3oulder detection | 
ADEe-MOC 
1 
EEE + 
Hı]} 
=xpert ı 
Number of boulders 
If 
LE 
© 
O0 150 300m 
Ju. 
0 
4 1-5 
| 53 
Difference in boulder 
class human experts 
— 1 
L.2 
| 
° 
Boulder 
Fig. 5: Top: Number of boulders identified with the raster approach by the slope-model and to human experts. 
Bottom: Individual detection of the slope-model are plotted on top of the expert Il interpretation. Coloured cell boundaries 
visualise the difference in interpretation between the human experts. A) Example of a potential boulder not noticed by 
the experts. B) A potential false positive detection of the model. €) Detections near the side-scan sonar nadir, where no 
judagment of the madel detections is possible. Far C, the slope map is shown in addition 
stratification on the eastern side, a nadir stripe in 
:he centre of the area and an overlap of two differ 
ent profiles recorded with different side-scan so- 
nars towards the west. The experts found 26 and 
54 boulders. No human misinterpreted the water 
column artefacts, nadir stripes or overlapping pro- 
Iles as boulders. A higher variability exists in the 
Juter parts of the swath near the overlapping pro 
les, where the appearance of potential boulders 
/aries. The same human experts interpreted boul 
der densities over a larger area using the raster ap- 
proach applied to 50 m x 50 m cells (Fig. 5). Dense 
Doulder assemblages were confirmed in the east 
“owards the outcropping glacial till, while boulders 
are sparse towards west. Corresponding to the dif- 
ferent number of individual boulders found in the 
test area, expert | identified a larger area covered 
by one to Ave haulders compared to expert Il. The 
ÜMSEER  SE 
ÜYSEEE 9 
MBES SLOPE 
MBES DEPTH SLOPE BACKSCATTER 
"555 BACKSCATTER large objects 
1555 BACKSCATTER small obiects 5 
MBES DEPTH 36 % 
MBFS BACKSCATTER 
Table 1: Overview of performance on the validation data set 
(measırred in AP} for the different models and data sets 
-, score, measuring the agreement between the 
Wo experts, is 0.61 based on 196 raster cells 
3.3 Automated boulder detection 
The Average Precision (AP) of the models on the 
/alidation data is shown in Table.1. The highest 
performance is 64 % by the slope-only model, fol 
owed by a model working on a 3-band data set 
comprising MBES backscatter, slope and depth 
with 61 % AP. The MBES backscatter-only mode 
achieves an AP of 18 %. The side-scan sonar per- 
formance is 37 % to 43 %, with the lower AP for the 
'raining data set with a focus on small objects. The 
detections of the best-performing slope-mode 
are plotted on top of boulder densities as deter 
mined by human experts (Fig. 5). 
The resulting detections of the models in the 
‚est area are shown in Fig. 6. The SSS models 
find a total of 35 boulders, all including a discern- 
‚ble shadow on visual inspection. One likely false 
positive occurs around water column stratifica 
tion artefacts and one false positive in the nadir 
region. The MBES backscatter model finds a tota' 
of 29 boulders. Of these, seven have no discern- 
ıble shadow, while the remaining display at least 
one pixel of acoustic shadows behind. The mod 
el working on the area-wide bathymetric grids 
detects 14 boulders with elevations of 6 cm to 
40 cm compared to the surrounding seafloor, 
albeit most boulders smaller than 15 cm are not 
recognised in the data set. The slope model finds 
Hyvdroaraphische Nachrichten