implementation of unsupervised classification. First, selecting the appropriate
number of classes is not always straightforward. Second, the segments and classes
found are not always useful, are not always possible to fit into a pre-determined
hierarchical system, or are not easy to identify why they were segmented out.
Third, available ground-truth data, which is often limited, must be divided into two
subsets for labeling and accuracy assessment. Fourth, clusters are usually sitespecific and sometimes sensitive to the size of the dataset. Nonetheless, unsupervised classification can still be an effective tool.
One method for labeling the clusters found in an acoustic dataset is to compare
the shapes of the average echoes for each cluster identified in the segmentation
step. Clusters comprised of echoes with long average duration are labeled hardbottom, and those with short average echoes are labeled sediment. The advantages
of this approach are, first, that class labeling can be performed with minimal to no
ground truth data, and, second, that the same classes can be mapped in different
areas or using different SONAR systems. The disadvantage of this approach is that
the result is simply a two-class map (i.e., a map with low thematic detail). This
section will illustrate the results of this technique applied to survey areas in the
Florida Keys and the Bahamas, illustrating that even a map with low thematic
detail can be effective and useful.
Gleason et al. (2009) describe four surveys using single-beam data from a
50 kHz QTCView Series V (QTCV): one from Lee Stocking Island (LSI),
Bahamas; a second from Carysfort Reef, in the Florida Keys; a third from Fowey
Rocks, also in the Florida Keys; and a fourth more recent survey from Andros
Island, Bahamas. The Andros survey consists of approximately 73 km of track
lines acquired in depths from 1 to 8 m. One portion of the bank was covered with a
grid of 19 approximately 1,500 m long transects spaced 100 m apart. A second
portion of the bank top was covered by nine widely spaced, cross-shelf transects
that, in turn, were connected by segments of an along-shelf transect formed while
transiting between the cross-shelf areas.
The processing steps for each dataset were threefold. First, cluster the data
using the IMPACT software package (Quester Tangent Corporation 2002). Second, use ancillary datasets such as satellite imagery, snorkeler observations, mean
echoes for each cluster, and previous seabed maps to label the largest clusters
(those comprising over 90 % of the dataset) as either hardbottom or sediment.
Third, assess the accuracy of the hardbottom/sediment classified map using
independent measurements from divers or towed video. Overall accuracy for the
hardbottom/sediment maps was 86, 78, 74, and 73 % for the Carysfort, Fowey,
LSI, and Andros survey areas, respectively.
The mean echoes for all clusters show a steep rise in amplitude corresponding
to the initial reflection from the seabed followed by a gradual decay (Fig. 9.5). The
amplitude of the mean hardbottom echoes decayed more slowly with time (i.e.,
they were longer) than the sediment echoes in all four areas (Fig. 9.5), which was
expected since rock has stronger off-nadir backscattering than sediment (APL-UW
1994). There was significant variation in the shapes of the hardbottom echoes even
though all of the mean echoes from the hardbottom clusters have slower decay
9 Acoustic Applications
235
number of classes is not always straightforward. Second, the segments and classes
found are not always useful, are not always possible to fit into a pre-determined
hierarchical system, or are not easy to identify why they were segmented out.
Third, available ground-truth data, which is often limited, must be divided into two
subsets for labeling and accuracy assessment. Fourth, clusters are usually sitespecific and sometimes sensitive to the size of the dataset. Nonetheless, unsupervised classification can still be an effective tool.
One method for labeling the clusters found in an acoustic dataset is to compare
the shapes of the average echoes for each cluster identified in the segmentation
step. Clusters comprised of echoes with long average duration are labeled hardbottom, and those with short average echoes are labeled sediment. The advantages
of this approach are, first, that class labeling can be performed with minimal to no
ground truth data, and, second, that the same classes can be mapped in different
areas or using different SONAR systems. The disadvantage of this approach is that
the result is simply a two-class map (i.e., a map with low thematic detail). This
section will illustrate the results of this technique applied to survey areas in the
Florida Keys and the Bahamas, illustrating that even a map with low thematic
detail can be effective and useful.
Gleason et al. (2009) describe four surveys using single-beam data from a
50 kHz QTCView Series V (QTCV): one from Lee Stocking Island (LSI),
Bahamas; a second from Carysfort Reef, in the Florida Keys; a third from Fowey
Rocks, also in the Florida Keys; and a fourth more recent survey from Andros
Island, Bahamas. The Andros survey consists of approximately 73 km of track
lines acquired in depths from 1 to 8 m. One portion of the bank was covered with a
grid of 19 approximately 1,500 m long transects spaced 100 m apart. A second
portion of the bank top was covered by nine widely spaced, cross-shelf transects
that, in turn, were connected by segments of an along-shelf transect formed while
transiting between the cross-shelf areas.
The processing steps for each dataset were threefold. First, cluster the data
using the IMPACT software package (Quester Tangent Corporation 2002). Second, use ancillary datasets such as satellite imagery, snorkeler observations, mean
echoes for each cluster, and previous seabed maps to label the largest clusters
(those comprising over 90 % of the dataset) as either hardbottom or sediment.
Third, assess the accuracy of the hardbottom/sediment classified map using
independent measurements from divers or towed video. Overall accuracy for the
hardbottom/sediment maps was 86, 78, 74, and 73 % for the Carysfort, Fowey,
LSI, and Andros survey areas, respectively.
The mean echoes for all clusters show a steep rise in amplitude corresponding
to the initial reflection from the seabed followed by a gradual decay (Fig. 9.5). The
amplitude of the mean hardbottom echoes decayed more slowly with time (i.e.,
they were longer) than the sediment echoes in all four areas (Fig. 9.5), which was
expected since rock has stronger off-nadir backscattering than sediment (APL-UW
1994). There was significant variation in the shapes of the hardbottom echoes even
though all of the mean echoes from the hardbottom clusters have slower decay
9 Acoustic Applications
235