collected in the trials area ontwo occasions. In October 2000, the area wassurveyed with a 150 kHz echosounder. This echo-sounder operated at a pulse length of �0.5msand at a full (�3 dB) beam width of 17.51. Theroute consisted of 10 horizontal legs with a lengthof 10 nautical miles spaced by 1 nautical mile(Fig. 2). Soon after, in November 2000, a secondacoustic expedition was undertaken with anothership. This ship sailed roughly the same legs, but its
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2277
echosounder operated at a frequency of 66 kHz. Forthis sounder the pulse duration and beam widthamount to �0.5ms and 151, respectively. All echosignals were tapped straight from the transducerand digitally stored for laboratory processing.Approximately 25,000 echoes were recorded at eachfrequency. A few recording interruptions occurredbecause of bad weather (150 kHz) or technicalproblems (66 kHz).
2.4. Ground truthing
In order to obtain recent information on thesurface sediment, 50 bottom grabs were collected inthe trials area during the first sea trials in October2000. Five samples per leg were collected at intervalsof 4 km (Fig. 2). Use was made of a so-called‘Hamon grab’ and a ‘Van Veen grab’. Owing to amore sophisticated grab bucket sealing, the sampleis better preserved during transport from theseafloor to the ship with the Hamon grab than withthe Van Veen grab. However, as the ship wasrelatively small and the Hamon device unwieldy, itcould only be used under calm weather conditions:24 bottom grabs were collected with the Hamongrab and the remaining 26 grabs were collected withthe smaller Van Veen grab.
The sampling depth was approximately 20 cmwith both devices, which is large compared with theacoustic wavelength l of the sonar systems. Inwater, l is 1.0 cm at 150 kHz and 2.3 cm at 66 kHz.Since the attenuation of the sound in the sediment istypically in the range of 0.1–1 dB/l (Hamilton,1972), only a thin layer near the surface contributesto volume backscattering. The largest penetrationdepth is expected at the lower of the two frequen-cies, and in the softest sediment. At 66 kHz, thepenetration depth (�6 dB, 2-way) is about half ametre for an attenuation of 0.1 dB/l. A more likelyvalue of 0.2 or 0.3 dB/l for sandy Mud alreadyrestricts the volume scattering to the thin layerprobed by the grabs. In the harder sandy andgravelly sediments, and at 150 kHz, the soundpenetration will be even smaller. Thus, it can beconcluded that the bottom grabs are representativeof the portion of the sediment probed by theechosounders.
The laboratory analysis of the bottom grabsconsisted of the following steps. The samples weredried and sieved with a 2mm mesh to separate thegravel and shell fragments from the smaller grains.Subsequently these smaller grains were analysed
with a laser-diffraction granulometer. In laser-diffraction granulometry, the illumination of dis-persed particles with a laser beam yields a lightscattering pattern that contains information aboutthe particle sizes. This diffraction pattern istranslated into a volumetric grain size distribution,from which the mud and sand fractions of theoriginal grab can be determined. Grab labelling isfinally accomplished in the top left graph of Fig. 3by equating the mud, sand, and gravel percentagesto the Folk triangle in Fig. 2. Although the overallimpression of the mud, sand, and gravel grounds isthe same between the Folk classes in Figs. 2 and 3,nearly half of the grabs deviate from the historicalmap by at least one section in the Folk triangle. Thiscan be due to the interpolation used for thehistorical map, differences in ground truthingmethodology, and to actual seabed changes. Forinstance, the wandering sand ribbons that overliethe gravel bed influence the sand content of the grabsamples. Up-to-date ground truthing is undoubt-edly of importance for the validation of acousticseafloor characterization techniques.
The grab analysis also returned three grain sizeparameters called D10, D50 (the median grain size),and D90, calculated from the volumetric grainsize distribution. Ten volume percent of the grabhas particle diameters smaller than D10, 50% issmaller than D50, and 90% is smaller than D90.A common logarithmic scale uses the so-calledphi unit: j ¼ �log2ðDÞ, with D the grain sizein mm. Using the j values for D10, D50 and D90,one can approximate the mean grain size Mz in phiunits by
Mz ¼3j10 þ 4j50 þ 3j90
10. (1)
With this definition j10 is considered representativeof the grain size interval j0 � j30, j50 of the intervalj30 � j70, and j90 of the interval j70 � j100. Otherreasonable weightings can be used, but the differ-ences for the mean grain size are small (p0.2 phiunits) for the present grab samples. For a compar-ison with the acoustic results in Sections 5 and 6, theFolk class, the mean grain size, the shell percentageand the gravel percentage of the 50 grab samples aredisplayed in Fig. 3.
2.5. Bathymetry
The depth in the trials area was derived from theecho return time of the 150 kHz sounder and is
ARTICLE IN PRESS
Fig. 3. Graphical representation of the ground truth. Top left graph: the Folk class of the bottom grabs plotted on top of the 150 kHz
echosounder tracks. Maps of the mean grain size (top right, in phi units), the shell percentage (bottom left), and the gravel percentage
(bottom right) were obtained by linear triangular interpolation.
Fig. 4. Bathymetry of the investigated area. The plot results from
a linear triangular interpolation of the depths determined with
the 150 kHz sounder. Notice that the scale does not refer to the
distance from the sea level, but between the echosounder and the
seafloor.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932278
plotted in Fig. 4. The scale differs from the nominalwater depth by a fixed offset of a few metres due tothe depth of the echosounder, and by a variableoffset of 71 m due to the tide. Fig. 4 is notcorrected for these offsets because it is the actualdistance of the echosounder to the seafloor that isrelevant for the sound propagation losses and thesonar footprint size. (In the 66 kHz processing, thedepth determined with the 66 kHz sounder wasused.) A comparison between Figs. 2 and 4 revealsthat there is a strong correlation between thebathymetry and the primary sediment types ofmud, sand, and gravel. Gravel and sand groundsare found at depths between 30 and 40m, while thepresence of mud is restricted to a trench (the BotneyCut valley) with a depth between 50 and 60m. It is,therefore, important to remove any influence of thesound propagation path length from the echosignals prior to feature extraction.
