The classified area is found on the upper slope, between 35 and 70m water depth (Figure 2). The multi-beam images indicate 3 different zones (large 2-D straight dunes, solitary dunes and area without features), organised by the sediment transport.

Previous Studies
General information
The Fraser River delta has semi-diurnal tide, averaging 3m but reaching a maximum 5m, generating tidal currents averaging 0.35m/s, with a maximum of 1.3m/s. Because of the dominant flood current, the Coriolis effect and the training effect of Steveston jetty (construction started in 1908), the southern Roberts Bank slope has a negligible sedimentation rate. The small amount of sediment present is mainly relict and is continually reworked.

Small-scale information
Dunes on the Roberts Bank slope have previously been studied by Luternauer and Liam Finn, (1983), von Deichmann (1992), Kostaschuk and Luternauer (1993), Kostaschuk et al. (1995) and Carle (2003, unpublished). Two different types of dunes are shown in this study area (1,500m length, 400m width). The first is a 2-D straight crested dune (27m wavelength and 0.35m height), mostly found in deeper water, though smaller forms are found in shallower water (Figure 2). These dunes are surrounded by two sets of solitary dunes that may be interpreted to be sediment starved due to the strong current and the lack of sediment supply. Between the solitary dunes and the northern edge of the study area there are no features evident on the image at 2m resolution.

Classification Methodology
The EM3000 multi-beam data (Kongsberg Simrad) was classified using QTC MULTIVIEW 2.40 software from Quester Tangent Corporation. The program generates features from rectangular patches of the backscatter image, performs principle component analysis of the features to extract three data points per patch and uses cluster analysis to split this data into classes. Ground truth is usually necessary to relate the classification to these seabed characters, but in this case samples were not available. Because of the dunes wavelength (30m), different sizes of rectangles, smaller than the wavelength, representing variable spatial resolution were tested to get the best classification possible.

Results
There are two classifications using different rectangle sizes, emphasising differing information. The first (Figure 3) using rectangular patches 257x17 pixels (15.9x7m) distinguishes, at large scale, between the rougher areas of dense dunes and the smooth area with few features. The second (Figure 4) using rectangular patches 257x9 pixels (15.9x3.7m) distinguishes the crest and the trough of the features.

First classification (257X17)
The presence of dunes in a strong current suggests sediment transport. Class 1 (red) is omnipresent and does not give any sediment transport information, but the two other principal classes are more informative, especially in the deeper part. Class 2 (blue) is well developed on the Southeast region and on the gentle slope from the northern solitary dunes. Occurrence of Class 2 diminishes in deeper water in the north, to be replaced by Class 3 (green). This class is well represented in areas between the solitary dunes and could represent smooth troughs.

Second classification (257X9)
Using narrower rectangles in the along-track direction gives a better distinction between the crests and the trough. The Class 3 (green) represents the crests of most of the predominant features, while Class 1 (red) represents the area without dunes or the troughs. The large solitary dunes are easier to define because of their regularity and size. The 2-D dunes are not as well defined because of their irregularity.

Discussion
The first classification is compatible with the decrease in sediment availability along the transport path, leaving solitary dunes and a changed backscatter return. This result shows that the seabed classification is able to provide information about sediment transport in an area covered with seabed features. The second seabed classification shows differences in backscatter at the feature scale. This classification may be related to the different grain size of the sediments composing the crests and troughs of the dunes.

Conclusions and Recommendations
Classification of multi-beam images can be used as a tool for mapping areas influenced by sediment transport. By segmenting the image into different sizes of rectangular patches, areas with dunes can be distinguished from areas where these are absent, and dune crests can be distinguished from troughs. Grab samples from the area are still required to confirm correlation between the classification and the inferred variations in grain size caused by transport of sediment. Research is still required to determine whether classification is dependent on orientation of features relative to the direction of the survey vessel.

Acknowledgements
Data and hydrographic map (3463) are courtesy of the Canadian Hydrographic Service (CHS) (Sidney). Thanks are due to Glenda J. Rathwell from Quester Tangent for her processing advice. The coastline map is courtesy of Dr Martin Scherwath, School of Earth and Ocean Sciences (SEOS), UVic.