7
Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data Yuri Rzhanov a,n , Luciano Fonseca a,b , Larry Mayer a a Center for Coastal and Ocean Mapping, University of New Hampshire, 24 Colovos Road, Durham, NH 30824, USA b Faculty of Engineering at Gama, University of Brasilia, Brazil article info Article history: Received 15 April 2011 Received in revised form 31 August 2011 Accepted 1 September 2011 Available online 21 September 2011 Keywords: Acoustic backscatter Seafloor characterization Segmentation Combinatorial optimization abstract An automatic approach for delineation of areas representing different acoustic facies is presented. A backscatter mosaic (a gray-scale image) is oversegmented, honoring all possible boundaries, both real and false (acquisition and construction artifacts), in order to separate relatively homogeneous and contiguous groups of pixels of the mosaic, which are called segments. The size of the segments is chosen such that each one is considered to represent a single acoustic facies. These segments are then joined together to end up with a predefined number of ‘‘acoustic themes,’’ in a process called coalescence. The difference between ‘‘facies’’ and ‘‘theme’’ is that the former represents an abstract type of seafloor, and the latter, an area (or areas) selected manually or automatically with common seafloor properties. Ideally, one theme would correspond to a single facies. Themes that are used for assignment to segments are chosen prior to coalescence, either manually or by one of the three automatic methods proposed in the paper. The process of coalescence of segments is based on all acoustic data associated with each segment (not just data used for mosaic construction), their proximity and shape. Results of segmentation are unbiased but depend on a specific goal that the user is trying to achieve. & 2011 Elsevier Ltd. All rights reserved. 1. Introduction Characterization of the seafloor is extremely important for a range of applications, including mapping of underwater habitats and delineation of geological and geotechnical properties of the seafloor. Due to difficulties in directly accessing the seafloor, remote acoustic methods using sidescan and multibeam sonars are normally employed for underwater surveys, which rely on echoes that are reflected or scattered from the seafloor. The amplitudes of these echoes are often used to estimate the back- scattering strength returning from the seafloor. Maps showing the spatial distribution of the backscatter strength recorded by side- scan or multibeam sonars (MBES) are normally presented as images or mosaicspaper in the early days, digital later. The analysis of these maps helps the interpreter draw conclusions about seafloor relief (via the distribution of highlights and shadows on sidescan records) and the composition and texture of the scattering material. As a rule of thumb, harder and rougher substrates (bedrock, gravel, etc.), produce stronger backscatter signals than softer and smoother substrates (clays, silts, etc.). However the relationship between substrate and scattering strength is complex and involves many other parameters, for instance, the seafloor roughness at spatial wavelengths compa- tible with the acoustic wavelength of the incident signal, the acoustic attenuation inside the sediment, and heterogeneities in the sediment volume. Following recent advances in acoustic theory, in seafloor mapping systems (higher spectral and tem- poral resolution, quantifiable systems), and in acquisition geo- metry, efforts have been focused on developing methodology for quantitative extraction of seafloor property information from MBES data (Jackson et al., 1986; Pace and Gao, 1988; Fonseca and Mayer, 2007). Building on these (and other) earlier approaches, herein we describe a procedure for the automatic delineation of regions of common seafloor properties based on all available backscatter information (both strength and angular response). The ultimate goal is to provide seafloor mappers with a robust and automatic tool for seafloor characterization. 2. Traditional processing The construction of maps of acoustic facies from MBES data begins with the acquired sonar records being used to assemble a geo-referenced bathymetric map with the resolution limited by the properties of the sonar device and water depth (de Moustier, 1988). Backscatter amplitudes, which are collected simultaneously with Contents lists available at SciVerse ScienceDirect journal homepage: www.elsevier.com/locate/cageo Computers & Geosciences 0098-3004/$ - see front matter & 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.cageo.2011.09.001 n Corresponding author. Tel.: þ1 603 8620075; fax: þ1 603 8620839. E-mail addresses: [email protected] (Y. Rzhanov), [email protected] (L. Fonseca), [email protected] (L. Mayer). Computers & Geosciences 41 (2012) 181–187

Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Embed Size (px)

Citation preview

Page 1: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Computers & Geosciences 41 (2012) 181–187

Contents lists available at SciVerse ScienceDirect

Computers & Geosciences

0098-30

doi:10.1

n Corr

E-m

lucianof

journal homepage: www.elsevier.com/locate/cageo

Construction of seafloor thematic maps from multibeam acoustic backscatterangular response data

Yuri Rzhanov a,n, Luciano Fonseca a,b, Larry Mayer a

a Center for Coastal and Ocean Mapping, University of New Hampshire, 24 Colovos Road, Durham, NH 30824, USAb Faculty of Engineering at Gama, University of Brasilia, Brazil

a r t i c l e i n f o

Article history:

Received 15 April 2011

Received in revised form

31 August 2011

Accepted 1 September 2011Available online 21 September 2011

Keywords:

Acoustic backscatter

Seafloor characterization

Segmentation

Combinatorial optimization

04/$ - see front matter & 2011 Elsevier Ltd. A

016/j.cageo.2011.09.001

esponding author. Tel.: þ1 603 8620075; fax

ail addresses: [email protected] (Y. Rzhanov

[email protected] (L. Fonseca), [email protected]

a b s t r a c t

An automatic approach for delineation of areas representing different acoustic facies is presented.