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2279
3. Physical processes behind the echo formation
In order to understand the behaviour of echofeatures as they vary with the bottom type it isinstructive to consider the interaction of sound withthe seafloor. The geometry of echo formation hasbeen described by, for instance, Caughey and Kirlin(1996) and Pouliquen et al. (1999). Starting outfrom the transducer, a spherically expanding waveimpinges upon the seafloor. The seafloor area that iseffectively illuminated by the echosounder is knownas the footprint. After initial reflection in the centreof the footprint, annuli of increasing radii areensonified by the spherically expanding wave. Thebackscatter from these annuli is delayed relative tothe initial bottom return, affecting the shape andduration of the echo. A measure of the footprintdimensions is obtained by considering the (�3 dB)beam width of the echosounder main lobe. Thesound cone thus delineated is characterized by ahalf-angle b and a height h that equals the distancebetween the transducer and the seabed. The amountof time spreading experienced by the echo isdetermined by one of two competing processes,depending on which dominates. On the one handthere is a propagation delay for scattering from thefootprint margins
t1 ¼2h
v�
1
cos b� 1
� �, (2)
where v is the sound speed in water. On the otherhand there is the delay associated with volumescattering, which amounts to
t2 ¼2d
vs, (3)
with d the penetration distance into the sedimentand vs the sediment sound speed. Both delays aresmall, of order 1ms or smaller for the parameters ofthe current experiments. The importance of scatter-ing contributions relative to the initial bottomreturn of specular reflection depends not only onthe beam geometry, but also on the frequency andangular dependence of the scattering strengths andon the sediment type. Thus, the echo signal willgenerally depend on the echosounder frequency andon the sediment surface texture and bulk constitu-tion. It also depends on the shape and length of theoutput ping, which can vary from sounder tosounder.
The shape of acoustic signals reflected from theseafloor is influenced by three main mechanisms.
First, there is a contribution due to coherentreflection at normal incidence. This contributionnormally comes from the central portion of thefootprint equivalent to the first Fresnel zone(Brekhovskikh, 1973), although the presence ofripples or dunes can cause specular contributionsfrom the entire footprint. Second, there is anincoherent surface backscatter signal from the outerparts of the footprint. Here the sound is at obliqueincidence and the specular reflection is directedaway from the transducer. The only possibility forthe footprint periphery to contribute to the echo isby backscattering at surface irregularities andparticles such as gravel and shells. The strength ofthis scattering tends to increase with increasingfrequency. Third, there is a contribution due tovolume scattering in the sediment, from bulkinhomogeneities (pebbles, gas bubbles) or layering.The effect of volume scattering on the echo shapebecomes important when the penetration depth ofthe sound into the sediment is comparable with, orlarger than, the spatial extent of the transmittedpulse (i.e., pulse duration multiplied by soundspeed). In contrast with surface backscatter, theimportance of volume scattering is expected todecrease with an increasing sounder frequencybecause of the enhanced absorption. Roughlyspeaking, the shape of an echo signal depends onthe relative effectiveness of coherent reflection,surface scattering, and volume scattering.
4. Data processing
4.1. Echo extraction and compensation for
propagation losses
The recorded echo signals were digitally filteredto remove noise outside the frequency band of thesonars. Envelopes were determined by a Hilberttransform of the time series. The envelopes werecropped to an interval of T0 ¼ 10ms around thepeak echo intensity, which is long enough to collectall backscattered energy. Subsequently, the echosignals were corrected for sound propagation losses.Spherical wavefront spreading was taken intoaccount and absorption losses were compensatedfor by use of the Francois and Garrison (1982)relationships. Temperature and salinity, necessaryingredients in their description of absorption, weremeasured in situ. At the propagation distances andfrequencies involved in the echosounder experi-ments, absorption losses are smaller than the losses
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932280
due to wave expansion. Absorption losses amountto 0.02 dB/m at 66 kHz and 0.04 dB/m at 150 kHz.This results in a difference of 1.2 dB between thedeep (�60m) and shallow (�30m) areas at 66 kHz,versus 2.4 dB at 150 kHz. In contrast, sphericalexpansion results in a difference of 6 dB for thedepth ratio of two between the mud and sand/gravelgrounds, independent of the frequency. All echoenvelopes were corrected for both intensity falloffmechanisms.
4.2. Feature calculation
4.2.1. Statistical moments in the time domain
Echo energies follow from the intensity I(t) thatvaries quadratically with the recorded voltage,which is proportional to the sound pressure. Thetotal energy of an echo is defined here as
E ¼
Z T0
0
IðtÞdt, (4)
where the integral runs over the duration T0 ¼
10ms of the selected crop. This energy E does notneed to have the proper dimension of energy,because we are only concerned with relative valuesin this paper. E is an important parameter because itrelates directly to the hardness and roughness of theseabed. The hardness is associated with the acousticimpedance mismatch between the water and thesediment and affects the coefficient for specularreflection. The roughness is instrumental in surfacebackscattering and influences the echo energy forsufficiently strong scattering strengths.
Let the echo centre of gravity t0 be defined as
t0 ¼1
E
Z T0
0
IðtÞtdt. (5)
As the first shape parameter, the timespread T iscalculated as
The above definitions resemble the moments asso-ciated with statistical distributions (Press et al.,1989). In the current context there is no deeperthought behind the use of these moments other thanto find discriminating echo features. The timespread
is equivalent to the second central moment and is ameasure of the temporal extent of the echo. Thefactor 4 in the nominator is inserted to arrive at T
values of the same order as the full width at halfmaximum of the echo intensity envelope.
