Boulder detection |

MBES Deptr

VMBES combined

MBES Sie

BES. BS

AEF

&

Fig. 8: Test area composed of fine sediments with a marked impact of bottom trawling activity.
Because of the fine sediment composition, it can be assumed no boulders are present in this area.
he model working on slope data detects several false positives in the area, while models running
on backscatter, depth and the combined multi-band image report no false positives. Refer to the
northwest of Fig. 2 for location

0 75 15m
X

a

in acoustic data, and - except for obvious instancas
 - the interpretation of a human interpreter of
what is and what is not a boulder varies based on
nis/her experience, with no possibility to Judge
what is the correct interpretation. The appearance
and visibility of boulders in backscatter data can
Change with swath width and incidence angle (Paogenmeier
 et al. 2020; von Rönn et al. 2019). While a
methodological description on how to assess geogenic
 reefs exists (Heinicke et al., in press), it de
fines no sufficient criteria to decide which objects
are to be identified as boulders in acoustic data.
Still, our case study allows qualitative insight
'nto the advantages and disadvantages of 5SS
and MBES-based boulder mapping by neural net
works. To mitigate the impact on AP for the dif
ferent models, a single person confirmed all samoles
 in the training database used for this study.
"herefore, model performance is only compared
relative to the interpretation of the acoustic data
py one human expert and not to the true seafloor
zonditions. Both SSS and MBES systems supply
packscatter information. A problem of 5SS-based
boulder detection are artefacts (Wilken et al. 2012),
2.g., near the nadir or in areas of water column
stratification that can in their structure resemble
small boulders (Fig. 6). Due to the requirements to
detect tiny objects comprising only 7 to 9 pixels in
the examples shown here and even less if objects
af 25 cm in size are to be detected in acoustic data
(von Rönn et al. 2019), there is limited information
to differentiate between artefacts and real objects.
This causes a trade-off during the training of sidescan
 sonar-based models: if the sensitivitv of the

mNodel to detect small boulders — as required by
egulations — is increased, the amount of false
Dositive identifications increases as well. Because
af the absence of well ground-truthed reference
;ites, a calculation of meaningful precision-recall
zurves to find optimal threshold values is not possible.
 Tuning the threshold level of the model to
'he local conditions (e.g., the number of artefacts
n the data) is done manually, which is a subjec-“ive
 procedure. A possible solution is to include
ıadir and water column stratification effects as
Jistinct classes and define these areas as insuffizient
 for boulder detection. While MBES snippet
Jderived backscatter data is not affected by water
zolumn stratification and is used for object detec
ion (e.g., Kunde et al. 2018), individual boulders
are not displayed in the specular regime (Fig. 7) at
ıear-vertical incidence angle and are resolved in
ess detail compared to side-scan sonar images in
"he data (Fig. 2). The loss of detail may be caused
>»y a different along-track resolution due to dif
‘erent opening angles of the used systems (0.5‘
at 400 kHz for the Reson 7125, CSS-2000: 0.26°
at 600 kHz, respectively 0.29° at 410 kHz for the
4300 MPX). Combined with the less pronounced
3coustic shadows, the AP of the MBES backscatter
nodel data set, therefore, is worse compared to
“he model trained on side-scan sonar backscatter
data (Table 1). MBES-based backscatter maps can
10t be recommended as the principal data source
‘or boulder detection based on our case study.
An obvious problem related to the use of MBES
Dathymetry and derived slope values is the required
 thorough cleaning of the data, with outli-A



1719 — 06/2027