A backscatter mosaic (a gray-scale image) is oversegmented, honoring all possible boundaries, both real

and false (acquisition and construction artifacts), in order to separate relatively homogeneous and

contiguous groups of pixels of the mosaic, which are called segments. The size of the segments is

chosen such that each one is considered to represent a single acoustic facies. These segments are

then joined together to end up with a predefined number of ‘‘acoustic themes,’’ in a process called

coalescence. The difference between ‘‘facies’’ and ‘‘theme’’ is that the former represents an abstract type

of seafloor, and the latter, an area (or areas) selected manually or automatically with common seafloor

properties. Ideally, one theme would correspond to a single facies. Themes that are used for assignment

to segments are chosen prior to coalescence, either manually or by one of the three automatic methods

proposed in the paper. The process of coalescence of segments is based on all acoustic data associated

with each segment (not just data used for mosaic construction), their proximity and shape. Results of

segmentation are unbiased but depend on a specific goal that the user is trying to achieve.

& 2011 Elsevier Ltd. All rights reserved.

1. Introduction

Characterization of the seafloor is extremely important for arange of applications, including mapping of underwater habitatsand delineation of geological and geotechnical properties of theseafloor. Due to difficulties in directly accessing the seafloor,remote acoustic methods using sidescan and multibeam sonarsare normally employed for underwater surveys, which rely onechoes that are reflected or scattered from the seafloor. Theamplitudes of these echoes are often used to estimate the back-scattering strength returning from the seafloor. Maps showing thespatial distribution of the backscatter strength recorded by side-scan or multibeam sonars (MBES) are normally presented asimages or mosaics—paper in the early days, digital later. Theanalysis of these maps helps the interpreter draw conclusionsabout seafloor relief (via the distribution of highlights andshadows on sidescan records) and the composition and textureof the scattering material. As a rule of thumb, harder and roughersubstrates (bedrock, gravel, etc.), produce stronger backscattersignals than softer and smoother substrates (clays, silts, etc.).However the relationship between substrate and scattering

ll rights reserved.

: þ1 603 8620839.

),

.edu (L. Mayer).

strength is complex and involves many other parameters, forinstance, the seafloor roughness at spatial wavelengths compa-tible with the acoustic wavelength of the incident signal, theacoustic attenuation inside the sediment, and heterogeneities inthe sediment volume. Following recent advances in acoustictheory, in seafloor mapping systems (higher spectral and tem-poral resolution, quantifiable systems), and in acquisition geo-metry, efforts have been focused on developing methodology forquantitative extraction of seafloor property information from MBESdata (Jackson et al., 1986; Pace and Gao, 1988; Fonseca and Mayer,2007). Building on these (and other) earlier approaches, herein wedescribe a procedure for the automatic delineation of regions ofcommon seafloor properties based on all available backscatterinformation (both strength and angular response). The ultimate goalis to provide seafloor mappers with a robust and automatic tool forseafloor characterization.

2. Traditional processing

The construction of maps of acoustic facies from MBES databegins with the acquired sonar records being used to assemble ageo-referenced bathymetric map with the resolution limited by theproperties of the sonar device and water depth (de Moustier, 1988).Backscatter amplitudes, which are collected simultaneously with

Page 2: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Y. Rzhanov et al. / Computers & Geosciences 41 (2012) 181–187182

the bathymetry, are geometrically and radiometrically corrected(Fonseca and Calder, 2005) and presented as geo-referencedraster gray-scale image maps, with the gray-level proportionalto the backscatter strength of the insonified area at a single angle(typically normalized to 451). An experienced interpreter thendelineates homogeneous regions on the image (acoustic facies)and labels them according to the task that needs to be accom-plished. The interpretation is typically based on available ground-truth information (analyzed grab samples and/or video footage),and existing knowledge about the regional geomorphology. Thenumber of labels may vary from two, as for a biologist, outliningpotential habitat for a certain type of species (i.e., areas wherethese species could exist and those where they might not), tomany more, often following the standard grain-size or habitatclassification schemes (see, for example, Wentworth, 1922;Krumbein and Sloss, 1963; Greene et al., 1999, 2007).

The visual interpretation approach described above has anumber of weaknesses. First, the backscatter image mosaic, whichis a composite of different survey lines and thus requires somesort of blending for seamlessness, shows only a single backscattervalue for each pixel, despite the fact that, in many cases, therecould be multiple acoustic measurements available per pixel, andthe measurements could come from different incident angles.Second, such a mosaic provides only a limited amount of theinformation available about the seafloor. In particular, the mosaicdoes not provide information about the change in backscatter as afunction of angle of incidence (angular dependence). The angulardependence of backscatter strength, which has been modeledbased on physical and acoustical parameters of the seafloor, is areliable descriptor that can be very useful in seafloor character-ization (Jackson et al., 1986; Pace and Gao, 1988; Hughes Clarke,1993; Fonseca et al., 2002). Fonseca and Mayer (2007) comparedthe angular response curves (ARC) for an area of the seafloor to amathematical model relating physical properties of the insonifiedsubstrate and sonar characteristics to backscatter strength. It isimportant to note that in this case the ARC represents the peak ofa distribution specific for a set of samples of backscatter strengthobtained from an acoustically homogeneous area (Rayleigh,K-distribution, or other). Instead of the peak, one may choose asthe ARC the average value for the distribution—the particularchoice does not affect the reported results.

