Peer-reviewed paper 
Boulder detection | 
10°70 10°30°0 11°0°0 11°300 
ME 0 AA MI 
———.. x. 
ABES Depth J-19 Imı 
491 
an - 
BES Slone #0 1 4 
A 
di 
Mr GaDüs 
OD 
0 10 20 km 
|___ | 
Je 
4 
Mr“ 
a 
J 
N 
Fig. 1: Location of the investigation site west of Fehmarn (left). Water depths in the area range between 16 m and 25 m (centre) 
dashed lines are the survey lines run during MBES data collection. Right: Slope calculated from the local bathymetry 
aids the detection of small objects. Both for man- 
Jal and automatic methods, boulder detection 
was found to be more reliable, with an increas- 
'ng number of pixels forming an object's repre- 
sentation in backscatter (BS) mosaics (Feldens 
at al. 2019). Acoustic shadows, which form be- 
hind boulders, increase the number of pixels of 
o0ulder representations in backscatter mosaics. 
Shadow sizes increase with grazing angle, thus 
“avouring towed sonar systems (Papenmeier et 
al. 2020). Therefore, while the spatial resolution 
af modern MBES derived backscatter information 
zan rival that of side-scan sonar systems in many 
relevant practical applications (depending on 
water depth), their survey geometry is unfavour- 
able for boulder detection in backscatter data. 
However, the pixel-perfect co-registration of 
depth and backscatter and derived data sets may 
offset this disadvantage and facilitate boulder 
detection based on MBES data. Considering the 
;nterpretation of extensive areas, human experts 
have difficulties in combining information of mul- 
ti-dimensional data sets, while machine learning 
algorithms are less limited by dimensionality and 
more efficient (Yokoya et al. 2017). 
In the last decade, object detection frame- 
works based on convolutional neural networks 
(CNN) were applied to different topics, including 
remote sensing In the earth sciences (Ghamisi et 
al. 2017; Zhu et al. 2017) with great success. CNNs 
were used to find boulders in side-scan sonar 
backscatter mosaics, showing performance com- 
parable to human experts in areas of moderate to 
J00d data quality (Feldens et al. 2019). It is the aim 
af this case study to compare the performance of 
mMultibeam echo-sounder and side-scan sonar to 
'mage boulders in single-band and multi-band 
data sets including depth, slope and backscatter 
intensity. An object detection framework based 
an a neural network is used to identify boulders 
in the data sets, and the results are compared with 
the interpretation of human experts 
AA 
119 — 06/2027 
2 Methods 
2.1 MBES 
Multibeam echo sounder data were collected in 
;ummer 2019 from the hydrographic survey ves- 
zel VWFS Deneb, operated by BSH, by a state-of- 
:he-art MBES system Teledyne-Reson Seabat 7125- 
52. The system operates at 400 kHz with a 140° 
2pening angle, a pulse length of 300 us and 512 
Jeams per swath. The seafloor of the study area 
Fig...1, left) was fully covered by 50 survey lines 
with 100 % overlap (Fig. 1, centre). The software 
"eledyne PDS was used for real-time data acquisi 
rion. A combined GNSS (Global Navigation Satellite 
Systems; good global but poor relative accuracy) 
and INS (Inertial Navigation System; good local ac 
zuracy but drifts without external reference) forms 
"he basis for an accurate and reliable real-time di 
‘ect georeferencing of MBES measurements. MBES 
nstruments require an accurate portrayal of the 
;ound speed structure of the water column. In this 
zampaign, the distribution of water sound velocity 
was determined by continuous profile measure 
Nents using the multi-parameter online probe 
sea & Sun Technology CTD 60Mc. Bathymetry data 
were processed using Teledyne CARIS HIPS & SIPS. 
he processing chain holds techniques for i.a. cor 
‚ection of sound velocity induced effects, calcula 
:ion of a georeferenced 3-D point cloud, genera 
tion of 3-D surface representation of the bottom 
topography, outlier detection and filtering. 
To create backscatter grids with a resolution of 
2.25 m based on the multibeam echo sounder 
data provided as s7k-Ailes, angular variations in 
ntensities were removed using the open-source 
arocessing toolbox MB-System (Caress and 
Zhayes 1996). A grazing angle of 40° (here, minor 
variations in incidence angle have little effect on 
jackscatter intensity) was used as a reference 
angle. A low pass Gaussian mean filter stretching 
Ave samples in the across-track and three samples 
n the along-track direction was applied once to