SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING Seabed type clusterng usng sngle-beam echo sounder tme seres data PETER HUNG, SEÁN MCLOONE Department of Electronc Engneerng, StratAG Natonal Unversty of Ireland Maynooth Maynooth, Co. Kldare IRELAND {phung, sean.mcloone}@eeng.num.e http://www.stratag.e XAVIER MONTEYS Geologcal Survey of Ireland Beggars Bush Haddngton Road, Dubln IRELAND Xaver.Monteys@gs.e http://www.gs.e Abstract: - Acoustc seabed charactersaton s known to make avalable very useful nformaton for varous marne related applcatons. Most exstng sngle-beam echo sounder (SBES) classfcatons requre ground truthng n order to have confdence n the fnal results and sometmes to relate labels to geologcal features. A statstcal feature extracton approach, wth an emphass on qualty control, s proposed as an alternatve to extensve ground truthng. Survey data collected from Maln Sea durng 2003 s used as a test bed. Prncpal component analyss (PCA) and k-means as well as qualty threshold (QT) local clusterng are employed to label the echo return samples wth an arbtrary number of classes. Results show that proper qualty control enables more geologcally realstc seabed type clusterng. Key-Words: - sngle-beam; echo sounder; seabed; clusterng; PCA; k-means; qualty threshold Introducton Charactersaton of marne seabed floor s known to make avalable very useful nformaton for varous applcatons such as, seabed mappng [], mpact assessment of human actvty [2], nature conversaton [3] and underwater plant speces classfcaton [4]. The tradtonal way of ganng knowledge on sedment type by corng, grab samplng and vsual nspecton s mpractcal due to the tme and cost nvolved. In addton, the spatal dstrbuton of mud, sands and gravel derved from the lmted amount of samples may be msleadng, especally n survey areas where sedment patterns are too complex [5]. The dffculty of drect seabed access, even n shallow regons, means that most seabed surveys are carred out remotely. By far the most popular and wdely used technque n seabed-related remote sensng s echo sonar, based on the measurement of sound energy usng a sonar transducer. Early applcatons of sonar nclude bathymetry or the measurement of seabed depth, based on echo return tmngs of a transducer that emts one sound wave. Ths s referred to as a sngle-beam echo sounder (SBES). Nowadays, most seabed type charactersatons are based on classfcatons of dgtal data from mult-beam echo sounders (MBES), where multple sounders are used smultaneously to allow a greater area of coverage durng seabed surveyng. Although SBES tme seres potentally contan valuable echo data n yeldng classfcaton nformaton, t typcally receves less attenton because t s not desgned for that purpose. Ths paper dscusses the vablty of such an approach. The paper s structured as follows. Secton 2 descrbes exstng approaches to SBES seabed classfcaton. The proposed approach, whch puts an emphass on qualty control, s detaled n Secton 3. Secton 4 then presents some results usng survey data from Maln Sea, Ireland, as a test bed. Fnally, conclusons are dscussed n Secton 5. 2 Background on SBES Classfcaton SBES echo tme seres data s collected durng a geologcal sea trp and analysed afterwards. Exstng SBES classfcaton nvolves data analyss from the dgtal acoustc data, wth or wthout hardware adustments at the tme of capture. In contrast to bathymetry, the data processng s more nvolved when the obectve s seabed type classfcaton. To acheve relable and consstent classfcaton results, each data processng step has to be carefully assessed by experenced marne geologsts. The task s complcated by the fact that the marne echo sgnal s nfluenced by varous factors, ncludng salnty, sea temperature, depth and slope. In addton, echo return characterstcs are frequency-dependent. The effect of depth has to be compensated n areas wth the same seabed type []. ISSN: 792-5088 308 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING Echo data from transcevers operatng at dfferent frequences needs to be consdered separately. Approaches to tackle the challenges nclude calbrated acoustc mpedance matchng [6], model nverson [] and recent statstcal methods, such as RoxAnn TM [7] and QTC Impact TM [9]. RoxAnn TM use an echo-ntegraton methodology to derve feature values from the tal secton of echo tme seres. QTC Impact TM uses only the frst part of the echo returns to extract features usng prncpal component analyss (PCA) followed by k-means clusterng. Both methods suffer from the nosy nature of sonar sgnals [8,0], contrbutng uncertanty to seabed classfcaton. 