Third, coverage density of a typical survey rarely exceeds 200%,meaning that each footprint (area insonified instantly by a MBESbeam) is insonified on average only twice, usually with differentgrazing angles. Denser surveys are not economically feasible andare conducted only when there is an explicit requirement for this(Hughes Clarke, 1994; Hughes Clarke et al., 2008).

Finally, the manual interpretation of a backscatter mosaicdepends dramatically on experience and even on the mood ofthe interpreter (Fonseca and Rzhanov, 2008). Manually drawnboundaries tend to be exceedingly smooth and do not delineateaccurately fine details even if they are clearly distinguishable inthe mosaic.

Early attempts at automatic quantitative characterization ofseafloor have been based on construction of patches (Pace andGao, 1988; Hughes Clarke, 1993, 1994; Hughes Clarke et al., 1997).A patch is directly related to the dimension of an MBES acquisitionline and consists of all soundings from a certain number (forexample, 30) of consecutive half-swaths, on the port or starboardside. By definition, a patch contains soundings from the range ofgrazing angles available to the MBES used. Averaging over severalconsecutive pings is done to reduce the impact of noise, which isinherently present in backscatter measurements. When a patch,by chance, covers a relatively homogeneous area, the quantitativecharacterization of this area is meaningful, and provides thesurveyor with useful information. However, more often than not,

patches cover acoustically heterogeneous areas, and then theiraveraged ARCs may be confusing and do not carry any newinformation (except for indication of their heterogeneity).

To overcome this shortcoming the notion of an acoustic theme

was introduced by Fonseca and Calder (2007) and later developedin Fonseca et al. (2009a, 2009b). An acoustic theme is a relativelylarge, usually manually delineated, area considered to be homo-geneous by an expert interpreter. An ARC averaged over such anarea allows for quantitative characterization similar to that in apatch-based technique. The main deficiencies of this approach are,as noted earlier, subjectivity in delineation, and decision-makingbased on incomplete information (only backscatter mosaic, i.e.,gray-scale image where pixel values represent backscatter strengthadjusted for a certain grazing angle). Thematic analysis is notlimited to the use of half-swath width area, as the techniqueaccounts for all the soundings that hit the chosen area, indepen-dently of the acquisition line or port/starboard geometry.

3. Automatic processing

We have developed an automated procedure permitting sub-division of a survey area into a predefined (but unlimited) numberof acoustic themes. In this procedure, all angular backscatter datafrom a survey, even those not depicted in a backscatter mosaic areutilized. First, an acoustic backscatter mosaic is assembled usingthe research version of computer program ‘‘Geocoder’’ (Fonsecaand Calder, 2005), which is now implemented in several commer-cial packages for MBES data processing. Backscatter data are usedat highest reasonable resolution—from time series within a singlebeam (‘‘snippets’’). Depending on chosen resolution, mosaic pixelmay contain many, one, or even no soundings at all (backscattervalue in the latter case is obtained by spatial interpolation). Then,the automatic delineation (which constitutes the core of this paperand is described in detail in this section) is done at the mosaicresolution, and can be performed by anyone without prior trainingin backscatter interpretation. Comparison of the mosaic delinea-tion results and ground-truth data from the underwater videofootage will be reported in the accompanying paper (Schimelet al., in preparation).

In a typical MBES survey each acoustic beam touches theseafloor at a known grazing angle, which depends on the positionof the transducers, their orientation, refraction, and local bathy-metry. All of the collected acoustic data can be presented as asparsely populated 3-dimensional matrix—backscatter strengthas a function of two spatial dimensions and a grazing angle (on apixel level this matrix is not different from that used in Parnum(2007)). Ideally, one would like to characterize each pixel in amosaic (the smallest area that is assumed to be homogeneous)independently. Pixel size is selected by a surveyor depending onsonar characteristics and bathymetry of surveyed area. This wouldbe possible only if the above noted matrix is densely populated,i.e., when survey density exceeds 500–600%, and when all the pixelsare insonified at a wide range of angles—from highly obliqueto those with almost normal incidence. Under these conditionsalmost all the elements of the 3D matrix contain at least severalsoundings. (Note that the experiment providing such data densityhas been reported in Hughes Clarke (1994) and Hughes Clarke et al.(2008)).

