3oulder detection |

ers or morphological features having similarities
to small boulders in slope maps. An example of
such morphological feature in the German Baltic
5ea is related to bottom trawling (Fig. 8). The trawl
doors create steep local, almost circular morpho
'ogical features when lifted off the seafloor. These
features are misinterpreted by the slope-only
model as boulders. The backscatter model cor
rectly ignores these features. The combination of
Dackscatter and slope data also prohibit false posi
ives in the combined model. Therefore, while the
AP of the slope model is the best on the valida
sion data overall (Table_1), it also produced unde
sirable false positives in areas where boulders are
very unlikely to appear (Fig. 7). The pixel-perfect
zoregistration of depth and backscatter informa
tion by multibeam echo sounders can mitigate
*his downside. Being the best model in our case
study, the slope-model results were compared
with the human raster-based interpretation of a
larger area. The F; score of the model compared tc
*he human experts is 0.75 and 0.62. Both scores are
nigher than the score for the direct comparison of
:he human experts, although the number of raster
zells counted is not identical due to the different
axtension of available SSS and MBES data. Position
al inaccuracies between the side-scan sonar and
Nultibeam echo sounder data of approximately
5m may negatively impact the comparison of cells
where boulders are situated close to the edges. In
naindsight interpretation of the model-human dif
r‚erences, potential errors on both sides were idenifed
 (examples shown in Fig. 5). In addition, the
slope data is less affected than backscatter intensity
 by survey geometry and finds boulders that
zould not be identified in the side-scan sonar data
because they are located close to the nadir.
"he poorer performance of the MBES depth-deived
 model compared to the slope model is not
surprising, given that the maximum resolution of
:he input image is the regional depth interval di
vided by the available discrete pixel values. In out
study, this is 9 m divided by 256 (28 bpp, bit per
pixel), artificially limiting the vertical resolution tc
ca. 0.035 m in the single band 8-bit image. Given
<hat many boulders have smaller elevations (Fig. 2)
and are visible in slope maps, the performance
af the depth model is good and may have great
ootential for models operating on point clouds
and derived statistics which became available in
the last years (Held and Schneider von Deimling
2019; Guo et al. 2020). The advantages and disadvantages
 of including absolute depths as an input
<hannel for neural networks must be considered,
nowever. In the Baltic Sea, for example, finding
Doulders in deeper muddy basins is unlikely, but

possible (Beisiegel et al. 2019), and changing depth
intervals between different sites (and thus chang
ing resolution in colour-coded depth images) may
be problematic. We suggest exploring the use of
“urther depth-derived information, such as the
bathymetric position index, or texture parameters
derived from backscatter mosaics in the future.
Models working on a combination of depth,
slope and backscatter data produced false negaüves
 In the near-nadir region, as boulders are not
imaged in the backscatter channel. They also show
fewer false positives and are less susceptible to re
maining outliers in bathymetric data. Therefore,
while the performance of the joint depth-slopebackscatter
 data set is worse than for the slope
only model (due to validation examples in the
nadir region) in our case study, its inherent robustness
 to false positives by combining independent
data sets makes it the method of choice for practical
 applications in the future. Ideally, and needed
for many commercial applications anyway, an over
100 % overlap would remove the near-vertical incidence
 backscatter data and is expected to im
prove model results. Multi-band images with cali
brated backscatter data collection (Lamarche and
urton 2018) would also allow for a quantitative
definition of boulders, e.g., by measuring increase
of backscatter intensity in addition to local slope
and local bathymetric position index.
5 Conclusion
Our case study shows that boulders are detected
with higher precision in bathymetric data com
Dared to backscatter mosaics recorded by either
Multibeam echo sounder or side-scan sonar. The
results of the best model are comparable to the
‚ange of results achieved by human interpret
ers. We recommend combining bathymetry and
hackscatter data into a multi-band image to limit
“alse positive detections. However, the limiting
“actor for the automated detection of boulders in
acoustic data is not the technology, but the do
main knowledge and the availability of accurately
annotated training images. Future activities should
ınvolve the careful choice of sites for ground-truth
ing and acoustic surveys, to create a high-quality
rraining data set. /
Acknowledgment
The authors thank Elham Al-Akrami for initial prep
aration of MBES-related training data set, and Mer-'e
 Hennig for support in digitising boulders on the
5$5 mosaics. We thank the crew of VWFS Deneb
for their great support during the measurement
campaigns, and the two reviewers who provided
helpful and constructive comments.

Hydrographische Nachrichten