3oulder detection | 
inaccuracy that required shifting the side-scan 
sonar mosaic location by a few metres — seems 
to show that the slope data is correct, and these 
objects should have been identified as boulders. 
'n contrast, in the northern test area (Fig. 8), circu 
ar elevated features are identified as boulders by 
"he slope-model. We find similar examples, not 
displayed here, in areas with remaining outliers in 
morphological data which have a similar appear- 
ance. Such outliers cause artificial slopes but do 
not affect backscatter data information. 
"he results of the raster approach using the 
model with the highest AP (the slope-model) are 
shown in Fig. 5. The slope-model identifies be- 
‘ween 0 and 42 boulders in the 50 x 50 m cells. 
"he agreement with the human experts | and Il as 
measured by the F, score for 182 cells (cells where 
both SSS5 and MBES data are available) is 0.75 and 
7.63, respectively. 
MBES Slope 
4 Discussion 
The high difference of boulder detection by very 
experienced human interpreters (Fig. 3) shows 
the need for an objective, automatic method for 
boulder detection. The different count of individ 
Jal boulders transfers to an agreement of 0.61 (Fı 
score) over 196 cells that were interpreted with the 
raster approach. This poses a significant challenge 
both for quantification of model performance 
and for the establishment of correctly annotated 
rraining images, a problem faced by many other 
applications of neural networks to remote sensing 
data (Zhu et al. 2017). The same person interpret 
ing the training database and the reference sites 
‘or boulder detection (Feldens et al. 2019) partially 
mitigates the problem. However, this approach 
does not scale to more than one involved person 
or to applications where objective results without 
interpreter bias are required. Almost no study in- 
cludes an extensive ground truthing for boulders 
MBES Backscatter 
A 
’rofile E 
Profile 
Overlap 
MBES combined 
On 
* 
‘Nadir 
BB 
„7“ 4 
Ba 
0: 
6 8465 
a 
Fig. 7: Boulders detected by the MBES-models are displayed. Boulders are verified in the side-scan 
sonar image, whose position was shifted to account for positional inaccuracies. Near the nadir, potential 
boulders are not imaged in MBES backscatter data, while present in the slope map (blue rectangle). 
Vice versa, the backscatter map displays increased backscatter intensities in areas where no increased 
slope exists (red rectangle). No boulders are detected in both areas by the combined model working or 
depth-slope and backscatter channels. Refer to Fig, 6 for colour scales 
EA _ = 
O0 5 10m 
EL 
Hydrographische Nachrichten