While this is ideal, 500–600% overlap is not often practical;thus we have to start the processing of a backscatter mosaic withsegmentation into larger areas that can still be consideredhomogeneous. These segments are often much smaller than thetypical swath, and do not contain soundings with the whole rangeof grazing angles. To be effective the segmentation proceduremust honor all boundaries found in the mosaic—some of them

Page 3: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Y. Rzhanov et al. / Computers & Geosciences 41 (2012) 181–187 183

reflecting real spatial changes in the seafloor content, while othermay be a result of the artifacts due to problems with datacollection (‘‘nadir artifact’’) or mosaic construction (‘‘feathering’’).We have found that the best results are obtained by applyingalgorithms based on color quantization techniques. These tech-niques have been extensively studied in computer graphicsliterature since the 1970s (see, for example, Schrader (1998) fora survey). The goal of color quantization is a reduction of thenumber of distinct colors used to represent given image, usuallyfor the purpose of displaying this image on a hardware device thatcan display simultaneously only a limited number of colors. Inour experiments we have employed the JSeg algorithm (Dengand Manjunath, 2001) or the RaveGrid algorithm (Prasad andSkourikhine, 2005). Note that any segmentation procedure pro-vides a pyramid of results—from the finest, where each pixelrepresents an individual segment, to the roughest, with a singlesegment comprising the whole survey area. For our purpose wechoose some intermediate level of segmentation, with typicalsegments consisting of tens to hundreds of pixels (Figs. 1 and 2show examples of backscatter mosaics and the segmentationresults). For our experiments we have used two data sets. Thefirst was acquired in 2005 by the Irish Marine Institute onboardthe survey vessel Celtic Explorer around at the Stanton Banks,west of Scotland, using a Kongsberg-Simrad EM1002S MBES(95 kHz). The survey area, which has area approximately

Fig. 1. (A) Backscatter mosaic of the Stanton

Fig. 2. (A) Backscatter mosaic of the HA

7.5�9 km2 with an average depth of 170 m, encompasses differ-ent underwater habitats in substrates including mud, sand,gravel, and rock outcrops (Fig. 1). The second dataset wasacquired in August 2005 by Science Applications InternationalCorporation (SAIC) onboard M/V Atlantic Surveyor at the HistoricArea Remediation site (HARS) using a Reson 8101 MBES (240 kHz,101 beams). The survey area located offshore New Jersey has areaapproximately 500�900 m2 with an average depth of 25 m(Fig. 2). Since the 19th century, this site has been a preferredarea for disposal of assorted material (Soares Rosa, 2007). Thebackscatter in this survey area shows a very high spatial varia-bility and consequently provides a good test site for the proposedsegmentation methodology.

During the process of mosaic segmentation, the structure ofthe 3D matrix changes so that the organization of the matrix inspatial dimensions is no longer based on equally spaced pixels;instead it becomes irregular and two-dimensional—matrix ele-ments represent now segment numbers and grazing angles. Theseelements have a variable number of neighbors. Segments (spatialrepresentation of matrix elements) are still too small to containnumber of soundings sufficient for characterization, i.e., to con-tain a complete angular response, from which seafloor propertiescan be derived. To deal with this problem we propose an alternativeapproach: we choose a reasonably diverse set of acoustical facies(we call this a catalogue), assign each facies a unique label, and then

Banks areas. (B) Oversegmented mosaic.

RS site. (B) Oversegmented mosaic.

Page 4: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Y. Rzhanov et al. / Computers & Geosciences 41 (2012) 181–187184

attribute each segment a label from this set. In other words, insteadof trying to answer the question: ‘‘What are the most probablephysical parameters for this segment?’’ we try to answer anotherquestion: ‘‘Which set of physical parameters (from the catalogue) ismost likely to describe backscatter data available for this segment?’’Thus the inversion procedure (computationally expensive andrequiring some minimal number of samples for each segment fora wide range of grazing angles) is replaced by the labeling process(fast and guaranteed to provide an answer for any segment contain-ing any number of samples).

The labeling process can be efficiently solved by means ofcombinatorial optimization (Boykov et al., 2001; Boykov andKolmogorov, 2004). The solution is achieved by minimizingcertain functional consisting of two terms—the ‘‘data’’ term(sum of data terms for each segment) and the ‘‘smoothness’’term. Smoothness represents how similar one neighbor is to thenext. We use the term ‘‘smoothness’’ here because it is thestandard term used in combinatorial optimization (see Boykovet al., 2001). The data term considers each segment independentlyand reflects the similarity between incomplete ARC for thissegment and known ARCs from the catalogue. It is estimated asfollows:

D¼XK

k ¼ 1

1

Nk

XNk

n ¼ 1

9BSk,nðaÞ�ARClðaÞ9( )

:

Label l is the current label of the kth segment containing Nk

samples. K is the total number of segments. The procedure doesnot impose any restrictions on Nk BSk,n(a) is the backscatterstrength of the nth sample (sounding) of kth segment which hasgrazing angle a, and ARCl(a) is a value of the ARC from thecatalogue with this label and the same grazing angle.

The data term for a segment having label l is minimal whenthe data available for this segment (Nk soundings) fully agree withthe ARC that has the same label.

The smoothness term consists of contributions from all spatiallyneighboring segments, favoring spatially close segments that havesimilar characteristics. Two segments sharing a boundary and havingthe same label have zero contribution to the smoothness term. If thelabels are different, the contribution depends on the actual meaningof these labels, and this in turn depends on the purpose or goal of thecharacterization. It is intuitively clear that if two neighboring seg-ments are labeled such that their expected physical parameters differinsignificantly, the corresponding contribution to the smoothnessterm must be small. For example, two neighboring segments labeled‘‘muddy silt’’ and ‘‘silty mud,’’ respectively, should have a relativelysmall contribution to the smoothness term. As for other possiblecombinations, it may not be as straightforward as the case above. Forexample, a rocky outcrop segment may happily neighbor the segmentwith sand or mud. The boundary between these segments should beclearly visible on a backscatter mosaic. This sort of neighborhoodmust not be penalized despite significant difference in labels. How-ever, in our experiments we have not specified any rules whichwould depend on known (or expected) geomorphology of thesurveyed area.