3 Proposed Approach The obectve of the proposed approach s to perform seabed type clusterng as an alternatve to extensve ground truthng, pror to label valdaton. In statstcal learnng, ths s equvalent to unsupervsed seabed type clusterng, whch does not nvolve the use of labelled samples pror to classfcaton. Fgure shows a flowchart of the overall process. Fgure. SBES data processng for seabed clusterng. 3. Geologcal and Overall Consderatons In prncple, t s the echo tme seres nterval mmedately beneath the detected seabed that s the best canddate for seabed classfcaton, due to the strong sgnal return and ease of detecton. The data processng exercse assumes that the rate of seabed type varaton s much smaller than the spatal rate of change, whle the characterstcs of dfferent seabed types are captured only n a lmted secton of the echo return sgnal. To acheve relable and consstent classfcaton results, each data processng step needs to be carefully assessed by experenced marne geologsts. Due to the hgh dmensonalty of tme seres data, the large number of samples collected n sea trps and the heteroscedastc nose contaned wthn, features should frst be extracted from the raw echo data. However, snce SBES contans less geologcal nformaton and redundancy for qualty assurance compared to MBES, addtonal measures need to be taken to ensure the qualty of the raw data before feature extracton. Here a sem-automatc statstcal approach along wth an mproved qualty check s proposed whch s descrbed below. It s common practce to enable automatc gan control n the echo sounder to acheve maxmal sgnal-to-nose rato wth varyng seabed stuatons. Other external factors such as drecton of shp movement, roll and ptch moton, cruse speed and weather condtons also have a sgnfcant nfluence on the SBES transcevers and cannot be gnored. Consequently, the ampltude of the echo envelopes cannot be reled upon and, hence, feature spatal normalsaton s recommended before classfcaton s performed. 3.2 Sonar Pre-processng The pre-processng stage nvolves further expert nspecton for data ntegrty, ntal cleanup to mtgate the effects of systematc devatons of sonar measurement, ncludng tdal movement, decbel normalsaton and converson to ndustry-standard format for compatblty and storage purposes, Next, spatal samplng s carred out to reduce data sze, followed by seabed depth determnaton targeted for feature extracton. In ths case, the am of dentfyng seabed depth s to facltate feature extracton by segmentng the sonar data nto two sectons, above and below seabed. It s noted that, unlke the tradtonal SBES applcaton, accurate bathymetry mappng employng specalsed flterng technques and consstency crtera are not sutable for seabed type clusterng. Nonetheless, proper calbraton of the locaton of the seabed surface s mportant n order to reduce the bas on the resultng feature extracton. Two bathymetrc approaches are dscussed. The frst method assumes the seabed to be located at the peak ampltude of each echo return from the lowest frequency sonar. However, small but regular bathymetrc fluctuatons (~ ±5 m) are commonly observed n ths data, possbly caused by the nherent measurement nose, whch are unlkely to be geologcal artefacts. In the second method, the locaton of peak ampltudes are spatally smoothed usng a second-order Butterworth low-pass flter. Method two was found to be a good compromse between mantanng both smplcty and statstcal consstency. Even wth ths pre-processng, stuatons exst where corrupted samples wth abnormal changes n bathymetry reman, causng possble data nconsstency. An automatc procedure to dentfy ISSN: 792-5088 309 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING such bad samples was developed as follows. Defne each sonar tme seres pror to pre-processng as column vector z where z = { }, = {, 2, K, m}, k = {, 2, K, n} () z k m s the number of spatal data samples (.e. number of tme seres collected), n s the number of (temporal) samples n each tme seres and z k s the kth temporal png of the th survey sample. The resultant matrx comprsng all spatal samples can then be wrtten as Z = { z}, ~ Z = { ~ z }, = {, 2, K, m} = {, 2, K, m} where ~ s used to denote data derved from spatally fltered seabed estmates. Ten-fold stackng s performed on the echo return data to mprove the sgnal-to-nose rato, at the expense of the spatal resoluton of the classfcaton, that s (2) 0 Y = { y } = z+, = {,0,20, K, m /0} 0 (3) = The fltered verson Y ~ s obtaned n a smlar fashon. It s essental to algn seabed pngs to the same matrx column for consstent feature extracton. To ndcate the segment of echo return used, the conventon A_B_ s employed. For example, A5B25 represents 5 m above and 25 m below the seabed. As such, the echo segment a m above and b m below the seabed (AaBb) from each tme seres n the Y matrces are grouped together to form correspondng X famly matrces wth X = { x} = YA abb, X Y, = {, 2, K, m} ~ ~ X = { ~ (4) x } = Y AaBb x = { }, k = {, 2, K, n'}, n' n (5) x k < where n' s the number of temporal samples n the tme seres subset. Note the seabed s algned to the same tme seres echo png n both X matrces. Defnng the varaton n mean spatally fltered echoes, v, as v = v } = {E( ~ x ) E( x )} (6) { where E s the expectaton operator, the followng automatc outler decson rule can be employed ' bad' f v > ασ ( v) b =. (7) 0 ' good' otherwse Here σ(v) s the standard devaton of all v on the same survey track and α s a threshold on the sze of the bad data regon. In other words, any tme seres sample vector found outsde α tmes the standard devaton of v wll be labelled as bad. For each echo sounder frequency, bad data detecton s performed usng the segments AaB0, A0Bb and AaBb, returnng a seres of Boolean vectors b = { }, k {AaB0, A0Bb, AaBb}. (8) k b = A bad echo sample s then defned as one wth at least one b = n any b. The three data subsets are selected to mprove the robustness of the outler detecton process whch s senstve to the locaton of the outlers relatve to the seabed peak. 3.3 Feature Extracton For each tme seres, a total of four statstcal features wll be used to capture the quanttatve echo characterstcs of X ~ for clusterng purposes. These are descrbed below. 3.3. Mean and standard devaton The temporal mean of the echo tme seres (m) s one of the most mportant statstcal measures n acoustc feature extracton. The nfluence of dfferent types of sedments on echo seres wll nevtably be reflected n the average echo return value. It s related to the mpedance contrast at the seafloor n geology. Smlarly, the assocated standard devaton (σ) of sonar tme seres usually contans nformaton that has drect relatonshp wth the seabed geology [] and tends to relate to the seabed roughness. 3.3.2 Measure of randomness Whle both mean and standard devaton are useful features that should be extracted, ther applcablty s generally restrcted to the statstcal propertes wthn an echo tme seres. In order to convey nformaton about the relatonshp between adacent tme seres, a measure of randomness (s) s proposed usng the standard devaton of the autocorrelaton of the echo dfference data, defned as ISSN: 792-5088 30 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING u d m x x = σ ( x ) σ ( x = m λ= m a = { u m d d a s( X) = σ a, λ..., u ), u,..., u m where d s the row vector of the dfference of two adacent echo tme seres, λ s the tme seres delay durng convoluton u, a s the modfed autocorrelaton row vector of d as dscussed below. To ensure that the autocorrelaton s unaffected by the magntude of the data, each correlaton sequence s normalsed by the largest value n the sequence as returned by the nfnty norm. It has been found that the randomness measure wll be more representatve f the largest value of a conventonal verson of a s removed before calculatng the standard devatons of the autocorrelaton sequence. In theory, a value of unty can only be found at the no lag pont (u 0 ) n the autocorrelaton unless the sequence s perfectly auto-correlated to tself. 3.3.3 Measure of correlaton nose The randomness measure s appled on adacent samples only. To establsh the relatonshp between neghbourng echo tme seres, a measure based on the sgnal-to-nose rato of the mean correlaton coeffcent nose (c) s proposed 0 r0( x ) = r( x, x+ ) 0 = r ( ~ 0 x ) r0( x ) c( X) = log r ( x ) 0 } (9) (0) where r s the correlaton coeffcent between two tme seres and r 0 s the average correlaton from 0 adacent tme seres. Correlaton nose s defned as the dfference between the r from the orgnal and temporally fltered verson of x, denoted wth superscrpt ~. It s another measure that quantfes the amount of dfferences between consecutve samples based on correlaton. However, t should convey more nformaton n a spatal context. 