The third moment or skewness is a measure of theecho asymmetry. Since T and S are normalized bythe energy E they are pure shape parameters,independent of the echo energy. Moreover, S isnormalized by the third power of T to ensure thatthe echo duration has no influence on the skewness,i.e., a compression or expansion of the time axis hasno influence on the skewness. Higher-order mo-ments such as the kurtosis were also tried out in theanalysis but observed to add no further discrimina-tion between sediments. The skewness S is typicallypositive for seafloor echoes, which is due to thepresence of postcursors in the echo signal. Follow-ing an initial peak of specular reflection at thewater-sediment interface in the centre of theechosounder footprint, the ‘tail’ of the echo isshaped by surface scattering from the footprint areaand/or volume scattering from within the bulk.Both scattering contributions lead to an echoasymmetry with a tendency towards a positiveskewness.
4.2.2. Fractal dimensions
A fractal set is defined as a scale-invariant (i.e.,self-similar) geometric object (Mandelbrot, 1982;Hastings and Sugihara, 1994), which means that theobject can be described as a union of rescaled copiesof itself. In nature fractal structures are found, forinstance, in clouds, snowflakes, leaves, and corru-gated seafloors. However, the rescaled smallerpieces of these natural structures are not strictlyidentical to the larger pieces, but statistically similar.Fractal analysis is nonetheless suitable to study suchelements, more so perhaps than Euclidean geome-try, which is more appropriate for describing simple,regular figures. The idea of using fractal dimensionsto characterize sonar echoes comes from theassumption that a corrugated seafloor leaves afractal fingerprint on the echo signal. Fractalstructures may be encountered not only on thesurface of the sediment by self-similarity of sandripples of various sizes, but also in the bulk as layersof varying acoustic impedance. The latter scenario isdescribed in detail by Yamamoto (1996). Thus, bothsurface roughness and volume scattering contribu-tions to the echo signal may bear evidence of fractalstructures.
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2281
In fractal analysis, a method to measure andcompare fractal object dimensions is needed. One ofthe possible measures is the Hausdorff dimension,which is defined as the limit (Mandelbrot, 1982;Hastings and Sugihara, 1994)
DH ¼ limr!0
� log NðrÞ
log r, (8)
where N(r) denotes the smallest number of openballs B(p, r) of position p and radius r needed tocompletely cover the object. B(p, r) ¼ {x: dist(x,p)or}, where dist(x, p) is the distance between thepoints x and p. It is readily verified that theHausdorff dimension is a measure of the complexityof the shape of a given figure.
In the case of an echo envelope, which consists ofa finite number of samples, the Hausdorff dimen-sion cannot be calculated as the limit at r! 0.Methods to estimate fractal dimensions for experi-mental data have been extensively developed (e.g.Rothrock and Thorndike, 1980; Wadhams andDavis, 1994). In this paper we first determine theso-called Hurst exponent via the averaged waveletcoefficient method described by Simonsen et al.(1998). A wavelet transform breaks up the signalinto shifted and scaled versions of the motherwavelet c(x; a, b):
cðx; a; bÞ ¼1ffiffiffiap c
x� b
a
� �. (9)
The wavelet transformation coefficient W(a, b) ofa signal y(x) is defined by the wavelet transform pair
W ða; bÞ ¼1ffiffiffiap
Z þ1�1
yðxÞcx� b
a
� �dx, (10)
yðxÞ ¼1
Cc
Z þ1�1
Z þ1�1
W ða; bÞffiffiffiap c
x� b
a
� �1
a2dadb,
(11)
where a is a scaling parameter that controls thewavelet function spread, b a translation parameterthat determines the shift of the mother wavelet, andCc a normalization constant determined by theFourier transform of the mother wavelet.
The Hurst exponent and the Hausdorff dimen-sion are derived from the slope of a log–logrelationship of averaged (with respect to thetranslation parameter b) wavelet transforms of theecho envelope y versus the scaling parameter a.Fig. 5 illustrates the procedure for one of the150 kHz echoes. The greyscale graph (top left)depicts the wavelet transformation magnitude
W ½y�ða; bÞ�� �� of the echo envelope y, as a functionof the parameters a and b. Let W ½y�ða; bÞ
�� ��� �b
denote the mean of the wavelet transform withrespect to the translation parameter b. The Hurstexponent H, then, is derived from the slope of
W ½y�ða; bÞ�� ��� �
bversus the scaling parameter a,
shown in the bottom left graph of Fig. 5.Mandelbrot (1982) proved that H is connected tothe Hausdorff dimension DH through DH ¼ 2�H.
For the calculation of H many different waveletswere tried (e.g. Haar, Symlets, Daubechies, MexicanHat, Meyer, Coiflets, Morlet, etc.). It was observedthat the wavelet choice somewhat influences themagnitude of the resulting Hurst exponents. ACoiflet wavelet (Dcoif3) was eventually selected asthe mother wavelet for the computations. The onlymotive for this choice is that the Hurst exponentobtained with this wavelet is in the centre of thespectrum of H values for all tested wavelets. Thechoice does not fundamentally affect the outcome ofthe analysis. A more detailed description of allcalculation steps can be found in Tegowski (2005).
Another member of the family of fractal dimen-sions, the so-called box dimension, is defined as(Barat et al., 1998)
Dbox ¼ limDs!0
� log NðDsÞ
log Ds(12)
for a two-dimensional object. A rectangular grid isconstructed around the object with boxes measuringDs� Ds. The number of boxes that enclose part ofthe object under investigation is denoted by N(Ds).The right half of Fig. 5 exemplifies the boxdimension for the same echo that was used toillustrate the Hausdorff dimension. Its envelope isembedded in a unit box with both axes normalizedto unity. As a first step, the box is divided into tensubboxes with Ds ¼ 0:1. Boxes traversed by the echoenvelope are shaded. In this example N(0.1)amounts to 31. In the continuing calculation bothaxes are successively divided into 11, 12, y up to100 divisions and the shaded boxes are counted. Theratio of logðNDsmÞ to logðDsmÞ is used to determineDbox, derived from the slope of a straight line fittedto the data in the least-squares sense (bottom rightillustration in Fig. 5).