For the proof of concept we have been utilizing only onephysical parameter, specifically the mean grain-size j. The largerthe difference between expected mean grain sizes of two neigh-boring segments, the larger the corresponding contribution to thesmoothness term. Grain-size distribution is one of the mostimportant descriptors of seafloor characteristics with mean grainsize (e.g., Wentworth, 1922) being one of the more commondescriptors used. For example, very fine sediments, like clays andsilts, have small grain sizes (from 0.06 to 62 mm); sand may havegrain sizes from 62 mm (very fine sand) to 2000 mm (coarse sand).Gravel corresponds to even larger particles. In this work we

ordered catalogue entries according to ascending mean grain-sizediameter.

The choice of the specific function used for mapping of meangrain-size difference to the numerical value of the contribution(data part of the functional) was found not to be critical to thefinal result. We have experimented with monotonic functionsvarying from logarithmic to weak exponential with nearly thesame final delineation of boundaries.

To make use of combinatorial optimization techniques it isnecessary to choose a set of labels that are going to be assigned tomosaic segments—the catalogue. This choice depends on theproblem that needs to be solved. In the case of the binarydelineation like the presence or absence of a particular species,the first label (‘‘Yes’’) corresponds to a facies which is likely to beinhabited by a studied organism, and the second label (‘‘No’’) tothe other facies. To give an example, consider Atlantic sea scallops(Placopecten magellanicus). It is known that these species prefer toinhabit areas covered by pea-size gravel (Smith et al., 2006). Toconstruct a catalogue for delineation of potential scallop habitatwe will need at least three entries: one that corresponds to theoptimal conditions for scallops, and two entries defining upperand lower bounds in the grain-size sense. The result of optimiza-tion will be the delineation of the survey area into segmentsmarked with the first label that indicates potential scallop habitatand those marked with two other labels—areas unsuitable forscallops (too small or too large grain sizes). Note that three entriesin the catalogue are needed because they are organized in a grain-size ascending order, and the ‘‘too small’’ entry cannot becombined with the ‘‘too large’’ one.

The user may choose classes manually, for example, followingthe classification from Krumbein and Sloss (1963) or any othergrain-size classification scheme. More often, however, seafloormappers may not have a clear idea of the content of the surveyedseafloor, and the choice of labels is pretty much arbitrary. In thiscase the completely automated choice of a catalogue is desirable.

Success of delineation crucially depends on the choice ofcatalogue entries, as the proposed algorithm does not provide acomplete set of physical parameters for a theme, but only selectsthe most probable entry from the catalogue. The coalescence processcurrently uses only one physical parameter of the seafloor—averagegrain size, because we consider it the most important. This restric-tion is used for simplicity only, as we only tried to prove theconcept. If this single-parameter approach proves to be insufficientfor successful delineation, the smoothness cost may be madedependent on as many parameters as needed (label, currentlyassigned to a segment, determines the full set of parameters usedin Jackson’s model (Jackson et al., 1986)).

It is important to note that the coalescence procedure usesonly original backscatter data and not values from the mosaicwhich can be normalized to a common grazing angle or even be aresult of interpolation.

4. Construction of the catalogue

We propose three different methods for the automatic construc-tion of a catalogue. The first (naive) approach is to generate the label(or theme) catalogue by quasi-random sampling of the physicalparameter space (parameters from a mathematical model of acous-tic backscatter, for instance, acoustic impedance, roughness, etc.).Each point in this space corresponds to a unique ARC. In thisapproach, however, any prior knowledge about the survey area isignored. Some themes from the catalogue will not be used at all;others may not well represent ARCs of existing segments.

A more sophisticated second approach requires the creation ofa full 2D (backscatter strength versus grazing angle) histogram for

Page 5: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Y. Rzhanov et al. / Computers & Geosciences 41 (2012) 181–187 185

all the survey data (Fig. 3A). This histogram lacks spatial informa-tion, but shows the range of ARCs that can be attributed to thesegments. In case of a homogeneous area a 2D histogram uniquelydefines an ARC (noise aside). When an area can be described by anumber of different ARCs, like in Fig. 5.8(d) of Parnum ’s Ph.D.dissertation (2007), the corresponding 2D histogram would be arandom sampling of these ARCs. Dominant ARCs would lead todenser areas in the histogram. Using an acoustic backscattermodel we generate a ‘‘dictionary’’ of the themes—similar to thecatalogue in the first approach, but a much more extensive one,comprising several thousand labels (types of ARCs). In order tochoose the required number of labels from the dictionary, weemploy the matching pursuit strategy (Mallat and Zhang, 1993).We consider a 2D histogram to be a sampled version of a sum oflarge number of ARCs. Histogram density reflects weights ofcertain ‘‘atoms.’’ This expansion is not unique, but our task is toconstruct an overcomplete catalogue of nonorthogonal functions,only few of which will be represented in a final result. First we findan ARC covering the most densely populated part of the 2Dhistogram (the most dominant). This ARC corresponds to a labelrepresenting the first entry in the theme library. Data correspondingto this ARC are excluded from the 2D histogram, and the process isrepeated until the required number of entries is obtained (Fig. 3B).The remaining data points we consider as representing either noiseor backscatter from insignificant facies.