3.3.4 Optmal depth nterval Choosng the correct tme seres nterval for feature extracton has mportant mplcatons on classfcaton performance. Too lttle and the feature loses certan dscrmnatng nformaton. Too much and the feature would be polluted by unwanted measurement nose and sdelobe backscatterng. One possblty s to determne the optmal range by gradually ncreasng the nterval depth and lookng for sudden and unreasonable change n the values of features, prmarly tme seres mean (m) due to ts clear geologcal relatonshp. Another possblty s by calculatng the adacent mean-square-error (AMSE) of the mean feature from consecutve ntervals of both the above and below seabed sectons, defned as follows. Denote m( a, b) = m( X A abb ), X AaBb X () where a and b represents the upper lmt of above seabed and lower lmt of below seabed depths, respectvely. A number of depth lmts are chosen that are approxmately evenly dstrbuted across the range of seabed depths consdered. Then, the AMSE, e, can be defned as e = S = { e } M = M M { S } = {[ a ; b] } = [ m( S ) m( S = n( S ) + )] 2 M = (2) where M s the number of spatal tme seres depth ntervals selected for optmal depth range determnaton and S s the augmented matrx storng the chosen above and below seabed depths. The value of e s normalzed by n(s ), the number of temporal tme seres measurements contaned n a partcular echo segment. As the tme seres nterval ncreases, the nfluence of nose wll gradually domnate the mean feature. Therefore, the optmal depth (S opt ) s located ether at a local mnmum or pont of nflecton of e, S opt = arg mn e = { S : e = nf e} (3) such that S opt contans the maxmal amount of nformaton for seabed type clusterng and mnmum amount of unwanted nose that mght dstort the fnal results. 3.4 Feature Clusterng 3.4. PCA and k-means Prncpal component analyss (PCA) and k-means are the de facto standard commonly employed n ths partcular applcaton both n research and commercal products. It s known that PCA facltates k-means by proectng to the subspace ISSN: 792-5088 3 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING where k-means clusterng le [2]. However, k- means requres the number of clusters to be specfed a pror, whch s unknown n ths case. Also, t s less effectve n dealng wth non- Gaussan data and may not return meanngful results for rregular clusters. 3.4.2 QT local Qualty threshold (QT) s a densty based clusterng algorthm. Frst proposed by Heyer [3], t has several advantages over the k-means, ncludng automatc cluster number determnaton and the ablty to deal wth rregularly shaped data dstrbutons. QT local s a closely-related varant of QT that dffers by how the qualty parameter s specfed. Whle conventonal QT defnes the qualty parameter usng the maxmum dameter of clusters, QT local uses the longest dstance between any two samples wthn a cluster. 4. Pre-processng Based on geologcal and echo sounder techncal consderatons, the secton of the echo returns selected for analyss was lmted to 5 m above and 25 m below the seabed. The fltered bathymetry s spatally smoothed usng a second-order Butterworth low-pass flter at a normalsed cut-off frequency of 0.. Fgure 3 shows a typcal example of a detected bad sample usng α=3.5 overlad on a good sample. Overall, a total of 35 unque bad samples (.98%) were found n the survey lne whch conssted of a total of 6800 samples, as shown n Fg. 3. Echo return (db) -0 - -0 0-0 Peak ampltude Bad bathymetry Good bathymetry 4 Results Data were collected by Geologcal Survey Ireland (GSI) durng a seres of Maln Sea survey n the year 2003. The man equpment used s the Kongsberg EA600 sngle-beam echo sounder [4], whch s capable of smultaneously emts and records echo returned at sonar frequences 2, 38 and 200 khz, respectvely. Maln Sea (Fg. 2) s located to the north of the Republc of Ireland and has been chosen as a case study due to ts known seabed type varatons. Fgure 2. SBES Locaton of survey tracks (red) n Maln Sea, Ireland. To llustrate the effect of optmal tme seres depth nterval, an example survey track, lne 63, s used throughout the secton. The track