As described in Section 4.1 a time interval of10ms was adopted for all echoes. The questionarises whether the far limit of the time axis is betterrestricted to the end of the echo, as determined forexample by a threshold value. To answer thisquestion the box dimension was computed twice,
ARTICLE IN PRESS
Fig. 5. Visualization of the computation of the Hausdorff dimension (left) and the box dimension (right) for one of the 150 kHz echo
envelopes (top right). The greyscale graph shows the wavelet transformation W ða; bÞ defined by Eq. (10). The translation parameter b runs
over the 100 available samples of the echo signal, whereas the scaling parameter a, which determines the spread of the wavelet, is restricted
to 80. Each point in the corresponding log–log plot corresponds to the mean value of a horizontal line in the greyscale plot, i.e., W ða; bÞ�� ��
averaged over the translation parameter b. A linear regression algorithm is used to determine the slope a of log10ðhjW ½y�ða; bÞjibÞ versuslog10ðaÞ. The Hurst exponent H is given by a�0.5 (Simonsen et al., 1998), and the Hausdorff dimension follows from DH ¼ 2�H. The
two graphs at the right illustrate the calculation of the box dimension (Eq. (12)). In the top right graph the echo envelope is embedded in a
unit box and divided into 100 squares with a size Ds ¼ 0:1. Note that the time axis, which is normalized to unity, corresponds to a 10ms
crop of the original echo signal. Boxes traversed by the echo envelope are shaded. A count N of shaded squares yields Nð0:1Þ ¼ 31. This
yields the leftmost point in the bottom right graph. Upon repeating the computation for decreasing box sizes, the box dimension ultimately
corresponds to the slope of log10(N(Ds)) against �log10(Ds). For the present echo DH ¼ 1:24 and Dbox ¼ 1:26.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932282
once with the fixed time interval and once forclipped echoes with a variable duration. The resultswere only slightly different, which hints at self-similarity of the echo envelopes. It also helps thatthe spread in durations among the examined echoesis not large to start with. It was decided to use thefixed-interval results for presentation in Section 5.
For a two-dimensional object, such as the echosignal in Fig. 5, the Hausdorff and box dimensionsalways have a value between one and two. Asmooth line yields values D � 1, whereas a wildlyfluctuating curve shifts the dimensions towards D �
2 since it will shade more boxes in case of the boxdimension, or host more open balls in case of theHausdorff dimension. Both fractal dimensions are ameasure of the complexity of the echo envelope.
4.2.3. Spectral moments
A third family of echo features is calculated in thefrequency domain. Spectral moments can be used todescribe the shape of the echo envelope spectrum.With an increasing order, these moments areincreasingly sensitive to the presence of highfrequency components. Together with the spectral
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2283
skewness, a measure of the spectral asymmetry, themoments provide condensed information on bottomroughness or sediment layering, which influencethe echo envelope spectrum. Their definition isanalogous to the moments in the time domain(Section 4.2.1). Let S(o) be the power spectraldensity of the echo envelope s(t). The spectralmoment M of order N is then defined as
MN ¼
Z 10
SðoÞoN do. (13)
For N ¼ 0 we obtain the variance of the signal,which equals the echo energy of Eq. (4). Thespectral moment of the second order (N ¼ 2)describes the concentration of spectral poweraround the mean frequency of the echo spectrum.Notice that the definition Eq. (13) assumes that themaximum energy is at or near zero frequency, whichis indeed the case for our echo envelopes. Finally,the spectral skewness g is defined by
g ¼ ~M3= ~M3=22 , (14)
where the ~Mi are the central moments of order i.
4.3. Depth correction
Apart from the energy correction for propagationlosses (Section 4.1), a shape parameter such as thetimespread T (Eq. (6)) requires an additionalcorrection for the sonar footprint dimensions. Thefootprint diameter increases in proportion to thewater depth, and hence the backscatter area andecho duration increase as well. According to Eq. (2)an increasing water depth comes with an increasingdelay of scattering from the footprint periphery.Therefore, the timespreading would depend on theheight of the water column even without sedimentchanges. A first-order correction is applied toremove the influence of the depth on the timespread.The timespread T in Eq. (6) is multiplied by a factorhref/h, where href is a reference depth and h the depthat the position of the individual echo captures. Thisprocedure is equivalent to the depth-dependentresampling rate described by Caughey and Kirlin(1996). For our data set we—arbitrarily—adoptedthe minimum depth in the data set (30m) as thereference depth. No such correction is necessary forthe other shape parameters, i.e. the skewness andthe Hausdorff and box dimensions. Indeed, if weconsider this depth correction as a time compressionor expansion of the echo signal, it follows from theirdefinitions that these parameters are not affected.
5. Comparison of individual echo features with the
ground truth
5.1. Feature mapping
The parameters described in the previous sectionwere calculated for the echo signals described inSection 2.3. In order to reduce echo to echofluctuations, it is desirable to perform some kindof averaging. The averaging may be performedbefore or after feature calculation. Fluctuations dueto ship movements can be mitigated by stacking anumber of echoes prior to feature calculation.Fluctuations due to actual seabed variations, how-ever, can be preserved if the features are extractedfor the individual echo signals. Since acousticsurveys are generally subject to both causes of echosignal variability, the preferred order of featureextraction and averaging is not always clear. In thepresent paper the shape parameters are extracted foreach echo, and a median filter of length 31 issubsequently applied to these features. Although themedian filter reduces the spatial resolution to some200m it preserves sharp transitions, and theresolution is still considerably finer than the 4 kmspacing between the bottom grabs.