Fig. 3. (A) 2D histogram for the whole dataset. (B) Extraction of first four most

significant ARCs (red curves). Note that histograms in parts A and B have slightly

different density-to-greyscale mapping, and images in part B show density of

soundings at oblique angles in more detail. (For interpretation of the references to

color in this figure legend, the reader is referred to the web version of this article.)

Fig. 4. (A) Coarse segmentation of the original mosaic. Large segments are construc

dominant facies. (B) Average ARCs (red curves) and fitted ‘‘ideal’’ ARCs (blue curves) f

nevertheless allows for extraction of the dominant ARC.

This approach has advantages in comparison with the first one,because the choice of catalogue entries is not completely arbitrary,but is driven by the collected data. It is also important to note thatnumber of entries in the catalogue is not limited, and it is possibleto choose virtually all ARCs covering the data points in 2Dhistogram. Those ARCs that do not have support in the originaldata will not appear in a final result.

The third approach utilizes some coarse segmentation (from asegmentation pyramid noted above) of the backscatter mosaic(gray-scale image) into roughly as many segments as expectedentries in the catalogue. These large segments cannot be consid-ered homogeneous. Nevertheless each has some dominant themewhich can be extracted from the accompanying noise (contribu-tions from all other facies present in this segment) for inclusion inthe catalogue (Fig. 4).

Each entry in the catalogue is characterized by a set of physicalparameters and a complete ARC. Each label is associated with atraditional description based on grain size, accepted among themarine geologists and sedimentologists, such as ‘‘muddy silt,’’‘‘coarse sand,’’ ‘‘rocky outcrop,’’ etc. (see Table 1).

A practical strategy for the catalogue construction would be asfollows: (1) The catalogue entries should be chosen manually ifthe immediate goal of the delineation is obvious (as in mappingscallops’ habitat on the seafloor). (2) If delineation of severaldominant facies is necessary, we recommend the coarse segmen-tation approach. Dominant facies are the most represented in thespecified survey area. Exact meaning may depend on the parti-cular goal of the researcher.

5. Algorithm

The shape of the segments and the way they are positioned withrespect to each other may affect the corresponding components ofthe smoothness term. For example, long narrow segments sharinglong boundaries are more likely to have close labels than the samesegments sharing boundaries of a short extent.

All these considerations are included as free parameters in thecalculation of the smoothness term. Another free parameter,the coalescence parameter, is the most important of all. It weighs

ted using only partial information and may have either a different or the same

or themes 1, 2, and 9. Note that theme 9 demonstrates strong heterogeneity but

Page 6: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Table 1Properties of catalogue entries obtained from coarse segmentation.

Theme

number

Grain size

(j scale)

Height variance

(cm2)

Description

1 �1.00 6.70 Gravel

2 2.66 4.53 Silty sand

3 �1.00 6.70 Gravel

4 �1.00 6.70 Gravel

5 3.63 3.63 Very fine sand

6 �1.00 6.70 Gravel

7 2.84 5.33 Silty sand

8 3.36 5.56 Muddy sand

9 3.55 4.57 Very fine sand

10 2.62 6.10 Silty sand

11 2.42 6.70 Fine sand

Fig. 5. Result of final labeling. Original backscatter values shine through same

colors as used in coarse segmentation. Number of entries in the catalogue is equal

to number of coarse segments (in this case). Hence number of labels in the final

result is always equal or less than number of coarse segments. (A) Stanton Banks,

(B) HARS.

Y. Rzhanov et al. / Computers & Geosciences 41 (2012) 181–187186

the relative importance of the smoothness term as compared tothe data term. When this parameter is set to zero, each segment isconsidered independently of other segments, even if they areadjacent. Setting it to infinity assigns all segments the same label,implicitly assuming that all the survey area is acousticallyhomogeneous.

Our experiments have shown that there is no ‘‘optimal’’ valuefor the coalescence parameter. Results obtained with this para-meter varying over a relatively wide range were all accepted byhuman interpreters as reasonable (Fonseca and Rzhanov, 2008). Ifsome ground truth information is available (e.g., grab samples,seafloor photographs) then it makes sense to choose the para-meter that corresponds to the result best in agreement with theground truth (Fig. 5).

Ground-truth information is relatively straightforward toinclude in the algorithm described above. Each grab sample islinked to a certain segment (via its geographic location) and acertain label (via the sample content). This segment is assignedthe determined label and is excluded from the optimizationprocedure.

Due to the scarcity of ground-truth information, we have con-ducted a study comparing the results of automated delineation withthose performed independently by a number of interpreters (Fonsecaand Rzhanov, 2008). The general conclusion was that the automaticprocedure mostly agreed with the human interpretation. The dis-agreement was detected in several areas—each was investigatedseparately and it was confirmed that the automated procedure wasmore accurate. The main reason for discrepancies was the inability ofthe backscatter mosaic to convey all the complexity of the

corresponding ARC. The recent paper by Hamilton and Parnum(2011) considers a full set of ARCs and clusters them in anappropriate space. This approach shows good potential but assumesthat the insonified area is homogeneous on at least half-swath scale,while the method proposed in this paper has no such limitation.

