SEAFLOOR PROPERTIES AND SEGMENTATION G. CANEPA, N. G. PACE, E. POULIQUEN, P. FRANCHI, R. LOMBARDI, C. SISTI NATO Undersea Research Centre Viale San Bartolomeo 400, I-19138 La Spezia, Italy E-mail: canepa@saclantc.nato.int T. AKAL TUBITAK-MAM, Marmara Research Center Earth and Marine Sciences Research Institute, P.K. 21 Gebze, Kocaeli 41470 Turkey The interest here is in using acoustic signals to divide the seabed into regions that are relevant in the context of MCM operations and extracting those features, which drive the process. This is termed segmentation followed by classification. A recently developed seabed segmentation algorithm is presented: SESAM (SEafloor Segmentation AlgorithM). It is based on the analysis of angular dependence of the acoustic backscatter. Such information is readily acquired with multibeam systems. The algorithm will be described together with illustrations of segmentation results from an area that has been very well characterized. Comparison of segmentations with ground truth data (grabs, penetrometers, stereo photography, sidescan sonar) is used to endorse the algorithm and confirm that the identified regions are related to seabed features of relevance. 1 Introduction We present further validation results of the SESAM algorithm 1,2 supported by acoustic and non-acoustic data acquired during the CLUTTER 03 cruise, off Marciana Marina, Elba Island, Italy. Data from a 120 beam SIMRAD EM3000 operating at 300 khz was used; contemporaneously, ground truth data were acquired using high frequency sidescan, grabs, expendable bottom penetrometers (XBP) and stereo-photos. When the segmentation maps produced by SESAM were compared with ground truth data there was good agreement. However, the Marciana Marina site presented three features that challenged the segmentation algorithm and are described in the following. Two important aspects of the segmentation algorithm have been improved. The accuracy of the Backscattering Strength (BS) data has been improved by calibrating the multibeam echosounder (EM3000) with respect to the transmit beam 2 and then taking into account the way the supplied echosounder software calculates the BS (the calculation technique changes when the beam-seafloor incident angle is, approximately, 30 ). The algorithm that filters the processed BS data has also been improved to allow the treatment of a high number of samples in a relatively short time while concurrently increasing the stability of the results. The following word definitions are used: Segmentation: separation into N categories of the studied area. Each category corresponds to a particular acoustical and/or geo-acoustical characteristics. paper: submitted to World Scientific on August 27, 2003 1
Figure 1. The segmentation of Marciana Marina obtained using SESAM is shown together with an artificially illuminated bathymetric map. With the help of ground truth information, it is possible to obtain a classification map of the site. Blue is used for muddy/silty classes, green for presence of sea grass while gravel/rocky classes are red. Bathymetric contour lines (yellow) are indicated. paper: submitted to World Scientific on August 27, 2003 2
Classification: assignment to each category a class name (e.g., clay, silt, sand, or soft, medium hard, hard, or low clutter, high clutter, etc). Characterization: assignment of a physical and quantitative description to each category or class (e.g., average grain size, RMS roughness, average sound speed, average density, amount of clutter, etc). 2 SESAM basics SESAM uses the variation of BS Versus beam-seafloor incident Angle (BSVA, Fig. 3) to segment different seafloor regions by comparing the behaviour of the couple BS/incident angle with a set of typical BSVA variations selected by the operator 1,2. The first step of the segmentation process consists of combining the BS and the beam-seafloor incident angle data from the multibeam data. The BS data are calibrated, outliers removed, corrected for local slope and geo-referenced. The BS data are then geographically displayed for the operator to select areas of typical BSVA characteristics. The information contained in the BSVA curves is affected by numerous geoacoustic and geometrical parameters: At a given frequency, water-sediment interface scattering and scattering from sediment volume heterogeneity compete to produce a complex BSVA, with frequent dominance of interface scattering at very high frequency. At normal incidence (0 ) the BS is strongly dependent on the water seafloor impedence contrast (i.e., water-sediment interface dominance). At very high frequency (e.g., 300 khz) and away from specular, the BSVA decay is primarily dependent on seafloor roughness: typically the lower the decay, the higher the RMS roughness. At 300 khz, if seafloor roughness is low, a slight specular reflection peak is observable on the BSVA curve. It is possible to obtain similar BSVA curves for different geological compositions of the seafloor: the same impedance can be obtained using mixtures of sand, gravel, silt and clay; roughness can be modified by the presence of