Six features are shown for each sounder fre-quency: the echo energy (Eq. (4)), the timespread(Eq. (6)), the skewness (Eq. (7)), the Hausdorffdimension (Eq. (8)), the box dimension (Eq. (12)),and the spectral skewness (Eq. (14)). Fig. 6illustrates the features for the 66 kHz sounder, andFig. 7 does the same for the 150 kHz sounder. In thefollowing the six parameters will be discussed andcompared with the ground truth, phenomenologi-cally and quantitatively. It is not expected before-hand that the six features all behave very differently,and similarities can be observed. For the sake ofcompleteness, graphs of all parameters are none-theless shown.
5.1.1. Echo energy
As it appears, the echo energy E is an excellentparameter to discriminate the various sedimenttypes (Figs. 6 and 7). The gravel, sand and mudareas stand out clearly against one another,especially at 150 kHz, and compare well with thehistorical map in Fig. 2 and the ground truth inFig. 3. The most conspicuous difference betweenthe three representations is found in the lowerleft-hand corner of the square. According to the1987 map this is a sand area, but both the
ARTICLE IN PRESS
Fig. 6. Feature maps for the 66 kHz echosounder. To aid the comparison between features, the orientation of the colour bars is
consistently chosen such that it runs from dark for the mud trench to bright for the gravel areas.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932284
ground truth and the energy plots indicate that ithas become a gravel area. Indeed, the acousticresults provide supporting evidence for theground truth and show that the single sandy Gravelgrab is representative of a larger area. Althoughthe energy is a good seafloor discriminator atboth echosounder frequencies, the dynamic rangediffers. At 66 kHz the difference between the mudand gravel is some 9 dB, whereas at 150 kHz it
amounts to 16 dB. Apparently there are frequency-dependent echo formation mechanisms at playthat cannot be accounted for by the Rayleighreflection coefficient. A likely explanation is theexistence of a nonuniform sediment top layer, with athickness of only a few cm, which gives rise to afrequency dependent reflection coefficient (Mouradand Jackson, 1989; Lyons and Orsi, 1998; Ainslie,2005).
ARTICLE IN PRESS
Fig. 7. Feature maps for the 150 kHz echosounder. To aid the comparison between features, the orientation of the colour bars is
consistently chosen such that it runs from dark for the mud trench to bright for the gravel areas.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2285
5.1.2. Timespread and skewness
A measure of the temporal extent of the echo, thetimespread T is a pure shape parameter independentof the energy. From Figs. 6 and 7 it appears that thetimespread discriminates better between sand andgravel at 66 kHz than at 150 kHz. Overall thetimespread values are small —of order 1ms— andthey are larger for the 150 kHz sounder, whichsuggests that surface roughness scattering is a more
efficient mechanism than volume scattering. Thetimespreading is highest for seafloor segmentsenriched in gravel. Since there is a strong correlationbetween water depth and sediment type in theCleaver Bank survey area, the question arises towhat extent the timespread is affected by the depthcorrection described in Section 4.3. Notice in thisregard that no depth correction would be requiredin the hypothetical case of a purely specular
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932286
reflection from a mirror-like seafloor. For the echoreturned by a mirror always has the same shape asthe transmitted pulse, unlike echoes shaped bysurface scattering and volume reverberation. Sinceit is unclear to what degree specular reflection onthe one hand, and surface and volume scattering onthe other hand contribute to the echo duration, the‘truth’ will be somewhere in between. Thus, to acertain extent the depth correction described inSection 4.3 may overcompensate. However, it wasverified that the timespread still offers discriminat-ing power when the depth correction is switched off.
The skewness S is shown in the bottom leftgraphs of Figs. 6 and 7. At 66 kHz the skewness is areasonable parameter to discriminate between thethree main sediment types of mud, sand, and gravel,whereas it does not discriminate well between sandand gravel at 150 kHz. Its discriminating power isroughly similar to that of the timespread.
5.1.3. Fractal dimensions
In situations where specular reflection dwarfsscattering contributions to the echo energy, the echois relatively smooth and the fractal dimensions willbe close to (but greater than) one. For acousticallyhard sediments such as mixtures of gravel and sandthe scattering process takes place exclusively at thesurface. The shape and corrugation of the water–-bottom interface then influence small scale fluctua-tions of the echo signal. For acoustically softsediments, i.e. mixtures of mud and sand, volumescattering at inhomogeneities trapped in the sedi-ment must be considered. If there are sedimentlayers near the surface with a fractal structure, thatstructure may be transferred onto a corrugated echoshape. The latter mechanism is more likely tomanifest itself at 66 kHz than at 150 kHz becauseof the increased penetration of the sound into thesediment.
The Hausdorff dimension computed via thewavelet method, and the box dimension (in theright columns of Figs. 6 and 7) show that the echoesfrom muddy sediments are more corrugated thanthose from sand and gravel. For muddy Sand andsandy Mud the Hausdorff dimension achievesvalues up to 1.4 (66 kHz), whereas the box dimen-sion achieves a value of 1.32 at 150 kHz. For sandand mixtures of gravel and sand (S, gS, sG) the boxdimension drops to 1.2 and the Hausdorff dimen-sion almost to one. The Hausdorff and boxdimensions clearly separate the soft sediments fromthe harder ones at both sounder frequencies.
Discrimination between sand and gravel is uncon-vincing at either frequency.