runs almost horzontally at lattude of 56N and covers a survey dstance of about 0 km. Seabed depth (m) -0 2-5 0 5 0 5 20 25 Dstance from seabed (m) Survey dstance covered: 09.6 km -80-90 seabed peak -00 fltered -0-20 -30-40 -50-60 -70-80 -9-8.8-8.6-8.4-8.2-8 -7.8-7.6-7.4 Longtude Fgure 3. Plots of temporal 2 khz echo sample returns. The dotted lne represents the seabed surface detected usng peak ampltude; fltered Maln sea survey lne Bathymetry marked wth bad samples. Both bathymetrc methods detect a number of bad samples that are only partally overlapped. The flterng bathymetry wll be the chosen method to segregate echo data nto above and below seabed segments. It s noted that only the lowest frequency s used for the depth determnatons. The ablty of the 2 khz sonar to penetrate deeper nto the seabed allows more relable detecton of abnormal data samples. A graphcal comparson between tme seres algned usng peak ampltude (X) and bathymetry flterng ( X ~ ) s shown n Fg. 4. Whle tme seres shft brought by bathymetrc spatal flterng only results n x-axs translaton, the fact that X ~ s obtaned by stackng 0 consecutve tme seres (yaxs averagng) means that X ~ s an approxmate, rather than an exact temporal translated X. ISSN: 792-5088 32 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING 0-0 -20 A B peak fltered seabed Echo return (db) -30-40 -50-60 -70-80 -5 0 5 0 5 20 25 Dstance from seabed (m) 0-0 -20 A B peak fltered seabed mean Echo return (db) -30-40 -50-60 -70-80 -5 0 5 0 5 20 25 Dstance from seabed (m) Fgure 4. Typcal 2 khz, and; 200 khz sonar tme seres X before and after spatal flterng. 4.2 Feature Extracton and Clusterng The optmal depths at varous sonar frequences are shown n Fg. 5. The optmal depth selectons are dvded nto above and below seabed segments. 80 60 40 30 2 khz 38 khz 25 200 khz 2 khz 38 khz 200 khz std. dev. Mean Ad. MSE 20 00 80 60 20 5 0 40 20 5 0 0 0.5.5 2 2.5 3 3.5 4 4.5 5 Above seabed (m) 0 0 5 0 5 20 25 Below seabed (m) Fgure 5. Typcal 2 khz, and; 200 khz sonar tme seres X before and after spatal flterng. The results suggest that the optmum tme seres ntervals at lower frequences are larger, consstent wth the fact that lower frequency sonar penetrates deeper nto the sedments. To reduce hgh-frequency nose n feature space, each feature presented has been smoothed usng a second-order Butterworth low-pass spatal flter wth a normalsed cut-off frequency of 0.002. It s noted that the same type of flter s used n correlaton measure (Sect. 3.3.3) at a cut-off of 0.0. To mprove vsblty, all features are spatally normalsed such that they are zero-mean and have unt varance (Fg. 6). The clusterng results wth both algorthms nclude bad samples for completeness, although these samples can be easly marked as unclassfed n the fnal result. (c) (d) rand. corr. Fgure 6. Bathymetrc scatter plots of mean; standard devaton; (c) randomness, and; (d) correlaton features from 2 khz sonar at optmal depth segments and ther correlaton coeffcents wth bathymetry. ISSN: 792-5088 33 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING optmal maxmal Fgure 7. Geographcal scatter plots of 6-class prncpal component k-means usng optmal; and maxmal tme seres ntervals. The clusterng results obtaned usng PCA and k- means are vsualzed usng geographcal scatter plots as shown n Fg. 7. Dfferent colours are used to represent dfferent class labels. A k-means clusterng s performed on both subsets of data, one wth tme seres nterval qualty control and one wthout. An arbtrary number of classes (6) s chosen. The clusterng was repeated 0 tmes to avod poor results due to bad ntal centrods. In terms of the spatal dstrbuton of clusterng labels, t appears that those evaluated from optmally selected ntervals have more dstnctve clusters, whle clusters generated usng maxmal ntervals appear to be more cluttered. The spatal dstrbuton of classes from optmal ntervals s mostly contnuous, whch s geologcally acceptable. On the other hand, the classes from maxmal ntervals are dscontnuous, scattered around dfferent locatons and hence geologcally unrealstc. The spatal dstrbuton of clusters generated usng QT clusterng s shown n Fg. 8. The spatal dstrbuton of clusters generated usng QT clusterng s shown n Fg. 8. The results are generated for both optmal and maxmal tme seres ntervals usng a qualty threshold parameter of 0.055. In each case more than 250 clusters were returned. However, only the sx largest clusters are dsplayed wth the rest grouped as outlers (label ndex 7). The fgure suggests smlar qualty mprovement when features are extracted from optmally selected echo tme seres ntervals. The geographcal scatter plots of the clusterng classes hghlght the mportance of nterval qualty control durng seabed type clusterng. optmal maxmal Fgure 8. Geographcal scatter plots of 6-class QT local usng optmal; and maxmal tme seres ntervals. 