In addition to the natural statistical distribution of backscatterstrength (Rayleigh, Rice, or other), acoustic backscatter is inherentlycontaminated with noise (Lurton, 2002). This becomes clear when2D histograms (as described above) for individual segments areconstructed, instead of using averaged ARCs. These 2D histogramsnormally appear as clouds of points scattered in ‘‘backscatterstrength vs grazing angle’’ plane. Unfortunately there are no meansto determine definitively whether this dispersion is due to ‘‘natural’’causes (statistical distribution and noise) or inhomogeneity of theseafloor composition. Within the framework of this study we assumethat all segments are homogeneous, so that they can be character-ized by a unique set of physical parameters, and all the dispersion inthe sample population is due to inherent backscatter noise.

Real surveys often have areas with gradual change in content,without sharp pronounced boundaries. These situations can bebetter dealt with by the automatic delineation procedures thanmanual ones—when the same set of labels is used, the boundariesbetween segments are always drawn in the same place.

The above-described approach deals mainly with ARCs whichare measurable quantities and are inherent properties of theseafloor (Fonseca and Mayer, 2007). Physical parameters (suchas grain size) come into play only in calculation of the smoothnessterm (quantification of difference between catalogue entries) andinterpretation of catalogue entries. Obviously for many faciesthere are no adequate scattering models and the very term ‘‘grainsize’’ does not make sense (for example, vegetation or rhodolithbeds). These cases require modification of the proposed algorithm(introduction of different types of descriptors) which will be thesubject of further research.

6. Conclusions

The paper reports the approach for automatic delineation ofacoustic facies based on all available MBES data. As an initial step,a catalogue of expectedly dominant acoustic facies is created,where each entry corresponds to an ‘‘ideal’’ angular responsecurve. Each relatively small segment of a backscatter mosaic isthen assigned a label (entry) from this catalogue. The labeling isobtained by finding a solution for the optimization problem,which makes use of all collected MBES data, and not only partof data used for construction of the mosaic.

From our perspective the proposed technique is applicable toalmost any survey data. If the area is completely homogeneous,the result will reveal the only dominant ARC and correspondingset of physical parameters (if Jackson’s model (Jackson et al.,1986) is applicable for this type of facies or if ground-truth dataare available). If the area is heterogeneous (with typical sizes ofhomogeneous areas much larger than sounding footprint), thetechnique will delineate acoustic themes at least not worse thanan experienced human interpreter (Fonseca and Rzhanov, 2008).

The reported approach showed promising results, which arecomparable to those obtained by human experts. Moreover, theautomatic results showed better spatial resolution and moreaccurate results in certain special areas. Examples of these areashave been presented in Fonseca and Rzhanov (2008). Usuallythese are the areas where the mosaic (which is a ‘‘many-to-one’’mapping) fails to convey all the angular information collectedduring the survey. To the best of our knowledge this is the firsttime that combinatorial optimization technique has been applied

Page 7: Construction of seafloor thematic maps from multibeam acoustic backscatter angular response data

Y. Rzhanov et al. / Computers & Geosciences 41 (2012) 181–187 187

to the problem of delineation and interpretation of acoustic faciesfrom multibeam echosounders backscatter.

Acknowledgments

The authors thank Prof. B. Calder for valuable discussions andall the experts who participated in the experiment of visualinterpretation and manual delineation of the data. This workwas supported by NOAA Grant NA05NOS4001153. The authorsthank the crew and scientific personnel of the RV Corystes andCeltic Explorer, for data collection at Stanton Banks, which wasconducted under the Interreg IIIB Mapping European SeabedHa-bitats (MESH) Project and also SAIC for providing the multibeamdataset collected at the HARS site.

References

Boykov, Y., Kolmogorov, V., 2004. An experimental comparison on min-cut/max-flow algorithms for energy minimization in vision. IEEE Transactions on PAMI26, 1124–1137.

Boykov, Y., Veksler, O., Zabih, R., 2001. Fast approximate energy minimization viagraph cuts. IEEE Transactions on Pattern Analysis and Machine Intelligence23, 1222–1239.

de Moustier, C., 1988. State of the art in swath bathymetry survey systems.International Hydrographic Review 65, 29–38.

Deng, Y., Manjunath, B.S., 2001. Unsupervised segmentation of color-textureregions in images and video. IEEE Transactions on Pattern Analysis andMachine Intelligence 23, 800–810.

Fonseca, L., Brown, C., Calder, B.R., Mayer, L.A., Rzhanov, Y., 2009a. Angular rangeanalysis of acoustic themes from Stanton Banks, Ireland: a link between visualinterpretation and multibeam echosounder angular signatures. Applied Acous-tics 70, 1298–1304.

Fonseca, L., Calder, B., 2007. Clustering acoustic backscatter in the angularresponse space. In: Proceedings, US Hydro Conference.

Fonseca, L., Calder, B.R., 2005. Geocoder: An efficient backscatter map constructor.In: Proceedings of the U.S. Hydrographic Conference, San Diego, CA, USA,29–31 March.