5.1.4. Spectral skewness
The spectral moments (Eq. (13)) yielded resultsthat do not differ much from the echo energy inFig. 6. However, the spectral skewness (Eq. (14)), asa measure of spectral asymmetry, shows a differentpicture (bottom right graphs in Figs. 6 and 7). Alarge spectral skewness implies an asymmetry of thepower spectral density around its mean value.Power spectral densities that are symmetricalaround the mean have a zero skewness. At bothsounder frequencies, areas with acoustically soft,muddy sediments are characterized by largerspectral skewnesses than sand or gravel.
5.2. Correlation of echo features with the grain size
The preceding subsection offered a qualitativedescription of the echo features. In this subsection amore quantitative analysis is carried out to determineto what degree the features are actually correlatedwith the ground truth. The value of a feature at theposition of a bottom grab is calculated as thearithmetic mean of all echoes collected within a givensearch radius. For a small radius the grab sample islikely to be more representative for the echoes, but inorder to average out echo to echo fluctuations a largeradius is preferred. Since the tracks of the acousticsurveys miss the grab locations by distances of up to100m, a radius of 150m has been adopted. This valueyields between 20 and 40 echoes per grab.
Echo features versus the mean grain size areexhibited in Fig. 8 (66 kHz sounder) and Fig. 9(150 kHz sounder). The graphs include the correla-tion coefficient r, whose magnitude ranges from 0.57for the spectral skewness at 66 kHz to 0.86 for theenergy at 150 kHz. For all correlations the asso-ciated P value is smaller than 0.0001, which tells usthat the probability that the obtained correlationcoefficients arise by chance is close to zero(o0.01%). The high confidences (100� ð1� PÞ
� 100%) are due to the relatively large number ofmeasurements. It is highly unlikely that E50measurements yield a correlation coefficient of0.57 or beyond by chance. The shape parametersare also correlated with one another. For instance,at 150 kHz the correlation coefficients among thefeatures range from r ¼ �0:23 (timespread2boxdimension, P ¼ 0:11) to r ¼ 0:94 (Hausdorffdimension2box dimension, P ¼ 0). The strong
ARTICLE IN PRESS
Fig. 8. A quantitative comparison between the mean grain size (in phi units) and the echo features at 66 kHz. Folk classes of the grabs are
distinguished by the use of different markers, for which a legend is presented as an inset in the top left graph. Correlation coefficients r are
included in the graphs. The dashed lines are linear regressions and serve as a guide to the eye. Three crosses are additionally plotted for the
main sediment types mud, sand, and gravel, here represented by sandy Mud, Sand, and sandy Gravel. The crosses are positioned at (X,
Y) ¼ (mean value of the grain size, mean value of the feature) and the length of the bars gives the standard deviation in either direction.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2287
and significant correlation between the two fractaldimensions is not surprising as they are both ameasure of the echo envelope complexity. Redun-dant information in the set of six echo features willbe removed in a natural fashion by the principalcomponent analysis of the following section.
Crosses are included in Figs. 8 and 9 to assess theecho features for the primary sediment types mud,sand, and gravel, for which the folk classes sandyMud, Sand, and sandy Gravel are the bestrepresentatives. The crosses are positioned at thecentre of gravity of each of these populations, with
ARTICLE IN PRESS
Fig. 9. A quantitative comparison between the mean grain size (in phi units) and the echo features at 150 kHz. Folk classes of the grabs are
distinguished by the use of different markers, for which a legend is presented as an inset in the top left graph. Correlation coefficients r are
included in the graphs. The dashed lines are linear regressions and serve as a guide to the eye. Three crosses are additionally plotted for the
main sediment types mud, sand, and gravel, here represented by sandy Mud, Sand, and sandy Gravel. The crosses are positioned at (X,
Y) ¼ (mean value of the grain size, mean value of the feature) and the length of the bars gives the standard deviation in either direction.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932288
horizontal and vertical error bars corresponding tothe standard deviations of the mean grain size andthe feature under examination. In this manner it isseen that the skewness yields the best discriminationbetween mud, sand, and gravel at 66 kHz. For this
frequency the timespread does a reasonable job inseparating sand and gravel. In contrast, the echoenergy is the best discriminator between mud, sand,and gravel at 150 kHz, at which frequency thespectral skewness convincingly separates sand and
ARTICLE IN PRESS
Table 3
Distribution of the standardized echo features over the first three
principal components (Z1, Z2, Z3)
Z1 Z2 Z3
66 kHz
Echo energy 0.438 0.200 �0.473
Timespread 0.389 0.367 0.601
Skewness �0.444 �0.054 �0.428
Hausdorff dimension �0.433 0.409 0.020
Box dimension �0.326 0.714 0.037
Spectral skewness �0.408 �0.382 0.480
Variance (%) 69.0 17.1 8.8
150 kHz
Echo energy 0.408 0.020 0.718
Timespread 0.299 �0.842 0.056
Skewness �0.408 0.208 0.614
Hausdorff dimension �0.440 �0.282 0.234
Box dimension �0.425 �0.405 0.107
Spectral skewness �0.451 �0.072 �0.196
Variance (%) 73.9 13.4 7.4
The tabulated coefficients (a1;k, a2;k, a2;k, see Eq. (15)) denote the
contribution of the six echo features to the respective compo-
nents. Also given is the percentage of the variance accounted for
by each principal component.
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2289
gravel. Notice that for both frequencies there areseveral features that discriminate mud from sandand gravel, but a clear distinction between sand andgravel is less common. No features other than theabovementioned offer this discrimination.
