3oulder detection | 
WO, MO WG 
ide Scan Sonar 
z S | hull 
la 3] = | mounted 
NAC © 
DIS B 
<'E 
MBES backscatter 
. 
n 
g.6 
> 
A 
oö 
CO 
ELLE 
ZOO 
Lad =» 
m 
Yan 
1 1 
EREnEnE Da 
CC 
LE 
MALTE 
VL 
ELLLLITE 
MENAaNE 1 
zZ 
‚ig. 2: Left: Side-scan sonar backscatter mosaic (0.25 cm pixel size) used for boulder 
classification in this study. Centre: Multibeam echo sounder backscatter mosaic. 
Right: Location of the boulders (black dots) and empty image examples (red dots) used for training of the models 
based on multibeam echo sounder data. The red box represents the test area for the manual identification of 
individual boulders in Fig, 6 and false positives examples in Fig, 8, respectively. The raster grid used for the manual 
determination of boulders densities is indicated 
low 
han] 9 200 400m 
vr 
the data to remove high frequency speckle noise. 
Data gaps of up to 1.25 m were interpolated. The 
grid was built by applying a Gaussian Weighted 
Mean. As available profiles are overlapping, sam- 
Dles of higher grazing angles were given an in- 
creased priority during gridding. Profiles were run 
in both N-S and 5-N directions on the same pro 
Ile track. For the backscatter maps, one of these 
directions was used, the other line was discarded. 
3ackscatter intensities were clipped at the 0.2 % 
and 99.8 % percentile to improve image contrast 
'n this study, higher backscatter intensities are dis- 
dlayed in darker colours. All backscatter intensities 
are uncalibrated, relative values (Lamarche and 
Lurton 2018) and were exported as 8-bit greyscale 
mMosaics following processing. Multi-band images 
af MBES-derived grids of backscatter, slope and 
depth were created using the open source GDAL 
Utilities (GDAL/OGR contributors 2021), by plac- 
‘ng slope information in the green image channel 
Fig..1, right), backscatter information in the red im- 
age channel (Fig. 2, centre) and depth values in the 
Dlue image channel (Fig. 1, centre). 
2.2 555 
"he side-scan sonar data were recorded in May 
2019 during cruise #164 with the vessel VWFS 
Deneb. The Edgetech CSS-2000 was towed at an 
altitude of approximately 12 + 3 m above the sea 
bed. Due to technical problems with the CSS-2000 
a Change to the hull-mounted side-scan sonar 
(Edgetech 4300 MPX) became necessary dur- 
ing the cruise (Fig. 2 shows the coverage of both 
data sets). The vessel speed varied between 4 and 
4.5 kn. Using a swath-width of 200 m the profile 
distance was set to 180 m to allow an overlap of 
approximately 10 % at the edges. 
Processing of the backscatter amplitudes was 
done with the software package SonarWiz 7.3. 
Only the higher frequency of the CSS-2000 has 
been used (600 kHz). The 4300 MPX used a fre 
quency of 410 kHz. After bottom tracking and em 
pirical gain normalisation, the data of the towed 
system additionally required a correction of the 
navigation data. The sheave offset was adjusted, 
and a layback correction was executed basing on 
data of a cable counter and a pressure sensor. To 
generate a final backscatter mosaic both data sets 
were merged. The overlapping profiles were cut at 
-he edges as far as possible without causing gaps 
-inally, a mosaic (8-bit greyscale) with a spatial res- 
olution of 25 cm was exported (Fig. 2, left). 
2.3 Manual boulder count 
Two experienced human interpreters did a manua' 
count of individual boulders in a test area (Fig. 2, 
red box) based on the side-scan sonar mosaics. 
Human iInterpreters generally recognise boulders 
boy an increased backscatter intensity facing to- 
wards the side-scan sonar, followed by an acoustic 
shadow forming behind. The human interpret- 
ers were not involved in picking the training data 
‘or the neural networks. To interpret larger areas, 
a raster approach is used. For 50 m x 50 m cells 
(Fig. 2, black raster grid), the same human experts 
decided whether it includes no boulders, one to 
Ave boulders, or over five boulders. This procedure 
is in line with currently published recammenda 
Hydroaraphische Nachrichten