Fonseca, L., Mayer, L., Orange, D., Driscoll, N., 2002. The high-frequency back-scattering angular response of gassy sediments: model/data comparison fromthe Eel River margin, California. Journal of the Acoustical Society of America111, 2621–2631.

Fonseca, L, Mayer, L.A., 2007. Remote estimation of surficial seafloor propertiesthrough the application angular range analysis to multibeam sonar data.Marine Geophysical Researches 28, 119–126.

Fonseca, L., Rzhanov, Y., 2008. Automatic construction of acoustic themes frommultibeam backscatter data. In: Proceedings of Shallow Survey Conference,Newcastle, NH.

Fonseca, L., Rzhanov, Y., McGonigle, C., Brown, C., 2009b. Automatic constructionof acoustic themes for benthic habitat mapping at Stanton Banks, UK.In: GeoHab’09 Conference Proceedings, Trondheim, Norway.

Greene, H.G., Bizzarro, J.J., O’Connell, V.M., Yoklavich, M.M., Brylinsky, C.K.,Reynolds, J., 2007. Application of a bottom-up marine benthic habitat schemealong the west coast of the United States. In: Proceedings of Coastal Zone’07,Portland, OR.

Greene, H.G., Yoklavich, M.M., Starr, R.M., O’Connell, V.M., Wakefield, W.W.,Sullivan, D.E., McRea Jr., J.E., Cailliet, G.M., 1999. A classification scheme fordeep seafloor habitats. Oceanologica Acta 22, 663–678.

Hamilton, L.J., Parnum, I., 2011. Acoustic seabed segmentation from directstatistical clustering of entire multibeam sonar backscatter curves. ContinentalShelf Research 31, 138–148.

Hughes Clarke, J.E., 1993. The potential for seabed classification using backscatterfrom shallow water multibeam sonars. In: Pace, N., Langhorne, D.N. (Eds.),Acoustic Classification and Mapping of the Seafloor, Proceedings of theInstitute of Acoustics, 15, pp. 381–388.

Hughes Clarke, J.E., 1994. Toward remote seafloor classification using the angularresponse of acoustic backscattering: a case study from overlapping GLORIAdata. IEEE Journal of Oceanic Engineering 19, 112–127.

Hughes Clarke, J.E., Danforth, B.W., Valentine, P., 1997. Areal seabed classificationusing backscatter angular response at 95 kHz. NATO SACLANTCEN ConferenceProceedings Series CP-45, High Frequency Acoustics in Shallow Water, Lerici,Italy, pp. 243–250.

Hughes Clarke, J.E., Iwanowska, K.K., Parrott, R., Duffy, G., Lamplugh, M., Griffin, J.,2008. Inter-calibrating multi-source, multi-platform backscatter data sets toassist in compiling regional sediment type maps: Bay of Fundy. In: Proceed-ings of Canadian Hydrographic Conference and National Surveyors Conference.

Jackson, D.R., Winebrenner, D.P., Ishimaru, A., 1986. Application of the compositeroughness model to high-frequency bottom backscattering. The Journal of theAcoustical Society of America 79, 1410–1422.

Krumbein, W.C., Sloss, L.L, 1963. Stratigraphy and Sedimentation, 2nd edn. W.H.Freeman and Company, pp. 159, 660 pp.

Lurton, X., 2002. An Introduction to Underwater Acoustics: Principles andApplications. Praxis Publishing, Chichester, UK, 680 pp.

Mallat, S.G., Zhang, Z., 1993. Matching pursuits with time–frequency dictionaries.IEEE Transactions on Signal Processing 41, 3397–3415.

Pace, N.G., Gao, H., 1988. Swathe seabed classification. IEEE Journal of OceanicEngineering 13, 83–90.

Parnum, I. M., 2007. Benthic habitat mapping using multibeam sonar systems.Ph.D. Dissertation, Curtin University of Technology, Perth, Australia, 213 pp.

Prasad, L., Skourikhine, A.N., 2005. Vectorized image segmentation via trixelagglomeration. In: Proceedings of the 5th IAPR Workshop on Graph-basedRepresentations in Pattern Recognition, Poitiers, France, LNCS, vol. 3434,pp. 12–22.

Schimel, A., Rzhanov, Y., Fonseca, L., Mayer, L., Healy, T., Immenga, D., Unsuper-vised acoustic seabed classification using both angular and spatial informationfrom multibeam backscatter data, in preparation.

Schrader, A., 1998. Evolutionare Algorithmen zur Farbquantisierung und asymme-trischen Codierung digitaler Farbbilder. Wissenschaftlicher Verlag, Berlin(in German).

Smith, S.J., Costello, G., Kostylev, V.E., Lundy, M.J., Todd, B.J., 2006. Application ofmultibeam bathymetry and surficial geology to the spatial management ofscallops (Placopecten magellanicus) in southwest Nova Scotia. 15th AnnualPectenid Workshop. Journal of Shellfish Research 25, 308.

Soares Rosa, L., 2007. Seafloor characterization of the historic area remediation siteusing angular range analysis. M.Sc. Thesis in Ocean Engineering, University ofNew Hampshire.

Wentworth, C.K., 1922. A scale of grade and class terms for clastic sediments.Journal of Geology 30, 377–392.