6. Multivariate statistics and seafloor discrimination
So far six individual echo features have beenconsidered. It was shown that these features havevarying degrees of discriminating power betweensediments and also that there is an overlap in theperformance of the individual features, i.e., the sixfeatures carry redundant information among them-selves. A multivariate statistics method has beenapplied to the data to remove this redundancy andto identify sediment types with distinctly differentacoustic properties. Manly (1994) provides anexcellent introduction to this matter. In the firststep the echo features are given a zero mean andunit variance. This standardization ensures that thesix features have equal weights in the next steps ofthe calculation. The standardized variables X areput through a principal component calculation thatgenerates a new set of orthogonal variables Z. Thesevariables, known as the principal components, arelinear combinations of the input variables accordingto
Zn ¼X6k¼1
an;kX k, (15)
where the coefficients, computed from the covar-iance matrix of the input parameters X, satisfy
X6k¼1
a2n;k ¼ 1 (16)
for each n. Table 3 lists the coefficients an;k for thefirst three principal components. It is observed thatthe relative contributions of the echo features to thefirst principal component are of comparable magni-tude. This component is not far from being just thesign-corrected average of the input parameters andremoves most of the redundancy. The secondprincipal component is dominated by the boxdimension at 66 kHz, and by the timespread at150 kHz. Subsequent principal components becomeless important, which is reflected by the variance inthe bottom row. At 66 kHz the first principalcomponent accounts for 69.0% of the variabilityin the standardized ratings. The second componentcarries 17.1% of the variability and the third one
8.8%. Together, the first three principal compo-nents account for 94.9% of the variance. At150 kHz this is 94.7%. Little information is lost,therefore, by removing the last three principalcomponents from the subsequent analysis. Thenumber of components that can be removedgenerally depends on correlations between the inputvariables. The stronger these correlations, thegreater the allowed reduction in the number ofprincipal components (Manly, 1994).
As a next step the remaining three componentsare fed to a cluster analysis based on the widely usedK-means clustering algorithm (Spath, 1980). Thealgorithm initially adopts a number of randomcentres in the parameter space. In the next steps, thealgorithm moves objects between clusters while itminimizes the variability within clusters and max-imizes the variability between different clusters,using Euclidean distances. At the end of theiteration procedure, clusters remain that are to beidentified with acoustically distinct sediments. Thequestion as to the definition and determination ofthe optimum number of clusters, or how manyprincipal components should be used for thatmatter, is involved (Legendre, 2003, and referencescontained therein). There is no generally accepted
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932290
‘‘best’’ method and there is usually a large subjectivecomponent in the assessment of the results from anyparticular method (Manly, 1994). The issue isavoided here by considering a fixed number of fourclusters (Fig. 10). This number was adopted afterinspection of the principal component space fromvarious viewpoints. To account for all visual cloudseparations, which are more convincing at 66 kHzthan at 150 kHz, a minimum of four clusters isrequired. Thus, the differences between the twoechosounders are examined for a fixed number ofclusters, which leaves the sounder frequency as theonly variable to influence the partitioning. Adissimilar number of clusters between the twofrequencies would make a very different compar-ison, with an outcome that depends on the criterionused to determine the ‘‘optimum’’ number ofclusters. The emphasis of the present paper is on
Fig. 10. Visualization of the acoustical classes in the three-
dimensional space set up by the first three principal components.
The four clusters partitioned by the K-means algorithm are
represented by different colours.
the presented echo features and behaviour differ-ences between the two frequencies, not on clusteringalgorithms.
Next, as the last step towards seafloor discrimina-tion, the spatial distributions of the sediment typesidentified by the principal component and clusteranalyses are plotted in Fig. 11. There is a goodoverall correspondence between the acoustic resultsin Fig. 11 and the historical map and the groundtruth (Figs. 2 and 3). Of course, one should notexpect to find the same number of acoustic sedimentclasses as there are Folk classes represented in theground truth (eight). The Folk triangle in Fig. 2 isone of many possible classification schemes anddoes not directly relate to the sediment acoustic
Fig. 11. Final outcome of the principal component and cluster
analyses. The colours of the four sediment types correspond to
the colours of the clusters shown in Fig. 10. For a comparison
with the ground truth the Folk class at the grab locations is
included.
ARTICLE IN PRESS
Table 4
Distribution of the grab sample Folk classes over the acoustic
sediment classes at the grab locations
Class I
(dark
green)
Class II
(light
green)
Class III
(yellow)
Class IV
(red)
66 kHz
sM 2 5 — —
(g)sM — — — 1
mS — 2 3 —
S — — 13 3
(g)S — — 1 —
gS — — 3 4
msG — — — 2
sG — — 2 8
150 kHz
sM 8 — — —
(g)sM — — — 1
mS 2 1 2 —
S — 2 13 —
(g)S — — 1 —
gS — — 1 6
msG — — � 2
sG — — � 10
The class numbering roughly ranges from soft (I) to hard (IV)
sediments, with colour designations that are in agreement with
Fig. 11. Notice that the grab samples add up to 49, and not 50,
which is due to interruptions in the acoustic data recording (cf.
Fig. 11).
P.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–2293 2291
properties as probed with the chosen echo features.The final acoustic results in Fig. 11 are to beinterpreted as a seafloor map based on thediscrimination of sediments with different acousticproperties.
In Fig. 11 there is little difference between the twofrequencies in the rendering of the sediments withlarger grain sizes, viz., msG, sG, gS, (g)S. However,the behaviour differs for the smaller grain sizes S,mS, sM, and (g)sM. As it appears, the transitionfrom mud to sand occurs more rapidly at 66 kHzthan at 150 kHz. Compared with the historical mapin Fig. 2 the width of the valley seems under-estimated by the lower frequency. An increase of thenumber of clusters used to initiate the K-meansalgorithm yields more sediment classes, but thedeviation in the width of the mud trench remains.Inspection of the feature maps in Figs. 6 and 7reveals that this phenomenon is present in mostfeatures, also (and notably) in the echo energy, theonly feature that is not a shape parameter.
