Boulder detection | 
ratio were also changed by +60 % for each image. 
The optimal anchor sizes for the YOLO network 
were Calculated. 15 % of the training samples were 
‚andomly selected for validation and used to cal- 
zulate the average precision for the boulder class 
(AP) of the different networks. After the image set 
for validation was separated, a Python script ro- 
tated every image in 45° steps to account for vari 
able survey directions. The training took place on 
a NVIDIA 2080 TI graphic card (11 GB RAM). Training 
af the MBES models required about twelve hours 
for the MBES models and 40 hours for the large 
5SS model. 
For model application, the training procedure is 
reversed. The (single or multi-band) mosaic is cut 
'nto small georeferenced image tiles of 64 x 64 
pixels. Threshold values for include objects were 
set to 0.2 for all models except the SSS model for 
small objects, which was set to 0.35. The model is 
run on these small tiles. The detection of objects 
an a single image requires about 10 ms on an 
NVIDIA 2080 TI. The pixel-coordinates of the result- 
'ng bounding boxes are converted to geographic 
zoordinates and displayed using QGIS. To emulate 
the raster approach used by human experts to 
cover large areas, detected boulders in each grid 
cell are counted. 
Jp also controls the local slope shown in Fig... 
Nhile high pixel-to-pixel slopes exceeding 60° at 
Maximum prevail in the areas of glacial lag depos- 
ts due to the presence of boulders and near the 
;rawl marks, the remaining area is flat with slope 
/alues below 2°. 
3ased on a visual inspection, we find most boul 
ders in the area composed of glacial lag deposits, 
with some also present in the sandy facies. The 
»o0ulders have different characteristics in the data 
sets that are displayed in Fig. 3. In the SSS-derived 
dackscatter mosaics, boulders can be recognised 
»y a high backscatter front, an intermediate in 
tensity signal behind and an acoustic shadow at 
"he back, relative to the side-scan sonar position. 
dowever, small boulders are often more difficult to 
nterpret. This is caused either by their small size 
ar their position in the outer part of the swath (a 
zombination of which is shown in Fig. 3B). In addi- 
tion, artefacts in side-scan sonar data can resem 
le smaller boulders. Such artefacts include scatter 
from water column stratification or areas near the 
side-scan sonar nadir. 
‚n MBES-derived backscatter, boulders are rec 
ognised by an increase in backscatter intensity 
z:ompared to the surrounding seafloor (Fig, 3) but 
are often lacking a pronounced acoustic shadow. 
The backscatter representation of boulders is less 
distinct compared to SSS imagery in close to inter 
mnediate distance to the nadir. Boulders are imaged 
3s circular to elliptic features in maps of the local 
slope. Slope values for boulders range from 3.5° to 
more than 60° degrees, related to the large vari- 
aty of boulder shapes in transported lag deposits 
transported by glaciers. Also, boulders may be par- 
tjally buried in the subsurface. However, not all cir 
zular features correspond to increased backscattet 
ntensities, for example in the areas of overlapping 
ırofiles. In MBES-derived maps of depth, boulders 
are displayed as circular features elevated 2.5 cm 
to over 50 cm compared to the adjacent seafloor. 
> Results 
3.1 Local geology and appearance of boulders 
Water depths in the investigation site (approxi- 
mately 2 km?) vary between 16 m and 25 m, with 
depths increasing towards the north. Backscatter 
maps derived from MBES and SSS show different 
zeafloor facies at the site (Fig. 2), with fine-grained 
deposits and intensive disturbance by bottom 
trawling activities in the north (low backscatter). 
nigh backscatter intensities characterise glacial 
1ag deposits towards the south and east. A high 
1umber of boulders are part of these deposits. In- 
termediate backscatter intensities towards south 
and west characterise fine to medium sands and 
Dartial outcrops of glacial lag deposits. In the side 
scan sonar mosaics, which cover a larger area, a 
series of elongated, elevated ridges exist in the 
southeast The general sedimentological build- 
3.2 Manual boulder identification 
zor a test area of about 30,000 m}, two experi 
anced human interpreters picked boulders on the 
side-scan sonar backscatter mosaic (Fig. 4). The 
test area showcases instances of water column 
Et 1 n- 26 
Zymart II a— 5 
"manıntari 
Owen 
"ig. 4: Manual interpretation of boulder occurrence in the test area hased on SSS backscatter data 
ıhe number of identified obiects is 26 and 54. Refer to Fig. 2 for locatior 
0 
25 
50 m 
a! 
AA 
119 — 06/2021