ripples or discrete scatterers. Similarly, identical geological compositions can give different BSVA depending on roughness: this is what happened at Marciana Marian site (see Section 3). The average seafloor grain diameter, obtained from ground truth grabs, is strongly correlated with BS at normal incidence (Tab. 1) an important parameter for subsequent classification. Iterations may be required to eliminate categories (e.g. for a rippled sandy seafloor surveyed along differing tracks) or to add one or more new categories. In particular, ripple direction seems to influence only the mid-angle part of the BSVA. The comparison between the BS/incident angle dataset and the set of typical BSVA patterns can be performed using different methods. At present a distance paper: submitted to World Scientific on August 27, 2003 3
criteria is applied by SESAM. Each ping is treated individually using Bayesian probability. The probability that the ping belongs to a given category is estimated. Then, the probability that each individual BS belongs to that category is estimated. The product of the two probability produces a spatially non-filtered segmentation. A spatial filtering technique that can be applied to the Bayesian probability associating the nearest data points has been implemented to produce more stable results 2. 3 Experimental validation This section presents the application of SESAM to data acquired during the CLUT- TER 03 cruise. Segmentation is usually performed without a priori knowledge: segmentation and bathymetric maps are produced for comparison with other acoustic and non-acoustic data acquired at the same site. This allows the improvement of the segmentation results and the transition to classification and/or characterization. Figures 1 2 show the segmentation and bathymetric results obtained at the site with super-imposed coloured dots corresponding to the positions of stereo-photos, grabs and expendable Bottom Penetrometers (XBP). Details of the ground truth information are given in Tab. 1. With the help of ground truth information, it is possible to obtain a classification map of each site. Blue is used for muddy/silty classes, green for presence of sea-grass and red for gravel/rocky classes. The Marciana Marina site has three challenging characteristics for the SESAM segmentation algorithm (see Figs. 1 2): The zone of silt at longitude between 250-1000 and latitude over 1700, is composed of two unmixed different materials, one of which is never found isolated. To separate the patch, it is necessary to manually trace the BSVAC over the data. The graphic user interface was improved to include this feature: for the patch zone of Fig. 1 the BSVA curve of Fig 4 was obtained. The hard zone found around the point (2200,1200) is composed of rocks over sandy silt (generated by the mouth of a small stream). The presence of sparse rocks over sandy silt produces a resulting BSVAC (class 5) which resembles to other sites BSVAC classified as gravel site (transition between rock and sand). The longitude meridian 2750 lays over two classes of seafloor: class 3 and 4. Class 3 is, near shore, a dense Posidonia seafloor class (Fig 3.a). Considering that dense Posidonia can be found only to 30 m, the seafloor classified as class 3 at depth greater than 30 m there cannot be Posidonia, even if it is equivalent from an acoustic point of view (at 300 khz). Stereo photos (Fig 3.c) and grabs show that this seafloor is sandy-silt (φ = 0.83 3.82) with patches of sea-grass. On the other hand, class 4 is also similar, from an acoustic point of view, to class 3 (Figs 3.b and 3.d). The problem is caused by very similar acoustic response of the silt present at Marciana Marina (the roughness of which is increased by patches of sea-grass), and the sea-grass of Figs 3.a and 3.b acoustic response. This is the first site to present such characteristics: to resolve this kind of uncertainty work is proceeding paper: submitted to World Scientific on August 27, 2003 4
Figure 2. Marciana Marina segmentation (on the left) and the corresponding 390 khz sidescan image (on the right). A cross with different symbols is drawn at each ground truth site: stereo photos (circles), grab (squares) and expendable bottom penetrometer (triangles) are overlaid on the segmentation map if acquired. paper: submitted to World Scientific on August 27, 2003 5
BSVA curve Photography a b c d Figure 3. The photos and the acoustic response of four different seafloor. All the BSVA curves are very similar but the first and the last are covered with thick sea-grass while the other two have the similar mean grain size and a high roughness due to the patch of sea-grass. paper: submitted to World Scientific on August 27, 2003 6