Fig. 11 is useful for an overall visual assessmentof the classification results, but it is not ideal for adetailed comparison of the acoustic bottom classesand the ground truth. All echoes are plotted inFig. 11, and it may occur that a single marker of aparticular class obscures ten or twenty underlyingmarkers belonging to other classes. A more directcomparison is presented in Table 4. The acousticclass at the grab location was determined by themajority vote of all echoes within a search radius of150m. Folk classes with a statistically reasonablepopulation include sandy Mud, Sand, and sandyGravel, which may be taken to represent theprimary sediment types mud, sand, and gravel.Each of these three Folk classes corresponds well toa particular acoustic sediment class in Table 4. Thisis particularly true at 150 kHz, where all sandy Mudgrabs are represented by Class I, all sandy Gravelsamples by Class IV, and 13 out of 15 sand grabs byClass III. Other Folk classes with a reasonablepopulation are muddy Sand, which is poorlydiscriminated at either frequency, and gravellySand, which is better distinguished at 150 kHz thanat 66 kHz.
A trend is observed for the grab samples devoidof gravel (sM, mS, S). At 150 kHz they populateacoustic classes more to the left in Table 4, and at66 kHz more to the right. This trend is reversed forgrab samples enriched in gravel ((g)sM, (g)S, gS,msG, sG), which overall are more to the right at150 kHz. The acoustic classification scheme thus
appears especially sensitive to the presence of gravelat 150 kHz, where 19 out of 21 grab samplescontaining gravel are a member of the same acousticclass. To some extent this phenomenon couldhave been enhanced by the presence of shellfragments, which occur in the same regions of thetrials area as gravel (Fig. 3). Goff et al. (2000) andGoff et al. (2004) showed that the presence ofonly a few percent of large grains (gravel and shellhash) has a disproportionate influence on theacoustic backscatter intensity of a 95 kHz side-scansonar. The same effect can also be responsible forthe gravel sensitivity of the present acousticclassification scheme. Reverting to Fig. 9, we notethat the echo energy indeed appears sensitive tothe presence of gravel. Nearly all samples holdinggravel are associated with a higher echo energythan the samples devoid of gravel. The presenceof gravel has a larger influence on the echo energythan the mean grain size, for which there is alarger overlap between the Sand grabs and thegravel(ly) grabs.
ARTICLE IN PRESSP.A. van Walree et al. / Continental Shelf Research 25 (2005) 2273–22932292
There is no definite answer to the question whichfrequency gives the best classification results, aseach frequency yields a representation of theseafloor in its own right. Nonetheless, if theobjective is to discriminate the main sediment typesmud, sand, and gravel, in the present surveyrepresented by the Folk classes sM, S, and sG, the150 kHz sounder is preferred. If the data had beencollected with a dual-frequency echosounder, acombined set of parameters could be fed to themultivariate statistics analysis in order to check howthis might affect the categorization. Unfortunatelythe data of the present paper were obtained duringtwo separate sea trials, with deviations between thesailed tracks that hinder a dual-frequency analysisand interpretation.
7. Summary and conclusions
In this paper an acoustic seafloor characterizationmethod depending on a set of six energetic,statistical, fractal, and spectral parameters, is testedagainst the ground truth in a North Sea areanorth–west of the Netherlands. The trials areaoffered a sedimentary diversity ranging from softand muddy (sandy Mud) to relatively hard andrough (sandy Gravel). A comparison is made fortwo echosounders, operating at 66 and 150 kHz,respectively. Significant correlations are foundbetween the echo features and the ground truth(mean grain size), but also between the echo featuresthemselves. The final analysis extracts the usefulinformation from the six parameters by means of aprincipal component analysis. A clustering algo-rithm is used to expose the influence of theechosounder frequency on a seabed partitioningwith four acoustic classes. Maps of the acousticsediment classes show that the acoustic transitionfrom the mud channel to neighbouring sand andgravel fields occurs more rapidly at the lower of thetwo examined frequencies. A direct comparisonwith the ground truth reveals that the mainsediment categories mud, sand, and gravel can bedistinguished at both sounder frequencies. The Folkclasses sandy Mud, Sand, and sandy Gravel areparticularly well separated at 150 kHz. At thisfrequency the acoustic classification scheme alsoappears sensitive to the presence of gravel. Thefrequency dependence of the echo energy as it variesacross the trials area is the subject of currentinvestigations, involving the addition of recent
surveys with three additional echosounder frequen-cies.
The performance of the seafloor characterizationscheme proposed in this paper is consideredsatisfactory. The last stage of the analysis, i.e. theprincipal component analysis in conjunction with acluster analysis, is similar to the operation of thecommercially available QTC bottom classificationsystem (Collins et al., 1996). A difference is thatQTC employs 166 echo parameters, largely un-known to the user, whereas the current analysisrelies on six documented parameters whose indivi-dual behaviour can be studied.
Finally, to expand the proposed method ofseafloor discrimination to fulfil the ultimate desireof bottom classification, a few bottom grabs wouldbe required for calibration. In a typical applicationtracks would be sailed in an area whose seafloor isto be mapped. After the analysis, a plot of thespatial distribution of the discriminated sedimenttypes would indicate suitable locations to collect thebottom grabs in order to label the acoustic classes.This procedure does not differ from the calibrationof currently available bottom classification systems.
Acknowledgements
The authors would like to thank the crews of thetwo measurement ships for their enthusiasticassistance and cooperation during the sea trials.Michael Ainslie is acknowledged for proofreadingthe manuscript and valuable comments.
References
Ainslie, M.A., 2005. The effect of centimetre scale impedance
layering on the normal-incidence seabed reflection coefficient.
Proceedings of ‘‘Boundary Influences in High Frequency,
Shallow Water Acoustics’’, Bath.
Alexandrou, D., Pantzartsis, D., 1993. Application of neural
networks to seafloor classification. IEEE Journal of Oceanic
Engineering 18 (2), 81–86.
Barat, P., De, S., Bandyopadhyay, S.K., 1998. Fractal property
of backscattered acoustic signals from polycrystalline alumi-