5 Conclusons A statstcal feature extracton approach, along wth an emphass on qualty control, s proposed to mprove the accuracy of seabed type clusterng from sngle-beam echo sounder data. The two key aspects of qualty control nvolve bathymetrc flterng for more accurate seabed detecton, as well as the determnaton of optmal seabed depth ntervals for feature extractons. Data fuson on mutl-frequency SBES data s employed to mprove the nformaton rchness and stackng mproves the sgnal-to-nose rato of the orgnal data at the expense of spatal resoluton. Results suggest the use of data qualty control mproves the relablty of clusterng results and provde a more geologcally realstc nterpretaton of the survey area. It s planned to take account of the depth of the echo wave s propagaton n soft sedments when ISSN: 792-5088 34 ISBN: 978-960-474-233-2
SELECTED TOPICS n POWER SYSTEMS and REMOTE SENSING determnng the optmal seabed segment depth ntervals durng feature extracton and statstcal clusterng. Acknowledgement Research presented n ths paper s supported by a Strategc Research Cluster grant (07/SRC/I68) StratAG, awarded to the Natonal Centre of Geocomputaton by Scence Foundaton Ireland under the Natonal Development Plan. Ths proect s a part of the collaboraton of StratAG wth the Geologcal Survey of Ireland (GSI) and INFOMAR, who generously provded the SBES data. References: [] B. R. Bffard, S. F. Bloomer, N. R. Chapman, and J. M. Preston, Sngle-beam seabed classfcaton: drect methods of classfcaton and the problem of slope, Proc. Boundary Influences n Hgh Frequency, Shallow Water Acoustcs, Bath, UK, 2005, pp. 227-232. [2] P. D. Eastwood, C. M. Mlls, J. N. Aldrdge, C. A. Houghton, and S. I. Rogers, Human actvtes n UK offshore waters: an assessment of drect, physcal pressure on the seabed, ICES Journal of Marne Scence, Vol. 64, No.3, 2007, pp. 453-463, do:0.093/cesms/fsm00. [3] L. J. Hamlton, Acoustc seabed classfcaton systems, Australa: DSTO Aeronautcal and Martme Research Laboratory, 200, avalable at: http://catalogue.nla.gov.au/record/629875. [4] C. J. Brown, et al., Acoustc Seabed Classfcaton of Marne Physcal and Bologcal Landscapes, Copenhagen, Denmark: ICES. 2007. [5] K. E. Ellngsen, J. S. Gray, and E. Bornbom, Acoustc classfcaton of seabed habtats usng the QTC VIEW(TM) system, ICES Journal of Marne Scence, Vol. 59, No.4, 2002, pp. 825-835. [6] P. A. van Walree, M. A. Ansle, and D. G. Smons, Mean gran sze mappng wth snglebeam echo sounders, Journal of Acoustcal Socety of Amerca, Vol. 20, No.5, 2006, pp. 2555-2566. [7] Sonavson, RoxAnn Seabed Classfcaton System,Aberdeen, UK: Sonavson, 2004, avalable at: www.sonavson.co.uk/pages/ seabed_classfcaton_menu.html. [8] Y. Satyanarayana, S. Nathan, and R. Anu, Seafloor sedment classfcaton from sngle beam echo sounder data usng LVQ network, Marne Geophyscal Researches, Vol. 28, No.2, 2007, pp. 95-99. [9] Quester Tangent Corporaton, Operator s Manual & Reference. Seabed Classfcaton. BC, Canada: Quester Tangent Corporaton, 997. [0] M. Zmmermann and C. Rooper, Comparson of echogram measurements aganst data expectatons and assumptons for dstngushng seafloor substrates, Fshery Bulletn, Vol. 06, No.3, 2008, pp. 293-304. [] R. Chapman, S. Bloomer, X. Monteys, X. Garca, Applcaton of mproved sngle beam echosounder classfcaton and characterzaton methods to multfrequency INFORMAR data, Scentfc report INFORMAR research call, Dec. 2009. [2] C. Dng and X. He, K-means Clusterng va Prncpal Component Analyss, Proc. Int'l Conf. Machne Learnng (ICML 2004), pp. 225 232, July 2004. [3] L. J. Heyer, S. Kruglyak, and S. Yooseph, S., Explorng Expresson Data: Identfcaton and Analyss of Coexpressed Genes, Genome Research, Vol. 9, No., 999, pp. 06-5. [4] Kongsberg EA-600 echo sounder seres. Webste at http://www.km.kongsberg.com ISSN: 792-5088 35 ISBN: 978-960-474-233-2