Stienessen et al.: Comparison of model types for prediction of seafloor trawlability in the Gulf of Alaska 195 
of seafloor trawlability. This finding may encourage steering 
a scientific echo sounder’s single beam between 35° and 50° 
to collect the necessary data. Single-beam echo sounders 
are fairly ubiquitous on both scientific and fishing vessels. 
Research results indicate that, when added to models that 
help explain distribution of benthic animals relative to their 
habitat, data collected with echo sounders increase the per- 
centage of explained variance (McConnaughey and Syrjala, 
2009; Somerton et al., 2017). However, S;, oblique is an inci- 
dence angle measurement, and is not dependent on steering 
angle. Theoretically, steering a single-beam echo sounder at 
35—-50° over flat terrain would yield the appropriate inci- 
dence angles. Yet as soon as there was any change in sea- 
floor bathymetry, the incidence angles would also change, 
likely outside of the range of the steered single beam. There- 
fore, multibeam echo sounders are well-suited to collecting 
S;, oblique data. Having beams steered at a range of angles 
(from —66° to 66° in our study) increases the odds that at 
any given time one or more of the beams will insonify the 
seafloor at angles between 35° and 50°. 
A limitation to using seafloor characteristics derived 
from multibeam sonar data to predict trawlability is that 
such data are restricted to the area within the width of the 
swath of the sonar and not necessarily applicable to the 
area outside of the swath. Additionally, seafloor character- 
istics cannot be calculated near the boundary of the swath, 
because of the prerequisite spatial scale of the analysis 
window needed to calculate them. The effect of this lim- 
itation is less for multibeam surveys with 100% bottom 
coverage (Pirtle et al., 2015). However, the “effective area” 
for extracting data on seafloor characteristics is further 
reduced by the inclusion of S, oblique. This predictor 
can be derived from only small areas of the sonar swath 
(i.e., incidence angles of 35-50°), effectively prohibiting 
full bottom coverage. That is, data collected at incidence 
angles <35° and >50° are not useful in predicting seafloor 
trawlability. 
Conclusions 
The GLM that combines S,, oblique values with VRM and 
BPI values produces reliable classification (on the basis of 
AUC results) of seafloor trawlability, even when applied 
to a larger geographic extent than that in previous work 
on seafloor trawlability in the GOA (Weber et al., 2013; 
Pirtle et al., 2015). However, VRM and BPI are not always 
critical predictors to the GLM—they are only beneficial in 
specific areas within the GOA (e.g., over the extent of the 
area sampled by Pirtle et al., 2015). The nonlinear models 
are not better at predicting seafloor trawlability, but the 
inclusion of VRM or BPI is more beneficial to these models 
and requires minimal computation effort. 
These results indicate that, to predict seafloor traw- 
lability in the GOA, it is important to use a model that 
has been developed with data from the entire extent of 
the area to which the model is being applied. Although 
results indicate that S;, oblique is a robust predictor of 
trawlability over both small and large extents of the GOA, 
the usefulness of the other seafloor characteristics may 
be limited to specific areas of the GOA. Consequentially, 
the use of seafloor characteristics derived from multibeam 
sonar data in models to predict seafloor trawlability is 
essential because the models are dependent on spatial 
representation across the specific area of interest within 
the GOA. 
The strength of the case for using models, rather than 
analysis of video images from camera deployments, to pre- 
dict seafloor trawlaiblity is the fact that multibeam sonar 
data can be collected relatively quickly and efficiently. A 
camera deployment on bottom for 30 min collects data only 
over an area that is approximately 1 m wide and 0.9 km 
long on the seafloor, whereas the Simrad ME70 echo 
sounder collects multibeam data across a swath width of 
335-1345 m (which corresponds to the range of camera 
stations at depths between 75-300 m) and along a track of 
10.2 km (assuming the ship is moving at 11 kt) during the 
same time period (i.e., 30 min). Reliable estimates of the 
extent of trawlable and untrawlable seafloor can be used 
to improve the accuracy of acoustic backscatter attributed 
to rockfishes in grid cells for areas classified as untraw- 
lable (Jones et al., 2021), in turn improving the accuracy 
of stock assessments for these species. 
Acknowledgments 
We thank the scientists and crew of the NOAA Ship Oscar 
Dyson for their hard work in helping with the data collec- 
tion. We thank T. Honkalehto and C. O’Leary for valuable 
comments and remarks on earlier drafts of this paper. 
Literature cited 
Aydin, K., S. Barbeaux, M. Bryan, S. Cleaver, C. Cunningham, 
O. Davis, M. Dorn, A. Dreary, K. Echave, C. Faunce, et al. 
2019. Stock assessment and fishery evaluation report for 
the groundfish resources of the Gulf of Alaska, 43 p. North 
Pac. Fish. Manage. Counc., Anchorage, AK. [Available from 
website.] 
Breiman, L. 
2001. Random forests. Mach. Learn. 45:5-32. 
Brown, C. J., and P. Blondel. 
2009. Developments in the application of multibeam sonar 
backscatter for seafloor habitat mapping. Appl. Acoust. 
70:1242-1247. 
Burnham, K. P., and D. R. Anderson. 
2002. Model selection and multi-model inference: a practical 
information-theoretic approach, 2nd ed., 488 p. Springer, 
New York. 
Cordue, P. L. 
2007. A note on non-random error structure in trawl survey 
abundance indices. ICES J. Mar. Sci. 64:1333-1337. 
Cutter, G. R., Jr., L. Berger, and D. A. Demer. 
2010. A comparison of bathymetry mapped with the Simrad 
ME70 multibeam echosounder operated in bathymetric and 
fisheries modes. ICES J. Mar. Sci. 67:1301—1309. 
Diaz-Uriarte, R., and S. Alvarez de Andrés. 
2006. Gene selection and classification of microarray data 
using random forest. BMC Bioinf. 7:3. 