Figure 4. Mouse constructed BSVA curves for a two distinct zones seafloor. on segmentation of lower frequency multibeam sonar (EM300 and EM1000 Simrad sonar, at 30 and 95 khz respectively). Except for class 3-4 composition of three different seafloors, the results of the segmentation are spatially consistent. In particular, they show the patches of class 0 seafloor inside class 1 and the presence of gravel where rock is protruding as shown by illuminated bathymetry (right of Fig. 1) and sidescan sonar image (right of Fig. 2). SESAM classification also shows the evolution of soft silt to harder bottom of class 0 to class 2. In Fig. 1 the bathymetry is overlaid on the segmentation map: in places the segmentation map is correlated with bathymetry. The dynamics of sedimentation being dependent on currents and grain size may explain the presence of harder/coarser sediments in the shallower areas and softer/finer sediments in the deepest areas. Also, Posidonia is only found at depths between 12 and 30 m. 3.1 Ground truth A summary table of the class analysis is given in Tab. 1. The geo-acoustic properties and stereo-photos are consistent with the underlying segmentation categories. Similarly, XBP results (Tab. 1) are in accordance with the segmentation. Stereophotographs were acquired during CLUTTER 03 by two cameras mounted on a frame to provide a digital height field 3, which allows the quantification of 2D seafloor roughness. Results from this system are not displayed in this paper but are consistent with the segmentation. Ground truth data within the spatial limits of each paper: submitted to World Scientific on August 27, 2003 7
Class Ground truth Appearance & Comment 0 No grab 1 2 3 4 5 gr 1.4%, sa 36.7%, si 9.1%, cl 52.8%, φ =7.79. XBP class II, 49g gr 0%, sa 57.2%, si 30.5%, cl 12.4%, φ =4.63. XBP class II, 57 94g gr 1.8 16.7%, sa 45.6 88%, si 3.1 50.8%, cl 1.7 7%, φ =0.83 3.82. Class I, 108 191g gr 0.7 9.1%, sa 27.5 42.6%, si 3 70.9%, cl 1 48.5%, φ =4.42 6.42. Class I, 68 91g gr 0.7%, sa 29%, si 19.4%, cl 50.8%, φ =7. XBP class I, 136 146g No Photo. The lower BSVA return: soft bottom. Silty with some worm activity. BSVA 10 db higher than the previous: still soft. Silty with intense worm activity. BSVA 5 db higher than the previous: silt. Two classes: thick posidonia and sandy with patches of sea-weed. Two classes: thick short sea-weed and sandy silt with patches of sea-weed. Sand and clay with large stones (12 cm) ; The resulting seafloor is hard and rough. Table 1. Summary of the six classes of Fig. 1 2 with corresponding ground truth, stereo photo interpretation and associated comments. In the table, gr, sa, si, and cl stand for gravel, sand, silt, clay. φ = log 2 D m where D m is the mean grain diameter in millimeters. XBP class I to III are, respectively, hard, medium and soft bottom. Penetrometer deceleration (units of g) is an indicator of seafloor hardness.) SESAM category have similar properties. SESAM. This is a point-to-point validation of 3.2 Sidescan images A comparison between geo-referenced sidescan sonar images acquired during CLUT- TER 03 (using an EdgeTech DF-1000 at 390 khz) and segmention/classification map was made (Fig. 2). The SESAM segmentation conforms well to the sidescan image patterns and provides additional information in areas where the sidescan image does not display appreciable textural differences: in particular the stronger return on sidescan signal over class 5, and the lower return where class 1 begins. Also the correspondence of grey level and class type is coherent over the whole site. Some black spot is due to shadowing. The comparison between sidescan sonar images and the SESAM segmentation maps demonstrate the spatial consistency of the technique. 4 Summary This work reports further experimental validation of SESAM, a segmentation algorithm, using data acquired by a multibeam echosounder. The methodology, based on the observation of the variation of the seafloor backscattering strength with the incident angle, has been improved by following acquisition of further echosounder system detail and enhancing the filtering applied to the segmentation data. The results are stable and create potential for further improvement of the algorithm for classification/characterization. Analysis of lower frequency multibeam sonar may paper: submitted to World Scientific on August 27, 2003 8
be appropriate for discrimination of sea-grass type and high backscattering seafloor (gravel versus rock). Aknowledgment We would like to thank the captain and the crew of RV Leonardo for the help and the nice collaboration during the acquisition of the data during CLUTTER 03. References 1. G. Canepa and N. G. Pace. Seafloor segmentation from multibeam bathymetric sonar. In M. E. Zakharia, editor, Proc. of the fifth European Conf. on Underwater Acoustics, ECUA 2000, pages 361 7. European Commission, Lyon, France, July 2000. 2. G. Canepa, E. Pouliquen, N. G. Pace, A. Figoli, and P. Franchi. Validation of a seafloor segmentation algorithm for multibeam data. In Proc. of the sixth European Conf. on Underwater Acoustics, ECUA 2002, pages 89 94, Gdansk, Poland, June 2002. 3. E. Pouliquen, M. Trevorrow, P. Blondel, G. Canepa, F. Cernich, and R. Hollet. Multi-sensor analysis of the seabed in shallow water areas: overview of the MAPLE 2001 experiment. In Proc. of the sixth European Conf. on Underwater Acoustics, ECUA 2002, pages 21 29, Gdansk, Poland, June 2002. paper: submitted to World Scientific on August 27, 2003 9